You are on page 1of 4

Designs of Bhattacharyya Parameter in the

Construction of Polar Codes

Shengmei Zhao Peng Shi Bei Wang


Institute of Signal Processing & Institute of Signal Processing & Institute of Signal Processing &
Transmission Transmission Transmission
Nanjing University of P& T Nanjing University of P& T Nanjing University of P & T
Nanjing, China Nanjing, China Nanjing, China

Abstract—Polar codes, introduced by Arikan recently, can It is the unique construction method of Polar codes so that
achieve capacity of symmetric channel with low complexity they could have the capacity-achieving property. This is named
successive cancellation decoding strategy. In this paper, we channel Polarization. There are two important parameters,
design efficient Bhattacharyya Parameters in Polar codes including the symmetric capacity and the Bhattacharyya
construction for common communications channels. We parameter for channel Polarization, are used as measures of
present the satisfactory recursive formula for the four rate and reliability respectively. Between them, the
conventional channels and discuss the initial value of each Bhattacharyya parameter Z(W) is an upper bound on the
recursive equations. Numerical simulations show that Polar probability of maximum-likelihood (ML) decision error when
codes are channel-specific designing codes; they can approach the channel is used only once to transmit a 0 and 1. Intuitively,
their capacities by selecting the frozen bits in our proposed channels with Z(W)≤ε are almost noiseless, while channel with
Bhattacharyya Parameter calculating method. Z(W)≥1-ε are almost pure-noise channel for a given 0≤ε≤ 1 [1].
Keywords-Polar codes; channel polarization; capacity; Therefore, the Bhattacharyya parameter of channel determines
Bhattacharyya parameter. the construction of Polar code, it plays an important part in the
construction of Polar codes.
I. INTRODUCTION In this paper, we design different Bhattacharyya parameter
for common communication channels, and compare the
Polar codes were proposed by Arikan [1] in 2007, they
performance of Polar codes in the same communication
have been proven to be capacity-achieving codes for Binary
channel with different formula. Besides, we discuss the effect
Symmetric Channel (BSC), and then for arbitrary binary-input
of initial value of Bhattacharyya parameter on system’s
discrete memoryless channels (DMCs) [2], with low-
performance, and propose a scheme for the construction of
complexity encoding and decoding method. Since Polar codes
Polar codes for different communication channels.
demonstrate their versatility, they attract a lot of attention after
they are proposed [3-13].
II. POLAR CODES AND THEIR CONSTRUCTION
Polar codes take advantages of channel Polarization, which
is a method, for constructing codes of capacity-achieving This section provides a concise overview of the
channel for DMCs [1]. Based on this conception, Korada and groundbreaking work of Arikan and others on Polar codes and
Urbanke then established the suitability for source coding to channel Polarization. Let
achieve the fundamental limits in lossy source
compression with binary reproduction alphabets [3,4]. Later, a ⎡1 0 ⎤
G2 = ⎢ ⎥ (1)
generalization to a broader class of reproduction alphabets was ⎣1 1 ⎦
established [5]. In addition, non-binary input alphabets of Polar
codes are investigated, such that Reed-Solomon codes and and G2⊗m denotes the m-th Kronecker power of G2. For
algebraic geometry core are selected as the kernel of Polar simplicity and generality, let W be a symmetric binary-input
codes [6]. Considering both source and channel coding, the DMC, and V=(V1,V2,…Vn) be a block of n=2m bits chosen
design of Polar codes for channel and source coding was randomly from {0,1}n. The encoding is an operation as,
addressed in [7-9]. Following these results, the interest on
X = VPn G2⊗ m (2)
channel Polarization and Polar codes has been extended to
problems with multiple sources and channels, for example, where Pn is a n×n bit-reversal permutation matrix. Finally, X is
lossless compression of correlated sources, source compression transmitted through n independent copies of W. Arikan [1]
with side information at the decoder, transmission over considered that the n channels could be “seen” by each of the n
broadcast channels [4], multiple access channels [4,10], or the individual bits V1,V2,…Vn as they undergo the above
relay channel [11]. Furthermore, Polar codes find their transformation, called the bit-channels. It is shown that as m
application in wiretap channels to achieve the secrecy capacity grows, the bit-channels start polarizing, they approach either a
[12,13].

