{"id":"5a2285a5-7ecf-4265-8f7c-9ee7dcdd1728","arxiv_id":"2602.20953","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An IF-TEM communication receiver can estimate symbol timing offset and detect multi-level symbols directly from spike times using a two-stage maximum-likelihood and zero-forcing scheme.","lead":"This paper develops a timing-recovery and data-detection method for receivers that sample signals with integrate-and-fire neurons, recording spike times instead of clock samples. It derives a maximum-likelihood framework, then uses simulations to show the receiver can estimate symbol timing offset and detect 4-PAM symbols more accurately than prior spike-count approaches.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's Gaussian likelihood conditions on random firing times without accounting for their stopping-time nature, so the 'ML' estimator may be misspecified.","rationale":"The reader identified exactly the same load-bearing concern: the likelihood derivation treats firing times as fixed observables and assumes Gaussian noise integrals, ignoring that firing times are random hitting times. This is the theoretical core of the paper. I examined Appendix A and confirmed the issue: the proof writes Z_k as a Gaussian but does not justify why the integral over a stopping-time interval retains a Gaussian distribution. Because the ML label and both algorithms rest on Eq. (4), this is a genuine correctness risk. However, the paper is an engineering contribution with simulations suggesting the method works, so a conditional acceptance is appropriate. I do not see a second concern that would change the verdict; the missing error bars and reproducibility issues are secondary. Hence UNCHANGED relative to the reader's CONDITIONAL verdict.","tokens_in":9370,"tokens_out":2479,"duration_ms":26820,"concrete_test":"Simulate the exact IF-TEM with a known signal and AWGN sample path (e.g., using a fine-time-step Euler–Maruyama integration of the integrator state and threshold crossings). For a fixed true τ, record firing times t_k and compute residuals e_k = y_k − [P(τ)s_p + G(τ)s_d]_k. Standardize by sqrt(N0 (t_k−t_{k−1})/2) and test (a) normality via Kolmogorov–Smirnov, and (b) correlation of |e_k| with the interval length. Also compare the timing estimator from Eq. (13) against the true τ over many trials: if the standardized residuals are non-Gaussian or if the estimator shows bias growing at low SNR, then the Gaussian likelihood in Proposition 1 is a misspecification and the central claim requires a corrected likelihood.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Proposition 1 (Eq. 4), which asserts a Gaussian log-likelihood for the observed firing times. Appendix A writes y_k as a signal term plus Z_k = ∫_{t_{k-1}}^{t_k} Z(t)dt and states Z_k ~ N(0, N0 T_k/2). This is only valid if the interval [t_{k-1}, t_k] is fixed or independent of the noise process. But in an IF-TEM, t_k is a hitting time determined by the accumulated noisy integral, i.e., a stopping time adapted to Z. The integral of white Gaussian noise over a stopping-time interval is not Gaussian in general; its distribution depends on the hitting event. Therefore the derivation replaces the true likelihood of the firing times with a Gaussian pseudo-likelihood that is neither stated nor justified as an approximation. This misspecification propagates into the timing estimator (Eq. 13) and the ZF detector (Eq. 17): the estimator is not actually ML, and its bias/consistency are unproven. At low firing rates (b=1.5) the intervals are long and the hitting times are strongly noise-dependent, so the error could be significant. The paper provides no theoretical or empirical check of the Gaussianity assumption, making the correctness of the central statistical framework unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses symbol timing recovery and data detection for integrate-and-fire time encoding receivers (IF-TEM). The received waveform is modeled as a pulse-shaped PAM signal with an unknown timing offset and additive white Gaussian noise, then encoded by an IF-TEM into firing times. The paper derives a Gaussian log-likelihood (Proposition 1, Eq. 4) for the firing times, from which it formulates a joint ML problem. To keep computation tractable, the receiver first estimates the timing offset from pilot firing times using Newton's method with multiple initial guesses (Algorithm 1), then detects the data sequence with a zero-forcing detector operating on the firing-time-derived observations (Eq. 17). Simulation results report NMSE for timing estimation and SER comparisons against a spike-count-based ZF receiver from prior work.","tokens_in":9608,"tokens_out":6314,"duration_ms":66773,"significance":"If the statistical framework is valid, the