Abstract
We propose multi-mode spectrally efficient frequency division multiplexing with index modulation (MM-SEFDM), an enhanced non-orthogonal waveform that embeds distinct constellations on every subcarrier and conveys additional bits through subcarrier-index patterns. By compressing the subcarrier spacing of orthogonal frequency division multiplexing (OFDM), MM-SEFDM significantly improves spectral efficiency at the expense of controlled inter-carrier and inter-symbol interference. This joint non-orthogonal and multi-mode index modulation structure introduces significant detection complexity, necessitating a low-complexity, yet effective receiver capable of handling interference and permutation-based index-symbol decoding. To this end, we propose a two-phase Transformer-based decoder: in Phase 1, the network learns joint index–symbol constellations over AWGN channels, while in Phase 2, it continues training over Rayleigh fading channels to enhance robustness under realistic propagation conditions. Furthermore, to enhance model generalization, the detector is trained under domain randomization across different non-orthogonality levels and channel estimation errors (CSE), ensuring robust and SNR-agnostic performance under diverse interference and operating conditions. Compared to brute-force maximum-likelihood detection, the proposed learning-based decoder reduces inference complexity from exponential to polynomial time, while achieving near-optimal bit-error rate performance.
Funding Information
  • Natural Science and Engineering Research Council (NSERC), Canada