Professional Documents
Culture Documents
9, SEPTEMBER 2021
Abstract— In this letter, a double polar code (D-Polar) for joint ity of any symmetric binary-input discrete memoryless
source and channel coding (JSCC) is proposed, in which the channel (B-DMC) with the efficient encoding and decoding
source compress is implemented by a polar code, the channel algorithms. Polar codes have been well studied in the field of
error correction is performed by a systematic polar code.
Furthermore, a turbo-like belief propagation (TL-BP) decoder
channel coding [8]–[10]. For a joint source-channel decoding
consisted of a channel BP decoder and a source BP decoder is (JSCD) problem, polar codes can effectively exploit the resid-
proposed for joint source and channel decoding in the receiver. In ual redundancy in the output of the source encoder [11], e.g.,
this TL-BP decoder, the soft information is exchanged between for language-based sources, polar list decoder can improve its
the channel BP decoder and the source BP decoder so as to performance by judging the reliability of the decoded words
improve the efficiency of channel decoding in utilizing the source through a dictionary [12]. In the joint decoding scheme for
information residual in the compressed bits. Simulation results
show that the performance of the proposed D-Polar JSCC scheme correlated sources, the use of systematic polar codes facilitates
with TL-BP decoder is significantly improved compared with the decoders to exploit the correlation among the sources [13].
existing source-channel joint polarization scheme. Since the polarization phenomenon also exists in the source,
Index Terms— Joint source and channel coding, double polar polar codes are applied in the field of source coding, including
codes, turbo-like belief propagation decoder. lossless source polar coding [14], [15] and lossy source polar
coding [16]. Considering joint source-channel polarization,
a quasi-uniform systematic polar code (qu-SPC) is proposed
I. I NTRODUCTION for JSCC, whose construction process changes the regular
Authorized licensed use limited to: ANURAG GROUP OF INSTITUTIONS. Downloaded on February 15,2022 at 04:50:39 UTC from IEEE Xplore. Restrictions apply.
DONG et al.: JSCC USING D-POLAR CODES 2811
Authorized licensed use limited to: ANURAG GROUP OF INSTITUTIONS. Downloaded on February 15,2022 at 04:50:39 UTC from IEEE Xplore. Restrictions apply.
2812 IEEE COMMUNICATIONS LETTERS, VOL. 25, NO. 9, SEPTEMBER 2021
Algorithm 1 TL-BP Decoding for D-Polar Codes 2) Source BP Decoding: In an external iteration, the source
Input: the received vector y1Ncc BP decoder is triggered after the termination of the internal
iteration of the channel BP decoder, as presented in Lines 9-13
Output: a decoded source sequence ûN 1
sc
of Algorithm 1. Similar to the left half of Fig. 2, the source
1 for t = 1, 2, . . . , T do
polar code with message length Nsc can also be represented
//Channel BP decoder
0 as a factor graph with m + 1 stages and Nsc levels, so the
2 Ri,1 ← +∞ for i ∈ A
source BP decoder is performed on a factor graph with m + 1
3 Li,n+1 ← L(yi ) + Lsa (yi ) for i ∈ B
0
stages, each stages consisting of Nsc nodes. For the sake of
4 L0i,n+1 ← L(yi ) for i ∈ B c distinction, in the t-th iteration, let li,jt t
and ri,j denote the
5 for tc = 1, 2, . . . , Tcc do messages propagating from left to right and from right to left
6 Update Lti,j and Ri,j t
based on (8) in the (i, j)-index node, respectively.
