{"id":"8b424770-369b-4e4c-bf32-b3e8a0161123","arxiv_id":"2411.17064","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"VQNS reconstructs noise power spectra from dynamical-decoupling coherence decays by variationally fitting a positive sum of Lorentzians, with confidence intervals and a sensitivity metric to guide experiment selection.","lead":"Researchers propose a variational method that extracts a quantum system's noise spectrum from standard coherence decay measurements by fitting a sum of Lorentzian peaks to the data. It runs in minutes, tolerates noisy or low-resolution data, and identifies which extra measurements would tighten the reconstruction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Low-frequency order-of-magnitude claim is not identified from these CPMG data; VQNS confidence bands exclude model and regularization error, so the central experimental conclusion is unverified.","rationale":"The synthetic benchmarks are genuine positive evidence: Ohmic and 1/f reconstructions show the method performs well when the target is close to the Lorentzian manifold, and the analytical filter-function expressions in SI Sec. I make the computational core credible. The concern is not the optimization machinery, but identifiability of the central experimental conclusion. The paper is candid at one point, noting that Ramsey measurements would be required for zero-frequency spectral weight, and the SI shows regularization sensitivity for high-frequency features; however, the abstract and main text present the order-of-magnitude low-frequency revision as an established result without carrying that caveat through. This matters because the headline experimental finding would collapse if an alternative admissible spectrum with a 10x higher low-frequency plateau fits the same coherence data equally well. The proposed profile scan tests exactly that null-space question, going beyond the reader's basis-representability concern by probing whether the data themselves, rather than the Lorentzian basis alone, can support the claimed precision. The method remains plausible and useful, so no verdict change is needed; the experimental conclusion should be conditionally stated until this identifiability check or an independent low-frequency measurement is performed.","tokens_in":24304,"tokens_out":7481,"duration_ms":78250,"concrete_test":"Run a profile-likelihood identifiability check on the experimental data: for low-frequency plateaus S0 spanning 0.1x to 10x the VQNS value, constrain the zero-frequency spectral weight (e.g., by fixing the integrated weight of the broadest Lorentzians or by using a more flexible positive spline representation), optimize all remaining parameters to minimize the coherence MSE in Eq. (1), and record the minimal loss for each S0. If any plateau within an order of magnitude of the DDNS value attains a loss below the experimental noise floor or below the xi = 1e-2 threshold used in Fig. 4, then the CPMG data do not identify the low-frequency amplitude and the order-of-magnitude revision should be downgraded to a modeling-dependent inference. Repeating the same scan with weight decay values 0.1, 0.2, 0.4, and 0.8 would directly show whether the claimed 10x factor is regularization-driven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing step is the inference from finite CPMG coherence data to the near-zero-frequency noise amplitude. The abstract's order-of-magnitude claim compares one VQNS reconstruction (Fig. 4e) with a DDNS reconstruction of the same data, but the inversion of Eq. (1) is not unique: the filter functions for the experimental set (no FID; SE is the lowest-order sequence) have weak low-frequency sensitivity, and the paper itself says that Ramsey measurements would be needed to assess zero-frequency spectral weight. The VQNS solution is selected by a finite positive-Lorentzian ansatz (Eqs. 2-3) plus AdamW weight decay 0.4 for Fig. 4 (SI Sec. II). This regularization pulls poorly constrained amplitudes toward zero, so a lower low-frequency plateau than DDNS is expected even without new physical information. The reported confidence intervals are standard deviations over 30 random initializations within the same basis; they do not include model error, regularization strength, or the null space of the S-to-C map. The SI (Sec. V, Fig. S1) shows that other spectral features are regularization- and subsampling-dependent, so the same test must be applied to the low-frequency plateau before the order-of-magnitude statement is accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces VQNS, a post-processing variational method to reconstruct Gaussian pure-dephasing noise spectra S(ω) from coherence decays under FID, Hahn echo, and CPMG pulse sequences. The trial spectrum is written as a sum of Nbasis symmetrized Lorentzians, for which the attenuation integral in Eq. (1) has closed-form expressions for CPMG-type filter functions (SI Sec. I). Parameters are optimized with Adam/AdamW to minimize the mean-squared error between measured and