Signature-Guided Hybrid Sequence Estimation with Sparsely-Activated MLSE in Severely Bandwidth-Limited Coherent Links

Scaling coherent optical interconnects towards 800G and beyond is primarily constrained by severe inter-symbol interference (ISI) induced by aggressive bandwidth limitations. While maximum likelihood sequence estimation (MLSE) offers strong sequence-level robustness against such impairments, its exponential complexity prohibits practical deployment in symbol rate sampled (SRS) receivers. To break this performance complexity trade-off, we propose a signature-guided hybrid sequence estimation (SG-HSE) scheme that transforms continuous sequence detection into a highly efficient, on-demand localized decoding process. The core contribution lies in a two-stage localization and arbitration engine: a multiplier-free sign switched sliding-accumulation matched filtering (MF-SAMF) extracts unique error signatures to accurately bypass noise induced pseudo-peaks, while a dual-hypothesis decision logic (DH DL) dynamically evaluates decision reliability. This ensures that the computationally intensive MLSE is triggered with high selectivity at actual error propagation hotspots, substantially suppressing unnecessary MLSE activations. Additionally, an intra-window metric sharing technique is devised to recycle branch calculations, further reducing the processing burden during active MLSE phases. We experimentally validate SG-HSE on an 80/90/100-GBaud symbol-rate sampling coherent transmission platform. The results reveal that the proposed architecture approaches the BER performance of full-state MLSE under the tested experimental conditions, yet requires an exceptionally sparse MLSE activation rate of merely 7% to 11%. Coupled with a branch metric sharing efficiency of ∼79%, SG HSE significantly reduces the overall computational overhead. This breakthrough demonstrates a viable path toward ultra-low power, high-performance digital signal processing for power constrained data centers.