{"id":"4bbba6ff-693b-4b59-ae4f-59d50adab584","arxiv_id":"2608.04863","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Prior-based MVT sampling is recast as an inverse problem with Fisher-information threshold designs, validated on 10,000 measured pulses, with model mismatch limiting the predictive regime.","lead":"This paper recasts prior-based multi-voltage threshold sampling as a structured inverse problem, deriving identifiability, timing-error, and threshold-design theory for strictly unimodal pulse families. It instantiates the theory for bi-exponential pulses and reports improved amplitude reconstruction on 10,000 measured LYSO/SiPM pulses, while exposing a regime where model mismatch limits predictive power.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical validation does not test the claimed design equation (Eq. 110): the recipe actually deployed is D-optimal log det I, not the nuisance-profiled effective-information stationarity, and the well-specified precondition fails at median rho_bias ≈ 0.60.","rationale":"The reader's weakest_assumption already identifies the core issue: the derived stationarity recipe Eq. (109) is not the configuration actually validated, and the paper's own diagnostic gives median rho_bias ≈ 0.60. My analysis sharpens this: the experiments use D-optimality (log det I) as the design objective, which is a different scalar from the effective-information objective Phi_A that yields Eq. (110). The D-optimal design is introduced specifically to fix ill-conditioning that the Phi_A objective creates, so the experimental comparison cannot be read as validating Eq. (110). This is load-bearing because the paper's central assertion is that threshold design is governed by the effective-information equation, not merely by 'some Fisher-guided criterion.' The deterministic identifiability analysis and the bi-exponential Chebyshev-system argument are independent and appear sound; the formal derivation of Eq. (110) under its stated assumptions is also plausible. But the empirical support for the central design claim is missing at exactly the point where the theory and validation should meet. A direct rerun with the Eq. (109)/(151) thresholds would settle whether the mathematically derived recipe actually delivers the claimed improvement, or whether the observed gains come from the D-optimal regularization and coverage constraints. The CONDITIONAL verdict remains appropriate, with this additional validation condition; hence unchanged.","tokens_in":55450,"tokens_out":4759,"duration_ms":52623,"concrete_test":"Run the single-event recipe of Section IX-B3 on the same 3,304 photopeak pulses: initialize uniformly, iterate the gamma-update (Eq. (148)) and the V-update (Eq. (151)) under the same spacing and pair-gap constraints used for D-optimality, and reconstruct with the same Gauss-Newton solver and A_ref. Compare the median relative amplitude error, the <5% pass rate, and the empirical MSE distribution against both uniform thresholds and the D-optimal set V*. Also compare the predicted B*_A from Eq. (110), computed with the measured beta_A,tot from Eq. (136), against the empirical MSE. If the Eq. (109)/(151) thresholds do not match or improve on the D-optimal result, the validated improvement is attributable to the D-optimal regularization and coverage constraints, not to the paper's claimed effective-information design equation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that threshold design is governed by the nuisance-profiled effective-information equation, Eq. (110). But the validation in Section X never implements that design. The recipe derived from Eq. (110) is the single-event fixed point in Section IX-B3, whose threshold update maximizes the per-pair projected Fisher contribution Phi_A (Eq. (151)) with gamma* from Eq. (148). Section X-C explicitly abandons this objective: it replaces Phi_A by log det I (Eq. (192)) with a coverage constraint (Eq. (197)), because the Phi_A objective clusters thresholds and makes J^T W J ill-conditioned (Section X-C1). The D-optimal set V*={54,105,213,264,315,664,714,765} is therefore not a solution of Eq. (109)/(151), and the reported improvement over uniform thresholds demonstrates only that some Fisher-guided placement with conditioning regularization helps on this dataset. It does not test Eq. (110). This is compounded by Section IX-D: median rho_bias = 0.598 implies (beta_direct_A)^2 is comparable to the variance floor, so the well-specified assumption epsilon ≈ 0 behind the bias-free recipe is violated; the bias term in Eq. (110) is not accounted for in the validation. Without a direct test of Eq. (109)/(151) thresholds, the headline 'recast as principled inferential framework' rests on an unvalidated design rule.