BCH码的低时延分阶统计译码

Low-Latency Ordered Statistics Decoding of BCH Codes

◼ 陈立 教授

◼ 中山大学 电子与信息工程学院

1
Outline

◼ Background

◼ BCH Codes

◼ Ordered Statistics Decoding

◼ Low-Latency OSD

◼ Hybrid Soft Decoding

2
Background
◼ Future scenarios featured by URLLC

Unmanned Driving Wise Information Technology of Med (WITMED)

Factory Automation Extended Reality (XR)

Short-to-medium length codes

3
Background
◼ Capacity-approaching codes

Turbo codes LDPC codes Polar codes

𝑛 is large → randomness & iterative decoding, channel polarization

◼ Good short-to-medium length codes

BCH codes TBCC Polar codes PAC codes

(OSD) （Circular Viterbi） (SCL) (Fano)

𝑛 is short-to-medium → near ML decoding, list decoding

4
Background
◼ Near ML decoding performance: rate 𝑅 = 1/2 & length 𝑛 = 128 [1]

0.1dB

[1] M. Shirvanimoghaddam et al., “Short block-length codes for ultrareliable low latency communications,” IEEE Commun. Mag., vol. 57, no. 2, pp.
130–137, Feb. 2019. 5
BCH Codes
◼ Encoding of BCH codes
Primitive element of 𝔽2𝑚 : 𝜎
𝑛, 𝑘 binary primitive BCH code Codeword length: 𝑛 = 2𝑚 − 1
𝒞BCH [𝑛, 𝑘] Designed distance: 𝑑 = 2𝑡 + 1

Generator polynomial: 𝑔 𝑥 = 𝑔0 + 𝑔1 𝑥 + ⋯ + 𝑔𝑛−𝑘 𝑥 𝑛−𝑘 ∈ 𝔽2 [𝑥] is the minimal (deg 𝑔 𝑥 )

nonzero polynomial such that
𝑔 𝜎 = 𝑔 𝜎 2 = ⋯ = 𝑔 𝜎 2𝑡 = 0

Message 𝒇 = 𝑓0 , 𝑓1 , … , 𝑓𝑘−1 ∈ 𝔽𝑘2 , in poly. 𝑓(𝑥) = 𝑓0 + 𝑓1 𝑥 + ⋯ + 𝑓𝑘−1 𝑥 𝑘−1 ∈ 𝔽2 [𝑥]

Codeword 𝒄 = (𝑐0 , 𝑐1 , … , 𝑐𝑛−1 ) ∈ 𝔽𝑛2 , in poly. 𝑐 𝑥 = 𝑐0 + 𝑐1 𝑥 + ⋯ + 𝑐𝑛−1 𝑥 𝑛−1 = 𝑓(𝑥)𝑔 𝑥 ∈ 𝔽2 [𝑥]

Parity-check condition: 𝑐(𝜎) = 𝑐(𝜎 2 ) = ⋯ = 𝑐(𝜎 2𝑡 ) = 0

BCH Codes
◼ Encoding of BCH codes
Generator matrix:
𝑔0 𝑔1 … 𝑔𝑛−𝑘 0 … 0
0 𝑔0 … … 𝑔𝑛−𝑘 ⋱ ⋮
𝐆= ∈ 𝔽𝑘×𝑛
2
⋮ ⋱ ⋱ ⋱ ⋱ ⋱ 0
0 … 0 𝑔0 𝑔1 … 𝑔𝑛−𝑘

Encoding process: 𝒄 = 𝒇 ⋅ 𝐆
Parity-check matrix:
(𝜎 )0 (𝜎 )1 ⋯ (𝜎 )𝑛−1
2 0
𝐇 = (𝜎 ) (𝜎 2 )1 ⋯ (𝜎 2 )𝑛−1 ∈ 𝔽2𝑡×𝑛 Parity-check matrix of
2𝑚
⋮ ⋮ ⋱ ⋮ the 𝑛, 𝑛 − 2𝑡 RS code
(𝜎 )0
2𝑡 (𝜎 )1
2𝑡 ⋯ (𝜎 )𝑛−1
2𝑡

Parity-check equation:

𝑐 𝜎 = 𝑐 𝜎 2 = ⋯ = 𝑐 𝜎 2𝑡 = 0 ⇔ 𝒄 ⋅ 𝐇T = 𝟎

7
BCH Codes
◼ Decoding of BCH codes

Hard-decision decoding (utilizing algebraic structure)

Berlekamp-Massey (BM)
Efficient – algebra fluency & hardware friendly
Euclidean
Limited performance
Guruswami-Sudan (GS)

Soft-decision decoding (utilizing soft information)

Ordered Statistics Decoding (OSD) High performance (~ ML)

Algebraic Chase decoding (Chase-BM, Chase-GS) High complexity & latency

8
Soft-Decision Decoding
Unreliable information Reliable information
Reliability
(LLR)
Decoding Re-encoding

Algebraic structure Information set

Equation system Systematic generator matrix

(Syndrome-based or interpolation-based)

Low-latency OSD: use 𝐆RS to replace 𝐆BCH

Hybrid soft decoding: use algebraic Chase decoding to identify test error patterns (TEPs)

9
Ordered Statistics Decoding
◼ Overview of OSD

Systematic Generator matrix 𝐆BCH

Codeword Optimal
LLRs GE Encoding
candidates codeword

MRIPs Test messages

TEPs
• GE: Gaussian elimination
• LLRs: log likelihood ratios
• MRIPs: most reliable independent positions
• TEPs: test error patterns

10
Ordered Statistics Decoding
◼ Construct the systematic generator matrix
Channel

• Received LLR sequence: 𝑳 = (𝐿0 , 𝐿1 , … , 𝐿𝑛−1 ) ∈ ℝ𝑛

• Hard-decision received word: 𝒓 = (𝑟0 , 𝑟1 , … , 𝑟𝑛−1 ) ∈ 𝔽𝑛2

Sort based on reliability and determine the most reliable independent positions (MRIPs)

