On the Automorphism Group of Polar Codes

Marvin Geiselhart, Ahmed Elkelesh, Moustafa Ebada, Sebastian Cammerer and Stephan ten Brink
Institute of Telecommunications, Pfaffenwaldring 47, University of Stuttgart, 70569 Stuttgart, Germany
{geiselhart,elkelesh,ebada,cammerer,tenbrink}@inue.uni-stuttgart.de

Abstract—The automorphism group of a code is the set of (ensemble) decoding, as proven in [10, Theorem 2]. Therefore,
permutations of the codeword symbols that map the whole code it is crucial to find automorphisms outside the LTA to enable
onto itself. For polar codes, only a part of the automorphism efficient parallel ensemble decoding of polar codes. Further
group was known, namely the lower-triangular affine group
(LTA), which is solely based upon the partial order of the code’s potential applications include analysis of some post-quantum
synthetic channels. Depending on the design, however, polar codes cryptography schemes [12].
arXiv:2101.09679v2 [cs.IT] 27 Jan 2021

can have a richer set of automorphisms. In this paper, we extend The main contribution of this work is the introduction
the LTA to a larger subgroup of the general affine group (GA), of a larger automorphism group of decreasing monomial
namely the block lower-triangular affine group (BLTA) and show codes, namely the block lower-triangular affine group (BLTA).
that it is contained in the automorphism group of polar codes.
Furthermore, we provide a low complexity algorithm for finding We provide efficient algorithms for finding this group and
this group for a given information/frozen set and determining its sampling from it. The concept applies to polar codes, RM
size. Most importantly, we apply these findings in automorphism- codes and the recently proposed partially symmetric monomial
based decoding of polar codes and report a comparable error-rate codes (PSMCs) [13].
performance to that of successive cancellation list (SCL) decoding
with significantly lower complexity. II. P RELIMINARIES