trial coherences, with non-negativity constraints and a convergence threshold ξ; independent stochastic restarts yield pointwise means and standard deviations used as confidence intervals. The method is tested on synthetic three-Lorentzian, Ohmic, and 1/f spectra, compared favorably with FTNS under measurement noise and low temporal resolution (SI Sec. VI), and a sensitivity measure Gx(ω) is proposed to guide pulse-sequence selection. The paper applies VQNS to published NV-center CPMG data, reporting a peak at the hydrogen Larmor frequency and a low-frequency plateau about an order of magnitude below the DDNS reconstruction of Ref. 53.","tokens_in":24506,"tokens_out":5264,"duration_ms":47926,"significance":"The core algorithmic idea is attractive and, if the experimental claims are properly supported, would be a useful addition to the noise-spectroscopy toolbox. The analytical Lorentzian filter responses make the optimization fast and portable, and the stochastic-restart confidence intervals plus the sensitivity heuristic are practical tools. The synthetic benchmarks in Figs. 2–3 and the SI demonstrate real robustness to additive coherence noise, low temporal resolution, and non-Lorentzian target spectra, and the SI provides enough hyperparameter detail to reproduce the method. The hydrogen-peak recovery from CPMG data alone, once confirmed by the robustness tests requested below, would be a valuable demonstration. However, the abstract's flagship quantitative claim—that previous measurements overestimated low-frequency noise by an order of magnitude—is not currently established by the evidence in the manuscript, because the low-frequency inversion is weakly constrained by the available data and is entangled with regularization choices.","major_comments":[{"comment":"The order-of-magnitude low-frequency statement is not supported by the CPMG data set used. Because no FID measurement is included and the lowest-order sequence is spin echo, the filter functions have very weak sensitivity near ω → 0; the paper itself notes (p. 10) that \"a more accurate assessment of the zero-frequency spectral weight would require Ramsey measurements.\" The VQNS result in Fig. 4(e) is obtained with AdamW weight decay 0.4 (SI Sec. II), which systematically pulls poorly constrained amplitudes toward zero. The observed factor-of-ten reduction relative to the DDNS reconstruction is therefore expected even in the absence of new spectral information. To retain the abstract claim, the authors must show that the low-frequency plateau is stable under the subsampling and regularization-strength tests of SI Sec. V, as was done for the hydrogen peak.","section":"Fig. 4 and surrounding text; abstract"},{"comment":"The reported confidence intervals are standard deviations over Nruns = 30 random initializations with a fixed basis size, fixed optimizer, and fixed regularization. They therefore quantify only initialization variance within the chosen Lorentzian ansatz; they do not include model error from basis incompleteness, uncertainty in the regularization strength, or the null space of the S → C map arising from the lossy filtering in Eq. (1). The claim that \"narrow confidence intervals ... suggest a high level of accuracy\" (p. 11) is not justified. The authors should either provide a synthetic inversion study demonstrating that these intervals have reasonable coverage for the experimental sequence set, or temper the language to \"precision\" rather than \"accuracy.\"","section":"Fig. 4(e) and 'confidence intervals', p. 11"},{"comment":"SI Sec. V applies two consistency tests to the experimental data—subsampling the available CPMG sequences and varying the weight decay—and concludes that high-frequency features and the feature near the 13C Larmor frequency are not robust, while the hydrogen peak persists. The low-frequency plateau, which is the basis of the abstract's central quantitative claim, is not subjected to the same tests. Without this control, the possibility remains that the plateau height is a regularization artifact rather than a property of the data. The authors should extend Fig. S1 to display the low-frequency region under the same variations and report the resulting range of plateau values.","section":"SI Sec. V, Fig. S1"}],"minor_comments":[{"comment":"The phrase \"previously undetected nuclear species at the diamond surface\" overstates the novelty: Ref. 53 already reported the hydrogen signal, although it required a separate XY8 measurement; the new achievement is recovering it from CPMG data alone.","section":"Abstract"},{"comment":"The caption contains a doubled period: \"set ω0 = 105/(2π) Hz. .\" should read \"set ω0 = 105/(2π) Hz.