𝒓′ = Λ(𝒓) = (𝑟𝑗0 , 𝑟𝑗1 , … , 𝑟𝑗𝑛−1 ൯ |𝐿𝑗0 | ≥ |𝐿𝑗1 | ≥ ⋯ ≥ |𝐿𝑗𝑛−1 |

Gaussian
Construct the systematic generator matrix 𝐆BCH elimination

𝐆 𝐆′ = Λ(𝐆) 𝐆BCH

11
Ordered Statistics Decoding
◼ Generate BCH candidate codewords
Re-encoding process
Initial message: 𝒇(0) = (𝑟𝑗0 , 𝑟𝑗1 , … , 𝑟𝑗𝑘−1 ) ∈ 𝔽𝑘2
(𝜔) (𝜔) (𝜔)
Test error pattern: 𝒆(𝜔) = (𝑒𝑗0 , 𝑒𝑗1 , … , 𝑒𝑗𝑘−1 ) ∈ 𝔽𝑘2 𝜔 = 0,1, … , 𝑁TEPs − 1

(𝜔) (𝜔) (𝜔)

Test message: 𝒇(𝜔) = (𝑓𝑗0 , 𝑓𝑗1 , … , 𝑓𝑗𝑘−1 ) ∈ 𝔽𝑘2

𝜔 𝜔 𝜔
Repetitive
BCH codeword candidate: 𝒄ො (𝜔) = (𝑐0Ƹ , 𝑐1Ƹ , … , 𝑐𝑛−1
Ƹ ) = Λ−1 (𝒇 𝜔 ⋅ 𝐆BCH) ∈ 𝔽𝑛2 processing
Number of TEPs (= number of BCH codeword candidates) (parallel)
𝑑H 𝒆 𝜔 ,𝟎 ≤ 𝜏
OSD with order 𝜏

𝑘 𝑘 𝑘
𝑁TEPs = + + ⋯+
0 1 𝜏
𝜔 𝜔
Number of TEPs 𝒆 s.t. 𝑑H 𝒆 ,𝟎 = 1

12
Ordered Statistics Decoding
◼ Challenges of OSD

Reduction of TEPs: skipping rule; stopping rule

Exponential complexity Trade off with performance: multiple bases; validation rule

𝜏–OSD complexity: 𝑂(𝑘 𝜏 ) Trade off with storage: box-and-match algorithm

Offline computed systematic generator matrices

Decoding latency bottleneck Only an auxiliary method

GE
Alternative solution? Utilize algebra of the code

13
Low-Latency OSD
◼ BCH codes and RS codes
Subfield subcode

Given two linear codes 𝒞 ⊂ 𝔽𝑛2 and 𝒞 ′ ⊂ 𝔽𝑛2𝑚 , if 𝒞 = 𝒞 ′ ∩ 𝔽𝑛2 , 𝒞 is called the subfield subcode
of 𝒞 ′ over 𝔽2 .

Lemma 1
An (𝑛, 𝑘) t-error-correcting BCH code defined over 𝔽2 is a subfield subcode of an 𝑛, 𝑘 ′ t-
error-correcting RS code defined over 𝔽2𝑚 . i.e., 𝒞BCH [𝑛, 𝑘, 2𝑡 + 1] = 𝒞RS [𝑛, 𝑘 ′ , 2𝑡 + 1] ∩ 𝔽𝑛2

BCH code 𝑑 = 2𝑡 + 1 < 𝑛 − 𝑘 + 1

Dimension: 𝑘 < 𝑘 ′
RS code 𝑑 = 2𝑡 + 1 = 𝑛 − 𝑘′ + 1 (MDS)
𝒞RS 𝒞BCH 𝔽𝑛2
Example 1
The (7, 4) BCH code is the binary subcode of the 7, 5 RS code

14
Low-Latency OSD
◼ Basic idea of low-latency OSD (LLOSD)

GE (sequential)
𝐆BCH
Test message Candidates

OSD

LLOSD

Test message 𝐆RS Candidates

Lagrange interpolation (parallel)

15
Low-Latency OSD
◼ Generation of GRS
Encoding of an 𝑛, 𝑘 ′ RS code
𝑘𝑚′ ′ −1
Message 𝒖 = 𝑢0 , 𝑢1 , … , 𝑢𝑘 ′ −1 ∈ 𝔽2 , in poly. 𝑢(𝑥) = 𝑢0 + 𝑢1 𝑥 + ⋯ + 𝑢𝑘 ′ −1 𝑥 𝑘 ∈ 𝔽2𝑚 [𝑥]

Codeword 𝒗 = 𝑢(𝛼0 ), 𝑢(𝛼1 ), … , 𝑢(𝛼𝑛−1 ) ∈ 𝔽𝑛2𝑚

𝑣0 𝑣1 … 𝑣𝑛−1
𝛼0 , 𝛼1 , …, 𝛼𝑛−1 ∈ 𝔽2𝑚 \{0} are the code locators

◼ Example 2 For a (7, 5) RS code over 𝔽8

Message 𝒖 = 2, 3, 5, 0, 1 , in poly. 𝑢 𝑥 = 2 + 3𝑥 + 5𝑥2 + 𝑥4

Codeword 𝒗 = 𝑢 1 , 𝑢(2), 𝑢(4), 𝑢(3), 𝑢(6), 𝑢(7), 𝑢(5) = (5, 0, 4, 7, 6, 1, 3)

