Abstract
Spatially Coupled LDPC (SC-LDPC) codes are renowned for their capacity-approaching performance under iterative decoding. However, accurately predicting the performance of finite-length SC-LDPC codes in the waterfall region remains a significant challenge. This paper introduces a refined mathematical scaling law that models the decoding process as two independent waves propagating from the boundaries of a code chain. By incorporating an Erlang distribution to describe failure times and extending the model to sliding window decoding, the authors achieve highly accurate predictions for Frame, Bit, and Block Error Rates, effectively bridging the gap between theoretical models and practical simulations.
Introduction
Spatially Coupled LDPC (SC-LDPC) codes have become a cornerstone of modern coding theory due to their ability to approach Shannon capacity using simple iterative decoding. While their asymptotic performance is well-understood, real-world applications utilize codes of finite length. Predicting the error probability (the "waterfall" region) for these finite-length codes is notoriously difficult.
Previous models, such as the scaling law proposed by Olmos and Urbanke, were designed for standard LDPC codes and assumed a single decoding wave. When applied to "terminated" SC-LDPC codes, these models showed a significant mismatch with experimental data. This paper identifies that the discrepancy arises because terminated SC-LDPC codes feature two decoding waves propagating simultaneously from both ends of the code chain.
The Two-Wave Model and Window Decoding
The core contribution of this work is the transition from a single-wave stochastic model to a two-wave model. In a terminated SC-LDPC code, decoding begins at both the left and right boundaries where the check node degrees are lower. These two waves move toward the center of the chain.
This physical reality necessitates a change in the underlying probability theory. The authors model the "first hit time" (the moment decoding fails) using an Erlang distribution with a shape parameter of 2, reflecting the sum of two independent failure processes.
Furthermore, the paper adapts this framework for sliding window decoding, a common technique used to reduce latency. Window decoding is modeled as a hybrid process: it behaves like a single-wave "unterminated" code while the window slides through the middle of the chain, and transitions into a two-wave "terminated" code once the window reaches the end of the coupling chain.
Performance Prediction and Results
The refined scaling law was evaluated against extensive Monte Carlo simulations of various SC-LDPC ensembles with varying component code lengths ($N$) and node degrees. The results demonstrate that the new model provides a near-perfect match for both Frame Error Rate (FER) and Bit Error Rate (BER) curves, particularly for codes with variable node degrees $d_v \ge 4$.
Key insights from the results include:
- Accuracy: The Erlang-based model corrects the optimistic bias of previous single-wave models.
- Efficiency: The analytical model allows for rapid performance prediction, bypassing the need for computationally intensive simulations that can take weeks to complete.
- Design Guidance: The framework enables engineers to jointly optimize window size ($W$) and code length ($N$) to balance latency requirements with target error rates.
Conclusion
By accounting for the dual-wave nature of terminated SC-LDPC decoding, this paper provides a robust analytical tool for predicting finite-length code performance. While currently focused on the Binary Erasure Channel (BEC), this work lays the foundation for future extensions to more complex channel models like AWGN. The refined scaling law offers a practical path for designing high-performance, low-latency communication systems that leverage the full potential of spatial coupling.
Summary updated