1

Frozen Set Design for Precoded Polar Codes

Vera Miloslavskaya, Yonghui Li, Fellow, IEEE, and Branka Vucetic, Fellow, IEEE

Abstract—This paper focuses on the frozen set design for list size required by SCL to approach the ML performance can
precoded polar codes decoded by the successive cancellation be characterized by the information-theoretical quantities [8].
list (SCL) algorithm. We propose a novel frozen set design The ML performance may also be approached by the Fano
method, whose computational complexity is low due to the use of
analytical bounds and constrained frozen set structure. We derive decoding [17], whose complexity is connected with the cutoff
new bounds based on the recently published complexity analysis rate [18]. We focus on the SCL decoder as the most widely
arXiv:2311.10047v1 [cs.IT] 16 Nov 2023

of SCL with near maximum-likelihood (ML) performance. To used decoder for precoded polar codes.
predict the ML performance, we employ the state-of-the-art In this paper, we propose a novel low-complexity frozen set
bounds relying on the code weight distribution. The bounds and design method for precoded polar codes with various trade-
constrained frozen set structure are incorporated into the genetic
algorithm to generate optimized frozen sets with low complexity. offs between the FER performance and decoding complexity.
Our simulation results show that the constructed precoded polar The main contributions are as follows. First, we explore the
codes of length 512 have a superior frame error rate (FER) SCL list size lower bound from [8] and identify the factors
performance compared to the state-of-the-art codes under SCL limiting its effectiveness as the predictive measure for near
decoding with various list sizes. ML decoding complexity. Second, we improve the prediction
Index Terms—Polar codes, complexity prediction, maximum- accuracy by tightening the lower bound from [8]. Third, we
likelihood decoding, successive cancellation list decoding, sequen- introduce an approximate lower bound that facilitates a fair
tial decoding.
comparison of various frozen sets. This approximation com-
bines our tightened lower bound with the upper bound from
I. I NTRODUCTION
[8]. Fourth, we propose to solve the frozen set optimization
The polar codes [1] have frozen bits that are all set to zeros problem by minimizing the ML decoding error probability
or other fixed values. The polar code generalizations such as estimate under the decoding complexity constraint, which is
the CRC-aided polar codes [2], polar subcodes [3], parity- given by the proposed approximate lower bound. The resulting
check-concatenated polar codes [4], polarization-adjusted con- frozen sets are intended for precoded polar codes utilizing
volutional (PAC) codes [5] and precoded polar codes [6] frozen bit expressions with near-uniformly distributed binary
involve frozen bits with non-fixed values, whose computation coefficients. Fifth, we impose constraints on the frozen set
may be specified by linear combinations of information bits structure to reduce the search space size and consequently
with lower indices. These combinations are referred to as the reduce the optimization complexity. Our simulation results
frozen bit expressions. Since polar codes with near-uniformly show that the constructed precoded polar codes of length 512
distributed frozen bit expressions are known to perform well have a superior FER performance compared to the state-of-
[7], [8], we limit our consideration to such codes. Their design the-art codes under SCL decoding with various list sizes.
problem reduces to the frozen set design problem. This confirms the efficiency of the proposed approximate
We treat the frozen set design problem as an optimization lower bound as the ML decoding complexity measure for
problem with the objectives of minimizing the decoding error comparing various frozen sets. Given an approximate lower
probability and complexity. For any particular decoder, the bound value, the frozen set optimization complexity is low due
frozen set may be optimized by using the genetic algorithm to the constraints on the frozen set structure and the absence
[9], where the code performance is evaluated via decod- of decoding simulations. For example, the genetic algorithm
ing simulations. However, the inherent high computational requires less than a minute to solve this problem for the code
complexity of these simulations necessitates a shift towards length 512.
analytical methods for code evaluation to ensure computational The paper is organized as follows. Section II provides a
efficiency. The state-of-the-art analytical methods for the polar background on the polar codes and relevant frozen set design
code evaluation are as follows. The frame error rate (FER) of criteria. In Section III, we derive the proposed bounds and
polar codes under the successive cancellation (SC) decoding specify the corresponding frozen set optimization process. In
[1] can be predicted using [10, Eq. (3)]. For the maximum- Section IV, we present the numerical results on the frozen
likelihood (ML) decoding, there are the FER bounds [11] set design complexity and the FER performance of precoded
parameterized by the weight distribution that can be computed polar codes with the proposed frozen sets and compare them
using [12]–[15]. Although there is no analytical bound predict- with the state-of-the-art.
ing the FER under the SC list (SCL) decoder [16], the average
The authors are with the School of Electrical and Computer II. P RELIMINARIES
Engineering, the University of Sydney, Sydney, NSW 2006, Aus-
tralia (e-mail: vera.d.miloslavskaya@gmail.com, yonghui.li@sydney.edu.au, This section provides a background on the polar codes, the
branka.vucetic@sydney.edu.au).
This research was supported by the Australian Research Council under ML performance of precoded polar codes, and the complexity
Grants FL160100032, DP190101988 and DP210103410. of near ML decoding using SCL.
 2

A. Polar Codes size as pointed out in [8, Remark 2]. To overcome this issue,
n
An (N = 2 , K) polar code [1] is a binary linear block [8, Remark 2] suggested to characterize the decoding list size

using the lower bound on D̄m that is derived
code consisting of codewords1 c = u·G⊗n , where G = 11 01 , Pin [8, Section
⊗n denotes the n-fold Kronecker product, the input vector u III-A].
P This lower bound is defined as D̄ m ≥ i∈A(m) Hn,i −
has K information bits ui , i ∈ A, and N − K frozen bits i∈F (m) (1−Hn,i ) by [8, Eq. (6a)], where Hn,i is the entropy
ui , i ∈ F = [N ] \ A, and [N ] , {0, . . . , N − 1}. Note that of the i-th bit-channel, i ∈ [2n ]. However, it follows from
[N ] = ∅ for N ≤ 0. The sets A and F are referred to as the the numerical results [8, Fig. 1] that the actual lower bound
low
information and frozen sets, respectively. In the case of the on D̄m , denoted by us as D̄m , takes into account the non-
original polar codes [1], all frozen bits have fixed values, e.g., negativity of entropy in Eq. (1) as
(
zeros. low
D̄m−1 + Hn,m , m ∈ A,
low
In a more general case, the frozen bits are equal to linear D̄m = low
(2)
max(D̄m−1 − (1 − Hn,m ), 0), m ∈ F ,
combinations of the other input bits with lower indices [19],
known as the frozen bit expressions. The resulting polar codes where m ∈ [2n ], and D̄−1 low
= 0. Note that Hn,i can be
are referred to as the polar codes with dynamic frozen bits, represented as 1 − In,i , where In,i is the mutual information
parity-check concatenated polar codes, precoded polar codes of the i-th bit-channel that can be recursively computed using
and pre-transformed polar codes in the literature. We use the [22, Eqs. (9), (10) and (26)] for the AWGN channel with BPSK
term “precoded polar codes” as in our previous works [6], modulation.
[20].

D. Frozen Bit Expressions

B. Weight Distribution of Precoded Polar Codes and Their
ML Performance It has been shown that codes with randomly generated
frozen bit expressions can perform well [7], [8], [13]. How-
The precoded polar codes are linear codes and, therefore,
ever, the random generation limits the reproducibility of the
can be characterized by their weight distributions. However,
results. Following [23], we ensure the reproducibility by
the complexity of computing the exact weight distribution is
using the deterministic binary sequence ω produced from the
high, except for very short codes and well-structured codes.
rational approximation of the π number: π ≈ 104348 33215 . Thus,
In this paper, we employ the average weight distribution [13],
ω = (ω0 , ω1 , ω2 , ω3 , . . . ) is equal to the binary expansion
whose computational complexity scales as O(N 3 ). To estimate
of 104348
33215 that can be easily computed. Given ω and the
the ML decoding error probability, we substitute the average
information bits ui , i ∈ A, we calculate the values of the
weight distribution into the union bound [11], known for its
frozen bits ui , i ∈ F , as follows:
simplicity, and the tangential-sphere bound (TSB) [21], known
for its tightness. l ← mini∈A wt(i)
b←0
for i ∈ F do ui ← 0
C. Complexity of SCL Decoding with Near ML Performance
for i ∈ F , wt(i) ≥ l do
It has been shown in [16] that the time complexity of SCL is for j ∈ A, j < i do
O(LN log(N )) and its space complexity is O(LN ), where L u i ← u i + ωb · u j
is the decoding list size. The FER performance of SCL decod- b←b+1
ing was experimentally shown to improve with increasing L at
where wt(i) is the Hamming weight of the binary expansion
the expense of increasing complexity. Recently, [8] provided
of the integer i. Note that all input bits ui with wt(i) < l are
ground-breaking results on the list size L such that SCL
set to zero, and therefore the resulting code is a subcode of
has a near ML performance. For general binary memoryless
the Reed-Muller code with the minimum distance 2l .
symmetric (BMS) channels, [8, Theorem 1] proved that the
mean value of the binary logarithm of L required at the m-th
stage of SCL to achieve the ML performance is upper bounded III. P ROPOSED F ROZEN S ET D ESIGN FOR P RECODED
by the conditional entropy D̄m P OLAR C ODES