𝔽8 is an extension field of 𝔽2 , defined by 𝑝(𝑥) = 𝑥 3 + 𝑥 + 1

In 𝔽8 , {0, 𝜎 0 , 𝜎 1 , 𝜎 2 , 𝜎 3 , 𝜎 4 , 𝜎 5 , 𝜎 6 } = {0, 1, 2, 4, 3, 6, 7, 5}

16
Low-Latency OSD
◼ Generation of GRS
Determine the MRIPs

Permuted received word: 𝒓′ = Λ (𝒓) = (𝑟𝑗0 , 𝑟𝑗1 , … , 𝑟𝑗𝑛−1 ) |𝐿𝑗0 | ≥ |𝐿𝑗1 | ≥ ⋯ ≥ |𝐿𝑗𝑛−1 |

Most reliable positions (MRPs): MRIPs

Θ = {𝑗0 , 𝑗1 , … , 𝑗𝑘 ′ −1 ൟ
𝑗0 𝑗1 … 𝑗𝑘−1 … 𝑗𝑘′ −1 𝑗𝑘 ′ … 𝑗𝑛−1

Θc = {𝑗𝑘 ′ , 𝑗𝑘 ′ +1 , … , 𝑗𝑛−1 ൟ
Θ (MRPs) Θc

MDS property MRPs are linearly independent

For RS codes: MRPs MRIPs

17
Low-Latency OSD
◼ Generation of GRS
′
Message 𝒖 = (𝑟𝑗0 , 𝑟𝑗1 , … , 𝑟𝑗′
) ∈ 𝔽𝑘2𝑚 , defined by Θ = {𝑗0 , 𝑗1 , … , 𝑗𝑘 ′ −1 ൟ
𝑘 −1

Construct the Lagrange interpolation polynomials

𝑥 − 𝛼𝑗 ′
𝑇𝑗 𝑥 = ෑ
𝛼𝑗 − 𝛼𝑗 ′
𝑗 ′ ∈Θ,𝑗 ′ ≠𝑗

where 𝑇𝑗 (𝛼𝑗 ) = 1, and 𝑇𝑗 (𝛼𝑗 ′ ) = 0 if 𝑗 ′ ∈ Θ and 𝑗 ′ ≠ 𝑗.

Form the systematic message polynomial of 𝒖 = (𝑟𝑗0 , 𝑟𝑗1 , … , 𝑟𝑗𝑘′−1 )

ℋ𝒖 (𝑥) = ෍ 𝑟𝑗 𝑇𝑗 (𝑥)
𝑗∈Θ

systematic message symbols 𝑟𝑗 , if 𝑗 ∈ Θ;

ℋ𝒖 (𝛼𝑗 ) generates
parity-check symbols, if 𝑗 ∈ Θc.

18
Low-Latency OSD
◼ Example 3
Given a (7, 4) BCH code and an LLR sequence 𝑳 = (−2.447, 5.115, −4.771, −1.349, −7.096, 0.443, −3.485)

Mother code: (7, 5) RS code

Hard-decision received word: 𝒓 = (1, 0, 1, 1, 1, 0, 1)

MRPs: Θ = 4, 1, 2, 6, 0
Lagrange interpolation polynomials:
𝑥 − 𝛼1 𝑥 − 𝛼2 𝑥 − 𝛼6 𝑥 − 𝛼0 𝑥 − 𝛼4 𝑥 − 𝛼2 𝑥 − 𝛼6 𝑥 − 𝛼0
𝑇4 (𝑥) = ⋅ ⋅ ⋅ 𝑇1 (𝑥) = ⋅ ⋅ ⋅
𝛼4 − 𝛼1 𝛼4 − 𝛼2 𝛼4 − 𝛼6 𝛼4 − 𝛼0 𝛼1 − 𝛼4 𝛼1 − 𝛼2 𝛼1 − 𝛼6 𝛼1 − 𝛼0
𝑥 − 𝛼4 𝑥 − 𝛼1 𝑥 − 𝛼6 𝑥 − 𝛼0 𝑥 − 𝛼4 𝑥 − 𝛼1 𝑥 − 𝛼2 𝑥 − 𝛼0
𝑇2 (𝑥) = ⋅ ⋅ ⋅ 𝑇6 (𝑥) = ⋅ ⋅ ⋅
𝛼2 − 𝛼4 𝛼1 − 𝛼2 𝛼2 − 𝛼6 𝛼2 − 𝛼0 𝛼6 − 𝛼4 𝛼6 − 𝛼1 𝛼6 − 𝛼2 𝛼6 − 𝛼0

𝑥 − 𝛼4 𝑥 − 𝛼1 𝑥 − 𝛼2 𝑥 − 𝛼6
𝑇0 (𝑥) = ⋅ ⋅ ⋅
𝛼0 − 𝛼4 𝛼0 − 𝛼1 𝛼0 − 𝛼2 𝛼0 − 𝛼6

19
Low-Latency OSD
◼ Generation of GRS

𝒖𝑗0 = (1, 0, 0, … , 0) codeword of 𝒖𝑗0

𝒖𝑗1 = (0, 1, 0, … , 0) codeword of 𝒖𝑗1
𝐆RS =
⋮
⋮
codeword of 𝒖𝑗 ′
𝒖𝑗 = (0, 0, 0, … , 1) 𝑘 −1
𝑘′ −1

1 0 ⋯ 0 ℋ𝒖𝑗 (𝛼𝑗 ′ ） ⋯ ℋ𝒖𝑗 (𝛼𝑗𝑛−1 ）

0 𝑘 0
0 1 ⋯ 0 ℋ𝒖𝑗 (𝛼𝑗 ′ ） ⋯ ℋ𝒖𝑗 (𝛼𝑗𝑛−1 ）
Λ(𝐆RS ) = 1 𝑘 1
⋮ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮
0 0 ⋯ 1 ℋ𝒖𝑗 (𝛼𝑗 ′ ） ⋯ ℋ𝒖𝑗 (𝛼𝑗𝑛−1 ）
𝑘′ −1 𝑘 𝑘′ −1

