{"id":"7995bf64-d7a6-4a17-bdab-74d045fedd6e","arxiv_id":"2506.04780","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"An FPGA-based frequency counter using a Hilbert transform and weighted linear regression achieves real-time frequency readout of free-induction-decay signals with sensitivity comparable to offline nonlinear fitting.","lead":"Researchers built a real-time frequency counter that extracts the oscillation frequency of pulsed, decaying signals using a Hilbert transform, and showed it matches the accuracy of slow offline curve fitting. The device reads out 200 to 1000 frequency values per second with noise near the theoretical limit, which could enable compact, high-bandwidth atomic magnetometers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline sensitivity is measured only on clean synthetic FID pulses; real-FID effects such as DC offset, envelope mismatch, and correlated noise are untested, so the generalization claim is conditionally supported.","rationale":"The paper's internal evidence is strong for what it directly measures: the HT-LR implementation is benchmarked against LM on identical data, the truncation error is characterized, and clock-stability limits are explored. The measured sensitivity numbers and the LM comparison are internally consistent. The load-bearing gap is external validity: every hardware result uses a clean synthetic exponential FID from a signal generator, and the only robustness study, the weighted fit, uses simulated white noise. The reader's weakest assumption identifies exactly this gap. I do not find an internal inconsistency or a flaw that would overturn the measured claims. The appropriate verdict remains CONDITIONAL, because adoption as a general-purpose real-time FID frequency counter requires demonstration on signals with the offsets, envelope variations, and correlated noise present in real atomic magnetometer outputs. No code or raw data are provided, which further supports the conditional stance. My concern does not move the verdict; it sharpens the condition that needs to be met: a real-signal or realistically corrupted-signal validation of HT-LR against LM.","tokens_in":7829,"tokens_out":9763,"duration_ms":133183,"concrete_test":"Process a set of recorded real FID magnetometer pulses (or, if unavailable, synthetic pulses that add a 0.1% full-scale DC offset and a 20% colored-noise component to the existing exponential model) through the same FPGA counter and offline LM fit, comparing the 10-Hz NSD over at least 3000 pulses. If HT-LR's NSD degrades relative to LM by more than the 15–40% margin quoted for the weighted fit, the robustness/generalization claim fails; if it remains comparable, the conditional can be relaxed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central performance claim—sub-100 µHz/√Hz at 10 Hz with HT-LR matching offline LM—is supported only for the single signal class A e^{-t/τ} sin(2π f t) with fixed τ=2.5 ms and A=2.5 V generated by a function generator (Sec. III). The robustness extension, the weighted fit, is demonstrated only in simulation with white additive noise (Fig. 6). The phase extraction φ[n] = atan2(y[n], x[n]) followed by weighted linear regression is known to be sensitive to DC offset/baseline, to envelope shapes that are not purely exponential, and to correlated noise; none of these is characterized in the manuscript. Since the abstract and conclusion generalize to FID atomic magnetometers and claim that the method 'does not require the pre-knowledge of the analytic expression of the input signals', the transfer of the claimed sensitivity and the 15–40% weighted-fit improvement to real sensor signals is the weakest load-bearing assumption. The manuscript contains no real FID sensor test, and no code or raw data are shipped, so the generalization cannot currently be independently reproduced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports an FPGA-based frequency counter that implements a Hilbert-transform linear-regression (HT-LR) algorithm for extracting the frequency of pulsed exponential-decay oscillations. The authors measure the noise spectral density (NSD) of the extracted frequency using function-generator signals with fixed amplitude 2.5 V and relaxation time 2.5 ms over 10 kHz to 500 kHz, reporting better than 100 µHz/√Hz at 10 Hz for a 200 Hz output rate and better than 400 µHz/√Hz at 1000 Hz. They compare the results with offline Levenberg-Marquardt fitting on the same data, study the effect of the Hilbert-transform truncation order K and of clock stability, and