\"","section":"Fig. 4 caption"},{"comment":"The constraint \"θ ≥ 0\" is ambiguous because θ is a vector containing Bi, di, and ωc,i; the authors should state explicitly which parameter classes are constrained to be non-negative (the SI initialization suggests di is drawn from positive intervals, but the body text should say so).","section":"p. 5, algorithmic step 4"},{"comment":"For the 1/f test, the input spectrum S(ω) = ζ/|ω| diverges at ω = 0, and the text explains that the zero-pulse sequence is omitted for this reason; it would be helpful to state in the main text that the finite plateau shown near ω → 0 is a regularized low-frequency representation rather than a faithful reconstruction of the divergent model.","section":"Eq. (3) and Fig. 3(d)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid methods paper with a promising technique, but the abstract overstates the experimental finding. I suggest the editor require the robustness analysis for the low-frequency plateau and a softened abstract claim; after those changes, the paper would be suitable for publication. The synthetic results and the hydrogen-peak consistency across subsamples are genuine strengths."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the core method is real and probably useful: writing the noise spectrum as a sum of symmetrized Lorentzians, using the known analytic CPMG/FID/SE attenuation functions, and optimizing with Adam/AdamW gives fast, stable reconstructions on synthetic Lorentzian, Ohmic, and 1/f spectra, even with 2% coherence errors and coarse time grids. The sensitivity heuristic (Eq. 4) is a nice, practical way to choose which pulse sequence to add next, and the stochastic-restart confidence bands are honest about run-to-run variability, if not about total uncertainty. Second, the headline experimental claim—that DDNS overestimated low-frequency NV noise by an order of magnitude—is not supported by the data shown. The main text concedes that zero-frequency weight would need Ramsey measurements. The CPMG filter functions have weak low-frequency sensitivity, and the chosen AdamW weight decay (0.4, picked from SI Fig. S1 as the 'most conservative' reconstruction) pulls poorly constrained amplitudes toward zero. So the low plateau relative to DDNS is largely a regularization choice. The SI's own subsampling and weight-decay tests show other features changing; the same test needs to be applied to the low-frequency plateau before that claim is accepted. Also, there is no code release and no quantitative benchmark against Sun–Cappellaro or Bayesian approaches, so the abstract's 'unprecedented accuracy' is not actually demonstrated by comparison.\n\nThe hydrogen Larmor peak is a genuine success: it appears across subsamples and at the known frequency, and it is physically expected from the immersion oil. That part is credible. The absence of a 13C peak is consistent with sample purity, though it is a null result.\n\nI agree with the reader's conditional verdict and with the stress-test: the central experimental conclusion is unverified. But the synthetic validation and the pipeline itself are strong enough to deserve a serious referee. The fixes are straightforward: release code, benchmark against the closest existing methods, and either add Ramsey/FID data or soften the low-frequency claim to 'consistent with limited sensitivity' rather than a factor-of-ten correction.","headline":"The core VQNS pipeline is fast and genuinely useful, but the headline low-frequency experimental claim outruns what the CPMG data can actually support.","tokens_in":25112,"tokens_out":2162,"would_cite":true,"duration_ms":21602,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Variational quantum noise spectroscopy (VQNS) reconstructs a quantum sensor's noise spectrum by fitting a sum of symmetrized Lorentzians to coherence decays measured under standard dynamical-decoupling pulse sequences, with closed-form…","keywords":["quantum noise spectroscopy","variational optimization","dynamical decoupling","Lorentzian basis","nitrogen-vacancy centers","decoherence","filter functions","confidence intervals"],"falsifier":"Generate synthetic coherence curves from a sharply peaked non-Lorentzian spectrum, for example a narrow Gaussian line or a two-level-fluctuator spectrum whose width is much smaller than the narrowest Lorentzian in the basis, and run VQNS on those curves: if the reconstructed peak position, height, or width is visibly biased despite excellent coherence fits, the Lorentzian basis is the limiting assumption. Alternatively, on the same NV sample studied here, perform a direct Ramsey free-induction measurement to probe the zero-frequency spectral weight and compare its initial decay slope with the VQNS spectrum; a slope implying several times more low-frequency power than VQNS reports would contradict the order-of-magnitude suppression claim.","tokens_in":24007,"feed_emoji":"🧲","tokens_out":4943,"duration_ms":50376,"temperature":0.7,"pith_summary":"This paper proposes a post-processing method, variational quantum noise spectroscopy (VQNS), that reconstructs a quantum sensor's environmental noise spectrum from coherence decays measured under ordinary dynamical-decoupling sequences. Instead of inverting the coherence-to-spectrum relation point-by-point, VQNS represents the trial spectrum as a sum of symmetrized Lorentzians and optimizes its