7 Apply early stopping criterion The initialization of the source BP decoder needs to be done
8 Lca (vF ) ← {Lti,n+1 |i ∈ B} before running, which is shown in Lines 9-10 of Algorithm 1.
//Source BP decoder Unlike the all-zero subvector wA , the compressed bits vF
9
0
ri,1 ← Lca (vi ) for i ∈ F are not fixed. The a priori information of vF , denoted by
10 li,m+1 ← ln 1−p
0 Lca (vF ), acts as the role of +∞ in (6), which is provided by
p
11 for ts = 1, 2, . . . , Tsc do the channel BP decoder in an external iteration. Therefore,
t
12
t
Update li,j and ri,jt
based on (13) the initialization of ri,1 is as follows
13 Apply early stopping criterion
0 Lca (vi ), if i ∈ F,
14 Lsa (xB ) ← {ri,m+1
t
|i ∈ F } ri,1 = (11)
0, otherwise.
15 Obtain the source estimate ûN
1
sc
t
return the decoded source sequence ûN sc Meanwhile, in a conventional BP decoder, ri,1 is initialized
16 1 .
as the channel received LLR, but for the source polar code
used in this letter, the source sequence is not transmitted
through the channel, instead, the decoder knows the probability
In the standard BP decoder, the right hand side left- distribution of the source symbols. Suppose Pr(u = 1) = p
propagating message Lti,n+1 is initialized to channel received t
in the source, then the initialization of li,m+1 is given by
LLR L(yi ). However, in the TL-BP decoder, since the chan-
nel BP decoder knows the priori information Lsa (xi ) of the 0 Pr(u = 0) 1−p
li,m+1 = ln = ln (12)
systematic bit xB sent from the source BP decoder in the Pr(u = 1) p
last external iteration, the initialization of Lti,n+1 (Lines 3-4, t t
Algorithm 1) should be written as follows: The message inside all nodes, except ri,1 and li,m+1 ,
is initialized to zero. Like the channel BP decoder, update
0 L(yi ) + Lsa (xi ), if i ∈ B, rules can be described as
Li,n+1 = (7) ⎧t
L(yi ), otherwise. t
li,j = f (ri+2 t t
⎪ j ,j + li+2j ,j+1 , li,j+1 )
⎪
⎪
⎨l t t t t
Other messages are initialized to 0 since no further a priori i+2j ,j = f (ri,j , li,j+1 ) + li+2j ,j+1
t t−1 t t
(13)
information is given. It should be noted that both messages ⎪
⎪ ri,j+1 = f (li+2 j ,j+1 + ri+2j ,j , ri,j )
Lti,n+1 and Ri,1 t
are not updated through the decoding. ⎪
⎩ t t−1 t t
ri+2j+1 ,j+1 = f (li,j+1 , ri,j ) + ri+2 j ,j
In t-th iteration, the channel BP decoder first updates Lti,j
serially in the order of j = n, n − 1, . . . , 1, and then serially Based on the above update rules, in the t-th iteration,
t
updates Ri,j in the order of j = 2, 3, . . . , n + 1. Update rules the decoder updates ri,jt
in the order j = 2, 3, . . . , m + 1
t
can be described as and then updates li,m+1 in the order j = m, m − 1, . . . , 1.
⎧ t t t t
⎪ Li,j = f (Ri+2 j ,j + Li+2j ,j+1 , Li,j+1 )
The maximum iteration number of the source BP decoder
⎪
⎪ is determined by a preset parameter Tsc . After the iteration
⎨L t t t t
i+2j ,j = f (Ri,j , Li,j+1 ) + Li+2j ,j+1 number t reaches Tsc , source BP decoding stops, at which
t−1 (8)
⎪ t t
⎪Ri,j+1 = f (Li+2j ,j+1 + Ri+2j ,j , Ri,j )
t
⎪
⎩ t point one external iteration ends. Same with (10), the posterior
Ri+2j+1 ,j+1 = f (Lt−1 t t
i,j+1 , Ri,j ) + Ri+2j ,j probability ratio in the logarithmic domain of vi is given by
where f (x, y) is as follows t t
Lp (vi ) = ri,m+1 + li,m+1 . (14)
x+y
1+e
f (x, y) log . (9) 3) Soft Information Exchange: As shown in Fig. 3, in an
ex + ey external iteration, the channel BP decoder supplies the extrin-
After reaching a preset number of iterations Tcc , channel sic information of xB to the source BP decoder, and the
BP decoding terminates. we can get the posterior probability extrinsic information of vF from the source BP decoder is fed
ratio in the logarithmic domain of xi to the channel BP decoder. The required extrinsic information
is obtained by subtracting the external input information from
t
Lp (xi ) = Ri,n+1 + Lti,n+1 . (10) the posterior information. Based on (7) and (10), the extrinsic
Authorized licensed use limited to: ANURAG GROUP OF INSTITUTIONS. Downloaded on February 15,2022 at 04:50:39 UTC from IEEE Xplore. Restrictions apply.