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a unified mathematical treatment of prior-based Multi-Voltage Threshold (MVT) sampling for strictly unimodal pulse families, treating the ordered threshold-crossing times as a structured inverse problem. It proves deterministic identifiability conditions via Chebyshev/Wronskian arguments, introduces a stochastic timing-error model with an affine voltage variance and a diagonal pseudo-Gaussian covariance, derives a nuisance-profiled effective-information quantity, and presents a threshold-design theory culminating in a closed-form amplitude-MSE proxy (Eq. (110)). The framework is instantiated for a five-parameter bi-exponential pulse model, and implementation recipes are given for single- and multi-event scenarios. The experimental part uses 10,000 measured 22Na/LYSO/SiPM pulses: it reports a model-mismatch diagnostic (median ρ_bias ≈ 0.60), and then compares uniform threshold placement with a D-optimal placement, finding a reduction in median amplitude bias from −4.8% to −1.2% and an increase in the fraction of pulses within 5% relative error from 51.3% to 82.0% on the photopeak population.","tokens_in":55932,"tokens_out":2979,"duration_ms":33652,"significance":"If the theoretical claims are correct, the paper provides a valuable conceptual advance: it recasts MVT threshold placement as a Fisher-information optimization problem rather than an empirical heuristic, and it contributes rigorous identifiability results for bi-exponential pulses from paired crossings. The Chebyshev/Wronskian-based identifiability proof (Section VIII-B) is coherent and self-contained, and the Schur-complement/Fisher-decomposition algebra (Sections VI–VII) is internally consistent. The paper also deserves credit for explicitly acknowledging the surrogate status of its noise model and for reporting a diagnostic that quantifies model mismatch rather than claiming perfect specification. However, the experimental validation tests an objective different from the one derived as the central design rule, which limits the evidentiary support for the paper's strongest claims.","major_comments":[{"comment":"The validation does not test the claimed design equation: after deriving the stationarity condition Eq. (109) and the recipe Eq. (151) based on the nuisance-profiled effective-information Φ_A, Section X-C1 explicitly abandons that objective in favor of log det I (Eq. (192)), because the Φ_A objective clusters thresholds and makes J^T W J ill-conditioned. The threshold set V* = {54,105,213,264,315,664,714,765} mV is therefore not the minimizer of Eq. (110), and the reported improvement over uniform thresholds demonstrates only that some Fisher-guided placement with conditioning regularization helps on this dataset. The abstract's claim that the experiments 'confirm' the framework's threshold designs is too strong; a direct test of the Eq. (109)/(151) configuration, or a clear statement that Eq. (110) itself is not validated, is needed.","section":"Section X-C, Eq. (192)"},{"comment":"The well-specified precondition behind the bias-free design recipe is not satisfied: Table I reports a median ρ_bias = 5.98×10^−1 and a 95th percentile of 4.49, so (β_direct_A)^2 is comparable to or larger than the variance floor B_0^(A). The recipe in Section IX-B3 and the simplified stationarity Eq. (132) rely on ε(t) ≈ 0, yet the paper's own diagnostic shows that this condition is not uniformly valid. Consequently, the validation results do not support the full misspecified MSE proxy in Eq. (110), which includes β^2_{A,tot}; the bias term is not accounted for in the experiments. The paper should either incorporate the bias projection into the validated design or explicitly restrict the validation claims to the well-specified (ε≈0) regime and explain why the D-optimal procedure remains useful despite the mismatch.","section":"Section IX-D, Table I"},{"comment":"The reference amplitude A_ref is obtained by fitting the same five-parameter bi-exponential model to the full 50,000-sample waveform, so the reported errors measure agreement with the model-based full-waveform estimate, not accuracy against an independent ground truth. This is acknowledged in the text, but it weakens the external-validity claim that the framework yields accurate amplitude reconstruction; any systematic bias shared by the bi-exponential model remains in A_ref. The paper should state more prominently that the validation is a model-internal comparison, and ideally include at least one independent check (e.g., against the directly observed peak voltage V_peak or a separately calibrated energy scale).","section":"Section X-B, Eqs. (184) and reference standard"}],"minor_comments":[{"comment":"The text refers to 'Eqs. (R.6a)–(R.6j)' and 'Eq. R.6d' without a main-text definition of these equations; the notation appears to point to an appendix that is not clearly identified. Please label the appendix explicitly or inline the relevant formulas.","section":"Section IX-D"},{"comment":"The dynamic-range coverage constraint ΔV_cov_min = (V_peak,min − b)/(2N) is introduced as an engineering heuristic; the remark contrasting it with the bandwidth diagnostic is useful, but the choice of the factor 1/(2N) is arbitrary. A short justification or citation would help.","section":"Section X-C3, Eq. (197)"},{"comment":"The paper