𝑘′× 𝑘′ 𝛼𝑗0 𝛼𝑗1 ⋯ 𝛼𝑗 ′ 𝛼𝑗 ⋯ 𝛼𝑗𝑛−1

𝑘 −1 𝑘′
identity submatrix
Θ (MRPs) Θc

20
Low-Latency OSD
◼ Generation of GRS

The row-𝑖 column-𝑗 entry of 𝐆RS ∶

𝑛−1

ෑ 𝛼𝑗 = 1
ෑ 𝛼𝑗 − 𝛼𝑗 ′ 𝑗=0 𝛼𝑖 ෑ 𝛼𝑖 − 𝛼𝑗 ′
𝑗 ′ ∈Θ 𝑗 ′ ∈Θc ,𝑗 ′ ≠𝑗
ℋ𝒖𝑖 (𝛼𝑗 ) = ℋ𝒖𝑖 (𝛼𝑗 ) =
𝛼𝑗 − 𝛼𝑖 ෑ 𝛼𝑖 − 𝛼𝑗 ′ 𝛼𝑗 ෑ 𝛼𝑗 − 𝛼𝑗 ′
𝑗 ′ ∈Θ,𝑗≠𝑖 𝑗 ′ ∈Θc ,𝑗 ′ ≠𝑗

where 𝑗 ∈ Θc
|Θ| = 𝑘 ′ |Θc | = 𝑛 − 𝑘 ′

Relationship: systematic generator matrix & systematic parity-check matrix

Complexity of the generation of 𝐆RS : 2𝑛 ⋅ min{𝑛 − 𝑘 ′ , 𝑘 ′ }

Low-Latency OSD
◼ Generation of GRS

Example 4 continues from Example 3 with MRPs being Θ = 4, 1, 2, 6, 0

𝑥 − 𝛼1 𝑥 − 𝛼2 𝑥 − 𝛼6 𝑥 − 𝛼0
𝒖4 = (1, 0, 0, 0, 0) ℋ𝒖4 (𝑥) = ⋅ ⋅ ⋅
𝛼4 − 𝛼1 𝛼4 − 𝛼2 𝛼4 − 𝛼6 𝛼4 − 𝛼0
𝑥 − 𝛼4 𝑥 − 𝛼2 𝑥 − 𝛼6 𝑥 − 𝛼0 0 0 0 5 1 3 0
𝒖1 = (0, 1, 0, 0, 0) ℋ𝒖1 (𝑥) = ⋅ ⋅ ⋅
𝛼1 − 𝛼4 𝛼1 − 𝛼2 𝛼1 − 𝛼6 𝛼1 − 𝛼0
𝑥 − 𝛼4 𝑥 − 𝛼1 𝑥 − 𝛼6 𝑥 − 𝛼0
0 1 0 4 0 2 0
𝒖2 = (0, 0, 1, 0, 0) ℋ𝒖2 (𝑥) = ⋅ ⋅ ⋅ 𝐆RS = 0 0 1 1 0 1 0
𝛼2 − 𝛼4 𝛼1 − 𝛼2 𝛼2 − 𝛼6 𝛼2 − 𝛼0
𝑥 − 𝛼4 𝑥 − 𝛼1 𝑥 − 𝛼2 𝑥 − 𝛼0
𝒖6 = (0, 0, 0, 1, 0) ℋ𝒖6 (𝑥) = ⋅ ⋅ ⋅
𝛼6 − 𝛼4 𝛼6 − 𝛼1 𝛼6 − 𝛼2 𝛼6 − 𝛼0
0 0 0 4 0 3 1
𝒖0 = (0, 0, 0, 0, 1) ℋ𝒖0 (𝑥) =
𝑥 − 𝛼4 𝑥 − 𝛼1 𝑥 − 𝛼2 𝑥 − 𝛼6
⋅ ⋅ ⋅
1 0 0 5 0 2 0
𝛼0 − 𝛼4 𝛼0 − 𝛼1 𝛼0 − 𝛼2 𝛼0 − 𝛼6

𝛼3 − 𝛼1 𝛼3 − 𝛼2 𝛼3 − 𝛼6 𝛼3 − 𝛼0
ℋ𝒖4 𝛼3 = ⋅ ⋅ ⋅ =5 Locators : 𝛼0 𝛼1 𝛼2 𝛼3 𝛼4 𝛼5 𝛼6
𝛼4 − 𝛼1 𝛼4 − 𝛼2 𝛼4 − 𝛼6 𝛼4 − 𝛼0
&
𝛼4 (𝛼4 −𝛼5 )
ℋ𝒖4 𝛼3 = =5
𝛼3 (𝛼3 −𝛼5 )

22
Low-Latency OSD
◼ Generation of BCH codeword candidates

Generate the initial codeword:

′
Initial message: 𝒖 = (𝑟𝑗0 , 𝑟𝑗1 , … , 𝑟𝑗 ) ∈ 𝔽𝑘2
𝑘′ −1
(0) (0) (0) (0)
Initial RS codeword: ෝ(0) = (𝑣ො0 , 𝑣ො1 , … , 𝑣ො𝑛−1 ) = 𝒖 ⋅ 𝐆RS
𝒗 𝑣ො𝑗 = 𝑟𝑗 , ∀𝑗 ∈ Θ

Generate any 𝑛, 𝑘 ′ systematic RS codeword:

𝜔 𝜔 𝜔 𝜔 ′
The 𝜔-th test error pattern: 𝒆′ = (𝑒 ′𝑗0 , 𝑒 ′𝑗1 , … , 𝑒 ′𝑗 ) ∈ 𝔽𝑘2
𝑘′ −1