D̄m , H(UA(m) |Y[N ] , UF (m) ), (1) This section presents our low-complexity frozen set design
method for precoded polar codes with various tradeoffs be-
where m ∈ [N ], A(m) , {i ∈ A | i ≤ m}, F (m) , {i ∈ tween the FER performance and decoding complexity. We
F | i ≤ m}, UT , {Ui | i ∈ T } for any set T , Ui is the focus on the problem of the complexity prediction for SCL
random variable corresponding to the i-th input bit, and Yi with a near ML performance, since this problem has been
is the random variable corresponding to the i-th output. Note partially solved by D̄mlow
from Eq. (2).
that we use the notation of [8] except for starting enumeration This section is organized as follows. We first consider
from zero instead of one. Unfortunately, the computation of limitations of D̄m low
as a decoding complexity measure in
D̄m requires performing decoding with a huge/unbounded list Section III-A and identify their source in Section III-B. To
1 We omit the multiplication by the bit-reversal permutation matrix B since
resolve the identified issues, we derive a new tightened lower
tight
u · B · G⊗n = u · G⊗n · B and the proposed techniques can be easily applied bound D̄m in Section III-C and alleviate the influence of
to permuted polar codes as well. the frozen set structure by combining the tightened lower
 3

bound with an upper bound in Section III-D. The result- that H(Um |YT , UΦ ) can be easily computed. The following
apx
ing approximate bound D̄m is further used as a decoding example illustrates the case of N = 4.
complexity measure during the frozen set optimization in
Example 1. For n = 2 and N = 2n = 4, the  N ×N  polar-
Section III-E. The optimization complexity is significantly 1000
reduced by imposing constraints on the frozen set structure. ization transformation2 is specified by G⊗n = 11 10 01 00 . Let
1111
Note that the proposed frozen design approach is intended for
us consider various cases of F (m−1) and calculate the corre-
precoded polar codes with near-uniform frozen bit expressions
sponding upper bounds on hm,F , H(Um |Y[N ] , UF (m−1) )
since both the performance and complexity criteria have been
Case m = 0 :
derived for such codes. (m−1)
• F = ∅ and then h0,F = H(U0 |Y[N ] ) = Hn,0 by
the definition of Hn,m .
low
A. Limitations of D̄m as a Decoding Complexity Measure
Case m = 1 :
low
The necessity to have a low D̄m for a precoded polar • if F
(m−1)
= {0}, then h1,F = H(U1 |Y[N ] , U0 ) = Hn,1
code to approach the ML performance under SCL with a low by the definition of Hn,m .
complexity has been proven in [8] for BMS channels. Besides, • if F
(m−1)
= ∅, then h1,F = H(U1 |Y[N ] ) ≤
[8, Appendix] specified three exemplary frozen sets for the H(U1 |Y1 , Y3 ) = Hn−1,0 since the received vector
code parameters (512, 256) and [8, Fig. 4] illustrated their (Y1 , Y3 ) corresponds to the transmitted (U1 , U3 )G.
remarkable performance. However, the following example
shows the limited applicability of D̄m low
for the frozen set Case m = 2 :
(m−1)
comparison. For the code parameters (512, 256), the frozen set • if F = {0, 1}, then h2,F = H(U2 |Y[N ] , U0 , U1 ) =
consisting of 256 less reliable bit-channels is characterized by Hn,2 by the definition of Hn,m .
maxm D̄m low
= 0.953, where the bit-channel reliabilities are • if F (m−1) ∈ {∅, {0}, {1}}, then h2,F =
calculated by the Gaussian approximation [24] for AWGN, H(U2 |Y[N ] , UF (m−1) ) ≤ H(U2 |Y2 , Y3 ) = Hn−1,0
BPSK, and Eb /N0 = 2 dB. The list size L = 23 suffices for since the received vector (Y2 , Y3 ) corresponds to the
the corresponding precoded polar code to achieve a near ML transmitted (U2 , U3 )G.
performance under SCL, e.g., FER = 10−3 at Eb /N0 = 2 Case m = 3 :
low
dB. The same maxm D̄m = 0.953 is provided by another • if F (m−1) = {0, 1, 2}, then h3,F =
(512, 256) frozen set that requires L > 214 to achieve a near H(U3 |Y[N ] , U0 , U1 , U2 ) = Hn,3 by the definition of
ML performance, e.g., FER = 5 · 10−5 at Eb /N0 = 2 dB. Hn,m .
The existence of (N, K) precoded polar codes with similar • if F (m−1) ∈ {∅, {0}}, then h3,F =
low
maxm D̄m but different complexities of near ML decoding H(U3 |Y[N ] , UF (m−1) ) ≤ H(U3 |Y3 ) = Hn−2,0 since
low
hinders the usage of D̄m as the decoding complexity measure the received Y3 corresponds to the transmitted U3 .
during the frozen set optimization for SCL. • if F
(m−1)
∈ {{2}, {0, 2}, {1, 2}}, then h3,F =
H(U3 |Y[N ] , UF (m−1) ) ≤ H(U3 |Y2 , Y3 , U2 ) = Hn−1,1
low
B. Derivation of D̄m in [8] since the received vector (Y2 , Y3 ) corresponds to the
low transmitted (U2 , U3 )G.
The source of the issues with D̄m follows from its deriva-
low • if F (m−1) ∈ {{1}, {0, 1}}, then h3,F =
tion in [8, Section III-A]. Specifically, the lower bound D̄m
H(U3 |Y[N ] , UF (m−1) ) ≤ H(U3 |Y1 , Y3 , U1 ) = Hn−1,1
on D̄m is obtained for BMS channels by introducing ∆m ,
since the received vector (Y1 , Y3 ) corresponds to the
D̄m − D̄m−1 and showing that ∆m = H(Um |Y[N ] , U[m] )
transmitted (U1 , U3 )G.
when m ∈ A and ∆m = H(Um |Y[N ] , U[m] ) −
H(Um |Y[N ] , UF (m−1) ) ≥ H(Um |Y[N ] , U[m] )−1 when m ∈ F . Example 1 specifies the upper bounds on
low
Thus, the gap between D̄m and its lower bound D̄m is H(Um |Y[N ] , UF (m−1) ) for N = 4. The following lemma
due to replacing H(Um |Y[N ] , UF (m−1) ) by its upper bound defines the upper bound on H(Um |Y[N ] , UF (m−1) ) for any
1 when m ∈ F . At the same time, H(Um |Y[N ] , UF (m−1) ) given N = 2n , m and F (m−1) . Let TI,J be a submatrix of
is lower bounded by H(Um |Y[N ] , U[m] ), which means that T consisting of the elements Ti,j , i ∈ I, j ∈ J.
∆m ≤P0 when m ∈ F and leads to the upper bound
D̄m ≤ m∈A(m) Hn,i [8, Eq. (6b)]. [8, Remark 2] explains Lemma 1. Let sets I, J ⊆ [2n ] and integer n e ≤ n satisfy the
the preferability of the lower bound on D̄m compared to the following conditions:
upper bound by the fact that the upper bound ignores the effect 1) |I| = |J| = 2ne ,
of the frozen bits. 2) (G⊗n )I,J = G⊗en ,
3) (G⊗n )I,J = 0,
tight 4) m ∈ I,
C. Proposed Tightened Lower Bound D̄m
5) I ∩ [m] ⊆ F (m−1) .
We propose to tighten the lower bound on D̄m by tighten- Then
ing the upper bound on H(Um |Y[N ] , UF (m−1) ). Observe that H(Um |Y[2n ] , UF (m−1) ) ≤ Hne,m
e, (3)
H(Um |Y[N ] , UF (m−1) ) is upper bounded by H(Um |YT , UΦ )
for any subsets Φ ⊆ F (m−1) and T ⊆ [N ]. In what follows 2 The bit reversal permutation matrix B can be easily incorporated by
below we show how to identify non-trivial sets Φ and T such permitting elements of Y[N] , i.e., by replacing Y[N] with Y[N] B.
 4

