{"id":"7d54d77b-b521-4a5a-9880-4233421ce482","arxiv_id":"2501.05875","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An adaptive sub-band search method recovered 79 new pulses from FRB 20190520B, doubling the known pulse count and improving completeness for faint narrow-band bursts.","lead":"BASSET is a new signal-processing toolkit that finds fast radio bursts by searching only the narrow frequency range where each burst appears, rather than averaging over the whole receiver band. Applied to the repeating burst FRB 20190520B observed by FAST, it recovered 79 extra pulses and doubled the known sample, which could make surveys more complete for faint narrow-band bursts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No trials correction is applied to the sub-band SNR that BASSET maximizes (Eq. 7), so the claimed sensitivity gain and 90% completeness threshold in §5.3 may be inflated by selection bias; this threatens the quantitative '79 additional pulses' claim.","rationale":"The paper's central claim has two parts: (1) BASSET improves detection sensitivity for narrow-band pulses, and (2) reprocessing FRB 20190520B yields 79 additional pulses. Part (2) is supported by manual inspection of dynamic spectra (Fig. 11, Table 4) and by the plausible reality of individual pulses. My concern targets the quantitative calibration of part (1): the algorithm selects the frequency range that maximizes Eq. (7) on the same data used to report the resulting SNR, and the search spans many trial widths and centers. This is a classic selection-bias problem, and no trials factor appears in the sensitivity analysis. The 90% completeness threshold and the SNR enhancement ratios are therefore upper bounds on true sensitivity. The paper's own stated limitations—Eq. (1) boxcar/Gaussian model, Triggering failure in §3.3, drift/scintillation residuals in §3.1—confirm that the idealized simulations are not conservative. A noise-only run and a split-sample or trials-corrected recomputation of §5.3 would settle whether the effect is real and how large it is. Until then, the reader's CONDITIONAL verdict is appropriate; I would not change it.","tokens_in":19512,"tokens_out":6956,"duration_ms":75717,"concrete_test":"Run BASSET on pure Gaussian-noise realizations or on time-scrambled/off-pulse FAST data with the same settings and count candidates passing SNRB>5 per unit time; compare with the standard pipeline's false-positive rate. Then, for the Section 5.3 injections, recompute the SNR gain and the 90% completeness threshold after applying a trials correction for the number of independent sub-band/time-range trials, or use a split-sample procedure in which the band is selected on the first half of the data and the SNR is measured on the second half. If the noise-only candidate rate is comparable to the 79-pulse rate, or if the bias-corrected completeness threshold moves back above 7.5, the sensitivity claims need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on SNR values obtained after BASSET selects the frequency range that maximizes Eq. (7) using the same data on which the SNR is then reported. The search spans many trial bandwidths L0..Lm, central frequencies, and adjacent time intervals, so SNRB is a max-statistic over a large family of sub-bands. No trials factor or false-positive calibration appears anywhere in the sensitivity analysis: Section 5.3 injects mock pulses, applies a fixed SNR=5 threshold, and reports a 90% completeness threshold lowered from 7.5 to 6.0 without correcting for the enlarged search space. In Gaussian noise, the maximum over M independent trials is biased upward by roughly sqrt(2 ln M); with thousands of sub-band trials this can be several sigma. Consequently, the 29 new FRB 20190520B pulses with SNRA<5, the SNR enhancement ratios in Fig. 10, and the completeness threshold may substantially overstate BASSET's real sensitivity. The paper itself flags adjacent limitations: Eq. (1) is an oversimplified boxcar/Gaussian model, §3.3 says the Triggering step fails for faint noise-contaminated pulses, and §3.1 attributes residual scatter to frequency drifting and scintillation. But the missing trials correction is more directly load-bearing because it affects every quantitative sensitivity claim, including the headline 'doubling' of the pulse sample.