{"id":"fce09ee6-bb72-45fb-aae4-e88666e95cfe","arxiv_id":"2608.04155","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A bidirectional Elman RNN with 25 to 65 trainable parameters matches M-BCJR bit error rate within 0.1 to 0.4 dB for FTN-BPSK at tau=0.8/0.9 while cutting LUT hardware cost by 38 to 67 percent.","lead":"This paper trains a small, simple recurrent neural network, an Elman RNN with 25 to 65 parameters, to decode faster-than-Nyquist signals, achieving bit error rates within a few tenths of a decibel of the optimal BCJR algorithm at lower hardware cost. It matters because it points to a very cheap, low-power receiver design for bandwidth-efficient communication.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training only on 7–10 dB SNR while evaluating at 0–10 dB leaves the low-SNR half of the BER curves unsupported; a matched-SNR retraining check is needed.","rationale":"The reader's weakest assumption is the sufficiency of a small gating-free Elman RNN for the FTN channel memory including colored noise. My concern is different but adjacent: the training SNR range (7–10 dB) does not cover the lower half of the evaluation range (0–10 dB), so the reported low-SNR agreement with M-BCJR is not directly supported by the described training protocol. This is load-bearing because the central claim is explicitly 'close to M-BCJR across 0–10 dB,' and a fixed-parameter RNN trained only at high SNR may mis-calibrate at low SNR. The concrete retraining check would settle whether this concern lands. I partially agree with the reader because the training/evaluation mismatch was listed among the addressable gaps, but the reader selected the capacity assumption as most load-bearing; my concern is more directly tied to the evidence for the reported curves. The reader's CONDITIONAL verdict is appropriate and I do not see a reason to move it: the paper is a plausible simulation study with clear reporting of the protocol, but the missing SNR-matched training check, code, and trial counts justify the conditionality. No ad hominem is intended; the issue is about the experimental protocol, not the authors. The proposed test is concrete, feasible, and would either confirm the low-SNR claims or localize them to the high-SNR training regime.","tokens_in":6738,"tokens_out":8241,"duration_ms":94067,"concrete_test":"Retrain the bidirectional Elman RNN from scratch with training SNR drawn uniformly from 0–10 dB (or per-SNR training for 0, 2, 4, 6, 8, 10 dB), keeping all other hyperparameters and the same evaluation setup as reported, and regenerate Fig. 3 and Table I. If the 0–4 dB BER curves shift by more than the Monte Carlo error, the reported near-M-BCJR performance is training-distribution-dependent and the claim must be restricted to 7–10 dB; if they are unchanged, the concern is resolved. Reporting the number of Monte Carlo trials and error bars in the same check would also settle the related reproducibility gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—near-M-BCJR BER across 0–10 dB—rests on simulations whose training distribution (Section III-C: 'training data are jointly generated from the 7–10 dB SNR range') excludes the lower half of the reported evaluation range (Fig. 3, Table I). A fixed-weight RNN trained only at high SNR has no explicit SNR input and cannot rescale its decision boundary with 1/N0; the optimal low-SNR detector must account for the SNR-dependent balance between prior and observation, and colored noise is proportionally more important at low SNR. If the weights were optimized on 7–10 dB, the 0–4 dB curves may reflect extrapolation rather than the claimed near-optimality. This is not a cosmetic issue: the paper provides no ablation demonstrating SNR robustness, no error bars, and no code, so the reader cannot distinguish a genuinely near-optimal low-SNR detector from an artifact of the training protocol. This concern is independent of the hidden-size sufficiency question: even a perfectly sufficient RNN could be mis-calibrated if never trained at the test SNR.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a bidirectional Elman recurrent neural network detector for BPSK faster-than-Nyquist signaling. The detector processes the matched-filter output once in each direction and maps the concatenated hidden states to bit decisions, with 25–65 trainable parameters. Simulations for tau=0.8 and tau=0.9 report BER within about 0.1–0.4 dB of an M-BCJR reference, with 38–46% lower LUT-based hardware cost (up to 67% for a smaller hidden size) and no explicit division or exponential operations. The authors claim this is the first Elman-RNN-based FTN detector.","tokens_in":6979,"tokens_out":4738,"duration_ms":50538,"significance":"If the results are reproducible, the contribution is a useful low-complexity detector for FTN-BPSK: the architecture is simple, the parameter count is explicitly quantified, the single-pass bidirectional structure is a natural analogue of the BCJR recursions, and the absence of division/exponential arithmetic is attractive for hardware. Strengths include the explicit parameter formula (2H^2+8H+1), the LUT costing methodology, and the numerical BER table that supplements the curves. The main load-bearing concerns are the mismatch between the training SNR range and the evaluation range, the unproven \"sufficient statistic\" claim for a small hidden state, the unspecified M-BCJR configuration, and the lack of error bars or trial counts for the low-BER region. These issues need to be resolved before the near-optimality and cost-reduction claims can be accepted.","major_comments":[{"comment":"The training data are generated from the 7–10 dB SNR range (Section III-C), while Fig. 3 and Table I report BER from 0–10 dB. Since the trained network has no SNR input, the weights cannot adjust the decision boundary to the noise level, so the low-SNR half of the BER curves is an extrapolation rather than a measured near-optimal behavior. Please provide matched-SNR training results (for example, training separately at each evaluation SNR, or training over the full 0–10 dB range) or an ablation demonstrating that the 0–4 dB performance is robust to the training SNR.","section":"Section III-C, Fig. 3, Table I"},{"comment":"The statement that the concatenated hidden state \"acts as a learned sufficient statistic of the BCJR α/β metrics\" is asserted without proof or citation. An Elman RNN with hidden size H=2–4 has very limited state capacity, and the colored noise at the matched-filter output may create memory beyond the nominal ISI length N. Please provide a formal argument or empirical evidence (for example, BER versus hidden size saturation, or a comparison against a detector with an explicit noise-whitening front end) to justify this load-bearing assumption.","section":"Section III-B, Eq. (10)"},{"comment":"The M-BCJR reference configuration is not specified: the number of survivors M, the truncation length, whether the colored-noise covariance is included in the branch metric, and the block length used for counting errors are all absent. Without these details, and without error bars or the number of channel realizations, the reported 0.1–0.4 dB gaps and the low-BER values around 1e-5 to 1e-6 cannot be distinguished from Monte Carlo noise. Please provide the reference configuration and confidence intervals.","section":"Section IV, Table I and Fig. 3"},{"comment":"The hardware-cost comparison uses LUT unit weights from the authors' prior paper [13], but the table does not break down the M-BCJR cost (for example, how many division and exponential operations are counted) or state whether the same LUT unit weights were applied to the M-BCJR operations. Since the claimed 38–67% reduction is a central result, please provide a per-operation breakdown and justify the unit weights.","section":"Table III"}],"minor_comments":[{"comment":"The term \"single pass\" is potentially confusing because Eqs. (8) and (9) require two traversals, one forward and one backward; please clarify that each direction is processed once.","section":"Section III-B"},{"comment":"The column formatting in Table III is hard to read (for example, \"yes594 750\" lacks a visible separator); please format the LUT counts as numeric entries with explicit column headings.","section":"Table III"},{"comment":"The noise autocorrelation is stated to be proportional to the ISI response X; please give the exact relation (for example, E[w(nτT)w(mτT)] = N0 x((n−m)τT)/2) to make the colored-noise model precise.","section":"Section II"},{"comment":"The test-set size or number of independent trials is not reported; please add this information to Table II or to the simulation text so that the BER values can be interpreted statistically.","section":"Section IV"},{"comment":"The statement that the per-symbol operation count is independent of block length refers to inference; during training with truncated backpropagation through time the cost may differ, so please clarify that the complexity claim is for inference.","section":"Section III-C"},{"comment":"The phrase \"using no explicit division or exponential operations\" should be clarified, because tanh and sigmoid are implemented with look-up tables; the claim is that the arithmetic datapath does not contain these operations, not that no nonlinear function approximations are used.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the authors' own prior work [13], [15], [18]–[20] for comparison and for the LUT cost model. This is not disqualifying, but independent validation of the LUT unit weights and an external reproduction of the M-BCJR reference would substantially increase confidence. The paper is within the scope of a communications or signal-processing venue, although the BPSK-only and τ≥0.8 results form a narrow slice; the authors should be encouraged to state clearly that the contribution is limited to this setting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Brief take: this is a competent, incremental simulation paper. The genuinely new piece is showing that a vanilla Elman RNN, no gates, is enough to approximate BCJR for FTN-BPSK when the ISI memory is only a few symbols. The authors report 25–65 trainable parameters getting within about 0.1–0.4 dB of M-BCJR for τ=0.8 and 0.9, with a consistent LUT-based complexity comparison. That's a useful result for hardware-constrained receivers, and it's honest about not having a direct same-environment comparison with CNN/GRU detectors yet.