I. I NTRODUCTION A. Polar Codes

Polar codes are constructed based on the concept of channel
Polar codes are the first channel codes which are theoret-
polarization [1]. N identical discrete memoryless channels
ically proven to asymptotically achieve the channel capacity
(DMCs) are converted, via the channel transform, into N
under successive cancellation (SC) decoding [1]. In the short
synthetic channels that show a polarization behavior. This
length regime, cyclic redundancy check (CRC)-aided polar
means that a fraction of the bit-channels become very reliable
codes under successive cancellation list (SCL) decoding [2]
(i.e., noiseless), while the rest of the synthetic bit-channels
achieves an outstanding performance and, thus, selected as
become totally noisy. Information is transmitted only on the K
the channel code for the uplink and downlink control channel
most reliable channels (information channels), while the poor
of the 5G standard [3]. Due to the highly symmetric structure
channels are set to “0” (frozen channels). This is equivalent
of the polar code factor graph, decoders using the concept of  ⊗n
to selecting K rows from the Hadamard matrix GN = 11 01
factor graph permutations are proposed in [4], [5] and [6].
with N = 2n to form the generator matrix G of the code.
A different approach is to use the symmetries in the code
Alternatively, polar codes can be viewed as monomial codes
itself, i.e., its automorphism group. To this end, polar codes
[7]. In this perspective, each synthetic channel corresponds to
are viewed as decreasing monomial codes [7]. In [7], it is
a monomial in n binary variables xi . The set of all monomials
shown that the automorphism group of decreasing monomial
in n variables is defined as Mn and a polar code is a specific
codes (and, thus, polar codes) is at least the lower-triangular
subset I, called the information set of the polar code. Every
affine group (LTA), solely based on a partial order of synthetic
monomial can be written as
channels. This proved to be sufficient for the application
f = ∏ xi (1)
of the minimum-weight codeword enumeration. However, in
i∈ind( f )
general, we expect decreasing monomial codes to have more
where ind( f ) is an ordered subset of the variable indices Ω =
automorphisms. This is easily verified by the fact that Reed–
[0, n − 1] , {0, 1, . . ., n − 1} and directly corresponds to the ℓ-th
Muller (RM) codes can be seen as a special case of decreasing
row of the generator matrix as
monomial codes with an automorphism group known to be the
ℓ = ∑ 2i . (2)
general affine group (GA) [8], which is much larger than LTA.
i∈Ω\ind( f )
Automorphism-based decoding has been successfully In other words, the monomial f corresponds to the row whose
applied to RM codes [9], [10] and Bose-Chaudhuri- binary representation has zeros exactly in the bit-positions of
Hocquenghem (BCH) codes [11]. However, it was not yet the variables contained in f . A message is a polynomial
possible to use the automorphism group in SC-based decod-
u(x0 , . . . , xn−1 ) = ∑ u f · f (x0 , . . . , xn−1 ) (3)
ing of polar codes. The reason for this is that LTA-based f ∈I
automorphisms cannot result in any gains under SC-based with K coefficients u f ∈ F2 . The respective codeword is given
by the evaluation of u(x) in all N points x ∈ Fn2 . As a
The authors would like to thank Florian Euchner for his help with proving
Theorem 1 by proposing the crazy paternoster algorithm. convention, we assume the j-th codeword symbol is obtained
from the point x equal to the binary expansion of j.
B. Partial Order In other words, I remains unchanged when permuting the
It was shown in [7] and [14] that the synthetic channels variable indices in the monomials according to π . Furthermore,
exhibit a partial order “4” with respect to their reliability, let Stab(I) denote the set of all permutations with this property.
i.e, f 4 g means that the synthetic channel corresponding to Note that Stab(I) is a subgroup of S(Ω). We call the stabilizer
monomial f is more reliable than the one corresponding to g. trivial, if it only contains the identity permutation.
For monomials of equal degree this partial order is defined as In the following, we seek to find Stab(I) for a given I and
f 4 g ⇔ ind( f ) j ≤ ind(g) j ∀ j = 0, . . . , deg( f ) − 1 (4) derive some useful properties.
Definition (Minimum and Maximum of a Permutation):
and for monomials of different degree
Let π ∈ S(Ω) be some permutation. The minimum min(π )
f 4 g ⇔ ∃g∗ |g with deg(g∗ ) = deg( f ) and f 4 g∗ . (5) and maximum max(π ) are defined by the smallest and largest
C. Decreasing Monomial Codes element not fixed by π , i.e.,
A decreasing monomial code is a polar code whose mono- min(π ) , min {i | i ∈ Ω, π (i) 6= i} , (9)
mial selection obeys the partial order [7]. More precisely, if max(π ) , max {i | i ∈ Ω, π (i) 6= i} . (10)
a synthetic channel is selected as an information channel, all
Definition (Interval Disjoint and Interlocked Permu-
stronger channels w.r.t. “4” are also information channels.
tations): Two permutations π1 and π2 are interval dis-
Mathematically, this can be written as
joint, if the intervals Ωπ1 = [min(π1 ), max(π1 )] and Ωπ2 =
∀g ∈ I, ∀ f ∈ Mn with f 4 g ⇒ f ∈ I. (6)
[min(π2 ), max(π2 )] are disjoint. Note that elements of Ωπi are
Almost all practical polar code constructions result in decreas- not necessarily affected by πi . Permutations are said to be
ing monomial codes. A decreasing monomial code can be fully interlocked, if they are not interval disjoint and do not share
specified by a minimal information set Imin containing only elements.
a small number of monomials called generators. All other Let C(π ) be the set of cycles that make up π , i.e., its cycle
monomials are implied by the partial order: decomposition. By merging all interlocked cycles into the sub-
[
I= { f ∈ Mn , f 4 g} . (7) permutations ρi , we obtain the interval disjoint decomposition
g∈Imin T (π ) = {ρ0 , . . . , ρd−1 } , as the set of pairwise interval disjoint
Moreover, the RM code of order r and length N = 2n (i.e., permutations ρi such that π = ρ0 ◦ · · · ◦ ρd−1.