where m e , |I ∩ [m]|, I , [2n ] \ I, and 0 is all-zero and therefore (G⊗n )I,J = G⊗(n−|Q|) = G⊗n and I = ∅. So,
matrix/vector. the statement holds for |Q| = 0. When |Q| = 1, Lemma
3 reduces to Lemma 2. We further proceed by induction.
Proof: For any such I and J, we have
Assume that the statement holds for Q, i.e., (G⊗n )S(Q),S(Q) =
H(Um |Y[2n ] , UF (m−1) ) ≤ H(Um |YJ , UI∩[m] ) due to J ⊆ [2n ]
G⊗(n−|Q|) and (G⊗n )[n]\S(Q),S(Q) = 0. By applying Lemma
and condition 5: I ∩ [m] ⊆ F (m−1) . By substituting the
e[2ne ] , UI and Ye[2ne ] , YJ , we b = n − |Q| and G⊗bn , we obtain that the statement holds
2 to n
random variable vectors U
(a) (b)
for Q ∪ {q} with any q ∈ [n] \ Q.
obtain H(Um |YJ , UI∩[m] ) = H(U em e e ], U
e |Y[2n
e[m]
e ) = Hn e.
e ,m The following theorem summarizes Lemmas 1–3.
Equality (a) holds since U em
e = U m and e
U [m]
e = U I∩[m] due to
Theorem 1. Let set Q ⊂ [n] satisfy mQ = 1 and iQ 6= 1 for
e and condition 4: m ∈ I. Equality (b) holds
the definition of m
all i ∈ A(m−1) . Then
since the received YJ = Ye[2ne ] corresponds to the transmitted
U (G⊗n )[2n ],J = UI (G⊗n )I,J ⊕UI (G⊗n )I,J = U e[2ne ] G⊗en H(Um |Y[2n ] , UF (m−1) ) ≤ Hn−|Q|,|S(Q)∩[m]| .
|{z} | {z } | {z }
e n e
G ⊗n
U[2 e ] 0 Proof: Let us show that such Q defines sets I = J =
due to conditions 1–3. This concludes the proof. S(Q) meeting all conditions of Lemma 1. By Lemma 3, the
The upper bound of Lemma 1 is non-constructive since it sets I = J = S(Q) with n e = n − |Q| satisfy conditions
does not specify how to find the sets I and J. The following 1–3 of Lemma 1. It follows from the restriction mQ = 1
two lemmas define sets I and J satisfying conditions 2–3 of and Eq. (4) that m ∈ S(Q), and therefore condition 4 of
Lemma 1: (G⊗n )I,J = G⊗en and (G⊗n )I,J = 0. Lemma 2 Lemma 1 is satisfied. Due to the restriction iQ 6= 1 for all
considers the case of |I| = |J| = 2n−1 , and then Lemma 3 i ∈ A(m−1) = [m] \ F (m−1) , we have S(Q) ∩ [m] ⊆ F (m−1) ,
generalizes the result for |I| = |J| = 2ne , n e ≤ n. Note that we which means that condition 5 of Lemma 1 is satisfied. There-
employ the binary representation
Pn−1 t (j0 , . . . , jn−1 ) ∈ {0, 1}n of e = n− |Q| and m
fore, by substituting n e = |S(Q)∩[m]| in Eq.
n
the integers j = t=0 jt 2 ∈ [2 ]. (3), we obtain H(Um |Y[2n ] , UF (m−1) ) ≤ Hn−|Q|,|S(Q)∩[m]| .
According to Theorem 1, there always exists at least one
Lemma 2. Given any integer q ∈ [n] and the corresponding
set set Q if m > 0. Specifically, it is easy to see that Q = {t ∈
n o
S(q) , j ∈ [2n ] | jq = 1 , [n] | mt = 1} satisfies the condition mQ = 1, as well as iQ 6=
1 for all i ∈ A(m−1) since iQ = 1 may be true only for i ≥ m.
where jq is the q-th bit in the binary expansion of the integer In this case, |Q| = wt(m) and |S(Q) ∩ [m]| = 0, leading to a
j. Then sets I = J = S(q) satisfy the conditions (G⊗n )I,J = simple upper bound H(Um |Y[2n ] , UF (m−1) ) ≤ Hn−wt(m),0 .
G⊗(n−1) and (G⊗n )I,J = 0. Note that there could exist several sets Q satisfying con-
Pn−1 t ditions of Theorem 1. It is desirable to find set Q that
Proof: As shown in [25], the j = t=0 jt 2 -th row provides the tightest upper bound H(Um |Y[2n ] , UF (m−1) ) ≤
⊗n
of ( 10 11 ) can be represented as the evaluation of poly- Hn−|Q|,|S(Q)∩[m]|. This requires to solve the following opti-
jn−1
nomial f (j, x) , xj00 xj11 · · · xn−1 over 2n elements x , mization problem:
Pn−1 nt n ⊗n
t=0 xt 2 ∈ [2 ]. Since the j-th column of G is equal to
⊗n Q∗ = min Hn−|Q|,|S(Q)∩[m]| , (5)
the transposed j-th row of ( 10 11 ) , it has the same polynomial Q∈Q
representation. Thus, columns of G⊗n with the indices j ∈
J = S(q) correspond to polynomials xj00 · · · xqq · · · xn−1
j =1 jn−1
. Q , {Q ⊂ [n] | mQ = 1, ∀i ∈ A(m−1) iQ 6= 1}. (6)
For all x ∈ I = [2n ] \ S(q), the multiplier xq = 0 due to the The number of sets Q to consider is upper bounded by 2n ,
definition of S(q). Consequently, we have fn (j, x) = 0 for i.e., by the code length N = 2n . Note that the condition
all x ∈ I, j ∈ J. Therefore, the condition (G⊗n )I,J = 0 is mQ = 1 reduces this number to 2wt(m) , where wt(m) is
satisfied. For all x ∈ I = S(q), the multiplier xq = 1 and the Hamming weight of the binary expansion of m. Since
jq−1 jq+1 jn−1
consequently fn (j, x)/xq = xj00 . . . xq−1 xq+1 . . . xn−1 = the cardinality of the set A(m−1) is upper bounded by m,
fn−1 (b j, xb), where the integers b j and x b are defined by we conclude that the time complexity of finding Q∗ scales
their binary expansions (j0 , . . . , jq−1 , jq+1 , . . . , jn−1 ) and as O(2wt(m) · m), assuming that the bit-channel entropies are
(x0 , . . . , xq−1 , xq+1 , . . . , xn−1 ), respectively. The evaluations pre-computed. Note that wt(m) ≤ n and m < 2n .
of polynomials fn−1 (b j, x b) over elements x b ∈ [2n−1 ] for The following lemma simplifies the search for Q∗ by show-
b
j ∈ [2 n−1
] give the matrix G ⊗(n−1)
. Therefore, the condition ing that Hn−|Q|,|S(Q)∩[m]| cannot be decreased by including
(G⊗n )I,J = G⊗(n−1) is satisfied. additional elements into Q.
Lemma 3. Given any set Q ⊂ [n] and the corresponding Lemma 4. For any sets Q ⊂ Q′ ⊂ [n],
n o
S(Q) , j ∈ [2n ] | jQ = 1 , (4) Hn−|Q|,|S(Q)∩[m]| ≤ Hn−|Q′ |,|S(Q′ )∩[m]| .
Proof: Using the notation of Lemma 1 and Theorem 1,
where 1 , (1, . . . , 1), and jQ = 1 means that jq = 1 for
n
e = n−|Q| and m e = |S(Q)∩[m]|. Let us denote δ , |Q′ \Q|
all q ∈ Q. Then sets I = J = S(Q) satisfy the conditions ′ ′
e , |S(Q ) ∩ [m]|. The entropy Hne,m
and m e corresponds to
(G⊗n )I,J = G⊗(n−|Q|) and (G⊗n )I,J = 0.
the m-th
e bit-channel of the polarization transformation G⊗en ,
Proof: When |Q| = 0, we have I = J = S(Q) = [2n ] denoted by Wne,me , while the entropy Hn e ′ characterizes the
e ′ ,m
 5