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"BASSET is a modified PRESTO-based single-pulse search toolkit that removes RFI with a wavelet mask and, after bandpass removal, applies an 'adaptive filter' to estimate the spectral band occupied by a candidate via an autocorrelation width and then forms a time series only from that band. The paper tests the method on FAST data, reports 79 additional pulses from FRB 20190520B (increasing the sample from 75 to 154), updates the energy and luminosity distribution, provides a parallelized implementation, and uses 2000 injected Gaussian mock pulses to claim a 90% completeness threshold lowered from fluence SNR 7.5 to 6.0. The claimed gain is attributed to excluding zero-detection frequency bands from the time-series summation.","tokens_in":19852,"tokens_out":13309,"duration_ms":127758,"significance":"The algorithmic idea is relevant and the experimental setup is unusually grounded: the paper ships code, tests on real FAST data, manually vets the new detections, and compares the BASSET time-frequency box against a data-defined 'best location' residual statistic. If the sensitivity claims survive the missing trial-factor calibration, BASSET would be a practical contribution to FRB search completeness. At present, however, the quantitative claims are not fully supported: the SNR statistic is a maximum over many sub-band trials and the energy normalization in Section 5.1 contains an apparent inconsistency. The 79 new pulses themselves are plausible, so the central discovery claim is credible; what needs work is the calibration of the sensitivity gain.","major_comments":[{"comment":"Equation (7) defines SNR_Cand as max(TS_Cand)/STD(TS_Bg) and the Searching Time-Frequency Range step also tests adjacent 1-ms intervals until Eq. (6) fails; along with the L0..Lm matched-filter lengths in Eq. (2), the reported SNR_B is a maximum over a large family of sub-bands and trial intervals. No null distribution, trials factor, or false-positive calibration for this max statistic is given anywhere in the paper. Section 5.3 then compares BASSET and the standard pipeline at the same fixed SNR=5 threshold and quotes the detection counts (1379 vs 1005) and the 90% completeness threshold (7.5 to 6.0) without establishing equal false-alarm rates. Because SNR_B is expected to be biased upward relative to SNR_A by selection over many trials, the claimed sensitivity gain and completeness threshold are not yet established. A noise-only injection set measuring the false-positive rate as a function of SNR_B would resolve this; alternatively the completeness curves should be reported at matched false-alarm probabilities.","section":"§2.3, Eq. (7); §5.3"},{"comment":"The text gives DL = 1.218 Mpc for z = 0.241, which cannot be correct and is presumably a typo for Gpc. Even reading it as 1.218 Gpc, the tabulated energies do not follow from Eq. (15) with the fluences in Table 4: for pulse 1, Fν = 193.14 mJy ms = 0.19314 Jy ms, and Eq. (15) gives roughly 3.4e38 erg, whereas the table lists 4.19e37 erg. Since Fig. 9 and Section 5.1's updated luminosity function are built on these energies, the normalization and units need to be audited and corrected.","section":"§5.1, Eq. (15); Table 4"},{"comment":"Section 3.3 states that the Triggering step 'will fail for faint pulses' when the calibration-noise diode contaminates the spectral intensity, and the three modes tested use relatively high injected SNRs (60, 10, 10). Section 4, however, claims 29 new pulses with SNRA < 5, i.e., exactly the faint regime in which the trigger is acknowledged to fail. The paper should characterize the trigger's recovery fraction as a function of pulse SNR and RFI contamination so that the completeness claims in Section 5.3 can be extended to the faint narrow-band population.","section":"§3.3 and §4"}],"minor_comments":[{"comment":"The text contains repeated typos and spacing artifacts ('T oolkit', 'F AST', 'T able'); 'trail lengths' in Eq. (2) should be 'trial lengths'. These should be cleaned up.","section":"Global"},{"comment":"The quantities 68.42% and 80.70% are labeled 'completeness' but appear to be the fraction of the 627-pulse sample satisfying a residual criterion; please define the denominator and explain why this is a completeness measure.","section":"§3.1, Eq. (14)"},{"comment":"The experiment draws 2000 parameter sets from specified distributions; this is a Monte Carlo injection study, not a Markov Chain Monte Carlo analysis, and the label 'MCMC simulated injection' should be changed.","section":"§5.3"},{"comment":"The percentages 79.13%, 58.57%, and 71.85% are ratios of SNR_B between RFI-contaminated and clean backgrounds, not detection-recovery fractions; the text should say so explicitly.","section":"§3.3"},{"comment":"The manual selection step should be described more fully (how many candidates exceeded SNR_B > 5, how many were rejected by eye, and the reproducibility criterion) so that the reprocessing is reproducible.