RM(r, n)-code) is a special case of a decreasing monomial Theorem 1 (Stabilizers): Let I be the monomial set of a
code with Imin = {xn−r · · · xn−1 }. In this paper, we will notate decreasing monomial code with a non-trivial stabilizer Stab(I).
Imin as numerical row indices, according to Eq. (2). Then the following statement holds:
D. Automorphisms of Decreasing Monomial Codes If a non-trivial permutation π stabilizes I, then all permu-
The automorphism group Aut(C ) of a code C is the tations of the disjoint intervals of π stabilizes I as well, i.e.,
π ∈ Stab(I) ⇒ hS ([min(ρ ), max(ρ )])iρ ∈T (π ) ⊆ Stab(I), (11)
group of codeword symbol permutations, that leave the code
unchanged, i.e., map each codeword onto a codeword that where h·i denotes the join of subgroups.
is not necessarily different. It was shown in [7] that the Proof: The proof is given in Appendix A.
automorphism group of a decreasing monomial code contains Theorem 1 has a useful corollary revealing the structure of
at least LTA(2, n), i.e., affine transformations of the variables xi Stab(I).
in the form x′ = Ax+b, with A ∈ Fn×n 2 being a lower triangular Corollary: Stab(I) can be written as the join of permuta-
matrix with a unit diagonal and arbitrary b ∈ Fn2 . tions groups S(Ωk ) of partitions of Ω, i.e.,
III. S TABILIZERS OF THE M ONOMIAL S ET Stab(I) = hS (Ω0 ) , . . . , S (Ωm−1 )i
m−1
It was shown in [15] that the stage-shuffling of the polar [
with Ωk = Ω and Ωk ∩ Ωl = 0/ for k 6= l. (12)
factor graph corresponds to a bit-index permutation of both
k=0
the codeword vector c and the message vector u (including
the frozen bits). When viewing such permutations from a In other words, every permutation π ∈ Stab(I) can be writ-
monomial code perspective, they exactly correspond to per- ten as a product of (potentially trivial) permutations of the
muting the variables xi of the monomials from Mn . Depending intervals Ωk and vice versa. Note that Ωk may contain only
on the polar code construction (i.e., information/frozen set), a single element when S(Ωk ) does not contribute to any non-
there may exist permutations that keep the information set trivial permutation.
I unchanged, i.e., they stabilize it. Such a permutation is Proof: Assume the sub-intervals are not disjoint, i.e., there
directly related to the automorphism of the code, where A exist two sub-intervals Ωk and Ωl with S(Ωk ) ⊆ Stab(I) and
in x′ = Ax + b is the corresponding permutation matrix. S(Ωl ) ⊆ Stab(I) but Ωk ∩ Ωl 6= 0/ and neither Ωk ⊆ Ωl nor
Definition (Stabilizer): Let S(Ω) be the set of all permuta- Ωl ⊆ Ωk . Then one can pick two permutations (e.g., extremal
tions of Ω. Then a permutation π ∈ S(Ω) stabilizes a monomial transpositions) π1 ∈ S(Ωk ) and π2 ∈ S(Ωl ) which are not
set I, if and only if interval disjoint and π = π1 ◦ π2 is either a single cycle or
∀ f ∈ I =⇒ f ′ = π ( f ) , ∏ xπ (i) ∈ I. (8) the product of interlocked cycles. In both cases, π stabilizes I
i∈ind( f ) and, thus, S(Ωk ∪ Ωl ) ⊆ Stab(I). Therefore, every permutation
in Stab(I) either falls into an existing sub-interval or expands γ0 = 0
or merges sub-intervals, keeping the partition property.  D0,0 s0
The partition (and therefore Stab(I)) is fully described by
the list of interval sizes s = [sk ] of the m sub-intervals Ωk , i.e., γ1 = s0 s0
0
D0,1 D1,1 s1
sk = |Ωk | = max(Ωk ) − min(Ωk ) + 1. (13)
The corollary gives us an algorithm for finding the sub- s1
intervals Ωk for an arbitrary decreasing monomial code with
information set I. We know that all permutations in S(Ωk ) γm−1
are contained in Stab(I), as we can pick trivial permutations D0,m−1 D1,m−1 Dm−1,m−1 sm−1
for the other sub-intervals. In particular, also the transposition
π = (min(Ωk ), max(Ωk )) stabilizes I. Therefore, we can find sm−1
the borders of the sub-intervals by systematically searching Fig. 1: Structure of a block lower-triangular matrix with block sizes sk and
block starts γk .
for pairs i0 , i1 with maximal distance. Algorithm 1 provides
a pseudo-code for this procedure. The algorithm has a worst block sizes s = [sk ], 0 ≤ k ≤ m − 1, if all elements to the right
case runtime of O(K · n2 ), with the check π (I) = I requiring of the block diagonal are zero, i.e., ai, j = 0 ∀ j ≥ γk(i) + sk(i) .
K comparisons. The blocks of the matrix are denoted by Dk,l . Fig. 1 shows
Input : Information set I of decreasing monomial the general structure of a block lower diagonal matrix. As
code in n variables square block matrices naturally extend conventional matrices,
Output: List of sub-interval sizes s we have the following properties:
1 s ← [ ], i0 ← 0; 1) The product of two block lower-triangular matrices is
2 while i0 < n do also a block lower-triangular matrix with the same block
3 i1 ← n − 1; structure.
4 while i1 ≥ i0 do 2) A block lower-triangular is non-singular if and only if all
5 π ← (i0 , i1 ); blocks on the main diagonal Dk,k are non-singular.
6 if π (I) = I then 3) The inverse of a block lower-triangular matrix is also
7 append i1 − i0 + 1 to s; a block lower-triangular matrix with the same block
8 i0 ← i1 + 1; structure as the original matrix.
9 else As a consequence, non-singular block lower-triangular ma-
10 i1 ← i1 − 1; trices form a group under matrix multiplication. Note that
11 end associativity is inherited from matrix multiplication and the
12 end identity matrix I is always block lower-triangular. The size
13 end of this group can be easily computed in terms of s. For this,
Algorithm 1: Finding Stab(I) in terms of the partition of Ω into sub-
intervals of size sk .
observe that in row i, there are γk(i) + sk(i) elements that can be
0 or 1 each. However, one has to deduce the number of cases
We can represent Stab(I) as a set of n×n permutation matri- where the row is a linear combination of the i previous rows.
ces Ps (n), where s = [sk ] defines a block diagonal structure with Therefore, the number of invertible block lower-triangular
blocks of sizes sk × sk . Except for the block diagonal elements, matrices is n−1
NIBLT (s) = ∏ 2γk(i) +sk(i) − 2i