\n\nWhat the paper does well: the architecture is simple and the reasoning is clear—FTN ISI is short, so gating is overkill. The complexity analysis is concrete: no division/exponential operations, and the LUT unit weights are stated. The authors acknowledge the main limitations and don't overclaim. The BER curves look consistent with the numbers in Table I.\n\nSoft spots, in order. The training/evaluation SNR mismatch is real. They train on 7–10 dB and evaluate on 0–10 dB, and a fixed-weight RNN can't rescale with 1/N0; the optimal symbol-wise MAP boundary does depend on SNR in general. But the reported numbers at 4 dB are still close (e.g., 1.9e-2 vs 1.6e-2 for τ=0.8), so the practical impact may be smaller than the stress-test implies. Still, they should train on the full SNR range or include an SNR input, and report an ablation. There are no error bars or Monte Carlo trial counts, which matters for a 0.1–0.4 dB claim. No code or full hyperparameters. The 'learned sufficient statistic' phrase in Sec III-B is hand-waving; it's not proven and probably just means the architecture mimics forward-backward recursion. M-BCJR's configuration (number of survivors M) isn't specified, which makes the reference comparison uncertain. These are all fixable in revision.\n\nMy take: the central claim is plausible. This paper is for people designing low-complexity FTN receivers and for anyone who wants a concrete example of a minimal RNN matching a trellis decoder on a short-memory channel. It deserves a serious referee; the methodology gaps are addressable, and the qualitative result is convincing. I'd want the SNR ablation and error bars before trusting the exact 0.1–0.4 dB gap, but this is a solid, honest contribution.","headline":"A tiny Elman RNN gets within 0.1–0.4 dB of M-BCJR for FTN-BPSK with just 25–65 parameters; the result is credible, but the training/evaluation SNR mismatch and missing reproducibility details need attention.","tokens_in":7501,"tokens_out":5858,"would_cite":true,"duration_ms":64113,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A 25-parameter Elman RNN detects faster-than-Nyquist signals within 0.4 dB of M-BCJR.","keywords":["faster-than-Nyquist signaling","Elman recurrent neural network","signal detection","intersymbol interference","low complexity","BPSK","bidirectional RNN","M-BCJR"],"falsifier":"Train the same bidirectional Elman detector on a root-raised-cosine faster-than-Nyquist channel with compression factor 0.7 or 0.6; if the bit error rate gap to M-BCJR at $E_b/N_0 = 10$ dB exceeds roughly 0.4 dB, the short-memory premise that justifies the small hidden size is falsified.","tokens_in":6526,"feed_emoji":"📡","tokens_out":11381,"duration_ms":94759,"temperature":0.7,"pith_summary":"The paper tries to establish that the simplest recurrent network, the Elman RNN with no gating, is the right-sized model for faster-than-Nyquist detection because FTN's intersymbol interference has short, finite memory. It proposes a bidirectional single-pass Elman detector for BPSK that mirrors the BCJR forward-backward recursion with 25–65 trainable parameters. Simulations over AWGN with a root-raised-cosine pulse show BER within 0.1–0.4 dB of the M-BCJR benchmark for compression factors 0.8 and 0.9, while cutting the look-up-table hardware cost by 38–46 percent (up to 67 percent in the smallest configuration) and avoiding explicit division and exponential operations. If correct, this gives near-optimal detection at a fraction of the hardware cost, with a shorter critical path and no noise-covariance estimation.","feed_headline":"A 25-parameter Elman RNN matches M-BCJR within 0.4 dB","feed_subtitle":"Bidirectional single-pass Elman detector cuts LUT cost up to 67 percent with no division or exponential operations.","key_machinery":"The central object is the Elman (plain) recurrent unit $h_t = \\tanh(W_x y_t + W_h h_{t-1} + b)$, run in both directions in a single pass over the received samples. At each symbol the forward and backward hidden states are concatenated into a $2H$-dimensional vector and passed through a fully connected layer with a sigmoid during training and a sign operation at inference; the paper treats this concatenated state as a learned sufficient statistic of the BCJR $\\alpha$ and $\\beta$ recursions. The mechanism works because the FTN interference is banded and short, so the same small hidden dimension can summarize all relevant past and future symbols; the parameter count is $2H^2 + 8H + 1$, giving 25, 43, and 65 parameters for $H=2,3,4$.","core_discovery":"The central claim is that a classical Elman RNN, with no gating, is sufficient as a standalone FTN detector because the FTN channel's ISI has a short finite memory of only a few symbols. The detector runs the Elman recurrence $h_t = \\tanh(W_x y_t + W_h h_{t-1} + b)$ forward and backward over the received sequence in a single pass, concatenates the two hidden-state sequences per symbol, and maps them through a fully connected layer to a BPSK decision. The concatenated state is presented as a learned sufficient statistic of BCJR's forward and backward metrics. On the tested settings the network, with 25 to 65 trainable parameters, stays within 0.1–0.4 dB of M-BCJR while using no explicit division or exponential operations and lowering the LUT-based hardware cost by 38–67 percent depending on hidden size.","pith_inferences":["The same short-memory argument suggests the detector should transfer to other channels with few-tap ISI, such