parameters so that the predicted coherence curves match all input measurements simultaneously and self-consistently. Because the coherence response to each Lorentzian is known analytically, the optimization is fast and needs no numerical quadrature, making it practical for routine experimental data. On experimental nitrogen-vacancy (NV) center data, VQNS claims to reveal that previous dynamical-decoupling noise spectroscopy overestimated low-frequency noise by an order of magnitude, and to resolve a hydrogen Larmor peak that earlier analysis missed. The method also provides confidence intervals and a sensitivity measure that identifies which additional pulse sequences would tighten the reconstruction at selected frequencies.","feed_headline":"New algorithm reads noise spectra from standard qubit-decay data","feed_subtitle":"Reanalyzing old NV-centre measurements, it cuts low-frequency noise estimates tenfold and resolves the hydrogen signal.","key_machinery":"The load-bearing object is the symmetrized Lorentzian basis expansion, $S_{\\text{trial}}(\\omega) = \\sum_i B_i(\\omega_{c,i}^2/(\\omega_{c,i}^2 + (\\omega-d_i)^2) + \\omega_{c,i}^2/(\\omega_{c,i}^2 + (\\omega+d_i)^2))$, together with the closed-form attenuation functions $\\chi(t)$ for CPMG pulse sequences. Because each Lorentzian's contribution to the coherence can be written analytically, the trial coherence $C^{\\text{trial}}(t;\\theta)$ is evaluated without expensive oscillatory quadrature, and a gradient-based optimizer minimizes the mean-squared error between measured and trial coherences subject to the positivity constraint $\\theta \\ge 0$. Stochastic initialization of the parameters yields slightly different optimized spectra across independent runs, and their pointwise spread defines confidence intervals. A time-integrated filter-function sensitivity $G_x(\\omega) = \\int_0^\\infty dt\\, F_x(\\omega t) e^{-\\chi_x(t)}$ identifies which additional pulse sequences tighten the reconstruction at selected frequencies, enabling an iterative experiment-theory loop.","core_discovery":"The paper claims that a single variational algorithm applied to standard FID, spin-echo, and CPMG coherence measurements can recover the full noise power spectrum $S(\\omega)$ with accuracy and precision that exceed pointwise DDNS and Fourier-transform noise spectroscopy, while remaining robust to measurement noise and sparse temporal sampling. Central to the claim is the empirical demonstration on shallow NV centers in diamond: VQNS, using the same CPMG coherence data that DDNS had analyzed, produces a spectrum whose low-frequency weight is roughly ten times smaller than the double-Lorentzian DDNS reconstruction, and which contains a clear peak at the hydrogen Larmor frequency (≈1.93 MHz at 454 G) that DDNS only revealed under a separate XY8 sequence. The paper treats this as evidence that VQNS extracts greater physical insight from the same experiment, setting up precision noise-spectroscopy-based quantum metrology.","pith_inferences":["Beyond the paper's specific NV demonstration, VQNS implies that the accuracy of any noise-spectrum reconstruction is limited by the match between the true spectrum and the chosen Lorentzian basis; for spectra with sharp resonances or long non-Lorentzian tails, the confidence intervals from random restarts may substantially understate the true reconstruction error.","The paper's claim that prior DDNS overestimated low-frequency noise is a strong physical conclusion that would be independently testable with a direct Ramsey free-induction measurement at the same NV center, which is sensitive to the zero-frequency spectral weight that VQNS could not directly probe from CPMG data alone.","The sensitivity measure suggests a closed-loop experiment-design protocol in which preliminary FID/SE measurements seed a trial spectrum, the measure identifies complementary pulses, and new data recursively tighten the confidence intervals; this protocol could be automated for future quantum-sensor calibration.","If the Lorentzian basis were replaced by another overcomplete set with closed-form coherence responses, such as Gaussian or algebraic-decay kernels, VQNS's basis sensitivity could be quantified directly; such a comparison would clarify how much of the reported low-frequency suppression is intrinsic to the data versus imposed by the basis."],"forward_implications":["Existing CPMG, spin-echo, and FID coherence datasets can be reanalyzed with VQNS without any new experiments, so previously published DDNS reconstructions may be worth revisiting for systematic low-frequency bias.","Spectral features such as nuclear Larmor peaks can be extracted directly from ordinary CPMG decays, without requiring dedicated pulse sequences or supplementary measurements.","The sensitivity measure provides a quantitative criterion for choosing the next pulse sequence to add, making noise-spectrum refinement a systematic iterative process rather than a heuristic one.","Because VQNS tolerates underconverged, noisy, and poorly time-resolved