me ′ -th bit-channel of the polarization transformation G⊗(en−δ) , puted as follows

denoted by Wne−δ,m e ′ . It follows from Eq. (4) that the binary
 tight

D̄m−1 + Hn,m , m ∈ A,
expansion of m e ′ can be obtained from the binary expansion tight tight
of m e by deleting δ bits equal to 1, whose indices are defined D̄m = max(0, D̄m−1 −


by Q′ \ Q. Therefore, by deleting the polarization layers with (Hn−|Q(m)|,|S(Q(m))∩[m]| − Hn,m )), m ∈ F ,
these indices from the polarization transformation G⊗en , the (8)
tight
bit-channel Wne,m e can be transformed into Wn e ′ . Since the
e −δ,m where m ∈ [2n ], D̄−1 = 0, and Q(m) means the set Q
deleted bits are all equal to 1, Wne−δ,m e ′ has lower symmetric calculated for a particular m by Algorithm 1 or Eq. (7). We
capacity and higher entropy than Wne,m e as follows from [1, used Algorithm 1 to produce numerical results for Section IV.
tight
Section III]. The complexity of computing D̄m in (8) is dominated
By Lemma 4, it suffices to explore only a subset of Q to by Hn−|Q(m)|,|S(Q(m))∩[m]|. Therefore, the overall time com-
find Q∗ . Specifically, plexity of computing D̄0tight , . . . , D̄2tight
n −1 using Algorithm 1
P2n −1
scales as O( m=0 wt(m)2 · m), which is upper bounded by
Q∗ = min Hn−|Q|,|S(Q)∩[m]|, (7)
b
Q∈Q O(n2 · 22n ) = O(log(N )2 · N 2 ).
b , {Q ∈ Q | ∀q ∈ Q
Q Q \ {q} ∈
/ Q}.
apx
Thus, set Q∗ should have a low cardinality compared to the D. Proposed Approximate Bound D̄m that Combines the
other Q ∈ Q. We further propose a low-complexity greedy Tightened Lower Bound and Upper Bound
approach aiming to find Q ∈ Q with the lowest cardinality.
Algorithm 1 specifies the proposed approach in which we Reference [8] proposed several frozen sets without the first
initialize set Q by the empty set and then add an element bit-channel, i.e., the less reliable bit-channel is used to transfer
q ∗ to Q at each iteration of the while loop until the condition information bits in [8]. To the best of the authors’ knowledge,
Q ∈ Q is satisfied. Each q ∗ is calculated at line 7 as the such frozen sets have not been used before. This motivated us
element q ∈ M \ Q minimizing the number of i ∈ A(m−1) to investigate why the first bit-channel is not frozen in [8].
violating the condition in Eq. (6), where q belongs to the set Observe that [8, Remark 2] suggests the frozen set design
low
M , defined at line 3, to ensure that Q satisfies to the condition criterion log2 (L) ≥ D̄m and the usage of D̄m as a proxy
mQ = 1 in Eq. (6). Note that the set Λ consists of i ∈ A(m−1) for D̄m , where L is the target SCL list size. Therefore,
low ∗
violating the condition in Eq. (6) for the current set Q. The we explore the gap between D̄m∗ and D̄m ∗ , where m ,
low
number of iterations in the while loop is upper bounded by arg maxm∈[2n ] D̄m . We further provide expressions in terms
low
|M | since |M | iterations result in Q = M and such Q satisfies of the information set A = [2n ]\F . Note that D̄m = D̄m =0
∀i ∈ A(m−1) iQ 6= 1 as explained right below the proof of holds for 0 ≤ m < min(A) and all information sets A.
Theorem 1. We focus on the information sets intended for moderate-to-
large L, in which case D̄m low
> 0 for min(A) ≤ m ≤ m∗ .
The corresponding examples can be found in [8, Figs. 1
Algorithm 1: Greedy approach to optimize set Q and 3]. ForPsuch information sets, the gap g(A) , D̄m −
∗
low
1 ConstructSetQ(n, m, A(m−1)) D̄m ∗ = m∈{min(A),...,m }\A
∗ (1 − H(U m |Y [N ] , U F (m−1) ))

2 begin as follows from Eq. (2). Thus, the gap g(A) substantially
3 M ← {t ∈ [n] | mt = 1} depends on the index of the lowest information bit min(A).
4 Q←∅ The lower min(A), the higher g(A). In particular, when
5 Λ ← A(m−1) min(A) = 0 as for the information sets proposed in [8]3 , the
6 while |Λ| > 0 do gap g(A)Pis especially large due to the maximized number of
7 q ∗ ← arg min |{j ∈ Λ | jq = 1}| terms in m∈{min(A)=0,...,m∗ }\A (1−H(Um|Y[N ] , UF (m−1) )).
low
q∈M\Q That is, the lower bound D̄m ∗ especially underestimates
8 Q ← Q ∪ {q ∗ } D̄m∗ when min(A) is close to zero. This means that the
9 Λ ← {i ∈ Λ | iq∗ = 1} code design criterion maxm∈[2n ] D̄m low
gives preference to
10 end information sets with a very low min(A). That is why the
11 return Q information sets constructed in [8] have min(A) = 0. The
12 end replacement of D̄m low
by our tightened lower bound D̄m tight

P the problem by reducing the gap D̄m −

partially solves ∗
tight
D̄m ∗ = m∈{min(A),...,m }\A∗ (H n−|Q(m)|,|S(Q(m))∩[m]| −
low
Although Algorithm 1 does not guarantee optimality, the H(Um |Y[N ] , UF (m−1) )) ≤ D̄m∗ − D̄m ∗.

resulting set Q is typically equal to Q∗ . The worst-case To eliminate the bias towards the information sets A having
time complexity of Algorithm 1 is O(wt(m)2 · m) since the low min(A), we propose to combine
tight up
P our tight lower bound
maximum number of the while loop iterations is |M | = wt(m) D̄m with the upper bound D̄m , m∈A(m) Hn,m from [8,
and the complexity of each iteration is dominated by line 7,
whose complexity is upper bounded by |M | · m. 3 In [8], the enumeration starts from 1, and therefore the lowest information
The proposed tightened lower bound on D̄m can be com- bit index is equal to 1.
 6

Eq. (6b)] as follows modify the genetic algorithm so that it solves the constrained
 minimization problem: minimize PeML subject to the constraint
apx

 D̄m−1 + Hn,m , m ∈ A,

 D̄apx ≤ TD , where the threshold TD is an input parameter of

 apx
m ∈ F ∩ [λ],
D̄m−1 , the genetic algorithm. To ensure that the frozen set population
apx apx
D̄m = max(0, D̄m−1 − satisfies this constraint, we discard frozen sets violating this


 (Hn−|Q(m)|,|S(Q(m))∩[m]|−
 constraint from the initial population and from the crossover


 H )), m ∈ F \ [λ], output. Besides, we allow the mutation operation to swap a
n,m
(9) frozen bit and a non-frozen bit only when this does not lead
where m ∈ [2n ], D̄−1 apx
= 0, and λ is an integer threshold. to the constraint violation. The resulting algorithm is referred
According to (9), D̄m apx
= D̄m up
for m ∈ [λ]. Thus, for to as GenAlgT.
the frozen bits with low indices m ∈ F ∩ [λ], we use Genetic algorithms are known to be suboptimal [27], i.e,
apx apx up up converge prematurely to local optima since genes of high-
D̄m − D̄m−1 = D̄m − D̄m−1 = 0. For the remaining frozen
bits m ∈ F \ [λ], we employ D̄m apx apx
− D̄m−1 = D̄m tight tight
− D̄m−1 . rated individuals (frozen sets) typically dominate the popula-
Note that all considered bounds process the information tion. According to our experimental results, GenAlgT returns
bits m ∈ A in the same way: D̄m apx
− D̄m−1apx
= different outputs when run multiple times. To leverage this
low low tight tight up
D̄m − D̄m−1 = D̄m − D̄m−1 = D̄m − D̄m−1 = Hn,m . up issue, we run GenAlgT algorithm ρ times for each TD . We use
Obviously, D̄m low
≤ D̄m tight
≤ D̄m apx
≤ D̄m up
. If ρ = 5 and consider various values of TD with the granularity
min(A) ≥ λ, then D̄m apx
= D̄m tight
for all m. It 0.1.
apx 2) Search Space Reduction. S-Constraint and GenAlgTS:
can be seen that the gap between D̄m∗ and D̄m ∗