paper fills a real gap: existing TEM-based receivers assume perfect symbol synchronization, while the proposed approach uses the fine structure of firing times for both timing and data recovery. The manuscript is clearly written, includes explicit algorithm pseudo-code, and uses a public simulation toolbox, which aids reproducibility. The two-stage architecture is practical, and the reported SER gains, if confirmed, are useful. However, the central theoretical claim — that Eq. (4) is the true log-likelihood of the observed firing times — is not established, and this undermines the 'ML' labeling and the optimality statements built on it.","major_comments":[{"comment":"The derivation of the Gaussian likelihood treats the firing times t_k as fixed and models Z_k = ∫_{t_{k-1}}^{t_k} Z(t)dt as Gaussian with variance N0 T_k/2. This is valid for a fixed integration interval. In the IF-TEM model (2), however, t_k is a hitting time: it is determined by the event that the accumulated integral of Y+b crosses κΔ. Thus t_k is a stopping time adapted to Z, and the event {t_k = τ} is informative about the noise path on [t_{k-1}, τ]. Consequently Z_k given t is not zero-mean Gaussian, and the joint density of the firing times is not of the form (4). The manuscript does not acknowledge this as an approximation; it calls (4) a log-likelihood. This is a load-bearing issue because the 'ML' timing estimator (13) and the data detectors built on (14)/(17) inherit the misspecification. Please either (i) derive the true likelihood via a change of variables from the continuou","section":"Appendix A / Proposition 1 (Eq. 4)"},{"comment":"The text in Proposition 1 and Appendix A describes f(s_d, τ_ϵ | t, s_p) as the conditional PDF of the parameters given the firing times. A likelihood should be a function of the parameters proportional to the density of the observed data given the parameters, f(t | s_d, τ_ϵ, s_p), not a posterior conditioned on t. This notation obscures the fact that the model conditions on the random firing times, which is the source of the misspecification. I recommend rewriting Eq. (4) as an approximate likelihood of t and avoiding posterior-style notation.","section":"Proposition 1, notation"},{"comment":"The Gaussianity assumption is never tested. In the low firing-rate mode (b=1.5), inter-spike intervals are long and the hitting-time dependence should be strongest, yet the paper provides no diagnostic. I request (i) a Monte Carlo check of the distribution of the normalized residuals (e.g., QQ plots or a Kolmogorov–Smirnov test) for the parameters used in Fig. 2, and (ii) a comparison of Algorithm 1 with a brute-force numerical evaluation of the true likelihood (or an importance-sampling approximation) to quantify the bias of the pseudo-ML estimator. Without such evidence, the 'accurate timing estimation' claim is only heuristic.","section":"Section IV / numerical validation"}],"minor_comments":[{"comment":"The SNR definition is not given. Please specify how the noise is generated and how SNR is computed (e.g., average signal power to N0, or E_s/N0). This is needed to reproduce Fig. 2 and Fig. 3.","section":"Section IV"},{"comment":"The update condition 'if ML Obj(ˆτ(ML)_ϵ) > ML Obj(ˆτ(ℓ)_candidate)' is ill-defined on the first iteration because ˆτ(ML)_ϵ is initialized to NULL. Initialize with the first candidate or restructure the logic.","section":"Algorithm 1, line 6"},{"comment":"The index range in the pilot-interference sum (l = L_p-1-L_f to L_p-1) is correct, but the sentence 'the final term ... accounts for interference' could be clearer: it should say it subtracts the known contribution of the last L_f+1 pilot symbols from the firing-time integrals.","section":"Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central statistical claim is questionable because firing times are stopping times, not fixed observation points. This is a fixable issue: the authors can reframe the contribution as a pseudo-likelihood-based receiver and add empirical validation. I do not see evidence of circularity or parameter fitting to target results. The simulation code appears to be based on a public toolbox, which is a point in the paper's favor. The main risk is that the 'ML' terminology is retained without addressing the stopping-time problem; the revision should either prove the claim or remove the optimality language."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper fills a real gap: nobody had explicitly addressed symbol timing recovery for IF-TEM communication receivers. The two-stage design—pilot-assisted timing estimation followed by a ZF detector that works on firing-time integrals rather than spike counts—is sensible and clearly motivated. The system model is clean, the notation is careful, and the proposed algorithm is implementable. That is genuine credit where it is due.