𝜔 𝜔
The 𝜔-th test message: 𝒖 = 𝒖 + 𝒆′
(𝜔) (𝜔) (𝜔)
The 𝜔-th systematic RS codeword: ෝ
𝒗 𝜔
= (𝑣ො0 , 𝑣ො1 , … , 𝑣ො𝑛−1 ) ∈ 𝔽𝑛2𝑚

𝜔 𝜔 𝜔
ෝ
𝒗 =𝒖 ෝ(0) + 𝒆′
⋅ 𝐆RS = 𝒗 ⋅ 𝐆RS

23
Low-Latency OSD
◼ Generation of BCH codeword candidates
Theorem 2 (Identify invalid TEPs)
(0)
If 𝑣ො𝑗 +σ ′(𝜔) ℋ𝒖𝑖 (𝛼𝑗 ) ∈ {0,1}, ∀𝑗 ∈ Θc , 𝒗
ෝ 𝜔 is a BCH codeword.
𝑖∈Θ,𝑒𝑖 ≠0
Θc

Example 5 continues from Example 4 0 0 0 5 1 3 0

𝒖 = (1, 0, 1, 1, 1) 0 1 0 4 0 2 0
Initial message : 𝐆RS = 0 0 1 1 0 1 0
Initial RS codeword: ෝ(0) = 𝒖 ⋅ 𝐆RS = (1, 0, 1, 5, 1, 3, 1)
𝒗 0 0 0 4 0 3 1
1 0 0 5 0 2 0
1
Test error pattern 1: 𝒆′ = (0, 0, 0, 0, 1)
+
Systematic RS codeword 1: ෝ(1) = 𝒗
𝒗 ෝ(0) + 𝒆′ 1
⋅ 𝐆RS = (0, 0, 1, 0, 1, 1, 1) BCH codeword
candidate
3
Test error pattern 3: 𝒆′ = (0, 0, 1, 0, 0)
ෝ(3) = 𝒗
Systematic RS codeword 3: 𝒗 ෝ(0) + 𝒆′ 3
⋅ 𝐆RS = (1, 0, 0,×, 1,×, 1) Invalid

24
Low-Latency OSD
◼ Overview of LLOSD

Sequential Parallel

Gaussian Codeword OSD

elimination candidates

The optimal
LLRs
codeword

Lagrange Codeword
polynomials candidates
LLOSD

Parallel Parallel

25
Low-Latency OSD
◼ Complexity comparison of LLOSD and OSD

Algorithms Operations Complexity

GE 𝑛 ⋅ (min{𝑛 − 𝑘, 𝑘})2 Key differences
Compute 𝒄ො (0) 𝑘 ⋅ (𝑛 − 𝑘)
OSD (𝜏) Operation type : binary → finite field
Compute 𝒄ො (𝜔) (𝑛 − 𝑘) ⋅ σ𝜏𝜆=1 𝜆 𝑘𝜆

Find 𝒄ො opt 𝑛 ⋅ σ𝜏𝜆=0 𝑘 Compute 𝐆BCH / 𝐆RS : 𝑂 𝑛3 → 𝑂 𝑛2

𝜆

Compute GRS 2𝑛 ⋅ min{𝑛 − 𝑘 ′ , 𝑘 ′ } sequential → parallel

ෝ(0)
Compute 𝒗 𝑘 ′ ⋅ (𝑛 − 𝑘 ′ )
𝑘
LLOSD (𝜏) Candidate list : σ𝜏𝜆=0 → 𝑁𝑛−𝑘 ′ = 𝑁BCH
𝑛−𝑘 ′ 𝜆
Compute 𝒄ො (𝜔) σ𝜏𝜆=1 𝑘𝜆 + 𝜏σ 𝑗 ′ =1 𝑁𝑗 ′
ෝopt
Find 𝒗 𝑛𝑁𝑛−𝑘 ′

𝑁𝑗 ′ is the number of TEPs that yield binary symbols after the 𝑗 ′ th judgement as in Theorem 2. 26
 Low-Latency OSD
◼ Performance of the (63, 45) BCH code, AWGN, BPSK
1.E+00
Complexity and latency comparison

1.E-01

1.E-02
FER

1.E-03 OSD (1)

LLOSD (1)
LLOSD (2)
1.E-04
LLOSD (3) Simulation environment: Intel core i7-10710U CPU
ML Stopping rule: ML criterion [2]
1.E-05
1 2 3 4 5 6
SNR (dB)

[2] T. Kaneko et al., “An efficient maximum-likelihood-decoding algorithm for linear block codes with algebraic decoder,” IEEE Trans. Inf. Theory, vol. 40, no. 2, pp.
320–327, 1994. 27
 Low-Latency OSD
◼ Re-encoding complexity distribution: LLOSD (3) of the (63,45) BCH code with ML criterion,
SNR = 5.0 dB

Reasons for Multiple Peaks

For the re-encoding phase of TEPs with
a larger weight
• The complexity is greater
• The ML criterion is more difficult to
satisfy
• The probability of finding a BCH
• Colors are used to distinguish the phases that the re-encoding terminates at
• Phase (𝑖): the re-encoding phase for weight-𝑖 TEPs codeword is lower

28
Low-Latency OSD
◼ Segmented variation of LLOSD

Analysis

• The error probability of the LLOSD is upper bounded by

𝑃e,LLOSD 𝜏 ≤ 𝑃e,ML + 𝑃list (𝜏)

• 𝑃e,ML : ML decoding error probability,

depends on the weight distribution of the code

• 𝑃list (𝜏): List decoding error probability,

= Prob{the channel introduces more than 𝜏 errors in MRPs}
𝑗0 𝑗1 … 𝑗𝑘−1 … 𝑗𝑘′ −1 𝑗𝑘 ′ … 𝑗𝑛−1

