{"id":"28b0d0b3-90de-43b1-b432-6bdcad0ab574","arxiv_id":"1908.08124","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A maximum likelihood SAR classifier for delayed versus instantaneous targets is extended with contrast-independent confidence thresholds, validated only on simulated speckle images.","lead":"This paper adds confidence levels to a maximum likelihood method that separates delayed from instantaneous scatterers in synthetic aperture radar images. The thresholds are computed from simulated image ensembles and are designed to keep misclassification rates below a preset level without knowing the target contrast.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The p-guarantee is definitional inside the model: the Monte Carlo generator hard-codes the same independence, Gaussianity, and known-delay-profile assumptions that the thresholds rely on, so the load-bearing, untested step is the transfer of these cdfs to real SAR speckle.","rationale":"The reader's CONDITIONAL verdict and weakest-assumption identification match my reading, and I find the model-to-reality transfer to be the single most load-bearing premise of the central claim. The in-model construction is correct: (58) sets the thresholds as p-quantiles of the cdfs that, by (59), define the error rates, so inside the model the bound is nearly definitional; (61) then extends it to arbitrary contrast mixtures because a min/max threshold bounds cdft(l-;q) and 1 − cdfs(l+;q) for every q separately. This is why I do not see an internal inconsistency: the failure mode is unvalidated modeling premises, not a broken derivation. The sharpest form of the concern, which reinforces the reader's point, is that the simulator and the likelihood are built on the same structural assumptions: the Figure 2 caption generates each ambiguity line independently from (31), exactly the independence that (46b) assumes, so the Monte Carlo experiments in Figures 3–6 verify consistency of the pipeline with itself, not with the imaging physics. The delta-correlated delay process (27), the known profile (34), and circular Gaussianity are likewise baked into both the data generator and the model used by the classifier; all are plausible but unmeasured. I also note the secondary gaps the reader listed: (61) is computed over the finite grid q in {0.0,…,0.9} with step 0.1 and validated at only two contrasts in Section 7, and Section 6.2 assumes cdf continuity and monotonicity without evidence. These do not by themselves defeat the claim, but they mean the advertised 'any probability distribution of contrasts' guarantee is demonstrated on a grid within a self-consistent simulation; a minor reproducibility blemish is that the simulation footnote names mvnpdf (a density evaluator) as the random generator. I credit the paper's independent support: the detailed derivation of the imaging kernel and correlation functions (10)–(32), the convex Gaussian ML formulation behind (49), the transparent scoping in the abstract and Section 8, and the honest discussion of what remains untested. The proposed concrete test — full-field simulation with cross-line correlations, and ultimately a measured scene with a known delay — would separate an in-model self-consistency result from a genuine error-rate guarantee for SAR.","tokens_in":17270,"tokens_out":35393,"duration_ms":333943,"concrete_test":"Regenerate test data without the product-structure assumption: sample the full coordinate-delay Gaussian field with the complete PSF correlation kernel (10), so that values on different ambiguity lines are correlated as the imaging physics requires, then apply algorithm (52) with the thresholds from (61) and measure r'_s and r'_t over at least 10^4 realizations per contrast (q = 0, 0.1, …, 0.95). If any measured rate exceeds p = 0.05, the Section 4 independence assumption is load-bearing and the in-model validation was blind to it; if rates stay below p, repeat the check with a non-flat delay profile (e.g., Gaussian Ft) to test assumption (ii). The definitive version of the same check is a controlled radar measurement of a target with a known delayed response.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The formal machinery is internally consistent: for fixed contrast q, equation (58) defines l- and l+ as quantiles of the very cdfs that, via (59), set r'_t and r'_s equal to p, and the extension (61) bounds any mixture of contrasts because the min/max threshold makes cdft(l-;q) ≤ p and 1 − cdfs(l+;q) ≤ p pointwise. The load-bearing premise is therefore not the algebra but the claim that the cdfs (55), estimated by Monte Carlo from the Gaussian white-noise models (23)–(29), are the true distributions of l for coordinate-delay SAR images. That claim rests on unvalidated assumptions: (i) delta-correlated, circular-Gaussian scattering with the delay process of (27); (ii) the known delay profile Ft = Fs = 1[0, zeta_max] of (34); and (iii) the Section 4 'simplified treatment' that ambiguity lines spaced pi in zeta are uncorrelated, which justifies the product likelihood (46b). The Monte Carlo validation in Figures 3–6 cannot detect a failure of any of these: Figure 2's caption generates each ambiguity line independently 'according to (31)', so the independence assumption is enforced by the simulator, not tested, and Gaussianity and the delay profile are likewise built into the generator. The p-bound is thus a self-consistency statement. Secondary gaps reinforce the conditionality: the min/max in (61) is evaluated only on the q-grid {0.0,…,0.9} with step 0.1 and validated at just two contrasts in Section 7, and Section 6.2 simply assumes cdf continuity and monotonicity. No measured SAR scene appears anywhere, and the abstract itself scopes the evidence to 'simple statistical models.' If real speckle, cross-line correlation, or the delay statistics differ from (23)–(29), the advertised error bound p need not hold on real images.