depends on λ instead of min(A), since D̄ m ∗ − D̄ ∗

apx
= Given the code length N = 2n and dimension K, the search
P m
space is given by all frozen sets F ⊂ [N ] of cardinality N −
Pm∈[λ]\A (H(Um |Y[N ] , U[m] ) − H(Um |Y[N ] , UF (m−1) )) + K. The number
m∈{λ,...,m∗ }\A (Hn−|Q(m)|,|S(Q(m))∩[m]| −
of such frozen sets is equal to the binomial
H(Um |Y[N ] , UF (m−1) )). coefficient N K , which grows rapidly with N and min(K, N −
It is very important that the value of λ is the same for K). We resolve this issue by introducing our constraints on
all information sets being compared during the code design the frozen set structure and incorporating them into the genetic
process to enable a fair comparison. We recommend setting λ algorithm.
close to min(A′ ), where A′ is the information set consisting Let r be the reliability sequence consisting of the bit-
of the most reliable bit-channels. For example, min(A′ ) is channel indices arranged in ascending order of their reliabil-
equal to 30 and 95 for the code parameters (128, 64) and ities. We construct r by using the Gaussian approximation
(512, 256), respectively. We use λ = 25 = 32 for (128, 64) [24] for the AWGN channel. Alternatively, the 5G reliability
and λ = 3 · 25 = 96 for (512, 256). Note that the information sequence [2] might be used. It has been shown in [28] that
sets A designed for SCL decoding with L ≥ 2 typically have both bit-channel reliabilities and index weights are of great
lower min(A). For example, (128, 64) and (512, 256) Reed- importance when designing polar codes for SCL decoding.
Muller codes have min(A) equal to 15 and 31, respectively. Following this direction, we characterize the closeness of a
given frozen set F to the reliability-based frozen set by the
E. Frozen Set Optimization first information bit indices αF v of all weights v
(
In this section, we consider the frozen set optimization prob- cv , 0 ≤ v < lF ,
lem with two objectives: minimize the decoding complexity αF
v ,
apx
/ F }, lF ≤ v ≤ n,
min{q ∈ [cv ] | τv,q ∈
characterized by D̄apx , maxm∈[2n ] D̄m , defined by Eq.
(9), and minimize the ML decoding error probability estimate lF , min wt(i), (10)
i∈[N ]\F
PeML , computed as in Section II-B. Since these two objectives
F
are conflicting, we are interested in constructing frozen sets where the binomial coefficient cv , nv , 2l is the minimum
leading to codes with various complexity-performance trade- Hamming distance of a pure polar code with the frozen set
offs. The best tradeoffs are provided by the Pareto front, which F , and the sequence τv consists of all bit-channel indices
is the set of all non-dominated (D̄apx , PeML ), i.e., the Pareto of weight v arranged in ascending order of their reliabilities.
front consists of pairs (D̄apx , PeML ) such that all other pairs That is, τv is the subsequence of the reliability sequence r
′
(D̄apx , PeML
′ ′
) satisfy D̄apx > D̄apx or PeML
′
> PeML , where the consisting of ri with wt(ri ) = v, 0 ≤ i < N . Using τv , the
computations are performed for fixed code length and rate. reliability-based frozen set RS of cardinality N − K − S
1) Optimization using the Genetic Algorithm GenAlgT: The
RS , {ri | i ∈ [N − K − S]}
computational complexity of finding the exact Pareto front
is huge since there are plenty of frozen sets to consider. can be represented as {τv,q | q ∈ [αR v ], 0 ≤ v ≤ n}. Let
S

Therefore, we find an approximate Pareto front using a vari- ℓ , lR0 . Since the bit-channels {τv,q | q ∈ [cv ] \ [αR
v ], ℓ +
0

ation of the genetic algorithm [9] with the hash table [26] to 2 ≤ v ≤ n} have high-weight indices and high reliabilities,
reduce time complexity and with the elimination of identical they are unlikely to generate codewords producing errors under
candidates from the population to preserve diversity. Since the SCL decoding. Therefore, the following constraint suggests
genetic algorithm [9] has only one objective of minimizing that these bit-channels are always non-frozen, while the least
the decoding FER/BER, we need to adjust it. Specifically, we reliable bit-channels are always frozen.
 7

20
(128,64) frozen sets, GenAlgTS
(512,256) frozen sets, GenAlgTS
Pareto front of (128,64) frozen sets, GenAlgTS
Pareto front of (512,256) frozen sets, GenAlgTS

15

10

∆
5

Fig. 1. S-constrained frozen set structure

350 0
Used S 1 1.5 2
Lowest S, (128,64) frozen sets, GenAlgT −2.5 3 3.5 4
Dapx
Lowest S, (512,256) frozen sets, GenAlgT
300 Lowest S, (128,64) frozen sets, GenAlgTS
Lowest S, (512,256) frozen sets, GenAlgTS
Lowest S, Pareto front of (128,64) frozen sets, GenAlgT Fig. 3. ∆ of the frozen sets generated by the genetic algorithm
250 Lowest S, Pareto front of (512,256) frozen sets, GenAlgT
Lowest S, Pareto front of (128,64) frozen sets, GenAlgTS
Lowest S, Pareto front of (512,256) frozen sets, GenAlgTS
200
S

F for various D̄apx , i.e., F belonging to the Pareto front as

150
defined at the beginning of Section III-E, are connected by
100
lines. It can be seen that the Pareto front is characterized by a
lower S than the average S. For (N, K) = (512, 256), almost
50 the whole Pareto front is located below S = 160, except for
one outlying frozen set. Note that we are interested in the
0 general trend and can ignore the outliers, since there exist
1 1.5 2 −2.5 3 3.5 4
Dapx plenty of frozen sets with almost identical (D̄apx , PeML ), and
it is very likely that one of them would have low S. Therefore,
Fig. 2. The lowest sufficient S for the frozen sets generated by the genetic we set S = 160. The S-constraint integration into the genetic
algorithms
algorithm GenAlgT is straightforward. The resulting algorithm
is referred to as the GenAlgTS. For (N, K) = (512, 256),
inf
we have ℓ =
4, N = 129, NSfr = 138, and therefore
Definition 1 (S-constraint). Given an integer parameter S, a 245
ΩS = 127 . For (N, K) = (128, 64), we have
ℓ = 3,
frozen set F satisfies the S-constraint iff
 NSfr = 29, N inf = 29, and consequently

ΩS = 70 35 . In both
F N

αv = cv , 0 ≤ v < ℓ, cases, ΩS is much lower than K . This complexity reduction
F
αv ≥ αv ,RS
ℓ ≤ v ≤ n, is achieved without reducing the Pareto front quality, as shown

 in Section IV-A. It can be seen from Fig. 2 that the lowest
∀q ≥ αR v
0
τv,q ∈
/ F , ℓ + 2 ≤ v ≤ n.
sufficient S of the GenAlgTS Pareto front gradually increases
with D̄apx , and it is actually much less than 160 for moderate
Fig. 1 illustrates the S-constraint on the frozen set structure. D̄apx .
The blue colour indicates
Pℓ−1 the permanently
Pn frozen bits, whose 3) Search Space Reduction. B-Constraint and GenAlgTB:
number is NSfr , ( v=0 cv ) + ( v=ℓ αR v ). The red colour
S
The gradual increase of S is due to the growing discrepancy
indicatesP the permanently non-frozen bits, whose number is between our frozen sets F and the reliability-based frozen
n
N inf , v=ℓ+2 (cv − αR fr
v ). These NS + N
0 inf
bits may be set R0 . This growing discrepancy can be characterized not
eliminated from consideration during the frozen set optimiza- only by the lowest sufficient S but also by the number ∆ of
tion. Thus, the frozen set optimization under the S-constraint frozen bit-channels with the highest indices of weight ℓ. The
reduces to the optimization over the green area consisting of importance of ∆ for increasing the minimum distance from
N − NSfr − N inf bits. The number of (N, K) S-constrained 2ℓ to 1.5 · 2ℓ in precoded polar codes being subcodes of Reed-
fr inf