\n\nNow the soft spots, in proportion. The central statistical claim is Proposition 1 / Eq. (4) and its proof in Appendix A. There, the noise integral Z_k = ∫_{t_{k-1}}^{t_k} Z(t)dt is declared Gaussian with variance N0 T_k/2. That is only valid if the integration interval is fixed or independent of the noise. But t_k is a hitting time determined by the accumulated noisy integral, i.e., a stopping time adapted to Z. The integral of white noise over a stopping-time interval is not Gaussian in general, and its distribution depends on the crossing event. So the 'log-likelihood' in (4) is a Gaussian pseudo-likelihood, not the true likelihood of the observed firing times. The stress-test note gets this right, and the paper never labels the step as an approximation or provides any check (residual Gaussianity, comparison against a true likelihood or a CRB).\n\nIs this fatal? Not necessarily. The estimator can be defended as a weighted least squares / pseudo-ML scheme, and the simulations suggest it behaves reasonably. But the abstract and Eq. (8) overclaim by calling it maximum likelihood, and at low firing rates (b=1.5), where noise strongly shapes the firing times, the misspecification could matter. This needs to be fixed or explicitly reframed.\n\nThe numerical section is also thinner than it should be: no trial counts, no error bars, no confidence intervals, and no code or data. The SER comparison to [22] is plausible but not statistically quantified. Minor: the invertibility condition for Gᵀ T_d G in (17) is not discussed.\n\nWho is this for? Researchers working on neuromorphic or event-driven communications, specifically IF-TEM receivers. It deserves a serious referee, but the revision should be major: confront the likelihood issue head-on, add Monte Carlo detail, and either prove or clearly state the pseudo-likelihood status of the estimator. I would send it to peer review rather than desk reject.","headline":"Useful first pass at IF-TEM timing recovery, but the core 'ML' likelihood is really a pseudo-likelihood because it conditions on random firing times without accounting for their stopping-time nature.","tokens_in":10124,"tokens_out":3646,"would_cite":true,"duration_ms":41033,"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":"This paper derives a maximum-likelihood framework that lets an integrate-and-fire time-encoding receiver estimate symbol timing offset and detect data directly from firing times.","keywords":["time encoding machine","integrate-and-fire","timing recovery","maximum likelihood estimation","sequence detection","zero-forcing detector","symbol timing offset","neuromorphic communications"],"falsifier":"Run an IF-TEM on a known pilot sequence with fixed timing offset and additive white Gaussian noise, record firing times, and compare the empirical residuals y_k − [P s_p]_k − [G s_d]_k against independent Gaussian random variables with variance N0 (t_k − t_{k−1})/2. If the residuals are correlated, non-Gaussian, or their variance does not scale with interval length, the derived log-likelihood is not the true likelihood of the observed firing times.","tokens_in":9159,"feed_emoji":"⚡","tokens_out":8758,"duration_ms":74843,"temperature":0.7,"pith_summary":"This paper takes on a gap in time-encoding receivers: prior designs assume perfect symbol timing, but an unknown timing offset breaks the mapping between transmitted symbols and observed firing times. The author's central claim is that the log-likelihood of the observed firing times can be written as a quadratic function of pilot and data contributions, so symbol timing and transmitted sequence can be estimated jointly by maximum likelihood. Since the joint problem is non-convex and NP-hard, the paper develops a practical two-stage receiver: an ML timing estimator driven by pilot firing times, followed by a zero-forcing sequence detector operating directly on data firing times. If the claim is right, an integrate-and-fire receiver can synchronize and decode without a separate clock, and can do so at lower firing rates than spike-count-based detectors.","feed_headline":"Spike times alone reveal symbol timing and data without a clock","feed_subtitle":"ML timing estimator plus zero-forcing detector uses inter-spike intervals, beating spike-count receivers.","key_machinery":"The load-bearing object is the per-interval measurement y_k = κΔ − b(t_k − t_{k−1}) taken from each firing interval, along with the matrices P(τ) and G(τ) whose entries are integrals of the pulse shape over that interval for pilot and data symbols. These convert asynchronous threshold crossings into a linear-plus-Gaussian model, with a diagonal weight T whose entries are the reciprocal inter-spike intervals, so shorter