as time-dispersive optical links, as long as the effective interference memory remains small.","A natural stress test is lowering the compression factor: the hidden size $H$ needed to stay within 0.1–0.4 dB of M-BCJR should grow roughly with the one-sided ISI length, mapping out a complexity-versus-spectral-efficiency tradeoff curve.","Since the network learns the colored-noise statistic from data, it may handle non-white noise without explicit covariance estimation; showing this outside AWGN would extend the claim beyond the current simulations.","Quantizing the weights, which the paper lists as future work, would most likely reduce the LUT cost further and make the approach competitive with fixed-kernel CNN detectors at equal precision."],"forward_implications":["For BPSK FTN at $\\tau=0.8$ and $\\tau=0.9$ over AWGN, the bidirectional Elman detector can replace the M-BCJR trellis with a BER penalty of only about 0.1–0.4 dB.","The LUT hardware cost drops by 38–46 percent for the main configurations and by up to 67 percent for $H=2$ at $\\tau=0.9$, so the receiver can trade a small performance loss for a large complexity saving.","Because the detector needs no explicit division or exponential functions, its critical path is shorter and its numerical behavior is more stable than the M-BCJR branch metric computation.","The single-pass bidirectional processing makes the per-symbol cost independent of block length, so the same network scales to arbitrarily long frames without windowing.","Gating mechanisms such as LSTM and GRU are unnecessary for FTN detection at these compression factors; a plain recurrent network captures the short-memory structure."],"supporting_citations":[{"why":"Defines faster-than-Nyquist signaling and the controlled intersymbol interference that motivates the detection problem.","marker":"[1]"},{"why":"Defines the BCJR forward-backward recursion that the proposed bidirectional single-pass architecture is designed to mimic.","marker":"[2]"},{"why":"Supplies the reduced-complexity M-BCJR algorithm used as the BER reference in all simulations.","marker":"[3]"},{"why":"Provides the 10-bit look-up-table unit costs and the gated-GRU comparison that make the hardware-cost reduction quantitative.","marker":"[13]"},{"why":"Provides the fixed-kernel CNN standalone detector whose LUT reduction is used as the comparison baseline.","marker":"[15]"},{"why":"Defines the classical Elman recurrent network that is the paper's proposed detector core.","marker":"[16]"}],"fun_headline_variants":["25-param Elman RNN matches M-BCJR within 0.4 dB","Elman RNN with 25 weights rivals M-BCJR on FTN","Single-pass Elman detector cuts LUT cost 67% vs M-BCJR","Tiny RNN detector for FTN nears BCJR accuracy","No-gate Elman RNN: BCJR-level BER for FTN signals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The design depends on the faster-than-Nyquist channel's interference having a memory short enough that two to four hidden states can capture all relevant past and future information, including the colored noise at the receiver.","fun_headline_variants_meta":{"raw":{"variants":["25-param Elman RNN matches M-BCJR within 0.4 dB","Elman RNN with 25 weights rivals M-BCJR on FTN","Single-pass Elman detector cuts LUT cost 67% vs M-BCJR","Tiny RNN detector for FTN nears BCJR accuracy","No-gate Elman RNN: BCJR-level BER for FTN signals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001227,"raw_usage":{"total_tokens":5048,"prompt_tokens":952,"completion_tokens":4096,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":3990}},"tokens_in":568,"tokens_out":4096,"duration_ms":30428,"temperature":1.0,"reasoning_tokens":3990,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:21:22.801920+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same bidirectional Elman detector on a root-raised-cosine faster-than-Nyquist channel with compression factor 0.7 or 0.6; if the bit error rate gap to M-BCJR at $E_b/N_0 = 10$ dB exceeds roughly 0.4 dB, the short-memory premise that justifies the small hidden size is falsified.","supporting_citations":[{"cited_title":"Faster-than-Nyquist signalling,","cited_arxiv_id":null,"evidence_quote":"Defines faster-than-Nyquist signaling and the controlled intersymbol interference that motivates the detection problem."},{"cited_title":"Optimal decoding of linear codes for minimizing symbol error rate,","cited_arxiv_id":null,"evidence_quote":"Defines the BCJR forward-backward recursion that the proposed bidirectional single-pass architecture is designed to mimic."},{"cited_title":"Reduced-complexity receivers for strongly narrow- band intersymbol interference introduced by faster-than-Nyquist signaling,","cited_arxiv_id":null,"evidence_quote":"Supplies the reduced-complexity M-BCJR algorithm used as the BER reference in all simulations."},{"cited_title":"GRU- Based Sequence Detection for Faster-than-Nyquist Signaling,","cited_arxiv_id":null,"evidence_quote":"Provides the 10-bit look-up-table unit costs and the gated-GRU comparison that make the hardware-cost reduction quantitative."},{"cited_title":"A novel CNN-based standalone detector for faster-than-Nyquist signaling,","cited_arxiv_id":null,"evidence_quote":"Provides the fixed-kernel CNN standalone detector whose LUT reduction is used as the comparison baseline."}],"review_version":1}