frozen sets is ΩS , N −N S −N
K−N inf
. Obviously, ΩS increases Muller codes with the minimum distance 2ℓ has been proven
with S from minS (ΩS ) = Ω0 to maxS (ΩS ) = ΩN −K . Since in [23]. Note that any precoded polar code being a subcode
the search space size rapidly grows with S, it is desirable to of the Reed-Muller code with the minimum distance 2ℓ must
identify the lowest S that preserves the best frozen sets. Such have αFv = cv for 0 ≤ v < ℓ, which is one of the S-constraint
S can be found experimentally4. In Fig. 2, the yellow/green requirements. Fig. 3 shows that ∆ of the S-constrained frozen
pluses indicate the lowest S such that ∀v αF v ≥ αv
RS
for each sets increases with D̄apx .
individual frozen set F generated by GenAlgT. The best sets
Motivated by the results on ∆, we propose to incorporate
4 We provide results for the bit-channel reliabilities computed using the
similar characteristics for the two lowest information index
Gaussian approximation [24] for AWGN, BPSK, Eb /N0 = 3.5 dB when weights into the frozen set structure. Specifically, we represent
(N, K) = (128, 64) and Eb /N0 = 2.75 dB when (N, K) = (512, 256). ∆ of a given frozen set F as ∆ = cℓ − 1 − χF ℓ and propose
 8

By imposing limits on αF v as in Definition 1, we obtain the

following upper bound on the number of F satisfying both
B-constraint and S-constraint with lF = ℓ
"ℓ+1 #" n #
Y Y
RS Bv R0 RS
(cv − αv − Bv )2 (αv − αv + 1) cℓ .
v=ℓ v=ℓ+2

This number is minimized when Bv = 0 and maximized when

Bv = cv − αR v
S
− 1.
Observe that in contrast with the S-constraint, the B-
constraint cannot be integrated into GenAlgT by simply
eliminating a subset of bit-channels from consideration. We
integrate the B-constraint into GenAlgT by eliminating the
crossover and adjusting the mutation as follows. To preserve
Fig. 4. B-constrained frozen set structure the B-constrained structure of a frozen set F , we allow the
mutation operation to perform only the following actions:
(i) increment χF v , l
F
≤ v ≤ lF + 1, (ii) decrement βvF ,
to optimize χF F
lF and χlF +1 , where
F
l + 2 ≤ v ≤ n, or (iii) modify bits within the flexible region
of size BlF + BlF +1 . Action (i) is due to the initial population
χF
v , max {zv,q ∈
/ F }, lF ≤ v ≤ n, consisting of the reliability-based frozen sets and frozen sets
q∈[cv ]
interpolating between the Reed-Muller and reliability-based
and zv is the subsequence of (0, 1, . . . , N − 1) consisting of frozen sets that have χF F
v = 0. Thus, χv gradually increases
all elements of weight v. Obviously, the length of zv is cv . with the increasing number of genetic algorithm iterations.
The value of χF F
lF and χlF +1 are expected to decrease with Action (ii) is since αF F
v and βv are expected to decrease with
increasing D̄apx due to their connection with ∆. Besides, we F
the increasing χv . Action (iii) makes use of the flexibility
propose to assume that highly reliable bit-channels having v- allowed by B. These actions are implemented as a random
weight indices less than χF F F
v are non-frozen, v ∈ {l , l + 1}. swap of a frozen bit and a non-frozen bit from the set
This assumption reduces the number of bit-channels allowed {τv,min(β F ,βvR0 )+Xv −i , χF
F
| i ∈ [Bv ]}lv=l+1 n
v F ∪ {τv,βv
F}
v=lF +2 ,
to be arbitrarily frozen or non-frozen. The resulting frozen set v
where Xv is an integer parameter, and τv,q with q ∈ / [cv ] are
structure is formalized as the B-constraint, where the flexibility
skipped. It is easy to see that the cardinality of this set is at
of frozen set F is controlled by integers BlF and BlF +1 .
most N flex , n−lF +1+BlF +BlF +1 , which is significantly
Definition 2 (B-constraint). Given an integer vector B, a lower than N − NSfr − N inf in GenAlgTS and N in [9]. Thus,
frozen set F satisfies the B-constraint iff only a small portion of bits are allowed to mutate at each
( iteration of the genetic algorithm. Although the search space
βvF < αF F F
v + Bv , l ≤ v ≤ l + 1, size can be minimized by using BlF = BlF +1 = 0, such a
F F F
βv = αv − 1, l + 2 ≤ v ≤ n, choice leads to a rigid frozen set structure and may eliminate
many good frozen sets from consideration. Therefore, the
where
values of BlF and BlF +1 are selected to balance the frozen
βvF , max{q ∈ [cv ] | τv,q ∈ F , τv,q < χF
v }, lF ≤ v ≤ n. set flexibility and its design complexity. We set XlF = 8,
XlF +1 = 6, BlF = 37, and BlF +1 = 8 for the code
The B-constraint on the frozen set structure is illustrated by parameters (128, 64) and BlF +1 = 23 for (512, 256). This
Fig. 4, where the upper index “F ” is omitted for simplicity. defines the number of bits allowed to mutate N flex = 50 for
Note that the dark and light blue frozen bits are arranged in the code parameters (128, 64) and N flex = 66 for (512, 256).
ascending order of their indices and reliabilities, respectively. The crossover operation does not preserve the B-constrained
That is, the v-weight dark and light blue frozen bits are frozen sets, and therefore, we replace the crossover by addi-
arranged as in the sequences zv and τv , respectively. The rest tional mutations. Since the truncated population size in [9] is
of the notation is as in Fig. 1. TPOP = 5, the crossover generates TPOP (TPOP − 1)/2 = 10
Remark 1. The B-constrained frozen set structure generalizes frozen sets. To preserve the maximum population size 20, we
our frozen set structure [23]. Specifically, the triplet-tuned apply 2 additional mutations to each frozen set from the trun-
frozen sets from [23] can be represented as special cases of cated population instead of the crossover since this generates
the B-constrained frozen sets with BlF = BlF +1 = 0 and 2TPOP = 10 frozen sets. The resulting B-constrained genetic
χF F algorithm is referred to as GenAlgTB.
v = zv,cv −1 for v = l + 1.

The number of B-constrained frozen sets F with fixed lF

is upper bounded by IV. N UMERICAL R ESULTS
F  
+1
lY Yn In this section, we evaluate the proposed frozen set design
 (cv − Bv )2Bv   c v  c lF . method and provide a comparison with the state-of-the-art for
v=lF v=lF +2 the AWGN channel with BPSK modulation.
 9

Pareto front of (128,64) frozen sets, GenAlgT Pareto front of (128,64) frozen sets, GenAlgT
Pareto front of (512,256) frozen sets, GenAlgT Pareto front of (512,256) frozen sets, GenAlgT
Pareto front of (128,64) frozen sets, GenAlgTS Pareto front of (128,64) frozen sets, GenAlgTS
Pareto front of (512,256) frozen sets, GenAlgTS 104 Pareto front of (512,256) frozen sets, GenAlgTS
Pareto front of (128,64) frozen sets, GenAlgTB Pareto front of (128,64) frozen sets, GenAlgTB
10−3 Pareto front of (512,256) frozen sets, GenAlgTB Pareto front of (512,256) frozen sets, GenAlgTB
(128,64) frozen sets, GenAlgT

Genetic algorithm iterations

(128,64) frozen sets, GenAlgT
(512,256) frozen sets, GenAlgT (512,256) frozen sets, GenAlgT
(128,64) frozen sets, GenAlgTS
(128,64) frozen sets, GenAlgTS (512,256) frozen sets, GenAlgTS
(512,256) frozen sets, GenAlgTS (128,64) frozen sets, GenAlgTB
(128,64) frozen sets, GenAlgTB 103 (512,256) frozen sets, GenAlgTB
PML

F1 (512,256) frozen sets, GenAlgTB

~

10−4
F2

102
Φ1
F3
F4

−5 Φ2
10 Φ3
Φ4
101
1 2 3 − 4 5 6 1 2 3 − 4 5 6
Dapx Dapx

Fig. 5. Frozen sets generated by the genetic algorithms Fig. 6. The number of iterations in GenAlgT, GenAlgTS and GenAlgTB