alternates between V_n and V_k for thresholds and between ordered-crossing and threshold-indexed notation. While the conventions are defined in Section II-A, the frequent switching (e.g., Eqs. (6), (85), (194)) makes the reading unnecessarily difficult; a consistent subscript convention would improve clarity.","section":"General notation"},{"comment":"Figures 1–3 report quantitative statistics (e.g., ρ_bias values) but the captions do not state the definitions of all plotted quantities; for example, Figure 2 and Figure 3 both show ρ_bias but with different axes and color scales. Please make the captions self-contained.","section":"Figure captions"}],"recommendation":"major_revision","confidential_remarks":"The core mathematical development appears sound, and the identifiability and Fisher-algebra sections are valuable. The main weakness is the mismatch between the derived design rule and the validated objective: the paper claims experimental confirmation of the effective-information design equation, but the implemented D-optimal design is a different criterion introduced precisely because the derived objective is numerically unstable. This gap should be addressed before publication by either validating the original design equation or explicitly re-scoping the claims. The model-internal reference standard and the non-negligible mismatch diagnostic further argue for a revised manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The formal part of this paper is the real contribution: they turn MVT sampling into a structured inverse problem with explicit identifiability conditions and a Fisher-information framework for threshold design. The stress-test note is on point: the validation does not test the headline design equation. Section X-C explicitly replaces the nuisance-profiled objective with D-optimal log det I because the former clusters thresholds and makes the normal matrix ill-conditioned. The improvement over uniform thresholds shows that some Fisher-guided placement helps, not that Eq. (110) works as stated.\n\nWhat is genuinely new and good: the bi-exponential identifiability proof via Chebyshev systems and the Wronskian is clean and self-contained. The general framework organizing deterministic identifiability, stochastic timing-error propagation, mismatch bias, and threshold design is a genuine unification—no prior work puts these together. The empirical comparison of D-optimal vs uniform thresholds is honestly reported: median amplitude bias improves from -4.8% to -1.2%, and the paper explicitly flags the misspecification diagnostic (median rho_bias ~0.6) and acknowledges that Aref is the same model's full-waveform fit, not ground truth.\n\nThe soft spots are real but addressable. The central claim overreaches: the abstract says the 'resulting predictions' are validated, but the design actually used in Section X is D-optimality, not the stationarity equation. The paper is transparent about this in Section X-C1, but the framing still implies the effective-information equation was tested. It wasn't. With median rho_bias ~0.6, the bias term in Eq. (110) is not negligible, yet the validation ignores it. The model-internal reference means the empirical improvement is a recovery-of-the-model test, not an absolute accuracy test. And no code or data is released, which compounds the reproducibility concern.\n\nWho will get value from this: detector physicists and PET instrumentation people. They should read the formal sections carefully and treat the validation as a demonstration that Fisher-guided placement improves conditioning, not as confirmation of the specific design rule. The identifiability result alone is worth citing.\n\nRecommendation: send it to peer review—the framework deserves referee time. But the authors should be asked to either validate the actual stationarity design or soften the claims to 'Fisher-guided threshold placement', and to release code and data. The empirical section as written overstates what is established.","headline":"Solid formal framework for MVT inverse problems, but the validation tests a D-optimal workaround rather than the claimed effective-information design equation.","tokens_in":56286,"tokens_out":4186,"would_cite":true,"duration_ms":43696,"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 recasts multi-voltage-threshold sampling as a structured inverse problem and derives a Fisher-information equation for optimal threshold placement.","keywords":["multi-voltage threshold sampling","structured inverse problem","Fisher information","identifiability","threshold design","scintillation detectors","partial triggering","bi-exponential pulse model"],"falsifier":"Measure the actual crossing-time covariance matrix of a comparator/TDC chain on real pulses and check it against the diagonal affine prediction; if off-diagonal correlations are significant or the variance deviates from $\\sigma_{\\mathrm{th}}^2 + \\kappa V$ at low thresholds, the Fisher-optimal ladders predicted by the theory will not reproduce the realized estimator variance. A simpler