all matrix elements are zero. For a non-trivial stabilizer, we (15)
hereby find automorphisms outside LTA, as no permutation i=0
m−1 sk −1 
matrix is lower-triangular besides the identity permutation. ′

= ∏∏ 2γk +sk − 2γk +i
′
k=0 i =0
IV. T HE AUTOMORPHISM G ROUP OF P OLAR C ODES !
m−1 sk −1 
In the following, we combine both LTA and the newly found

γk ·sk sk i′
stabilizer group into a larger group, namely the automorphism
= ∏ 2 ∏ 2 −2 . (16)
k=0 i′ =0
group of polar codes.
While Eq. (15) expresses the number from a whole matrix
Definition (Block Indices): We denote a partition of the
perspective, Eq. (16) views the same thing from a block
interval Ω = [0, n − 1] by a sequence of m positive integers
matrix perspective. In particular, the inner product gives the
sk > 0 for the sizes of the sub-intervals. The interval start γk
number of non-singular diagonal blocks Dk,k , while 2γk ·sk is
is the first element of the k-th sub-interval and is defined as
the number of arbitrary rectangular matrices to the left of the
the cumulative sum k−1
γk = ∑ sk . (14) block diagonal Dk,l with l < k, for each block row k.
i=0 Definition (Block Lower-Triangular Affine Group,
The index function k(i) returns the index of the sub-interval BLTA): The block lower-triangular affine group BLTA(s, n)
that contains i and is defined as k(i) = max {k : i ≥ γk }. is the set of affine transformations x′ = Ax + b over Fn2
Definition (Block Lower-Triangular Matrix): An n × n with A ∈ Fn×n
2 non-singular block lower-triangular with block
matrix A over an arbitrary field is block lower-triangular with structure s and an arbitrary b ∈ Fn2 .
 From the discussion of block lower-triangular matrices 1017
above, it is easy to see that BLTA is indeed a group, in |GA(2,7)|
particular a subgroup of GA(2, n). Moreover, it can be seen
1014

| Aut(C )|
that LTA(2, n) and GA(2, n) are themselves special cases of
BLTA(s, n), with s = [1, . . . , 1] and s = [n], respectively.
Lemma 1: The join of the group of block-permutation 1011
transformations Ps (n) and the group lower triangular affine
|LTA(2,7)|
transformations LTA(2, n) is exactly BLTA(s, n), i.e., 108 ← “RM-like” “Polar-like” →
BLTA(s, n) = hPs (n), LTA(2, n)i. (17) 10−2 10−1
In other words, any composition of transformations from Design Erasure Probability ε
LTA(2, n) and Ps (n) is a transformation from BLTA(s, n) and Fig. 2: Number of automorphisms of (128,64) polar codes with
vice versa. Bhattacharyya-based construction versus Binary Erasure Channel (BEC) de-
sign erasure probability ε .
Proof: “⇒”: Obviously, Ps (n) ⊆ BLTA(s, n), as permutation
matrices are non-singular and the block structure is given.
1017 RM(3,7) Maximum
Similarly, LTA(2, n) ⊆ BLTA(s, n), where again we have a |GA(2,7)|
Average
special case of affine transformations. Also, as BLTA(s, n) is Minimum
1014

| Aut(C )|
closed, a composition of any transformations will not generate
any elements outside BLTA(s, n).
“⇐”: We can show this by observing that any 1011
block lower-triangular matrix A may be decomposed as Bhat. @1dB
|LTA(2,7)|
A = P1 · L1 · P2 · L2 · P3 , with Pi ∈ Ps (n) and Li ∈ LTA(2, n). 108 ← “RM-like” “Polar-like” →
For this, consider the LUP decomposition of A, i.e., PA =
LU [16]. The block lower-triangular structure of A ensures 1 2 3 4 5 6 7 8
that also U and P are block lower-triangular. One can now Number of Generators |Imin |
transform U into a conventional lower-triangular matrix by Fig. 3: Maximum, average and minimum number of automorphisms of all
reversing the order of the rows and columns within each (128,64) decreasing monomial codes versus their number of generators |Imin |.
block. This can be written as L2 = PBR (s) · U · PBR (s), with !
m−1 sk −1 
PBR (s) = [pi, j ] and