Θ (MRPs) Θc

29
Low-Latency OSD
◼ Segmented variation of LLOSD
For OSD

𝑑
If 𝜏 ≥ min − 1 ,𝑘 The list error probability of OSD: 𝑃list 𝜏 ≪ 𝑃e,ML
4
MRIPs
The OSD can produce ~ ML decoding performance
𝑗0 𝑗1 … 𝑗𝑘−1 … 𝑗𝑘′ −1 𝑗𝑘 ′ … 𝑗𝑛−1
We only need the decoding order of 𝜏 for the MRIPs
Θ (MRPs) Θc
For LLOSD
𝜔 𝜔 𝜔 𝜔 𝜔 𝜔 𝜔
TEPs: 𝒆′ = (𝑒 ′𝑗0 , 𝑒 ′𝑗1 , … , 𝑒 ′𝑗𝑘−1 , 𝑒 ′𝑗𝑘 , 𝑒 ′𝑗𝑘+1 , … , 𝑒 ′𝑗 )
𝑘′ −1

𝑑
𝜏1 = min − 1 , 𝑘 is sufficient extra order 𝜏2
4

𝜏 𝑘′ 𝜏1 𝑘 𝜏2 𝑘′ − 𝑘
Number of TEPs 𝑁TEPs : ෍ ෍ ⋅෍
𝑖=0 𝑖 𝑖1 =0 𝑖1 𝑖2 =0 𝑖2

30
Low-Latency OSD
◼ Performance of the (63, 45) BCH code, AWGN, BPSK

1.E+00
Complexity and latency comparison

1.E-01

1.E-02
FER

OSD (1)

LLOSD (1)
1.E-03
LLOSD (2)

LLOSD (3)
1.E-04
Seg. LLOSD (1| 45, 3)

ML Simulation environment: Intel core i7-10710U CPU

1.E-05
1 2 3 4 5 6 Stopping rule: ML criterion
SNR (dB)

31
Low-Latency OSD
◼ Concatenated perspective

Systematic parity-check matrix 𝒞RS [𝑛, 𝑘 ′ ] Systematic generator matrix

Λ(𝐇RS ) = 𝐏 𝐈𝑛−𝑘 ′ Λ(𝐆RS ) = 𝐈𝑘 ′ 𝐏T
1. 𝔽2𝑚 ≅ 𝔽𝑚
2
(𝟎) systematic parity-check matrix of an 𝑛, 𝑘 ′
2. Row permutation 𝐏 𝐈𝑛−𝑘 ′ :
𝐏 0
𝐈𝑛−𝑘 ′ binary code 𝒞 ∗ [𝑛, 𝑘 ′ ]

Λ(𝐇BCH ) = 𝐏 1
𝐎 (over 𝔽2 ) 𝒞BCH [𝑛, 𝑘] punctured on the positions in Θc
1
⋮ ⋮ 𝐏 c
Θ
𝐏 𝑚−1
𝐎 𝐏∗ = ⋮ : parity-check matrix of 𝒞BCH [𝑘 ′ , 𝑘]
𝑚−1
𝐏
If 𝒖(𝜔) ⋅ 𝐏 ∗ T = 𝟎, let 𝒄ො (𝜔) = Λ−1 (𝒖 𝜔
⋅ 𝐈𝑘 ′ 0 T )
𝐏
𝜔 T
⟹ 𝒄ො ⋅ 𝐇BCH = 𝟎 ⟺ 𝒄ො (𝜔) ∈ 𝒞BCH
′ c
𝒖(𝜔) ∈ 𝔽𝑘2 Θ
𝒖(𝜔) ∈ 𝒞BCH 𝒄ො (𝜔) = 𝒖(𝜔) ⋅ 𝐆RS ∈ 𝒞BCH
Parity-checker of punctured Systematic encoder of RS code
c
Θ
BCH code 𝒞BCH [𝑘 ′ , 𝑘] (𝜔) ∗T 𝒞RS [𝑛, 𝑘 ′ ] with information set Θ
(𝒖 ⋅𝐏 = 𝟎)

32
 Low-Latency OSD
◼ Concatenated perspective
′ c
𝒄ො (𝜔) = 𝒖 𝜔
⋅ Λ−1 0 T
𝒖(𝜔) ∈ 𝔽𝑘2 Θ
𝒖(𝜔) ∈ 𝒞BCH 𝐈𝑘 ′ 𝐏
Parity-checker of punctured Systematic encoder of binary code
c
Θ
BCH code 𝒞BCH [𝑘 ′ , 𝑘] (𝜔) ∗T 𝒞 ∗ [𝑛, 𝑘 ′ ] with information set Θ
(𝒖 ⋅𝐏 = 𝟎)

Example 6
1 1 1 0 0 1 0 1 1 1 0 0 1 0
1 0 1 1 1 0 0 0 1 1 1 0 0 1
3
𝜎3 𝔽2𝑚 ≅𝔽𝑚 0 0 0 0 0 0 0 row perm. 1 0 1 1 1 0 0
= 𝜎4 1 𝜎 𝜎 1 0 2
𝐇RS = 𝐇BCH
𝜎 1 𝜎5 𝜎5 𝜎4 0 1 0 1 1 1 0 0 1 1 0 1 1 1 0 0
1 0 1 1 1 0 0 0 0 0 0 0 0 0
1 0 1 1 1 0 0 1 0 1 1 1 0 0
(1) 𝒖(𝜔) = 1, 0, 1, 1, 0 : 𝒖(𝜔) ⋅ 𝐏 ∗ T = (1,1,0,1)
(𝜔)
𝑐5Ƹ = 𝒖(𝜔) ⋅ 1, 1, 1, 0, 0 T
=0
(2) 𝒖 (𝜔)
= 1, 0, 1, 0, 0 : 𝒖 (𝜔)
⋅𝐏 ∗T
=𝟎 ቐ (𝜔)
⟹ 𝒄ො (𝜔) = (1, 0, 1, 1, 0, 0, 1)
𝑐6Ƹ = 𝒖(𝜔) ⋅ 0, 1, 1, 1, 0 T
=1