propose a weighted linear-regression variant tested in simulation. The conclusion is that the counter is suitable for FID atomic magnetometers and other pulsed-frequency applications.","tokens_in":8059,"tokens_out":7713,"duration_ms":92415,"significance":"If the reported sensitivity numbers hold, the HT-LR counter is a useful real-time alternative to offline nonlinear fitting for FID-type signals, with modest hardware requirements. The direct experimental comparison with the LM algorithm on identical recorded data is a strength, and the clock-stability check adds practical value. The main limitation is that the experimental evidence covers only one synthetic signal class with fixed amplitude and relaxation time; the weighted-fit robustness improvement is simulation-only, and the comparison with Ω-counters is dimensionally unclear. These gaps do not invalidate the central measurement, but they do limit the breadth of the conclusions as currently stated.","major_comments":[{"comment":"The truncation order K is selected using the same data set that is then used to report the headline sensitivity values. Because Fig. 2(b) shows that the NSD is still varying with K in the vicinity of K=20 for some of the tested frequencies, the reported <100 µHz/√Hz value may depend on this in-sample choice. The authors should either provide an out-of-sample rule for selecting K or present a sensitivity analysis showing that the headline number is stable over a range of K values.","section":"Sec. III, Fig. 2 and text around \"We choose K=20\""},{"comment":"The claimed 15-40% improvement from weighted fitting is demonstrated only on simulated white-noise data and is not verified experimentally. Because this improvement is presented as a robustness result with a quantitative range, the authors should confirm it on measured signals (for example, by adding controlled noise to the function-generator output) or explicitly label it as a simulation-based prediction that has not yet been experimentally validated.","section":"Sec. III, Fig. 6 and accompanying text"},{"comment":"The comparison with the Ω-counter is dimensionally inconsistent as written. Equation (5) evaluates to a frequency error in hertz (about 277 µHz for f=250 kHz, T=2.5 ms, δT=20 ps), not to a noise spectral density in Hz/√Hz. To obtain the quoted 20 µHz/√Hz one must apply an additional conversion factor, approximately √T, and this step is omitted. The comparison should be re-derived with consistent units or the claims should be restated accordingly.","section":"Sec. III, Eq. (5) and the following comparison"},{"comment":"The claim that the method \"does not require the pre-knowledge of the analytic expression of the input signals\" is broader than the evidence supports. All experimental tests use the single signal class A e^{-t/τ} sin(2π f t) with τ=2.5 ms and A=2.5 V; no tests cover DC offsets, non-exponential envelopes, detuning, multiple frequency components, or correlated noise. The abstract and conclusion should be restricted to the tested signal class, or additional experiments covering these cases should be reported.","section":"Abstract and Secs. I and IV"}],"minor_comments":[{"comment":"The text contains the typo \"Field-Prgrammable Gate Array\"; it should read \"Field-Programmable Gate Array.\"","section":"Sec. II"},{"comment":"The word \"determins\" should be \"determines\" in the sentence about crystal oscillator stability.","section":"Sec. III"},{"comment":"The caption refers to a CPU model \"i9-139000\"; this should be \"i9-13900\" to match the text.","section":"Fig. 2 caption"},{"comment":"The axis label \"NSD (10^-5 Hz/Hz^1/2 @ 10 Hz)\" would be clearer as \"NSD (10^-5 Hz/√Hz @ 10 Hz)\" to avoid confusion between Hz^1/2 and √Hz.","section":"Fig. 2(b) axis"},{"comment":"The phrase \"frequency range of 10 to 500 kHz\" should be written as \"10 kHz to 500 kHz\" for consistency and clarity.","section":"Abstract and Sec. III"},{"comment":"The truncation in Eq. (3) should state explicitly that the sum runs over odd values of k, as the surrounding text implies, so that the reader can reproduce the finite-sum implementation.","section":"Sec. II, Eqs. (1) and (3)"}],"recommendation":"major_revision","confidential_remarks":"The reader's conditional verdict is appropriate. The experimental core is sound and the LM benchmark is a genuine strength, but the K-selection procedure, the simulation-only weighted-fit result, and the dimensional issue in the Ω-counter comparison all need attention before the claims can be accepted. I would not reject: the headline sensitivity measurement is direct and externally benchmarked. I would ask the authors to either add experiments or carefully narrow the generalization claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere's my read on arXiv:2506.04780. It's a real hardware result: they implement a Hilbert-transform-plus-linear-regression frequency estimator on an FPGA and measure sensitivity below 100 µHz/√Hz at 10 Hz over 10–500 kHz, comparable to offline Levenberg-Marquardt fitting on the same data. The clock-stability study (Fig. 3) is a useful practical check, and the truncation-error analysis for the Hilbert series (Fig. 2) is honest—they show convergence and pick K=20 as a trade-off rather than hiding the tuning.\n\nWhat's new is not the Hilbert-transform idea (that's in [16,17,20]) but the specific FPGA implementation, the truncation-error characterization under real clock conditions, and the weighted-regression variant that gives 15–40% improvement in simulation. The top-level sensitivity numbers are directly measured against a benchmark, which is solid.\n\nSoft spots, in order of severity:\n\n1. The robustness claim is thin. All hardware tests use one signal class: A exp(-t/τ) sin(2πft) with fixed τ=2.5 ms and A=2.5 V. Real FID signals can have DC offset, envelope variations, detuning, multiple components, and correlated noise. The phase-unwrapping plus weighted linear regression is known to be sensitive to DC offset, and none of that is characterized. The weighted-fit improvement is simulation-only with white noise (Fig. 6). So the abstract's claim that it does not require pre-knowledge of the analytic expression is supported only for the tested envelope.\n\n2. K=20 is chosen post-hoc from the same data used for the headline sensitivity. The convergence curves are reasonable, but there's no independent validation on a held-out data set. Minor, but worth noting.\n\n3. The Omega-counter comparison is theoretical—they compute the sensitivity limit using δT=20 ps and compare to their measured numbers. That's fine as a benchmark, but it's not a measured Omega-counter performance.\n\n4. No code or raw data are shipped, so the FPGA implementation can't be independently reproduced. That's common for hardware papers but limits verification.\n\nI don't think these are fatal. The central sensitivity claim is directly measured and credible. The gaps are about generalization, not about whether the counter works on the signals it was tested on. The paper is honest about the signal-generator limitations and the improvement with HDAWG, which adds confidence.\n\nWho's this for? People building compact FID magnetometers for geophysics or space, and anyone needing real-time frequency readout of pulsed decaying sinusoids. It deserves a serious referee—the engineering is solid and the performance numbers are useful. I'd send it to review with a request to add at least one off-nominal envelope (e.g., different τ, DC offset) and to state explicitly that the weighted-fit benefit is simulation-only.\n\nMy recommendation: engage with it—send to peer review.\n\nBest,\n[You]","headline":"A solid FPGA frequency counter for FID signals that matches offline LM fitting; the robustness claims are the main caveat.","tokens_in":8567,"tokens_out":2124,"would_cite":true,"duration_ms":24170,"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":"An FPGA frequency counter extracts pulsed-signal frequencies with better than 100 µHz/√Hz sensitivity, matching offline nonlinear fitting.","keywords":["frequency counter","Hilbert transform","free-induction-decay","atomic magnetometer","FPGA","real-time frequency extraction","frequency sensitivity","linear regression"],"falsifier":"Feed the counter a signal with two close-frequency damped tones, an amplitude drift, or a different relaxation time, or a real FID magnetometer output, and measure the 10 Hz noise spectral density; if the sensitivity rises above 100 µHz/√Hz or departs from Levenberg-Marquardt fitting of the same data, the 'no analytic-form pre-knowledge' claim fails for realistic envelopes.","tokens_in":7659,"feed_emoji":"🧲","tokens_out":15048,"duration_ms":140577,"temperature":0.7,"pith_summary":"The paper aims to show that a frequency counter built around a Hilbert-transform linear-regression (HT-LR) algorithm can read out the oscillation frequency of pulsed free-induction-decay (FID) signals in real time, at a sensitivity that matches offline nonlinear fitting. Tested on synthetic damped oscillations between 10 kHz and 500 kHz, the counter reports better than 