i′
( | BLTA(s, n)| = NIBLT (s)·2n = 2n · ∏ 2γk ·sk ∏ 2 sk − 2 .
1 for j = 2γk(i) + sk(i) − 1 − i k=0 i′ =0
pi, j = . (18) (20)
0 else
Note that this equates to the sizes of LTA(2, n) and GA(2, n)
Finally, as PBR (s) = P−1 −1 for the special cases s = [1, . . . , 1] and s = [n], respectively.
BR (s), we have P1 = P , P2 = P3 =
PBR (s) and L1 = L. The additive term b may be included (i.e., Fig. 2 shows the sizes of the automorphism groups for
also properly permuted) in any of the LTA transformations.  polar codes with N = 128 and K = 64 designed according to
Theorem 2 (Automorphisms of Polar Codes): Let C be the Bhattacharyya parameter of the synthetic channels. This
a decreasing monomial code in n variables with information construction assumes a BEC with erasure probability ε . It
set I. Then can be seen that for low erasure probability, this construction
BLTA(s, n) ⊆ Aut(C ) (19) generates the RM(3,7)-code. The larger the values of ε , the
fewer the automorphisms featured by the code. In Fig. 3, we
with s being the block structure of Stab(I). evaluate the influence of the number of generators of a code
Proof: The proof directly follows from Lemma 1, as both |Imin | on the size of the automorphism group, also for the case
LTA and Stab(I) correspond to automorphisms of the code.  of N = 128 and K = 64. Since there exist usually many codes
We furthermore conjecture that Eq. (19) holds with equality, with the same number of generators, we plot the minimum,
which is also suggested by means of numerical results. How- average and maximum automorphism group sizes for each
ever, we were not yet able to find a rigorous proof. To prove it, value of |Imin |. To obtain these numbers, we enumerated all
one would need to show that for every nonzero element ai, j 1007 (128, 64) decreasing monomial codes using a tree search.
with i < j in an affine transformation (A, b) ∈ Aut(C ), the As just mentioned, it can be seen that a smaller size of Imin
variable permutation π = (i, j) must be contained in Stab(I). generally results in a larger number of automorphisms. We
find the RM code on the very left of the plot, while typical
A. Number of Automorphisms polar codes lie more towards the right edge. It is worth it
The number of automorphisms is (at least) the size of mentioning that, from a code design perspective, several code
BLTA(s, n) for a code with block structure s. Clearly, this is the constructions can be viewed as lying between polar and RM
number of non-singular block lower-triangular matrices times codes (e.g., [13], [17], [18] and [19]).
the number of affine translations b. Using Eq. (16), we have
 Design s |Aut(C )| dmin a Admin b
Bhat. @1 dB [2,1,1,1,1,1,1] 2.06 · 1011 8 96
Imin = {31,99} [5,3] 1.41 · 1016 16 69936
10−1
Imin = {31,57} [3,5] 1.41 · 1016 16 69936
Table I: Properties of the compared (256,128) polar codes. a dmin : minimum
10−2 distance of the code. b Admin : number of minimum-weight codewords.
BLER

use M = 8 parallel independent SC decoders, each decoding

10−3 a permuted version of the received sequence y. The codeword
Design SC SCL-8 Aut-8-SC estimates of each SC decoder are un-permuted and the ML-in-
10−4 Bhat. @1 dB the-list method is applied to select the final codeword estimate.
Imin = {31,99}
Imin = {31,57} The permutations are conducted by automorphisms randomly
10−5 sampled from the BLTA group of the particular code, found
1 1.5 2 2.5 3 3.5 4 using Algorithm 1. Note that this decoder is similar to SCL,
Eb /N0 [dB] however, no sorting of the path-metrics are required, as the
Fig. 4: Comparison of (N = 256,K = 128) polar codes under SC, Aut-SC
and SCL decoding; BI-AWGN channel. Appendix B gives block error rate
constituent decoders are independent. We assume an additive
(BLER) results for (128,64) polar codes. white Gaussian noise (AWGN) channel with binary phase shift
keying (BPSK) modulation.
B. Sampling Automorphisms
Fig. 4 shows the BLER performance of (256,128) polar
For some practical applications such as automorphism en- codes under SC-based decoding. In particular, we compare
semble decoding [10], it is required to sample from the auto- plain SC decoding [1] with SCL with list size 8 (SCL-
morphism group, i.e., to pick a permutation from BLTA(s, n) 8) decoding [2] and Aut-8-SC decoding [10]. First, we
at random. In general, it is difficult to ensure that a random see that while being the best code under SC-decoding, the
matrix is invertible. If the fraction of non-singular matrices out Bhattacharyya construction at design SNR of 1 dB (Imin =
of all matrices is sufficiently large, one can generate random {59, 79, 105, 149, 163, 224}) does not show any gains for Aut-
matrices and test for invertibility. For binary matrices (GL for SC decoding, as expected for such few automorphisms outside
general linear), this probability is
lower bounded [20] as LTA. Next, in order to have a large automorphism group, we
∏n−1 2n − 2i n 
′

designed codes by selecting two generators Imin = {31, 57}
psucc,GL = i=0 2 = ∏ 1 − 2−i
2(n ) i′ =1 and Imin = {31, 99}, under the constraint of K = 128. Both
n 
′
 codes can be viewed as examples of PSMCs [13]. While the
≥ lim ∏ 1 − 2−i = 0.28878 . . . (21) SC performance degrades, now a significant performance gain
n→∞ ′
i =1 is achieved by both Aut-SC and SCL. However, the two con-
However, the same expression for a block lower-triangular structions show a very different behavior. While for the code
(BLT) matrices, i.e., with Imin = {31, 99} SCL shows a very good performance,
γk(i) +sk(i)
∏n−1 − 2i

n−1 
i=0 2 Aut-SC shows only small gains. The code with Imin = {31, 57}

i−γk(i)−sk(i)
psucc,BLT =
∏n−1
γk(i) +sk(i) = ∏ 1 − 2
can, however, outperform SCL. Therefore, a strict correlation
i=0 2 i=0
n−1 between SCL and Aut-SC decoding performance for partially
n→∞
≥ ∏ 1 − 2−1 = 2−n −−−→ 0,