978-1-4244-6252-0/11/$26.00 ©2011 IEEE


noiseless channel (called good bit-channel) or a pure-noise III. DESIGNS OF BHATTACHARYYA PARAMETER FOR
channel (named bad bit-channel). The fraction of bit-channels DIFFERENT COMMUNICATION CHANNELS
with no noise approaches the capacity of W as n→∞. Here, the Arikan gave the formulas of Bhattacharyya parameter in the
good and bad for bit-channels is determined by the recursive properties of the channel Polarization in [1].
Bhattacharyya parameter of W, which is defined as, Bhattacharyya parameter is concluded as,
Z (W ) = ∑ W ( y | 0)W ( y | 1) (3) Z(WN(i) ) ≤ Z(W2(2Ni −1) ) ≤ 2Z(WN(i ) ) − Z(WN(i) )2 (9)
y

It is obvious Z (W ) is the erasure probability for binary erasure Z (W2(2Ni ) ) = Z (WN( i ) ) 2 (10)
channel (BEC). For simplicity, the above formula can be written as,
The basic idea of Polar codes is to create a coding system Z (W2(2Ni −1) ) = 2 Z (WN(i ) ) − Z (WN(i ) )2 (11)
where one can access each bit-channel individually and send
data only through those for its Bhattacharyya parameter is near Here, if W is a BEC, it is perfect. But for other
0. For a given N, where N = 2n for n ≥0, the encoding algorithm communications channel, it is not sure the formula will be
is expressed as, good. In this section, we will discuss the different designs of
Bhattacharyya parameter for different communication channels,
x1N = u1N GN (4) such as, binary symmetric channel (BSC), Gaussian channel
(AGNC) and Rayleigh channel.
where GN is the generator matrix of order N, G N = G2⊗n . For A
an arbitrary subset of {1,…,N}, the above formula may be According to the definition of the Bhattacharyya parameter,
written as, we modify the Bhattacharyya parameter formula to the
following three types, and analyze the effect of Bhattacharyya
x1N = u A GN ( A) ⊕ u Ac GN ( Ac ) (5) parameter to the performance of Polar code.

where GN(A) denotes the sub-matrix of GN formed by rows Type I: Z (W2(2Ni −1) ) = 2Z (WN(i ) ) − Z (WN(i ) )2 ,
with indices in A. Commonly, Polar code is identified by
(N,K,A,uAc), where K is the code dimension and specifies the Type II: Z (W2(2Ni −1) ) = Z (WN(i ) ) ,
size of A, the ratio K/N is code rate, A is information set and uAc
is named frozen bits. Polar codes will be specified by giving a Type III: Z(W2(2Ni−1) ) = 0.5×(2Z(WN(i) ) − Z(WN(i) )2 + Z(WN(i) )) .
particular rule for the selection of the information set A. We present the comparison results of three different types
Definitely, Polar codes are channel-specific designs that mean Bhattacharyya parameter for three common communication
a Polar code for one channel may not be a Polar code for channels. Figure 1 is the performance comparison of Polar
another. codes with three Z(W) types designed for BSC with code rate
On the other hand, the decoding is to generate an estimate 0.04 and block length 256. The abscissa shows the error
uˆ1N of u1N , given knowledge of A, uAc and y1N. Since the probability of BSC. Figure 2 is the comparison of AGNC for
transformation rate 0.14 and block length 256 with three types
decoder can avoid errors in the frozen part, the real decoding Z(W), the abscissa shows the signal-to-noise rate of AGNC.
task is to generate an estimate û A of u A . Successive Figure 3 is the comparison of Rayleigh channel which the
cancellation algorithm is usually employed. One can generate transformation rate is 0.14 and the block length is 256, the
the estimate by the following method, abscissa is also the signal-to-noise rate of channel. It is shown
that Type I Bhattacharyya parameter has the best performance
⎧⎪u if i ∈ Ac among three kinds of Z(W).
u i = ⎨ i N i−1 (6)
⎪⎩ hi ( y1 , uˆ1 ) if i∈A
and the decision function hi is defined as,

⎧0 if WN(i) ( y1N , u1i−1 | 0) >WN(i) ( y1N , u1i−1 |1)


hi = ⎨ (7)
⎩1 otherwise
For any BDMC W and any choice of Polar codes, the
probability of block error for a Polar code is proved to be,