100 µHz/√Hz sensitivity at 10 Hz when outputting at 200 Hz, and stays below 400 µHz/√Hz at a 1000 Hz output rate. The motivation is that FID atomic magnetometers convert magnetic-field measurements into frequency measurements, so a fast, precise, and algorithm-agnostic frequency readout would make these magnetometers practical for field use. The scheme requires no pre-knowledge of the signal's analytic form, which distinguishes it from fitting-based approaches.","feed_headline":"Real-time FID counter matches offline fitting at 100 µHz/√Hz","feed_subtitle":"Hilbert-transform linear regression reads pulsed FID frequencies at up to 1000 Hz output rate, no analytic form required.","key_machinery":"The engine of the work is the Hilbert-Transform Linear-Regression (HT-LR) algorithm: a discrete Hilbert transform truncated at $K=20$ converts the real signal $x[n]$ into its quadrature $y[n]$, forming the analytic signal $z[n]=x[n]+iy[n]$; the instantaneous phase $\\varphi[n]=\\arctan(y[n]/x[n])$ is unwrapped into a monotone cumulative phase $\\Phi[n]$, and a weighted linear fit of $\\Phi[n]$ versus time yields the frequency as its slope. The weights are the instantaneous amplitudes $|z[n]|$, which suppress low-SNR late-time samples. The hardware implementation uses an FPGA with a dual-core ARM processor and an 18-bit ADC at 1.53846 MSa/s, referenced to a 50 ppb oven-controlled crystal oscillator; one core acquires data while the other processes it. The truncation parameter $K$ trades computation time against precision, and $K=20$ keeps processing faster than an offline Levenberg-Marquardt fit of the same data.","core_discovery":"The central claim is that the HT-LR algorithm, implemented in an FPGA-based counter, turns nonlinear frequency fitting into a linear phase regression without losing accuracy. For each pulse, the device computes the discrete Hilbert transform of the acquired samples to build an analytic signal, extracts the unwrapped instantaneous phase, and performs a weighted linear regression of that phase versus time; the slope is the oscillation frequency, with sample amplitudes as weights to downweight the noisy tail of the decay. On 2.5 ms damped sinusoidal test pulses between 10 kHz and 500 kHz, the extracted frequency's noise spectral density at 10 Hz is better than 100 µHz/√Hz at a 200 Hz output rate, comparable to Levenberg-Marquardt fitting of the same data offline. The paper further shows that the truncation parameter $K$ of the Hilbert transform introduces a systematic frequency offset one order of magnitude smaller than typical heading errors of geomagnetic FID magnetometers, and that the weighted fit improves sensitivity by 15–40% in simulation depending on gate time.","pith_inferences":["Because the tests use a single damped tone from a signal generator, a natural next check is a real atomic magnetometer's FID output, where amplitude drift, detuning, and multi-component spectra would test the 'no analytic-form pre-knowledge' claim in the field.","The weighted fit was evaluated only in simulation, showing 15–40% improvement depending on gate time; a field implementation could verify whether real envelope noise follows the same weighting benefit.","The truncation-induced systematic frequency offset is nearly constant across pulses, so gradiometric or differential measurements that subtract two channels might cancel it; this is a testable consequence not explored in the paper.","The same phase-linear-regression idea could be extended to multi-frequency FID signals by first filtering into bands, since a weighted linear fit of unwrapped phase is model-free per band."],"forward_implications":["The counter's 100 µHz/√Hz floor at 200 Hz output corresponds to about 10 fT/√Hz for 87Rb magnetometers, bringing real-time field sensitivity to that level.","The chosen truncation K=20 keeps per-pulse processing faster than an offline LM fit of the same data on a high-end CPU, while adding a systematic frequency offset one order of magnitude below typical heading errors.","Increasing the output rate to 1000 Hz sacrifices less than a factor of four in sensitivity, which remains below 400 µHz/√Hz.","Within the 10–500 kHz band, HT-LR reaches the same sensitivity as an ideal 20 ps TDC Omega-counter, but with simpler hardware."],"supporting_citations":[{"why":"Provides the discrete-time Hilbert transform formula (Eq. 1) that the HT-LR algorithm starts from.","marker":"[19]"},{"why":"Demonstrated instantaneous-phase retrieval from FID-like signals in principle; this paper adapts it for fast real-time