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The main quantitative claims are all affected by the missing trial-factor calibration, but this is fixable with additional simulations, and the manually vetted new detections are credible. I therefore view this as a major revision rather than a rejection; the authors should also audit the energy normalization in Section 5.1 before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing about this one: BASSET is a genuinely practical contribution. The ACF-based adaptive sub-band selection integrated into PRESTO is new relative to Spitler, Kumar, and Men & Barr, and the authors show it on 627 real pulses, demonstrate a parallel version, ship code, and reprocess FRB 20190520B data to find 79 additional pulses. The residual analysis in Section 3.1, comparing BASSET's chosen time-frequency box to the best possible location, is a nice diagnostic and shows the method is doing something real. The manual vetting of the new pulses makes them plausible detections. I believe the qualitative conclusion: standard full-band searches miss faint narrow-band pulses, and BASSET can recover some of them.\n\nThe soft spots are in the quantified claims. Eq. (7) is a maximum over many trial sub-bands, trail lengths, and time intervals, and no trials factor or false-positive calibration appears anywhere. The 2000-pulse injection study applies a fixed SNR=5 threshold and reports recovery curves, but without correcting for the enlarged search space the 90% completeness threshold of 6.0 and the SNR enhancement ratios are optimistic. The comparison with Niu et al. is also not apples-to-apples: the old search recorded candidates at SNRA>7, while BASSET runs used SNRB>5. The paper itself notes that 50 of the 79 new pulses have SNRA>5, so those would likely have been found just by lowering the old threshold. The 29 pulses with SNRA<5 are the persuasive subset.\n\nThe authors are honest about some limits: Eq. (1) is an oversimplified boxcar/Gaussian model, Section 3.3 states that the Triggering step fails for faint noise-contaminated pulses, and Section 3.1 attributes residual scatter to frequency drifting and scintillation. These are real limitations, and they matter most at the faint end, which is exactly where BASSET claims its main gain. Still, the core method is sound enough to investigate further.\n\nWho should read this: FRB search software users, especially PRESTO/FAST users, and anyone building single-pulse pipelines for repeaters. It deserves serious peer review. A referee should ask for a trials correction or an empirical false-positive calibration, a like-for-like threshold comparison with the previous pipeline, explicit algorithm thresholds, and a break-down of how many of the 79 pulses are due to BASSET versus just a lower detection threshold.","headline":"BASSET is a practical, likely-useful adaptive sub-band search tool with real new pulse candidates, but its headline sensitivity numbers are inflated by a missing trials correction and an uneven comparison to the previous catalog.","tokens_in":20517,"tokens_out":2393,"would_cite":true,"duration_ms":27498,"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":"Standard full-band single-pulse searches for fast radio bursts are incomplete for narrow-band pulses, and BASSET—by averaging only the frequency channels a burst occupies—recovers 79 previously missed pulses from FRB 20190520B.","keywords":["fast radio bursts","single-pulse search","narrow-band pulses","adaptive filter","signal-to-noise ratio","FRB 20190520B","pulse detection completeness","wavelet RFI excision"],"falsifier":"Design a search with mock pulses that include linear frequency drifting or scintillation-induced spectral modulation, inject them into real noise, and compare BASSET's recovered bandwidth to the injected bandwidth and its SNR gain to the predicted beta ratio. If the recovered bandwidth is biased beyond the quoted uncertainties or the SNR gain falls well short of the predicted beta for a substantial fraction of injections, the boxcar/Gaussian model fails for realistic narrow-band pulses.","tokens_in":19347,"feed_emoji":"📡","tokens_out":8162,"duration_ms":70946,"temperature":0.7,"pith_summary":"This paper argues that standard full-band single-pulse searches for fast radio bursts are systematically incomplete for faint, narrow-band pulses, because the time series is formed by averaging over the