A. Frozen Set Design GenAlgT, GenAlgTS and GenAlgTB perform 1617, 1220 and
98 iterations on average for (512, 256), respectively. Note that
Fig. 5 characterizes the frozen sets generated by the pro- GenAlgT and GenAlgTS terminate if no improvement has
posed GenAlgT, GenAlgTS and GenAlgTB for (N, K) ∈ been observed for the last 50 and 200 iterations for the parame-
{(128, 64), (512, 256)}, TD ∈ {1.0, 1.1, 1.2, . . . } and ρ = 5. ters (128, 64) and (512, 256), respectively. In GenAlgTB, the
The pairs (D̄apx , PeML ) found by GenAlgT, GenAlgTS and numbers of such last iterations are 20 and 30 for (128, 64)
GenAlgTB are marked as “+”, “◦” and “△”, respectively, and (512, 256), respectively. So, GenAlgT, GenAlgTS and
where PeML is computed using the low-complexity union GenAlgTB found the resulting frozen sets in 189, 67 and
bound for intermediate iterations and the tight TSB bound for 22 iterations on average for (128, 64), respectively. GenAlgT,
the final output at Eb /N0 = 3.5 dB for the code parameters GenAlgTS and GenAlgTB found the resulting frozen sets
(128, 64) and Eb /N0 = 2.0 dB for (512, 256). It can be in 1417, 1020 and 68 iterations on average for (512, 256),
seen that GenAlgT, GenAlgTS and GenAlgTB provide similar respectively. Thus, GenAlgTB requires much less iterations
Pareto fronts, indicating that the search space reduction of than GenAlgT.
GenAlgTS does not deteriorate the frozen set performance. The execution of GenAlgTB required 0.2 and 10 seconds
Moreover, the outputs of GenAlgTS and GenAlgTB are con- on average for the parameters (128, 64) and (512, 256), re-
centrated closer to the Pareto front than that of GenAlgT. spectively, whereas the resulting frozen sets were found after
This is because GenAlgTS and GenAlgTB have fewer lo- 0.1 and 6 seconds on average for (128, 64) and (512, 256),
cal optima than GenAlgT due to the reduced search space respectively. The implementation is non-parallel and executed
size. As a result, GenAlgTS and GenAlgTB need a lower on a computer with i7 3.2GHz processor. Note that the
ρ to reach saturation than GenAlgT, where the saturation is complexity is independent of the design Eb /N0 , since the code
achieved if an increase in ρ does not provide any reduction performance is evaluated via theoretical bounds.
of mini∈[ρ] PeML,i , where PeML,i is the i-th run output of
GenAlgT/GenAlgTS/GenAlgTB. That is why the Pareto front B. Performance of Precoded Polar Codes
of GenAlgTS/GenAlgTB is slightly better on average than that In Section IV-A, we evaluated the proposed frozen set
of GenAlgT in the case of parameters (512, 256). In the case design. The produced frozen sets are further integrated with
of (128, 64), the Pareto front of GenAlgT is slightly better on the frozen bit expressions to yield the proposed precoded polar
average than that of GenAlgTS/GenAlgTB, since for short- codes. In this section, we compare the FER performances of
the proposed codes and the state-of-the-art codes. The codes
length codes, the search space of GenAlgT is small enough are labelled as follows:
to find near-optimal solutions. Note that the computational
complexities of GenAlgT and GenAlgTS grow rapidly with • Proposed F_ and Proposed Φ _ – precoded polar codes with
the code length N , while the complexity of GenAlgTB grows the proposed frozen sets from the Pareto front of Fig. 5 and
slowly with N , as follows from the description in Section frozen bit expressions from Section II-D.
III-E. Besides, it can be seen from Fig. 5 that the Pareto front • 5G polar CRC-11 – 5G polar codes with CRC-11 [2].
• eBCH subcode d=_ – eBCH polar subcodes [3] with the
of (128, 64) frozen sets has a more stepwise character than minimum distance d.
that of (512, 256) frozen sets. This implies that the Pareto • Code-0, Code-1 and Code-2 – (128, 64) code from [8, Fig.
front becomes smoother with increasing code length N . 2], (512, 256) Code-1 and Code-2 from [8, Figs. 4 and 6],
The computational complexity of genetic algorithms is respectively.
• PAC-RM – (128, 64) PAC code with the Reed-Muller frozen
often characterized by the number of iterations. According set [5].
to Fig. 6, GenAlgT, GenAlgTS and GenAlgTB perform 239, • Systematic PAC – (128, 64) systematic PAC code generated by
117 and 42 iterations on average for (128, 64), respectively. the genetic algorithm in which the minimum-weight codewords
 10

5G polar CRC−11, SCL L=32 design with the optimization criteria derived for average frozen
10−1 eBCH subcode d=24, SCL L=32
Code−2, SCL L=32
Number−polar s=(10,3,1) GA, SCL L=32 bit expressions. The problem of the frozen bit expression
Proposed F1, SCL L=32
Code−2, SCL L=128
Proposed F1, SCL L=128
optimization for a given frozen set is left for future work.
Code−1, SCL L=1024
10−2 Number−polar s=(14,2,0) GA, SCL L=1024
Proposed F2, SCL L=1024
Besides the excellent FER performance, the advantages of
Number−polar s=(14,2,0) GA, SQ L=16384
Proposed F3, SQ L=16384 the proposed frozen set design over the main competitors
Proposed F4, SQ L=16384
Normal approximation bound [8], [23], [5] and [29] are as follows. Our proposed design
FER

10−3 method is fully specified, providing a clear frozen set design

procedure. In contrast, [8] offers four exemplary frozen sets
but lacks a general frozen set design procedure. We use
10−4 deterministic frozen bit expressions, specified in Section II-D,
whereas [8] uses randomized frozen bit expressions. The
proposed B-constrained and S-constrained frozen set structures
10−5 are more flexible than the triplet-tuned frozen sets from [23].
1.4 1.6 1.8 2 2.2 2.4 2.6
Eb/N0, dB
This flexibility offers additional opportunities for optimization
at the expense of the increased number of evaluated frozen
Fig. 7. The performance comparison of (512, 256) codes sets. Fast frozen set evaluation is enabled by the use of
theoretical bounds instead of the decoding-based frozen set
5G polar CRC−11, SCL L=32
PAC−RM, SCL L=32
evaluation [23]. Thus, the computational complexity of the
Code−0, SCL L=32
eBCH subcode d=24, SCL L=32 proposed frozen set design is low as shown in Section IV-A.
RecursCode, SCL L=32
Proposed Φ1, SCL L=32 Note that [5] suggested only a single (128, 64) PAC-RM
10−3
code. Although the problem of designing PAC codes with
arbitrary parameters has been solved in [29] by using a genetic
algorithm, it involves the weight distribution computation via
FER

decoding at each iteration of the genetic algorithm, leading to

10−4
a large design complexity. Since the weight distribution is used
5G polar CRC−11, SCL L=128
PAC−RM, SCL L=128
as the optimization objective in [29], the corresponding codes
Code−0, SCL L=128
Proposed Φ2, SCL L=128 perform well only under Fano decoding or SCL with huge
PAC−RM, SCL L=256
Proposed Φ3, SCL L=256
Systematic PAC, Fano
L. In contrast, the proposed low-complexity frozen set design
10−5 Systematic PAC, SCL L=1024
method is suitable for various L and various code parameters.
Proposed Φ4, SCL L=1024
Normal approximation bound