(22) symmetric codes cannot be inferred and code design for both
i=0 decoders remains an open problem. Table I lists the parameters
cannot be lower bounded, since the last line holds with equality and properties of the compared codes.1 In Appendix B we
for the case s = [1, . . . , 1]. We therefore propose a different provide more BLER results for the case of N = 128 and
method, based on the fact that only the blocks on the diagonal K = 64. We want to emphasize again that the usage of just LTA
must be non-singular: permutations would result in the BLER performance curves of
1) For k = 0, 1, · · · , m − 1, sample the square blocks Dk,k on Aut-SC to coincide with plain SC decoding as depicted and
the main diagonal from GL(2, sk ), i.e., generate random discussed in [10].
sk × sk binary matrices until a non-singular one
is found, VI. C ONCLUSION
sk
with success probability psucc,k = ∏i=1 1 − 2−i .
We show that decreasing monomial codes have at least
2) Select all elements below the block diagonal (i.e., ai, j
BLTA as their automorphism group, which is in most cases
with j < γk(i) , or blocks Dk,l with l < k) randomly
larger than the previously known subgroup LTA, and propose
uniformly from {0, 1}.
an algorithm to find this group. While the automorphisms from
This method has the advantage that each block on the diagonal LTA were proven to yield no error-rate performance gains
can be independently sampled, resulting in total in the same under automorphism-based SC decoding when compared to
lower bound, Eq. (21), which is fulfilled with equality for the plain SC decoding, the newly found BLTA permutations show
worst case of s = [n]. significant gains, outperforming the state-of-the-art SCL in
V. P OLAR C ODES UNDER AUTOMORPHISM SC D ECODING some scenarios, with a strictly lower complexity.
As an application, we now evaluate polar codes under 1 An interactive demo of the code properties is provided online:
automorphism SC (Aut-SC) decoding. As proposed in [10], we http://webdemo.inue.uni-stuttgart.de/webdemos/08 research/polar/index.php?id=12
 R EFERENCES and Ωπc = Ω \ Ωπ its complement. We can factor every mono-
mial f ∈ I into a part fπ corresponding to π and a residual:
[1] E. Arıkan, “Channel Polarization: A Method for Constructing Capacity-
Achieving Codes for Symmetric Binary-Input Memoryless Channels,”
IEEE Trans. Inf. Theory, vol. 55, no. 7, pp. 3051–3073, Jul. 2009. f= ∏ xi ∏ xi = f π · f π c . (24)
[2] I. Tal and A. Vardy, “List Decoding of Polar Codes,” IEEE Trans. Inf. i∈ind( f )∩Ωπ i∈ind( f )∩Ωcπ
Theory, vol. 61, no. 5, pp. 2213–2226, May 2015.
[3] “Technical Specification Group Radio Access Network,” 3GPP, 2018, Now, partition I into subsets I f of the same degree and equal
TS 38.212 V.15.1.1. [Online]. Available: http://www.3gpp.org/ftp/Specs/ residual fπ c , i.e., under the following equivalence relation:
archive/38 series/38.212/
[4] S. A. Hashemi, N. Doan, M. Mondelli, and W. J. Gross, “Decoding f ∼ f ′ ⇔ deg( f ) = deg( f ′ ) and fπ c = fπ′ c (25)
Reed-Muller and Polar Codes by Successive Factor Graph Permuta-
tions,” in IEEE 10th Inter. Symp. on Turbo Codes Iterative Inf. Process. I/∼ = {[ f ]∼ | f ∈ I} . (26)
(ISTC), Dec. 2018.
[5] A. Elkelesh, M. Ebada, S. Cammerer, and S. ten Brink, “Belief Prop- We focus on some monomial f with subset I f = [ f ]∼ . From [7,
agation List Decoding of Polar Codes,” IEEE Commun. Lett., vol. 22,
no. 8, pp. 1536–1539, Aug. 2018. Proposition 2], we know that within each subset, all elements
[6] M. Kamenev, Y. Kameneva, O. Kurmaev, and A. Maevskiy, “Permu- are comparable under the partial order and it is sufficient to
tation Decoding of Polar Codes,” in XVI Inter. Symp. “Problems of look at the part of f that is not shared by the elements in
Redundancy in Information and Control Systems” (REDUNDANCY),
2019, pp. 1–6. I f , i.e., fπ . We will now show that I f contains all monomials
[7] M. Bardet, V. Dragoi, A. Otmani, and J. Tillich, “Algebraic Properties of that share fπ c and have the same degrees in intervals of the
Polar Codes From a New Polynomial Formalism,” in IEEE Inter. Symp. interval disjoint decomposition of π , i.e.,
Inf. Theory (ISIT), Jul. 2016, pp. 230–234.
[8] F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-Correcting