whole receiver band, including frequency channels where the burst has no emission. The authors introduce BASSET, a toolkit that estimates each candidate's occupied bandwidth and central frequency, then builds the time series from only those channels, raising the pulse's signal-to-noise ratio by the fraction of the band that is pure noise. Applied to 18.5 hours of observations of the repeating burst FRB 20190520B, BASSET recovered 79 additional pulses, doubling the known sample from 75 to 154, with 29 of the new pulses below a signal-to-noise of 5 in the standard pipeline. Injected-pulse simulations set the 90% completeness fluence threshold at 6.0 rather than 7.5. This claim matters because low-energy pulse censuses shape the inferred luminosity function of repeating FRBs, and this result indicates the census is incomplete at the faint end.","feed_headline":"Full-band searches miss narrow-band pulses; BASSET finds 79","feed_subtitle":"Summing only the channels a burst occupies lifts 79 faint pulses above the detection threshold.","key_machinery":"The load-bearing object is the adaptive filter, a three-stage modification to the extraction of the time series: Triggering (matched filtering the bandpass with boxcar trial bandwidths and removing false candidates whose autocorrelation width is small), Searching Time-Frequency Range (estimating the central frequency as the peak of the matched-filter output and extending the time range while adjacent 1 ms intervals keep consistent autocorrelation width and center), and Enhancement (averaging the data over the localised band and reporting an SNR boosted by beta = SNR_local / SNR_full). The autocorrelation function is fit with a Gaussian to measure W_ACF = 2 sigma as the bandwidth estimate. The paper models the ideal pulse spectrum as a boxcar or Gaussian of constant bandwidth, and the argument is that excluding channels outside this band removes only noise, so the reported SNR rises by the fraction of the band that was empty.","core_discovery":"The paper's central discovery is that a search which first localises the burst in frequency and then averages only over that band recovers narrow-band FRB pulses that a full-band pipeline misses. The authors call this bandpass-adaptive filtering: a matched-filter trigger finds candidate times, an autocorrelation of the candidate's bandpass estimates its bandwidth, and an enhancement step builds the time series over the localised frequency range. This raises the signal-to-noise ratio by roughly the fraction of the band that contains no burst, and on the FRB 20190520B data it doubles the number of known pulses, from 75 to 154. The authors' interpretation is that the standard full-band averaging step, not the underlying burst rate, is the limiting factor in the previous census.","pith_inferences":["If narrow emission is as common in other repeating FRBs as in FRB 20190520B, previously published repeater pulse counts and luminosity functions built from full-band searches are likely underestimates at the faint end, and the correction is not uniform across energy.","The bandpass-adaptive principle could be bolted onto other single-pulse search frameworks beyond the one BASSET extends, since their common step of full-band averaging creates the same incompleteness.","The per-event estimated bandwidth could become a useful observable in its own right: the distribution of burst bandwidths and its correlation with energy might constrain emission geometry or propagation effects, something the paper does not explore.","A direct test would be to run BASSET on archival data of other repeating FRBs observed with the same telescope and count how many sub-SNR=7 pulses appear; the paper's own 627-pulse sample suggests such a harvest is possible."],"forward_implications":["BASSET lowers the 90% completeness threshold from fluence SNR 7.5 to 6.0, so surveys using it can claim complete detection of fainter narrow-band pulses than full-band pipelines.","Reprocessing FRB 20190520B raises the known pulse count from 75 to 154, indicating that the burst rate of this repeater at low energies is roughly twice what the standard pipeline census suggested.","The energy distribution of FRB 20190520B remains single-modal after adding the 79 new pulses, so the shape of the luminosity function is not an artefact introduced by the new sample.","The GPU/OpenMP accelerated version achieves a 5.98x speedup on a 140-second test segment, making the adaptive search practical for large survey datasets.","The 90.91% of mock pulses detected by both methods show SNR