coherence data, it extends noise spectroscopy to systems where high-statistics measurements are difficult to obtain.","The method's computational cost is small enough (roughly minutes per run on a laptop, and trivially parallelizable) that confidence intervals and repeated analyses can be performed routinely in the experimental workflow."],"supporting_citations":[{"why":"Defines the pointwise dynamical-decoupling noise spectroscopy (DDNS) approach that VQNS improves on and whose low-frequency reconstruction is later compared against on NV data.","marker":"[28]"},{"why":"Provides the experimental NV-center coherence measurements and DDNS spectrum that VQNS reanalyzes, including the hydrogen Larmor peak that DDNS missed.","marker":"[53]"},{"why":"Introduces Fourier-transform noise spectroscopy, the complementary method in the opposite limit, which serves as a comparison baseline for noise resilience and temporal resolution.","marker":"[61]"},{"why":"Supplies the closed-form attenuation functions for Lorentzian spectra under CPMG pulse sequences, which make the variational optimization analytically tractable.","marker":"[89]"},{"why":"The stochastic gradient-based optimizer that drives the variational minimization of the coherence mismatch.","marker":"[78]"},{"why":"Grounds the physical motivation for the Lorentzian basis in the exponential decay of environmental correlation functions.","marker":"[90]"}],"fun_headline_variants":["Variational method cuts low-frequency noise estimate tenfold","Self-consistent noise spectra from standard qubit-decay data","Same CPMG data, tenfold better low-freq noise, plus hydrogen peak","Error-resilient VQNS extracts sharper NV noise spectrum","Reanalysis of NV data reveals hydrogen and fixes noise by 10x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reconstruction is valid only if the true noise spectrum can be well approximated by a finite sum of positive, frequency-symmetric Lorentzian bumps whose centers, widths, and heights are optimized; if the real spectrum contains sharper resonances or longer non-Lorentzian tails, the reported confidence intervals may be too narrow and the low-frequency estimate could be biased.","fun_headline_variants_meta":{"raw":{"variants":["Variational method cuts low-frequency noise estimate tenfold","Self-consistent noise spectra from standard qubit-decay data","Same CPMG data, tenfold better low-freq noise, plus hydrogen peak","Error-resilient VQNS extracts sharper NV noise spectrum","Reanalysis of NV data reveals hydrogen and fixes noise by 10x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000336,"raw_usage":{"total_tokens":1828,"prompt_tokens":881,"completion_tokens":947,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":856}},"tokens_in":497,"tokens_out":947,"duration_ms":9187,"temperature":1.0,"reasoning_tokens":856,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:36:47.111440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic coherence curves from a sharply peaked non-Lorentzian spectrum, for example a narrow Gaussian line or a two-level-fluctuator spectrum whose width is much smaller than the narrowest Lorentzian in the basis, and run VQNS on those curves: if the reconstructed peak position, height, or width is visibly biased despite excellent coherence fits, the Lorentzian basis is the limiting assumption. Alternatively, on the same NV sample studied here, perform a direct Ramsey free-induction measurement to probe the zero-frequency spectral weight and compare its initial decay slope with the VQNS spectrum; a slope implying several times more low-frequency power than VQNS reports would contradict the order-of-magnitude suppression claim.","supporting_citations":[{"cited_title":"A.; Suter, D","cited_arxiv_id":null,"evidence_quote":"Defines the pointwise dynamical-decoupling noise spectroscopy (DDNS) approach that VQNS improves on and whose low-frequency reconstruction is later compared against on NV data."},{"cited_title":"J.; Isoda, T.; Itoh, K","cited_arxiv_id":null,"evidence_quote":"Provides the experimental NV-center coherence measurements and DDNS spectrum that VQNS reanalyzes, including the hydrogen Larmor peak that DDNS missed."},{"cited_title":"Fourier transform noise spectroscopy","cited_arxiv_id":null,"evidence_quote":"Introduces Fourier-transform noise spectroscopy, the complementary method in the opposite limit, which serves as a comparison baseline for noise resilience and temporal resolution."},{"cited_title":"Accuracy of dynamical-decoupling-based spectroscopy of Gaussian noise","cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form attenuation functions for Lorentzian spectra under CPMG pulse sequences, which make the variational optimization analytically tractable."},{"cited_title":"Hydrodynamic fluctuations, broken symmetry, and correlation functions; CRC Press, 2018","cited_arxiv_id":null,"evidence_quote":"Grounds the physical motivation for the Lorentzian basis in the exponential decay of environmental correlation functions."}],"review_version":1}