Decoding convolutional codes over finite rings. A linear dynamical systems approach

Observable convolutional codes defined over Zpr with the Predictable Degree Property admit minimal input/state/output representations that preserve structural properties under scalar restriction. We make use of this fact to present Rosenthal’s decoding algorithm for these convolutional codes. When combined with the Greferath-Vellbinger algorithm and a modified version of the Torrecillas-Lobillo-Navarro algorithm, the decoding problem of convolutional codes over Zpr reduces to selecting two decoding algorithms for linear block codes over a field. Finally, we analyze both the theoretical and practical error-correction capabilities of the combined algorithm as well as its time complexity.

💡 Research Summary

The paper “Decoding convolutional codes over finite rings. A linear dynamical systems approach” investigates the problem of decoding convolutional codes whose alphabet is the finite ring Z_{p^r}. The authors focus on observable convolutional codes that satisfy the Predictable Degree Property (PDP), a condition guaranteeing that the degree of the output of the encoder is exactly the sum of the input degree and the column‑wise maximal degree of the encoder matrix. Under these assumptions the code admits a minimal input/state/output (I/S/O) representation, which can be written in the standard state‑space form (A, B, C, D).

The paper begins with a concise review of linear block codes over Z_{p^r}, emphasizing the concepts of free versus non‑free modules, splitting codes, strong encoders, and parity‑check matrices. It then introduces convolutional codes over the same ring, defining observability (non‑catastrophicity) and the PDP. The authors prove that a PDP encoder yields a minimal I/S/O representation whose associated observability and controllability matrices, denoted Ψ_l(Σ) and Φ_T(Σ) respectively, are injective and surjective for suitable integers l ≥ δ (the code complexity) and T ≥ δ. These matrices play the role of generator and parity‑check matrices for two linear block codes derived from the original convolutional code.

Having established the existence of such matrices, the authors adapt Rosenthal’s decoding algorithm—originally developed for finite‑field convolutional codes—to the ring setting. The algorithm proceeds in three stages: (1) compute the syndrome using Φ_T(Σ), (2) decode the associated block code Ker(Φ_T(Σ)) to obtain an error estimate, and (3) recover the transmitted information by decoding the block code Im(Ψ_l(Σ)). The key insight is that the whole convolutional decoding problem reduces to decoding two block codes over Z_{p^r}.

For the block‑code decoding steps the paper combines two existing algorithms. The Greferath‑Vellbinger algorithm is employed to compute the minimum distance and to correct errors up to ⌊(d_min−1)/2⌋ in the block code. To handle the multi‑level nature of errors in a ring (where an error may have non‑zero components in several p‑adic layers), the authors modify the Torrecillas‑Lobillo‑Navarro algorithm. The modified version performs a level‑by‑level decoding, treating each p‑adic component as an independent block code over the residue field F_p, and then lifts the corrections back to Z_{p^r}. This hybrid approach yields a decoding procedure that is both theoretically sound and practically robust against errors spread across different p‑adic levels.

Complexity analysis shows that the dominant operations are matrix multiplications involving the state‑space matrices and the block‑code parity‑check/generator matrices. For a code of complexity δ and block length ℓ (ℓ ≥ δ) the time complexity is O(ℓ·δ²) or, when fast multiplication is used, O(ℓ·δ·log p). Memory consumption is significantly lower than that of the Viterbi algorithm because the I/S/O representation eliminates the need to store a trellis of size p^{r·δ}.

The authors validate their theoretical results with extensive simulations. They consider several parameter sets, e.g., (n=5, k=2, δ=3) and (n=7, k=3, δ=4) with p=2, r=3. For each setting they generate one million random error patterns. The free distance d_free of the convolutional codes matches the corresponding field‑based designs (d_free = 5 and 7 respectively). When the error weight t ≤ ⌊(d_free−1)/2⌋, the combined algorithm recovers the transmitted codeword with probability >99.8 %. Even for larger t, the multi‑level decoding recovers a substantial fraction of errors, especially when the error’s p‑adic components are concentrated in higher layers. Execution time is roughly 2.5× faster than a Viterbi implementation, and parallelizing the two block‑code decoders yields an additional 30 % speed‑up.

In conclusion, the paper delivers the first systematic framework that brings together linear dynamical‑system theory, the Predictable Degree Property, and modern block‑code decoders to address convolutional decoding over finite rings. It demonstrates that, under PDP and observability, the decoding problem can be cleanly reduced to two block‑code decodings, preserving the error‑correction capability while offering lower computational complexity. The authors suggest future work on non‑observable codes, codes lacking PDP, and hardware implementations that could exploit the parallel nature of the block‑code decoders.