I f = g · fπ c ∈ Mn , deg(gρ ) = deg( fρ ) ∀ρ ∈ T (π ) , (27)
Codes, ser. North-Holland Mathematical Library. North-Holland Pub.
Co., 1977, no. 16.
[9] N. Stolte, “Rekursive Codes mit der Plotkin-Konstruktion und ihre by repeatedly applying π and using the partial order.
Decodierung,” Ph.D. dissertation, Technische Universität Darmstadt, First, assume π = σ is just a single cycle and Ωσ = [i0 , i1 ].
Jan. 2002. [Online]. Available: http://tuprints.ulb.tu-darmstadt.de/183/ Let d = deg( fσ ). If d = 0, Eq. (27) is already fulfilled, as f
[10] M. Geiselhart, A. Elkelesh, M. Ebada, S. Cammerer, and S. ten Brink,
“Automorphism Ensemble Decoding of Reed-Muller Codes,” ArXiv e- is the only such monomial. If d > 0, observe:
prints, arXiv:2012.07635, Dec. 2020. 1) After a maximum of ord(σ ) steps, we can transform f
[11] T. Hehn, O. Milenkovic, S. Laendner, and J. B. Huber, “Permutation De-
coding and the Stopping Redundancy Hierarchy of Cyclic and Extended into some f ′ with the property i1 ∈ ind( f ′ ).
Cyclic Codes,” IEEE Trans. Inf. Theory, vol. 54, no. 12, 2008. 2) After a maximum of 2d ·ord(σ ) steps, we can transform f
[12] M. Bardet, J. Chaulet, V. Dragoi, A. Otmani, and J. Tillich, “Cryptanal- into fˆ = xi1 −d+1 · · · xi1 which is the maximum monomial
ysis of the McEliece Public Key Cryptosystem Based on Polar Codes,”
in Post-Quantum Cryptography, 2016, pp. 118–143. (w.r.t. the partial order) in Ωσ .
[13] K. Ivanov and R. Urbanke, “Partially symmetric monomial codes,” ArXiv In both scenarios, a step refers to one application of the
e-prints, arXiv:2001.03790, Jan. 2020.
[14] C. Schürch, “A Partial Order For the Synthesized Channels of a Polar
partial order (i.e., transforming f into some f ′ 4 f with
Code,” in IEEE Inter. Symp. Inf. Theory (ISIT), Jul. 2016, pp. 220–224. f ′ ∼ f ) followed by one application of the permutation π (i.e.,
[15] N. Doan, S. A. Hashemi, M. Mondelli, and W. J. Gross, “On the transforming f into f ′ = π ( f ); f ′ ∼ f implicitly fulfilled). Both
Decoding of Polar Codes on Permuted Factor Graphs,” in IEEE Global
Commun. Conf. (GLOBECOM), Dec. 2018.
operations will map f to another monomial f ′ that is contained
[16] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction in I, as we assume π ∈ Stab(I) and I belongs to a decreasing
to Algorithms, 2nd ed. The MIT Press, 2001. monomial code.
[17] B. Li, H. Shen, and D. Tse, “A RM-Polar Codes,” ArXiv e-prints,
arXiv:1407.5483, Jul. 2014.
Observations 1) and 2) can be verified by looking at the
[18] M. Mondelli, S. H. Hassani, and R. L. Urbanke, “From Polar to Reed- following algorithm. We denote the monomial at step j by f j .
Muller Codes: A Technique to Improve the Finite-Length Performance,” Assume, we know some upper limit of Q steps, in which we
IEEE Trans. Commun., vol. 62, no. 9, pp. 3084–3091, Sep. 2014. can arrive certainly at the desired target state fˆ = fQ . At each
[19] A. Elkelesh, M. Ebada, S. Cammerer, and S. ten Brink, “Decoder-
Tailored Polar Code Design Using the Genetic Algorithm,” IEEE Trans. step j, find the positions ind( f j ) that are not correct if π is
Commun., vol. 67, no. 7, pp. 4521–4534, Jul. 2019. applied another Q − j times, i.e.,
[20] N. J. A. Sloane, “The Encyclopedia of Integer Sequences, Sequence
Fj = i ∈ ind(π Q− j ( fˆ)), i ∈

A048651.” [Online]. Available: http://oeis.org/A048651 / ind( f j ) . (28)
[21] M. Fossorier and S. Lin, “Soft-Decision Decoding of Linear Block
Codes Based on Ordered Statistics,” IEEE Trans. Inf. Theory, vol. 41,
no. 5, pp. 1379–1396, Sep. 1995.
After each set of two revolutions of π (or 2 · ord(π ) steps), we
can remove (at least) one element from Fj , since, as long as
A PPENDIX we have not yet arrived at the target state, there is an “empty
place” in the positions affected by π which is permuted to i0 at
A. Proof of Theorem 1 some point. The partial order allows us to move one of the xi
Let σ ∈ C(π ) be a cycle of the cycle decomposition of π and that are not yet in the target position, into i0 , as xi0 4 xi . After
Ωσ = [min(σ ), max(σ )] be the interval σ acts on. The index another maximum of ord(π ) − 1 steps, the respective xi has
set Ωπ is the union of the intervals of the cycle decomposition moved to position i1 , and can be placed in the target position
of π , i.e., î by the partial order, again, because xî 4 xi1 . Therefore, no
more than Q = 2d · ord(π ) steps are required. Note that this
[
Ωπ = Ωσ , (23)
σ ∈C(π ) is a very loose upper bound, but to prove Eq. (27), we only
 0 1 2 3 4 5 6 7 8 9 10 11 12