Pe ( N , K , A, u Ac ) ≤ ∑ Z (WN(i ) ) (8) Fig.1: The comparison of BSC for three types with R=0.04 N=256.
i∈ A
( y + 1) 2
1 −
W ( y | 0) = e 2σ 2
(15)
2π σ
( y − 1) 2
1 −
W ( y | 1) = e 2σ 2
(16)
2π σ
Therefore, Z0 of the Gaussian channel can be calculated by
the variance of Gaussian channel. In comparison, Rayleigh
channel is more complicated. Normally, Rayleigh fading can
be described as,
Fig.2: The comparison of AGNC for the three types, with R=0.14 N=256.
a = K x2 + y2 (17)
where x, y are independent Gaussian random variables with
expectation 0 and variance 1, K is the scaling factor. We
assume that information accepted in the receiver obeys
Gaussian distribution. The expectation is the peak value of
Rayleigh distribution, and the variance is related to the signal-
to-noise ratio S. On the basis of Rayleigh distribution’s
properties, we can get that information’s expectation E and
variance σ2 as follows,

Fig.3: The comparison of the Rayleigh channel for three types with R=0.14
4 (18)
E=± ln 4 K
N=256. 4−π
After we obtain the recursive formulas for the computation S

of Z(W), the initial value of Bhattacharyya parameter becomes


σ 2 = (10 10 −1
) (19)
an important element in Z(W) calculation. In [14], Arikan gave According to the definition, W ( y | 0) and W ( y | 1) can be
the initial value of Bhattacharyya parameter Z0 to 0.5 for any
channels. It is apparent this value is not suitable to all written as:
communication channels since Polar codes are channel-specific ( y+
4
ln 4 K ) 2
designs. Now, we will discuss the best initial value of Z(W) for 1 − 4 −π

communication channels. W ( y | 0) = e 2σ 2 (20)


2π σ
Let Pe denote the erasure rate of BEC. It is easy to get Z0 of 4
BEC by definition in (3) as, (y−
4 −π
ln 4 K ) 2
1 −
W ( y | 1) = e 2σ 2 (21)
Z 0 = pe (12) 2π σ
Simplicity, according to the definition, Z0 of BSC is , Now, we will give the bit error rate (BER) of Polar codes
by comparing Z0=0.5 with calculating Z0 for different
Z 0 = 2 pε (1 − pε ) (13) communication channels by above results. We consider the
length of Polar codes N = 256, and the transmission codeword
where Pε is the error probability of BSC. It is shown that the 1000. We employ BP decoding algorithm for the simulations.
initial value of Z(W) is dependent on the error probability of
discrete channel and could be calculated. Hence, we call this
initial value as “calculating Z0”. However, the initial value of
Bhattacharyya parameter is definitely difficult for Gaussian
channel and Rayleigh channel, because both channels are
continuous. First, we can extend the definition of
Bhattacharyya parameter from discrete to continuous channel.
The Bhattacharyya parameter for continuous channel will be ,

Z (W ) = ∫ W ( y | 0)W ( y | 1)dy (14)

Then, we analyze the initial value of Z(W) for Gaussian Fig.4. The BER of Polar codes for BEC with R=0.14 and N=256.
channel at first. Suppose there is a communication link with
Gaussian noise with expectation 0 and variance σ2. At the same
time, PSK is adopted as modulation mode, the output of
AGNC is a signal with expectation 1 or -1 and variance σ 2 ,
W ( y | 0) and W ( y | 1) can be written as,
IV. CONCLUSION
Polar codes are capacity-achieving channel codes based on
channel polarization. Bhattacharyya parameter plays an
important part in Polar codes construction. In this paper, we
present different formula of Bhattacharyya parameter for
different communication channels. Firstly, we discuss the
effects of the different formula in recursive properties of the
channel polarization for different channels. It is shown that
Type I Bhattacharyya parameter has the best performance
among three kinds of Z(W). Secondly, we deduce the
Fig.5:The BER of Poalr codes for BSC with R=0.04, N=256 expression of the initial value of Bhattacharyya parameter
Z 0 for different channels and verify them by numerical
simulations. For BEC, BSC, AGNC, we find our calculating Z0
is more satisfactory. But for Rayleigh channel, the case is more
complicated. The initial value of Z(W) is dependent on signal-
to-noise ratio of the channel.