2.6 2.8 3 3.2 3.4

Eb/N0, dB

V. C ONCLUSION
Fig. 8. The performance comparison of (128, 64) codes

In this paper, we proposed a new low-complexity frozen

are computed at each iteration [29, Fig. 5a].
set design for precoded polar codes with near-uniformly dis-
• RecursCode – (128, 64) precoded polar code, which is ob- tributed frozen bit expressions. The frozen set design criteria
tained by recursively optimizing the weight distribution of a are given by analytical bounds on the FER performance and
subcode of the Plotkin sum of shorter codes [20, Fig. 4d]. SCL complexity, where the proposed SCL complexity criterion
• Number-polar s=(_,_,_) GA – (512, 256) precoded polar codes is based on the recently published complexity analysis of SCL
specified by the triplets s and the Gaussian approximation-based
reliability sequence [23, Fig. 7a].
with near ML performance. These criteria define a frozen set
• Normal approximation bound [30]. optimization problem, whose solutions can be efficiently found
by the genetic algorithm. To reduce the search space size, we
We use the well-known SCL decoder [16] for moderate list
imposed constraints on the frozen set structure such that the
sizes L. When L is large, we employ the sequential (SQ)
number of the genetic algorithm iterations has been reduced
decoder [31], [32]. Note that SQ is a variation of SCL with
by 5 and 17 times for the code parameters (128, 64) and
a similar FER performance and time complexity approaching
(512, 256), respectively. The constructed precoded polar codes
O(N log(N )) in the high-SNR region [32], while the time
of length 512 have a superior FER performance compared
complexity of SCL scales as O(LN log(N )).
to the state-of-the-art codes under SCL-based decoding with
Fig. 7 shows the FER performance of (512, 256) pre-
various list sizes.
coded polar codes under SCL/SQ with the list sizes L =
32, 128, 1024, 16384. It can be seen that the proposed codes
outperform the state-of-the-art codes for all considered L.
In Fig. 8, the proposed (128, 64) codes exhibit similar per- ACKNOWLEDGMENT
formance to the state-of-the-art codes under SCL with the
list sizes L = 32, 128, 256, 1024. This is because (i) the The authors would like to thank Dr. Thibaud Tonnellier for
state-of-the-art (128, 64) frozen sets are near-optimal due to providing the frozen sets of systematic PAC codes [29] and Dr.
a moderate search space for short code lengths, and (ii) in Mustafa Cemil Coşkun for clarifying the bit-channel entropy
this paper, we have proposed a new low-complexity frozen set computation [8].
 11

R EFERENCES [20] V. Miloslavskaya, B. Vucetic, Y. Li, G. Park, and O.-S. Park, “Recursive
design of precoded polar codes for SCL decoding,” IEEE Transactions
[1] E. Arıkan, “Channel polarization: A method for constructing capacity- on Communications, vol. 69, no. 12, pp. 7945–7959, December 2021.
achieving codes for symmetric binary-input memoryless channels,” IEEE [21] G. Poltyrev, “Bounds on the decoding error probability of binary linear
Transactions on Information Theory, vol. 55, no. 7, pp. 3051–3073, July codes via their spectra,” IEEE Transactions on Information Theory,
2009. vol. 40, no. 4, pp. 1284–1292, July 1994.
[2] 3rd Generation Partnership Project (3GPP), “Multiplexing and channel [22] F. Brannstrom, L. Rasmussen, and A. Grant, “Convergence analysis and
coding,” 3GPP 38.212 V.15.3.0, September 2018. optimal scheduling for multiple concatenated codes,” IEEE Transactions
[3] P. Trifonov and V. Miloslavskaya, “Polar subcodes,” IEEE Journal on on Information Theory, vol. 51, no. 9, pp. 3354–3364, 2005.
Selected Areas in Communications, vol. 34, no. 2, pp. 254–266, February [23] V. Miloslavskaya, Y. Li, and B. Vucetic, “Design of compactly specified
2016. polar codes with dynamic frozen bits based on reinforcement learning,”
[4] T. Wang, D. Qu, and T. Jiang, “Parity-check-concatenated polar codes,” IEEE Transactions on Communications, Early Access, 2023.
IEEE Communications Letters, vol. 20, no. 12, December 2016. [24] P. Trifonov, “Efficient design and decoding of polar codes,” IEEE
[5] E. Arıkan, “From sequential decoding to channel polarization and back Transactions on Communications, vol. 60, no. 11, pp. 3221 – 3227,
again,” ArXiv, vol. abs/1908.09594, September 2019. November 2012.
[6] V. Miloslavskaya and B. Vucetic, “Design of short polar codes for SCL [25] M. Bardet, V. Dragoi, A. Otmani, and J. Tillich, “Algebraic properties of
decoding,” IEEE Transactions on Communications, vol. 68, no. 11, pp. polar codes from a new polynomial formalism,” in IEEE International
6657–6668, November 2020. Symposium on Information Theory (ISIT), 2016, pp. 230–234.
[26] H. Zhou, W. J. Gross, Z. Zhang, X. You, and C. Zhang, “Low-
[7] P. Trifonov and G. Trofimiuk, “A randomized construction of polar
complexity construction of polar codes based on genetic algorithm,”
subcodes,” in IEEE International Symposium on Information Theory,
IEEE Communications Letters, vol. 25, no. 10, pp. 3175–3179, 2021.
ISIT Aachen, Germany, June 25-30, 2017, pp. 1863–1867.
[27] Wikipedia contributors, “Premature convergence — Wikipedia, the
[8] M. C. Coşkun and H. D. Pfıster, “An information-theoretic perspective
free encyclopedia,” https://en.wikipedia.org/w/index.php?title=Prematu
on successive cancellation list decoding and polar code design,” IEEE
re_convergence&oldid=1169353656, 2023, [Online; accessed 5-
Transactions on Information Theory, vol. 68, no. 9, pp. 5779–5791,
September-2023].
September 2022.
[28] M. Mondelli, S. H. Hassani, and R. Urbanke, “From polar to Reed-
[9] A. Elkelesh, M. Ebada, S. Cammerer, and S. t. Brink, “Decoder-tailored
Muller codes: A technique to improve the finite-length performance,”
polar code design using the genetic algorithm,” IEEE Transactions on
IEEE Transactions on Communications, vol. 62, no. 9, September 2014.
Communications, vol. 67, no. 7, pp. 4521–4534, July 2019.
[29] T. Tonnellier and W. J. Gross, “On systematic polarization-adjusted
[10] D. Wu, Y. Li, and Y. Sun, “Construction and block error rate analysis
convolutional (PAC) codes,” IEEE Communications Letters, vol. 25,
of polar codes over AWGN channel based on Gaussian approximation,”
no. 7, pp. 2128–2132, July 2021.
IEEE Communications Letters, vol. 18, no. 7, pp. 1099–1102, July 2014.
[30] T. Erseghe, “Coding in the finite-blocklength regime: Bounds based on
[11] I. Sason and S. Shamai, “Performance analysis of linear codes under
laplace integrals and their asymptotic approximations,” IEEE Transac-
maximum-likelihood decoding: A tutorial,” Foundations and Trends®
tions on Information Theory, vol. 62, no. 12, pp. 6854–6883, 2016.
in Communications and Information Theory, vol. 3, no. 1–2, pp. 1–222,
[31] V. Miloslavskaya and P. Trifonov, “Sequential decoding of polar codes,”
2006. [Online]. Available: http://dx.doi.org/10.1561/0100000009
IEEE Communications Letters, vol. 18, no. 7, pp. 1127–1130, July 2014.
[12] A. Canteaut and F. Chabaud, “A new algorithm for finding minimum-
[32] P. Trifonov, “A score function for sequential decoding of polar codes,”
weight words in a linear code: Application to McEliece’s cryptosystem
in 2018 IEEE International Symposium on Information Theory (ISIT),
and to narrow-sense BCH codes of length 511,” IEEE Trans. Inf. Theory,
June 2018, pp. 1470–1474.
vol. 44, pp. 367–378, 1998.
[13] Y. Li, H. Zhang, R. Li, J. Wang, G. Yan, and Z. Ma, “On the
weight spectrum of pre-transformed polar codes,” IEEE International
Symposium on Information Theory (ISIT), pp. 1224–1229, July 2021.
[14] V. Miloslavskaya, B. Vucetic, and Y. Li, “Computing the partial weight
distribution of punctured, shortened, precoded polar codes,” IEEE Trans-
actions on Communications, vol. 70, no. 11, pp. 7146–7159, 2022.
[15] H. Yao, A. Fazeli, and A. Vardy, “A deterministic algorithm for
computing the weight distribution of polar code,” IEEE Transactions
on Information Theory, 2023, Early Access.
[16] I. Tal and A. Vardy, “List decoding of polar codes,” IEEE Transactions
on Information Theory, vol. 61, no. 5, May 2015.
[17] R. Fano, “A heuristic discussion of probabilistic decoding,” IEEE
Transactions on Information Theory, vol. 9, no. 2, pp. 64–74, 1963.
[18] E. Arıkan, “On the origin of polar coding,” IEEE Journal on Selected
Areas in Communications, vol. 34, no. 2, pp. 209–223, 2016.
[19] P. Trifonov and V. Miloslavskaya, “Polar codes with dynamic frozen
symbols and their decoding by directed search,” in Proceedings of IEEE
Information Theory Workshop, September 2013, pp. 1 – 5.