33
Low-Latency OSD
◼ Concatenated perspective
For LLOSD (𝜏) 𝜏
𝑘′
𝑁TEPs : number of TEPs 𝑁TEPs =෍
𝑖
𝑖=0
𝑁BCH : number of BCH codeword candidates:
Theorem 3
If the channel condition is sufficiently good (SNR ⟶ ∞)
𝜏

⟹ 𝑁BCH = ෍ 𝐴𝑖
𝑖=0
Θc
𝐴𝑖 : number of weight-𝑖 codewords in 𝒞BCH [𝑘 ′ , 𝑘]
• The set of punctured positions Θc varies for each decoding event, and 𝐴𝑖 varies accordingly

In general 𝜏 𝜏
𝑘′
෍ 𝐴𝑖 ≪ ෍
𝑖
𝑖=0 𝑖=0

34
Low-Latency OSD
◼ Concatenated perspective
The number of BCH codeword candidates 𝑁BCH in decoding the 𝒞BCH [63, 45] and 𝒞BCH [31, 21]

Theoretical line
Randomly puncturing 𝑛 − 𝑘 ′ positions of
Θ 𝑐
𝒞BCH [𝑛, 𝑘], obtaining 𝒞BCH [𝑘 ′ , 𝑘] and its
weight distribution {𝐴𝑖 }

𝑁TEPs = 30914

𝑁TEPs = 379

35
Low-Latency OSD
◼ The average number of BCH codeword candidates in LLOSD (3) of the (63,45) BCH code

Remarks
• The probability of finding a BCH codeword decreases as the TEP weight increases
• As the SNR increases, the likelihood of generating a BCH codeword using a weight-0 TEP
increases, while the likelihood decreases when using a TEP of a larger weight

36
 Hybrid Soft Decoding
◼ LLOSD vs. Chase - BM / GS
It remains challenging to decode longer BCH codes with LLOSD

GE-free and low-latency Efficient and hardware friendly

High decoding performance LLOSD Chase-BM/GS Good for long codes

Higher decoding complexity Difficult to approach ML

MRPs & re-encoding LRPs & test-vector decoding

LRPs: Least reliable positions 37

Hybrid Soft Decoding
◼ Integrating LLOSD and Chase-GS decoding
LLOSD
𝐆RS Yes Optimal
LLRs TEPs Re-encoding ML
codeword

TEPs No
Chase-GS decoding

Advantages

◼ LLOSD output list can be utilized

◼ Computation sharing
◼ Low complexity root-finding

38
Hybrid Soft Decoding
◼ Integrating LLOSD and Chase-GS decoding
LLOSD
𝐆RS Yes Optimal
LLRs TEPs Re-encoding ML
codeword

TEPs No
Chase-GS decoding

Test-vector Re-encoding
Skipping Interpolation Root-finding
formulation transform

candidate list ෝ(0)

𝒗 Module basis 𝐆RS
Module basis
construction reduction

Lagrange polynomials

39
Hybrid Soft Decoding
◼ Key steps of HSD
Ψ (LRPs)
Test-vector formulation
𝑗0 𝑗1 … 𝑗𝑘′ −1 𝑗𝑘 ′ … 𝑗𝑛−𝜂 𝑗𝑛−𝜂+1 … 𝑗𝑛−1
Ψ = { 𝑗𝑛−𝜂 , 𝑗𝑛−𝜂+1 , … , 𝑗𝑛−1 }: 𝜂 least reliable positions (LRPs)
Θ (MRPs) Θc
The 2𝜂 test-vectors can be formulated as
00…000
00…001
𝜔 𝜔
Λ(𝒓𝜔 ) = (𝑟𝑗0 , 𝑟𝑗1 , … , 𝑟𝑗𝑛−𝜂−1 , 𝑟𝑗𝑛−𝜂 , … , 𝑟𝒋𝑛−1 ) 00…010 0 ≤ 𝜔 ≤ 2𝜂 − 1
00…011
…
11…110
11…111 𝒓𝜔

Skipping (based on candidate list) ෝ

𝒗
𝒓𝜔 𝑡
If 𝑑H (ෝ ෝ
𝒗, 𝒓𝜔 ) ≤ 𝑡, 𝒓𝜔 will be decoded as 𝒗 𝒓𝜔 can be skipped Hamming
sphere

𝑑H (ෝ ෝ and test-vector 𝒓𝜔
𝒗, 𝒓𝜔 ): Hamming distance between 𝒗
40
Hybrid Soft Decoding
◼ Key steps of HSD

Re-encoding transform (based on the initial RS codeword)

0
ෝ
Initial RS codeword (from LLOSD): 𝒗 = 𝒖 ⋅ 𝐆RS ∈ 𝔽𝑛2𝑚
0 𝜔 𝜔
ෝ
Transformed test vector: 𝒛𝜔 = 𝒓𝜔 − 𝒗 Λ(𝒛𝜔 ) = (0, 0, … , 0, 𝑧𝑗 , … , 𝑧𝑗𝑛−1 )
𝑘′

Finding a Grö bner basis of the interpolation module

𝜔
ℳ𝜔 = {𝑄 ∈ 𝔽2𝑚 𝑥, 𝑦 ∣ deg 𝑦 𝑄 ≤ 1; 𝑄 𝛼𝑗 , 𝑧𝑗 /𝑉 𝛼𝑗 = 0, ∀𝑗 ∈ Θc }
Interpolation 一 module basis construction (based on the Lagrange polynomials)
Interpolation points： 𝑧𝑗
𝜔 𝜔
𝑘′
𝑧𝑗𝑛−1
（𝛼𝑗0 , 0), (𝛼𝑗1 , 0), … , (𝛼𝑗 , 0), (𝛼𝑗 ′ , ) … , (𝛼𝑗𝑛−1 , ）
𝑘′ −1 𝑘 𝑉 𝛼𝑗𝑛−1
𝑉 𝛼𝑗𝑘′

