SOGRAND Assisted Guesswork Reduction

Proposals have been made to reduce the guesswork of Guessing Random Additive Noise Decoding (GRAND) for binary linear codes by leveraging codebook structure at the expense of degraded block error rate (BLER). We establish one can preserve guesswork reduction while eliminating BLER degradation through dynamic list decoding terminated based on Soft Output GRAND’s error probability estimate. We illustrate the approach with a method inspired by published literature and compare performance with Guessing Codeword Decoding (GCD). We establish that it is possible to provide the same BLER performance as GCD while reducing guesswork by up to a factor of 32.

💡 Research Summary

The paper addresses a long‑standing trade‑off in Guessing Random Additive Noise Decoding (GRAND): reducing the number of noise‑pattern queries (guesswork) by exploiting codebook structure often degrades block error rate (BLER) because the query order becomes sub‑optimal. The authors propose to retain the guesswork savings while completely eliminating the BLER penalty by integrating Soft‑Output GRAND (SOGRAND) into a dynamic list‑decoding framework.

SOGRAND provides, after each query, an accurate estimate of the probability that the correct codeword is not among the candidates already collected. By terminating list decoding as soon as this probability falls below a pre‑selected threshold θ, the decoder stops further queries once it is sufficiently confident that the correct codeword has been found. This dynamic termination adapts the list size to each received word, avoiding unnecessary work while preserving ML‑like performance.

Building on this idea, the authors introduce a new soft‑input GRAND variant called Syndrome‑Enhanced GRAND (SyGRAND). SyGRAND augments the standard GRAND process with a syndrome‑based candidate generation step: for each guessed word y^(j) it computes the syndrome s^(j)=H·y^(j). If s^(j) matches any column of the parity‑check matrix H, flipping the corresponding bit yields a new candidate codeword ˆw = y^(j)⊕e_p. This operation is essentially a one‑step syndrome decoder performed on every guess, allowing early discovery of valid codewords without requiring the full ML ordering of noise patterns.

The algorithm proceeds as follows: (1) generate noise patterns in decreasing reliability order (using a 1‑line ORBGRAND generator); (2) form the tentative word y^(j)=y_hd⊕˜z^(j); (3) update the accumulated noise probability P_noise; (4) compute the syndrome; (5) if the syndrome is zero, output the word immediately; (6) otherwise, test the column‑match condition, create ˆw if applicable, add it to the list L (avoiding duplicates), and update the list probability P_L; (7) recompute the SOGRAND estimate ˆP(C∉L|ℓ) using both P_noise and P_L; (8) stop when ˆP ≤ θ or when a maximum list size L_max is reached, and finally select the most likely word in L as the decoded result.

Two hyper‑parameters control the trade‑off: the confidence threshold θ and the maximal list size L_max. The authors propose a two‑stage optimization: first find the smallest L_max that guarantees BLER no worse than a reference decoder (ORBGRAND) with θ=0; then, with this L_max fixed, increase θ as much as possible while still meeting the BLER constraint.

Performance is evaluated on three codes: extended BCH (256,239), extended BCH (32,21), and CRC‑assisted Polar (128,110) with a 5G‑NR 11‑bit CRC. All decoders use the same 1‑line ORBGRAND pattern generator, and for even‑weight codes they restrict queries to even‑parity patterns (a reduction unavailable to GCD). The results show:

• BLER: SyGRAND matches or improves upon ORBGRAND across the entire SNR range, confirming that the dynamic termination fully compensates for the sub‑optimal query order.
• Guesswork: Compared to ORBGRAND, SyGRAND reduces average queries by factors ranging from 2 to 32 (‑1 to ‑5 dB in log₂ scale). It consistently outperforms ORDEPT in guesswork reduction, and the advantage of ORDEPT disappears at higher SNRs.
• Compared with Guessing Codeword Decoding (GCD), SyGRAND achieves the same BLER while requiring up to 32× fewer queries, demonstrating that a noise‑centric approach that leverages syndrome information can be more efficient than a codebook‑centric approach.

The paper’s contributions are threefold: (i) a practical method to eliminate BLER loss for GRAND variants with non‑optimal query orders by using SOGRAND‑based dynamic list termination; (ii) the SyGRAND algorithm that exploits syndrome information to find candidate codewords early, achieving greater guesswork savings than prior codebook‑structure methods; (iii) a systematic parameter‑tuning procedure that balances BLER and complexity.

The authors suggest future work on extending the candidate generation to larger Hamming radii, applying the technique to non‑linear codes, and quantifying hardware resource requirements for ASIC/FPGA implementations. Overall, the paper demonstrates that integrating soft‑output reliability estimates with modest code‑structure exploitation can dramatically reduce decoding effort without sacrificing error‑rate performance, a result of high relevance for next‑generation wireless, low‑power IoT, and high‑reliability storage systems.