DONG et al.: JSCC USING D-POLAR CODES 2813
Authorized licensed use limited to: ANURAG GROUP OF INSTITUTIONS. Downloaded on February 15,2022 at 04:50:39 UTC from IEEE Xplore. Restrictions apply.
2814 IEEE COMMUNICATIONS LETTERS, VOL. 25, NO. 9, SEPTEMBER 2021
IV. C ONCLUSION
This letter proposed a D-Polar code for JSCC. The proposed
D-Polar code consists of a source polar code and a channel
systematic polar code. The source polar coding is directly
applied in the proposed D-Polar codes scheme. Meanwhile,
the soft iterative structure of the TL-BP decoder allows the
channel decoder to take full advantage of the residual source
information in the compressed bits. Simulation results indicate
that the D-Polar codes achieves 2.33 dB performance gain
under the TL-BP decoder compared with the existing source-
Fig. 5. BER comparison of the TL-BP decoder with Rsc = {3/5,
7/10, 3/4} and p = {0.07, 0.04, 0.02}.
channel joint polarization scheme, i.e., qu-SPC. Theoretical
analysis of error floor and construction of soft iterative SCL
algorithms applicable to D-Polar codes are future works to be
addressed.
R EFERENCES
[1] C. E. Shannon, “A mathematical theory of communication,” Bell Syst.
Tech. J., vol. 27, no. 3, pp. 379–423, 1948.
[2] M. Fresia, F. Perez-Cruz, and H. V. Poor, “Optimized concatenated
LDPC codes for joint source-channel coding,” in Proc. IEEE Int. Symp.
Inf. Theory (ISIT), Jun. 2009, pp. 2131–2135.
[3] M. Fresia, F. Perez-Cruz, H. V. Poor, and S. Verdu, “Joint source
and channel coding,” IEEE Signal Process. Mag., vol. 27, no. 6,
pp. 104–113, Nov. 2010.
[4] Q. Chen, L. Wang, S. Hong, and Y. Chen, “Integrated design of
JSCC scheme based on double protograph LDPC codes system,” IEEE
Commun. Lett., vol. 23, no. 2, pp. 218–221, Feb. 2019.
[5] S. Liu, C. Chen, L. Wang, and S. Hong, “Edge connection optimization
Fig. 6. BER comparison of the D-Polar codes with the DP-LDPC for JSCC system based on DP-LDPC codes,” IEEE Wireless Commun.
codes [4], [6]. Lett., vol. 8, no. 4, pp. 996–999, Aug. 2019.
[6] S. Liu, L. Wang, J. Chen, and S. Hong, “Joint component design for
the JSCC system based on DP-LDPC codes,” IEEE Trans. Commun.,
number of iterations for the channel and source BP decoders vol. 68, no. 9, pp. 5808–5818, Sep. 2020.
[7] E. Arıkan, “Channel polarization: A method for constructing capacity-
are 15.56 and 11.67, respectively. achieving codes for symmetric binary-input memoryless channels,” IEEE
The error floor of the D-Polar codes is due to residual Trans. Inf. Theory, vol. 55, no. 7, pp. 3051–3073, Jul. 2009.
decoding errors introduced by source code. The proposed [8] I. Tal and A. Vardy, “How to construct polar codes,” IEEE Trans. Inf.