Interpolated by 𝑉(𝑥) = ැ𝑗∈Θ (𝑥 − 𝛼𝑗 ) Lagrange interpolation

41
Hybrid Soft Decoding
◼ Key steps of HSD
Interpolation 一 module basis construction (based on the Lagrange polynomials)

Seed polynomials:

ς𝑗 ′ ∈Θc,𝑗 ′ ≠𝑗(𝑥 − 𝛼𝑗 ′ )
𝒢(𝑥) = ෑ (𝑥 − 𝛼𝑗 ) & 𝑅𝜔 (𝑥) = ෍ 𝑧𝑗 𝜔 𝑇𝑗′ (𝑥) 𝑇𝑗′ (𝑥) =
𝑗∈Θc 𝑗∈Θc ς𝑛−1
𝑗 ′ =0,𝑗 ′ ≠𝑗(𝛼𝑗 − 𝛼𝑗 ′ )

For 𝒓𝜔 (𝒛𝜔 ), module ℳ𝜔 can be generated by

Computed in generating 𝐆RS
𝑃𝜔,0 (𝑥, 𝑦) = 𝒢(𝑥) Basis reduction
The Gröbner basis of ℳ𝜔
𝑃𝜔,1 𝑥, 𝑦 = 𝑦 − 𝑅𝜔 (𝑥)
The interpolation polynomial
∗ 0 ∗ 1
𝑄𝜔 𝑥, 𝑦 = 𝑉 𝑥 𝑄𝜔 𝑥 + 𝑄𝜔 𝑥 𝑦
Hybrid Soft Decoding
◼ Key steps of HSD
Root-finding
∗ 0
𝑉(𝑥)𝑄𝜔 (𝑥)
Estimated message polynomial: 𝑢ො 𝜔 (𝑥) = ∗ 1
𝑄𝜔 (𝑥)
Lemma 4
The evaluation values of polynomial 𝑢ො 𝜔 (𝑥) over MRPs form a TEP of the LLOSD, i.e.,
𝒆ො 𝜔 = (𝑢ො 𝜔 (𝛼𝑗0 ), 𝑢ො 𝜔 (𝛼𝑗1 ), … , 𝑢ො 𝜔 (𝛼𝑗 ))
𝑘′ −1

𝜔 𝜔
ෝ
Codeword candidate of LLOSD: 𝒗 = 𝒆′ ෝ(0)
⋅ 𝐆RS + 𝒗

ෝ𝜔 = 𝑢ො 𝜔 𝛼0 , 𝑢ො 𝜔 𝛼1 , … , 𝑢ො 𝜔 𝛼𝑛−1 ෝ 0
Codeword candidate of Chase-GS: 𝒗 +𝒗

= 𝑢ො 𝜔 𝛼𝑗0 , 𝑢ො 𝜔 𝛼𝑗1 , … , 𝑢ො 𝜔 𝛼𝑗 ෝ
⋅ 𝐆RS + 𝒗 0
𝑘′ −1

= 𝒆ො 𝜔 ⋅ 𝐆RS +ෝ
𝒗 0

43
Hybrid Soft Decoding
◼ Key steps of HSD
Partial root-finding (based on 𝐆RS )
The estimated TEP 𝒆ො 𝜔 = (𝑢ො 𝜔 (𝛼𝑗0 ), 𝑢ො 𝜔 (𝛼𝑗1 ), … , 𝑢ො 𝜔 (𝛼𝑗 )) can be determined as
𝑘′ −1

∗ 1
1, if 𝑄𝜔 𝛼𝑗 = 0
𝑢ො 𝜔 (𝛼𝑗 ) = ∀𝑗 ∈ Θ
0, otherwise

ෝ𝜔 can be generated as
The estimated codeword 𝒗 Provided by Chase-GS

ෝ𝜔 = (𝑢ො 𝜔 (𝛼𝑗0 ), 𝑢ො 𝜔 (𝛼𝑗1 ), … , 𝑢ො 𝜔 (𝛼𝑗

𝒗 ෝ
)) ⋅ 𝐆RS + 𝒗 0 = 𝒆ො 𝜔 ⋅ 𝐆RS + 𝒗
ෝ 0
𝑘′ −1

LLOSD

44
Hybrid Soft Decoding
◼ Performance of the (63, 39) BCH code
1.E+00

1.E-01 OSD (1)

Complexity: 4.08 × 104
(BOPs)
1.E-02
LLOSD (3)
FER

OSD (1) Complexity: 4.45 × 103

1.E-03 OSD (2)
LLOSD (2)
HSD (1, 6)
LLOSD (3)
Complexity: 2.71 × 103
1.E-04 HSD （1, 4）
HSD （1, 6）
HSD （1, 8） SNR: 5 dB
1.E-05
1 2 3 4 5 6
SNR (dB)

45
 Hybrid Soft Decoding
◼ Performance of the (255, 223) BCH code
1.E+00
LLOSD (2)
Complexity: 1.27 × 104
1.E-01
OSD (1)
Complexity: 2.71 × 105
1.E-02 (BOPs)
FER

OSD (1) PLCC (6)

1.E-03 PLCC (6) Complexity: 4.67 × 104
PLCC (8)
LLOSD (2)
HSD (1, 6)
HSD (1, 4)
1.E-04 Complexity: 1.56 × 104
HSD (1, 6)
HSD (1, 8)
SNR: 5.5 dB
1.E-05
3 3.5 4 4.5 5 5.5 6 6.5
SNR (dB)

PLCC decoding is the progressive variant of the Chase-GS 46