V. ACKNOWLEDGMENT
This work was supported in part by the Foundation (No.
NJ210002).

Fig.6:The BER of Polar codes for AGNC with R=0.14, N=256 REFERENCES
[1] E.Arikan, “Channel Polarization: A method for constructing capacity-
achieving codes for symmetric bianry-input memoryless channels,”
IEEE Trans. Inf. Theory, vol. 55, pp. 3051-3073, 2009
[2] S.Korada, R.Urbanke, and E.Sasoglu, “Polar coes: Characterization of
exponent, bounds, and constructions,” Informaiton Theory, 2009, ISIT
2009. IEEE International Symposium on, pp. 1483-1487, 2009
[3] S.B. Korada and R.L. Urbanke, “Polar codes are optimal for lossy
source coding,” IEEE Transactions on Information Theory, pp.1751 –
1768, 2010
[4] S. B. Korada, “Polar Codes for Channel and Source Coding,” PhD
thesis, EPFL, Lausanne,Switzerland, 2009
[5] M. Karzand and E. Telatar, “Polar codes for Q-ary source coding,” In
Fig.7: The BER of Polar codes for Rayleigh channel with low signal-to-noise Proc. IEEE Int. Symp. Information Theory (ISIT), pp 909 –912, 2010
ratio for R=0.14,N=256,K=0.9
[6] R.Mori and T.Tanaka,“Non-binary Polar codes susing Reed-Solomon
codes and algebraic geometry codes,” IEEE Trans.Information Theory
Workshop(ITW), pp.1-5, 2010
[7] R. Mori and T. Tanaka, “Performance and construction of Polar codes
on symmetric binary-input memoryless channels,” In Proc. IEEE Int.
Symp. Information Theory (ISIT), pp 1496 –1500, 2009
[8] R. Mori and T. Tanaka, “Performance of Polar codes with the
construction using density evolution,” IEEE Communications Letters,
13(7):519 –521, 2009
[9] N. Hussami, S.B. Korada, and R. Urbanke, “Performance of Polar codes
for channel and source coding,” In Proc. IEEE Int. Symp. Information
Theory (ISIT), pp 1488 –1492, 2009
Fig.8: The BER of Polar codes for Rayleigh channel of high signal-to-noise [10] E. Abbe and E. Telatar, “MAC Polar codes and matroids,” In Proc.
Workshop on Information Theory and Applications (ITA), pp. 1–8, 2010
ratio, R=0.36,N=256,K=0.9
Figure 4, Figure 5 and Figure 6 are comparison results for [11] M. Andersson, V. Rathi, R. Thobaben, J. Kliewer, and M. Skoglund,
“Nested Polar codes for wiretap and relay channels,” IEEE
BEC, BSC and AGNC respectively. It’s clearly that the BER of Communications Letters, vol 14, no. 8, pp.752 –754, 2010
using calculating Z0 is lower than using Z0=0.5. In other words, [12] H.Mahdavifar and A.Vardy, “Achieving the secrecy capapcity of
we can select the calculating Z0 for these three communication Wiretap channels using Polar codes,” IEEE. Information Theory
channel in Polar codes designs. Figure 7 is the comparison of Proceedings(ISIT), pp.913-917, 2010
BER in Rayleigh channel of low signal-to-noise ratio with [13] E.Hof and S.Shamai, “Secrecy-achieveing Polar-coding,” IEEE Trans.
transformation rate 0.14, block length 256 and scaling factor Informtion. Theory. Workshop(ITW) , pp.1-5, 2010
0.9. It’s shown that using calculating Z0 is better than using [14] E.Arikan, “A performance Comparison of Polar codes and Reed-Muller
Z0=0.5. Figure 8 is the comparison result of BER in Rayleigh Codes,” IEEE Comm. Letters. vol.12, no.6, pp.447-449, 2008
channel of high signal-to-noise ratio with transformation rate
0.36, block length 256 and scaling factor 0.9. It is shown that at
this case Z0=0.5 has a better performance.

You might also like