decisive test: on held-out photopeak pulses, compute realized amplitude MSE for the D-optimal ladder and for uniform ladders; if the theory-guided ladder does not systematically beat uniform spacing under controlled initialization, the central design claim fails.","tokens_in":55231,"feed_emoji":"⚡","tokens_out":12878,"duration_ms":112126,"temperature":0.7,"pith_summary":"This paper is trying to give prior-based Multi-Voltage Threshold (MVT) sampling a unified mathematical foundation. Its central proposal is that, for strictly unimodal pulse families, the ordered set of paired threshold-crossing times should be treated as the observation map of a structured inverse problem, with identifiability, noise propagation, model mismatch, and threshold design all following from that map. On that foundation it derives a stochastic timing-error model, a leading-order bias formula for model mismatch, and a threshold-design theory in which optimal ladders maximize nuisance-profiled Fisher information; the closed-form amplitude-MSE proxy is $B^*_A = [\\sum_n \\Phi_A^{(n)}(V^*_n; \\gamma^*)]^{-1} + \\beta^2_{A,\\mathrm{tot}}$. The value of getting this right is that detector threshold placement stops being an empirical heuristic and becomes a computable optimization with explicit recipes and a diagnostic for when the model is misspecified.","feed_headline":"Sparse threshold sampling becomes a Fisher-information design problem","feed_subtitle":"Fisher-optimal thresholds cut median amplitude bias from -4.8% to -1.2% on photopeak events.","key_machinery":"The central object is the paired-crossing observation map $S: \\theta \\mapsto t$ defined by the level-set equations $f(t_{r,n};\\theta)=V_n$ and $f(t_{f,n};\\theta)=V_n$, whose Jacobian is $J_{kj} = -(\\partial f/\\partial \\theta_j)/(\\partial f/\\partial t)$. The identity that carries the design theory is the nuisance-profiled effective-information decomposition: for target $\\vartheta$ and nuisance $\\eta$, the total effective Fisher information equals $\\sum_n \\Phi_\\vartheta^{(n)}(V_n; \\gamma)$ with per-pair projected contributions $\\Phi_\\vartheta^{(n)} = \\tilde{g}_r^2/D_r + \\tilde{g}_f^2/D_f$ evaluated at the global coupling vector $\\gamma = (I_{\\eta\\eta}^{\\mathrm{tot}})^{-1} I_{\\eta\\vartheta}^{\\mathrm{tot}}$. Thresholds are fixed by the stationarity condition $d\\Phi/dV = 2(\\Delta I_{\\mathrm{eff}})^2 \\beta \\, \\partial\\beta/\\partial V$, and the closed-form amplitude-MSE proxy is $B^*_A = [\\sum_n \\Phi_A^{(n)}(V^*_n; \\gamma^*)]^{-1} + \\beta^2_{A,\\mathrm{tot}}$. A dimensionless D-optimal criterion $\\log \\det(\\tilde{I})$ is added as a regulariser to stop threshold clustering from making the full Fisher matrix near-singular.","core_discovery":"The central claim is that prior-based MVT is a well-posed structured inverse problem for strictly unimodal pulses. The forward map $S: \\theta \\mapsto t$, defined implicitly by $f(t_k; \\theta) = V_{\\ell(k)}$, is shown to be a $C^2$ embedding on the admissible domain when injectivity, full Jacobian rank, and properness hold, so the deterministic recovery problem is globally identifiable; for the bi-exponential family the paper verifies these hypotheses on a compact physical prior and shows three threshold pairs suffice. The stochastic layer adds a diagonal timing-error model from Campbell's theorem, a leading-order mismatch bias $\\beta = (J^\\top W J)^{-1} J^\\top W \\Delta t_{\\mathrm{bias}}$, and a profiled effective-information identity $\\Delta I_{\\mathrm{eff}}^{\\mathrm{tot}} = \\sum_n \\Phi_\\vartheta^{(n)}(V_n; \\gamma)$ that turns threshold design into a stationarity problem with the closed-form proxy $B^*_A = [\\sum_n \\Phi_A^{(n)}]^{-1} + \\beta^2_{A,\\mathrm{tot}}$. The paper's own experiments on 10,000 $^{22}$Na/LYSO/SiPM pulses show Fisher-guided (D-optimal) thresholds reduce median amplitude bias from -4.8% to -1.2% relative to uniform spacing in the photopeak regime, while the mismatch diagnostic $\\rho_{\\mathrm{bias}} \\approx 0.60$ marks the boundary of the well-specified regime.","pith_inferences":["If the diagonal noise model is the main source of error, replacing it with a measured full covariance matrix could shift the predicted optimal thresholds; this is a direct test of the framework's quantitative layer.","The same embedding-and-profiling machinery should transfer to timing-target designs and to other strictly unimodal pulse families, but the paper only instantiates amplitude-target design for the bi-exponential model.","The rho_bias diagnostic could serve as an online per-pulse mismatch monitor, flagging events for which closed-form threshold recipes should give way to a misspecification-aware estimator.","For pile-up or high-rate regimes, the single-event isolation assumption fails, so the identifiability and design results do not automatically extend there."],"forward_implications":["Threshold design for MVT becomes a computable optimization: given fitted pulse and noise parameters, the stationarity equations yield an executable ladder, so future