FPGA processing.","marker":"[16]"},{"why":"Showed a related Hilbert-transform phase-extraction scheme on a pulsed magnetometer, which HT-LR extends.","marker":"[17]"},{"why":"Established the theory of Hilbert-transform FID frequency extraction including systematic uncertainties, the error-analysis basis here.","marker":"[20]"},{"why":"Defines the Levenberg-Marquardt fitting algorithm used as the offline benchmark for sensitivity comparison.","marker":"[7]"},{"why":"Presents the Omega-counter, the state-of-the-art counting method whose sensitivity and hardware cost are compared with HT-LR.","marker":"[13]"},{"why":"Supplies the theoretical sensitivity limit for TDC-based Omega-counters (Eq. 5) used to put HT-LR's sensitivity in context.","marker":"[22]"},{"why":"Provides the 20 ps TDC timing-resolution value assumed in the Omega-counter sensitivity comparison.","marker":"[23]"}],"fun_headline_variants":["FID frequency counter reaches 100 µHz/√Hz sensitivity","Hilbert-transform counter rivals offline fitting for FID","No analytic form needed for FID frequency readout","FPGA Hilbert counter for FID at 0.1 mHz/√Hz"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The counter's tested signals are a single synthetic damped sinusoid with fixed amplitude and fixed 2.5 ms decay time, so the claimed sensitivity assumes real FID signals resemble that ideal pulsed tone.","fun_headline_variants_meta":{"raw":{"variants":["FID frequency counter reaches 100 µHz/√Hz sensitivity","Hilbert-transform counter rivals offline fitting for FID","No analytic form needed for FID frequency readout","FPGA Hilbert counter for FID at 0.1 mHz/√Hz"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000839,"raw_usage":{"total_tokens":3659,"prompt_tokens":946,"completion_tokens":2713,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2639}},"tokens_in":562,"tokens_out":2713,"duration_ms":22898,"temperature":1.0,"reasoning_tokens":2639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:32:53.453605+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the counter a signal with two close-frequency damped tones, an amplitude drift, or a different relaxation time, or a real FID magnetometer output, and measure the 10 Hz noise spectral density; if the sensitivity rises above 100 µHz/√Hz or departs from Levenberg-Marquardt fitting of the same data, the 'no analytic-form pre-knowledge' claim fails for realistic envelopes.","supporting_citations":[{"cited_title":"Willsky, and Syed Hamid Nawab","cited_arxiv_id":null,"evidence_quote":"Provides the discrete-time Hilbert transform formula (Eq. 1) that the HT-LR algorithm starts from."},{"cited_title":"”Wide-bandwidth atomic magnetometry via instantaneous-phase retrieval.” Physical Review Research 2.1 (2020): 013213","cited_arxiv_id":null,"evidence_quote":"Demonstrated instantaneous-phase retrieval from FID-like signals in principle; this paper adapts it for fast real-time FPGA processing."},{"cited_title":"Dyer, and Erling Riis","cited_arxiv_id":null,"evidence_quote":"Showed a related Hilbert-transform phase-extraction scheme on a pulsed magnetometer, which HT-LR extends."},{"cited_title":"”Systematic and statistical uncertainties of the Hilbert-transform based high-precision FID frequency extraction method.” Journal of Magnetic Resonance 329 (2021): 107020","cited_arxiv_id":null,"evidence_quote":"Established the theory of Hilbert-transform FID frequency extraction including systematic uncertainties, the error-analysis basis here."},{"cited_title":"”The Levenberg-Marquardt algorithm: implementation and theory.” Numerical analysis: proceedings of the biennial Con- ference held at Dundee, June 28–July 1, 1977","cited_arxiv_id":null,"evidence_quote":"Defines the Levenberg-Marquardt fitting algorithm used as the offline benchmark for sensitivity comparison."},{"cited_title":"”Hardware computing mod- ule for frequencyΩ-counter.” Measurement 229 (2024): 114404","cited_arxiv_id":null,"evidence_quote":"Presents the Omega-counter, the state-of-the-art counting method whose sensitivity and hardware cost are compared with HT-LR."},{"cited_title":"”New frequency counting principle improves reso- lution.” Proceedings of the 20th European Frequency and Time Forum","cited_arxiv_id":null,"evidence_quote":"Supplies the theoretical sensitivity limit for TDC-based Omega-counters (Eq. 5) used to put HT-LR's sensitivity in context."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the 20 ps TDC timing-resolution value assumed in the Omega-counter sensitivity comparison."}],"review_version":1}