Theory, vol. 59, no. 10, pp. 6562–6582, Oct. 2013.
D-Polar codes has a design parameter, Rsc , which can be [9] K. Niu and K. Chen, “CRC-aided decoding of polar codes,” IEEE
used to trade-off the coding gain and the error floor. As the Commun. Lett., vol. 16, no. 10, pp. 1668–1671, Oct. 2012.
parameter Rsc increases, the coding gain decreases and the [10] I. Tal and A. Vardy, “List decoding of polar codes,” IEEE Trans. Inf.
error floor is lower. To illustrate this point, we show the effect Theory, vol. 61, no. 5, pp. 2213–2226, May 2015.
[11] Y. Wang, K. R. Narayanan, and A. A. Jiang, “Exploiting source
of different Rsc on the our proposed D-Polar JSCC scheme redundancy to improve the rate of polar codes,” in Proc. IEEE Int. Symp.
in Fig. 5. When Rsc = 7/10, the error floor reaches 10−5 , Inf. Theory (ISIT), Jun. 2017, pp. 864–868.
and when Rsc = 3/4, the error floor reaches 10−6 . The [12] Y. Wang, M. Qin, K. R. Narayanan, A. Jiang, and Z. Bandic, “Joint
source-channel decoding of polar codes for language-based sources,” in
reduction in error floor comes at the cost of performance Proc. IEEE Global Commun. Conf., Dec. 2016, pp. 1–6.
loss. At BER = 10−3 , the performance losses of Rsc = [13] L. Jin, P. Yang, and H. Yang, “Distributed joint source-channel decoding
7/10 and Rsc = 3/4 compared to Rsc = 3/5 are 0.4 dB using systematic polar codes,” IEEE Commun. Lett., vol. 22, no. 1,
pp. 49–52, Jan. 2018.
and 0.76 dB, respectively. In addition, we also provide the [14] E. Arikan, “Source polarization,” in Proc. IEEE Int. Symp. Inf. Theory
influence of different source statistics on the proposed JSCC (ISIT), Jun. 2010, pp. 899–903.
system. As shown in Fig. 6, the lower the entropy is, the better [15] H. S. Cronie and S. B. Korada, “Lossless source coding with polar
codes,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Jun. 2010,
the BER performance is. The gains that can be obtained for pp. 904–908.
p = 0.04 and p = 0.02 compared to p = 0.07 are 0.69 dB [16] S. B. Korada and R. Urbanke, “Polar codes are optimal for lossy
and 1.04 dB at BER = 10−3 , respectively. source coding,” in Proc. IEEE Inf. Theory Workshop (ITW), Oct. 2009,
Furthermore, we provide a comparison of the DP-LDPC pp. 149–153.
[17] L. Jin and H. Yang, “Joint source-channel polarization with side infor-
JSCC scheme and the D-Polar JSCC scheme under short codes mation,” IEEE Access, vol. 6, pp. 7340–7349, 2018.
in Fig. 6. For facilitate comparison with [4], [6], we change [18] B. Yuan and K. K. Parhi, “Early stopping criteria for energy-efficient
Nsc to 1024 and the overall code rate to 1, where Rsc and Rcc low-latency belief-propagation polar code decoders,” IEEE Trans. Signal
Process., vol. 62, no. 24, pp. 6496–6506, Dec. 2014.
are both 1/2. At BER = 10−2 , the D-Polar codes has 0.5 dB [19] P. Trifonov, “Efficient design and decoding of polar codes,” IEEE Trans.
gain compared to [4] but there is still a 1 dB performance gap Commun., vol. 60, no. 11, pp. 3221–3227, Nov. 2012.
Authorized licensed use limited to: ANURAG GROUP OF INSTITUTIONS. Downloaded on February 15,2022 at 04:50:39 UTC from IEEE Xplore. Restrictions apply.