detectors can be designed from the theory rather than from ad hoc spacing rules.","Identifiability has a concrete hardware consequence: for the five-parameter bi-exponential model, at least three active threshold pairs are needed for finite amplitude or timing bounds; fewer thresholds leave the inverse problem underdetermined in the nuisance directions.","Heterogeneous threshold placements are superadditive: multiple pairs jointly disentangle amplitude-time nuisance coupling better than any single pair, so the optimal design is not simply a sum of per-pair information scores.","Because single-parameter Fisher optimization can concentrate thresholds at support voltages and make the Jacobian near-singular, the paper shows that D-optimal regularization is needed in practice; on the photopeak dataset it improved conditioning and cut median amplitude bias roughly fourfold.","The framework exposes its own regime boundary: partial triggering and model mismatch erode the predictive power of Fisher-guided optimization, and the bias-to-variance diagnostic near 0.60 marks where closed-form well-specified recipes should be replaced by misspecification-aware fitting."],"supporting_citations":[{"why":"Introduced MVT sampling as sparse threshold-crossing acquisition; defines the hardware observation regime this paper formalizes.","marker":"[1]"},{"why":"Hadamard well-posedness criterion used to define non-identifiability of the deterministic inverse problem.","marker":"[15]"},{"why":"Modern inverse-problems formulation of uniqueness and stability that the global embedding theorem is built on.","marker":"[17]"},{"why":"Campbell's theorem supplies the exact first- and second-moment formulas from which the affine noise law sigma_th^2 + kappa V is derived.","marker":"[25]"},{"why":"Gaussian log-likelihood and weighted least-squares formulation underlying the Gauss-Newton estimator.","marker":"[35]"},{"why":"Slepian-Bangs high-SNR reduction that justifies treating noise covariance as locally constant and makes the Fisher matrix the design criterion.","marker":"[36]–[38]"},{"why":"Classical optimal experimental design result that a p-parameter model needs at most ceil(p/2) support voltages; this motivates the D-optimal regularization against threshold clustering.","marker":"[63]"}],"fun_headline_variants":["MVT sampling gets a unified inverse-problem theory","Fisher-optimal thresholds cut pulse amplitude bias 75%","From threshold heuristic to principled inference framework","Structured inverse problem unifies MVT threshold design","Identifiability, error, and design unified for MVT sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a simple formula — thermal noise plus shot noise proportional to threshold voltage plus timing-quantization jitter, treated as independent — accurately describes real crossing-time errors; the paper's own validation finds this well-specified condition only partly holds, with median bias-to-variance ratio near 0.60.","fun_headline_variants_meta":{"raw":{"variants":["MVT sampling gets a unified inverse-problem theory","Fisher-optimal thresholds cut pulse amplitude bias 75%","From threshold heuristic to principled inference framework","Structured inverse problem unifies MVT threshold design","Identifiability, error, and design unified for MVT sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000736,"raw_usage":{"total_tokens":3371,"prompt_tokens":1105,"completion_tokens":2266,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":721,"completion_tokens_details":{"reasoning_tokens":2188}},"tokens_in":721,"tokens_out":2266,"duration_ms":17289,"temperature":1.0,"reasoning_tokens":2188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:26:38.170466+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual crossing-time covariance matrix of a comparator/TDC chain on real pulses and check it against the diagonal affine prediction; if off-diagonal correlations are significant or the variance deviates from $\\sigma_{\\mathrm{th}}^2 + \\kappa V$ at low thresholds, the Fisher-optimal ladders predicted by the theory will not reproduce the realized estimator variance. A simpler decisive test: on held-out photopeak pulses, compute realized amplitude MSE for the D-optimal ladder and for uniform ladders; if the theory-guided ladder does not systematically beat uniform spacing under controlled initialization, the central design claim fails.","supporting_citations":[{"cited_title":"Hadamard,Lectures on Cauchy’s Problem in Linear Partial Differential Equations","cited_arxiv_id":null,"evidence_quote":"Hadamard well-posedness criterion used to define non-identifiability of the deterministic inverse problem."},{"cited_title":"Comparison of energy window choice and parameter implementation in dual energy window scatter correction performance in 3d pet,","cited_arxiv_id":null,"evidence_quote":"Classical optimal experimental design result that a p-parameter model needs at most ceil(p/2) support voltages; this motivates the D-optimal regularization against threshold clustering."}],"review_version":2}