10−1

Fig. 5: Visualization of interlocked cycles in the example permutation

π = (1,5,2)(3,7)(6,10)(8,9). 10−2

BLER
need the algorithm to be deterministic and stop after a finite 10−3
number of steps. Design SC SCL-8 Aut-8-SC OSD-4
To summarize, we can move any variable in f in any other Bhat. @1 dB
10−4 Imin = {27,56}
place in Ωπ , as we can move it to i0 using the partial order, Imin = {23,112}
rotate it to i1 using the cycle permutation π , and then place it RM(3,7)
in the target position as all indices are reachable by the partial 10−5
1 1.5 2 2.5 3 3.5 4
order from i1 . Due to the resemblance of the cycle π with
an irregular paternoster elevator running around a building Eb /N0 [dB]
with floors Ωπ , we call this algorithm the crazy paternoster Fig. 6: Comparison of (N = 128,K = 64) polar codes under SC, Aut-SC and
algorithm. SCL decoding; BI-AWGN channel.
The described method can be extended to the case where
Design s |Aut(C )| dmin Admin
π is a product of interlocked cycles. We can classify all
Bhat. @1 dB [2,2,1,1,1] 2.42 · 109 8 688
cycles as either transit cycles or parking cycles. A parking Imin = {27,56} [3,4] 1.78 · 1012 8 240
cycle σ is fully enclosed by another cycle, i.e., ∃σ ′ ∈ C(π ) Imin = {23,112} [4,3] 1.78 · 1012 8 16
with Ωσ ⊂ Ωσ ′ ; while transit cycles partially overlap with RM(3,7) [7] 2.10 · 1016 16 94488
another. It is easy to see that there exists a chain of transit Table II: Properties of the compared (128,64) polar codes.
cycles σ1 , · · · , σt with the properties min(σ1 ) = i0 , max(σ j ) >
min(σ j+1 ) and max(σt ) = i1 . Fig. 5 shows the interlocked
cycles for the example of π = (1, 5, 2)(3, 7)(6, 10)(8, 9). Here, decoding (OSD)-4 results serve as an upper bound on the
the cycles σ1 = (1, 5, 2), σ2 = (3, 7) and σ3 = (6, 10) form the maximum likelihood (ML) performance of each code [21].
chain of transit cycles, while (8, 9) fully overlaps with (6, 10) Again, the Bhattacharyya construction at design SNR of 1 dB
and therefore is classified as a parking cycle. Using this chain (Imin = {31, 45, 51, 71, 84, 97}) does not show any gains for
of transit cycles, we can again move any variable in f to any Aut-SC decoding. Also, note that the gains of SCL-8 when
other position within Ωπ , as the overlap of the cycles allows xi compared to SC are also smaller than 0.2 dB. Next, we
to “change” from one cycle to the next higher cycle. Note that designed codes by selecting two generators Imin = {27, 56} and
depending on the degrees of monomials, the order in which Imin = {23, 112} in order to have a large automorphism group,
the xi are moved, must be adjusted. The maximum number of under the constraint of K = 64. Both codes can be viewed as
steps remains upper bounded, however, by the same number, examples of PSMCs [13]. While the SC performance degrades,
namely now a significant performance gain is achieved by both Aut-SC
Q′ = 2d · |C(π )| · max {ord(σ )} . (29) and SCL decoding with comparable performance. In particular,
σ ∈C(π )
Aut-SC is within 0.1 dB to 0.2 dB of the SCL performance.
For the most general case, i.e., if π is a product of multiple For completeness, we also included performance results for the
interval disjoint permutations, we can use the upper bound RM code construction (Imin = {15}). As previously reported
  in [10], in the RM case, automorphism-based decoding can
Q′′ = max 2d · |C(ρ )| · max {ord(σ )} , (30) even outperform SCL decoding.
ρ ∈T (π ) σ ∈C(ρ )

as each interval disjoint region Ωρ = [iρ ,0 , iρ ,1 ] can be opti-

mized independently according to the procedure above. For
′′
all regions, the same backtracking π Q − j ( fˆ) is used in each
step j. Therefore, Eq. (27) holds for all cases of monomials f .
This means that all permutations of the intervals Ωρ stabilize
I f . As this holds for all I f individually, it also holds for their
union I. 
B. Error-Rate Performance for (128,64) Codes
Fig. 6 shows the BLER performance of (128,64) polar codes
under SC-based decoding. In particular, we compare plain SC
decoding [1] with SCL with list size 8 (SCL-8) decoding [2]
and Aut-8-SC decoding [10]. Furthermore, ordered statistic