intervals carry noisier measurements. The first-order optimality condition Eq. (9) and the zero-forcing estimator Eq. (17) are both built on this representation.","core_discovery":"On the paper's own terms, the central discovery is that each firing interval of the IF-TEM carries a linear measurement of the received waveform: the quantity y_k = κΔ − b(t_k − t_{k−1}) equals the integral of the delayed pulse-shaped signal over that interval plus a Gaussian noise term whose variance is proportional to the interval length. This turns timing recovery and sequence detection into a weighted linear regression, giving a log-likelihood of the form C − ½ ‖T^{1/2}(y − P(τ)s_p − G(τ)s_d)‖². The paper then uses a first-order optimality condition on the timing parameter to build a Newton-based ML timing estimator and uses the estimated offset in a zero-forcing detector on the firing t","pith_inferences":["The likelihood in Proposition 1 treats firing times as fixed, but since they are noise-driven hitting times of the integrator, the stated Gaussian model is an approximation; a rigorous version would condition on the stopping times or use an unbiased pseudo-likelihood, and performance should be re-checked under that correction.","The same framework could support online or adaptive timing tracking by sliding a short pilot window through the firing stream, since the ML objective uses only local intervals.","Replacing the zero-forcing detector with a decision-feedback or ML sequence detector in Eq. (14) would likely improve symbol error rate further at higher complexity, because ZF discards some temporal information.","Adding slight random jitter to the threshold or integrator constant would test the robustness of the linear-Gaussian model; that extension is directly simulable within the paper's setup."],"forward_implications":["An IF-TEM receiver can estimate symbol timing offset from pilot firing times alone, removing the perfect-synchronization assumption in earlier time-encoding receivers.","Using the fine structure of firing times rather than spike counts per symbol period yields better symbol error rate, allowing a low-firing-rate receiver to approach the performance of a higher-firing-rate spike-count receiver.","The derived likelihood provides a unified statistical basis for jointly handling timing and data, extending time-encoding receiver design to unsynchronized settings.","Timing estimation error drops with longer effective pilot sequences and higher SNR, so the receiver can trade pilot overhead against synchronization accuracy.","The zero-forcing detector's pilot-interference subtraction mitigates inter-symbol interference from the pulse shape, enabling decoding with Gaussian pulse shaping at low firing rates."],"fun_headline_variants":["Inter-spike intervals decode timing and data","Spike timing alone recovers symbols and clock offset","Joint ML timing and detection from firing intervals","No fixed clock: IF-TEM receiver estimates timing from spikes","Timing offset and data recovery from IF-TEM spike times"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise, introduced in the proof of Proposition 1 in Appendix A, is that each firing interval yields an independent Gaussian observation of the integral of the received signal with variance proportional to that interval's length, even though the firing times themselves are random stopping times of the noisy integrator; conditioning on them would generally change the noise distribution, so the stated log-likelihood is an approximation the paper does not prove.","fun_headline_variants_meta":{"raw":{"variants":["Inter-spike intervals decode timing and data","Spike timing alone recovers symbols and clock offset","Joint ML timing and detection from firing intervals","No fixed clock: IF-TEM receiver estimates timing from spikes","Timing offset and data recovery from IF-TEM spike times"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000785,"raw_usage":{"total_tokens":3290,"prompt_tokens":719,"completion_tokens":2571,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":2495}},"tokens_in":463,"tokens_out":2571,"duration_ms":17941,"temperature":1.0,"reasoning_tokens":2495,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:11:25.754213+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an IF-TEM on a known pilot sequence with fixed timing offset and additive white Gaussian noise, record firing times, and compare the empirical residuals y_k − [P s_p]_k − [G s_d]_k against independent Gaussian random variables with variance N0 (t_k − t_{k−1})/2. If the residuals are correlated, non-Gaussian, or their variance does not scale with interval length, the derived log-likelihood is not the true likelihood of the observed firing times.","supporting_citations":[],"review_version":1}