enhancement, with the median enhancement ratio of about 2.28 consistent with the injected 150 MHz bandwidth of the 500 MHz band."],"supporting_citations":[{"why":"Supplies the previous detection of 75 pulses from FRB 20190520B with a Heimdall-based pipeline; BASSET's reprocessing doubles that sample, so this is the baseline the central claim improves on.","marker":"Niu et al. 2022a"},{"why":"CHIME result that repeating FRBs have narrower bandwidths than one-off bursts; motivates the premise that narrow-band pulses are common and worth targeting.","marker":"Pleunis et al. 2021"},{"why":"Review that defines the standard single-pulse pipeline steps, including full-band averaging; the paper's argument is that this step is where narrow-band pulses lose SNR.","marker":"Petroff et al. 2019"},{"why":"Detected 134 pulses from the same FRB 20121102A data where an earlier search found 41, demonstrating that standard pipeline choices create large detection incompleteness.","marker":"Aggarwal et al. 2021"},{"why":"The 41-pulse search on FRB 20121102A data, used with Aggarwal et al. to show pipeline-dependent sensitivity for repeating FRBs.","marker":"Gourdji et al. 2019"},{"why":"Defines the standard single-pulse search toolkit that BASSET extends; the paper's modifications target the bandpass and time-series extraction steps within this pipeline.","marker":"Ransom 2001"},{"why":"Provides the energy equation used to convert the newly detected pulses' fluences into isotropic energies and update the luminosity function.","marker":"Zhang 2018"}],"fun_headline_variants":["BASSET finds 79 more pulses by summing only the burst's band","Bandpass-adaptive search doubles FRB 20190520B pulse count","New toolkit recovers 79 narrow-band FRB pulses full-band missed","Adaptive sub-band search lifts faint pulses: 79 new FRBs","BASSET: match burst bandpass, detect double the pulses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole gain rests on the assumption that a narrow-band FRB pulse looks like a boxcar or Gaussian bump of roughly constant bandwidth, so every frequency channel outside that bump contains only independent Gaussian noise; real pulses that drift in frequency or scintillate violate this, and the paper itself notes the Triggering step fails for faint pulses when the spectral intensity is contaminated by noise.","fun_headline_variants_meta":{"raw":{"variants":["BASSET finds 79 more pulses by summing only the burst's band","Bandpass-adaptive search doubles FRB 20190520B pulse count","New toolkit recovers 79 narrow-band FRB pulses full-band missed","Adaptive sub-band search lifts faint pulses: 79 new FRBs","BASSET: match burst bandpass, detect double the pulses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000731,"raw_usage":{"total_tokens":3291,"prompt_tokens":985,"completion_tokens":2306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":2209}},"tokens_in":601,"tokens_out":2306,"duration_ms":17549,"temperature":1.0,"reasoning_tokens":2209,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:52.205107+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Design a search with mock pulses that include linear frequency drifting or scintillation-induced spectral modulation, inject them into real noise, and compare BASSET's recovered bandwidth to the injected bandwidth and its SNR gain to the predicted beta ratio. If the recovered bandwidth is biased beyond the quoted uncertainties or the SNR gain falls well short of the predicted beta for a substantial fraction of injections, the boxcar/Gaussian model fails for realistic narrow-band pulses.","supporting_citations":[{"cited_title":"C., Kaspi, V","cited_arxiv_id":null,"evidence_quote":"CHIME result that repeating FRBs have narrower bandwidths than one-off bursts; motivates the premise that narrow-band pulses are common and worth targeting."},{"cited_title":"F., et al","cited_arxiv_id":null,"evidence_quote":"Detected 134 pulses from the same FRB 20121102A data where an earlier search found 41, demonstrating that standard pipeline choices create large detection incompleteness."},{"cited_title":"2019, The Astrophysical Journal Letters, 877, L19","cited_arxiv_id":null,"evidence_quote":"The 41-pulse search on FRB 20121102A data, used with Aggarwal et al. to show pipeline-dependent sensitivity for repeating FRBs."},{"cited_title":"2018, The Astrophysical Journal Letters, 867, L21 —","cited_arxiv_id":null,"evidence_quote":"Provides the energy equation used to convert the newly detected pulses' fluences into isotropic energies and update the luminosity function."}],"review_version":1}