Abstract
We propose a refined scaling law to predict the finite-length performance in the waterfall region of spatially coupled low-density parity-check (SC-LDPC) codes over the binary erasure channel (BEC). By introducing improvements to the scaling law previously proposed by Olmos and Urbanke, we achieve significantly better agreement between predicted and simulated frame error rates (FER). Furthermore, we extend the scaling law to analytically predict bit error rate (BER) performance.
Introduction
In coding theory, spatially coupled LDPC codes are renowned for their "threshold saturation" property, allowing them to approach the theoretical limits of a channel. While density evolution provides insights into their infinite-length behavior, real-world applications rely on finite-length codes. Predicting performance in the "waterfall region"—where error rates drop sharply—is critical for engineering these systems.
A previous model by Olmos and Urbanke (2015) utilized an Ornstein-Uhlenbeck (OU) process to predict finite-length performance. However, a persistent mismatch remained between the model and simulation data. This research identifies the source of this discrepancy: the original model assumed decoding occurs as a single stochastic process, whereas terminated SC-LDPC codes actually feature two simultaneous decoding waves originating from both ends of the coupled chain and moving toward the center.
Methods and Insights
The study focuses on the peeling decoder over the Binary Erasure Channel. The core of the refinement lies in the "Two-Wave Modeling" approach. In terminated codes, decoding starts at both boundaries simultaneously because these areas are "cleaner" than the middle. Failure occurs when stochastic noise in either wave causes the number of available decoding steps to hit zero.
Mathematically, modeling the decoding failure as the sum of two independent processes changes the distribution of the "first hit time" from a standard exponential model to an Erlang distribution (with shape parameter 2).
Additionally, the authors introduced channel-dependent parameters. Rather than treating scaling constants as fixed, they utilized linear interpolation to calculate parameters based on the specific channel erasure probability ($\epsilon$), further improving the model's precision.
Results and Evaluation
The refined scaling law was evaluated against Monte-Carlo simulations for various regular ensembles, including (5, 10), (4, 8), and (3, 6) codes with block lengths ($N$) ranging from 500 to 10,000.
For ensembles with a variable node degree $d_v \ge 4$, the refined law provides a near-perfect match with simulation results for both FER and BER. This correction demonstrates that the previous model's mismatch was a structural omission of the second decoding wave rather than a simple parameter estimation error.
Conclusion and Future Work
The proposed refined scaling law offers a powerful tool for engineers, accurately capturing how error rates scale with block length $N$. This allows for the determination of required code lengths to meet target error rates without the need for extensive, time-consuming simulations.
While the model is highly accurate for $d_v \ge 4$, a small gap remains for $d_v = 3$ codes, likely because failures in these ensembles occur outside the steady-state region. Future research will focus on closing this gap and extending the refined laws to more complex channels, such as the Additive White Gaussian Noise (AWGN) channel.
Summary updated