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a statistical method for discriminating between instantaneous and delayed scatterers in coordinate-delay synthetic aperture radar (SAR) images. Building on the authors' prior work, the manuscript derives the point spread function and correlation structure of coordinate-delay images under circular Gaussian white-noise scattering models, formulates a maximum likelihood classifier based on the log-likelihood ratio l, and proposes to set classification thresholds l- and l+ as quantiles of simulated cumulative distribution functions. The authors claim that with these thresholds, the misclassification rates are kept at or below a preset level p for any probability distribution of target contrast when the thresholds are computed via the min/max procedure of Eq. (61). The paper includes Monte Carlo simulations showing the dependence of classification quality on parameters kappa and zeta_max.","tokens_in":17680,"tokens_out":10372,"duration_ms":98890,"significance":"The derivations in Sections 2-6 are internally coherent, and the paper provides a useful, explicit recipe for attaching confidence labels to SAR target classifications without knowing the target contrast. The point-spread function analysis and the closed-form correlation expressions are valuable and appear correct. The Monte Carlo experiments are described in sufficient detail to be reproduced, and the authors are explicit about the modeling assumptions. However, the headline error-rate guarantee is conditional on the stochastic scattering model (23)-(29) and the independence assumption of Section 4; these assumptions are not validated against measured SAR data or alternative generative models. As a result, the significance of the contribution rests on whether the model captures real speckle.","major_comments":[{"comment":"The error-rate guarantee is definitional. Equations (58) define l- and l+ as the p-quantile and (1-p)-quantile of the cdfs cdft and cdfs, and equations (59) then state that the misclassification rates r'_t and r'_s equal p. Consequently, the Monte Carlo results in Figures 5 and 6, which are produced from the same generators used to estimate those cdfs, demonstrate self-consistency rather than independent predictive performance. To support the advertised confidence levels, the method should be evaluated on data generated from a different statistical model (for example, with non-Gaussian clutter or correlated ambiguity lines) or on real coordinate-delay SAR images.","section":"Section 6.2, Eqs. (58)-(59)"},{"comment":"The key independence assumption is both built into the likelihood and enforced by the simulator. The text states that ambiguity lines spaced pi in zeta_d can be considered uncorrelated, and the product likelihood (46b) relies on this. The Figure 2 caption describes generating each ambiguity line independently, so the simulations cannot detect a failure of this assumption. The authors should provide a quantitative justification of the pi-spacing based on the correlation functions in Eqs. (31)-(32), or test sensitivity to correlated ambiguity lines.","section":"Section 4 and Eq. (46b)"},{"comment":"The claim that error rates are bounded for any probability distribution of target contrasts is not fully supported. The minimum and maximum in Eq. (61) are evaluated on the finite grid q in {0.0, 0.1, ..., 0.9} with step 0.1, and no monotonicity of the quantiles with respect to q is demonstrated. If the worst-case contrast lies above 0.9, the computed thresholds will not satisfy cdft(l-;q) <= p and 1 - cdfs(l+;q) <= p for all q. Additionally, Section 6.2 assumes the cdfs are continuous and monotonic; this is not verified, and if the cdfs have flat regions or jumps, the quantile equations (58) may not have unique solutions.","section":"Section 6.3, Eq. (61)"},{"comment":"No measured SAR data are used in the paper. All quantitative results are obtained from the synthetic ensemble generator based on the Gaussian white-noise models (23)-(29) and the known delay profile (34). As a consequence, the transfer of the p-guarantee to real radar imagery rests entirely on the unvalidated adequacy of these models. This limitation should be stated explicitly in the abstract and conclusion, and, ideally, a demonstration on a real or at least a strongly perturbed dataset should be included.","section":"Section 7"}],"minor_comments":[{"comment":"Footnote 3 uses the symbol 'xi_d' where 'zeta_d' is clearly intended; this should be corrected.","section":"Footnote 3"},{"comment":"Equation (40) writes the correlation as <I_mj,s-model I_mj',s-model> without a complex conjugate; to match Eq. (31), the second factor should be the conjugate, and the notation should be made consistent with the real-vector formalism in (42)-(44).","section":"Eq. (40)"},{"comment":"The number of Monte Carlo realizations used to estimate the cdfs is not reported, and no error bars or confidence bands are given for the estimated quantiles l- and l+. This information would help the reader assess the statistical precision of the thresholds.","section":"Figures 3-6"},{"comment":"The term 'monotonic' is used without qualification; since cdfs are nondecreasing by definition, the authors should state whether they assume strict monotonicity, and in the flat-case they should note that the solutions to (58) may be non-unique.","section":"Section 6.2"},{"comment":"The captions describe the panels as 'flipped vertically' and refer to 'lower half' and 'upper half', which makes the presentation of the confusion-matrix entries difficult to follow; a clearer labeling scheme would improve readability.","section":"Figures 5 and 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid extension of the authors' previous work [7], and the mathematical machinery is correct within its stated model. My main concern is the gap between the model-conditional guarantee and the wording in the abstract and Section 6.3, which suggests unconditional confidence levels. I recommend asking for a substantial revision that (i) clearly scopes the claims to the assumed statistical model, (ii) adds an independent or real-data validation, and (iii) tightens the contrast-grid and cdf-smoothness assumptions. If the limitations are honestly framed, the paper could be a good fit for a signal-processing or radar imaging venue even without real data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this paper: the confidence-level construction is sound and clearly written, and the advertised error-rate guarantee is real only inside the simulation model. If you take the paper as a methods piece for coordinate-delay SAR, it is a legitimate advance over the authors' prior ML detector; if you take the headline as a statement about real radar images, there is no evidence yet.\n\nWhat's new: the reject option in algorithm (52), the threshold selection via (58), and especially the contrast-independent version (61) that removes the need to know target contrast beforehand. That last bit is genuinely useful for automatic target recognition. The derivations in Sections 2-6 are internally coherent, the notation is heavy but the logic is traceable, and the figures honestly show that even moderate speckle makes classification hard. The paper does not hide its scaffolding: the abstract says 'simple statistical models', and the method is explicitly Monte-Carlo-based.\n\nWhere it is soft: the central guarantee is definitional. The thresholds are quantiles of the simulated cdfs, so r'_s and r'_t equal p by construction (eqs. 58-59). That is not a flaw in the algebra—it is exactly how confidence levels work—but it means the only load-bearing question is whether the Gaussian white-noise scattering model (23)-(29), the known delay profile (34), and the independence of ambiguity lines survive contact with real SAR speckle. The paper never tests that. Also, the contrast-independent min/max in (61) is computed over a finite q-grid (0.0 to 0.9, step 0.1), while the text says 'entire range'; that is a small overstatement. Section 6.2 assumes cdf continuity and monotonicity, which is standard but unverified. No code or data are shipped, which makes the Monte Carlo hard to reproduce.\n\nNone of this sinks the paper. The authors are honest about their scope, and within that scope the statistics are correct. But a revision should either add a real or measured SAR scene (or at least a convincing validation against measured speckle), or explicitly state that the p-bound applies to the model, not to any particular physical scene.\n\nRecommendation: send it out. A serious referee can push on the model-to-reality transfer and the grid/monotonicity caveats, but the core methodological contribution deserves attention in the SAR imaging community.","headline":"Solid methodological paper: the confidence-level construction is correct and clearly presented, but the headline error-rate guarantee is proven only for the authors' simulation model, not for real SAR images.","tokens_in":18224,"tokens_out":2461,"would_cite":false,"duration_ms":23736,"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":"The paper claims that a two-threshold maximum-likelihood classifier, with thresholds set from speckle-simulation cumulative distributions and made contrast-independent by worst-case extremization, can keep delayed-versus-instantaneous SAR…","keywords":["synthetic aperture radar","coordinate-delay imaging","scattering delay","range-delay ambiguity","speckle statistics","maximum likelihood classification","confidence levels","target discrimination"],"falsifier":"Run algorithm (52) with the contrast-independent thresholds from (61) on real coordinate-delay SAR images of a scene with independently calibrated instantaneous and delayed scatterers, and compare the decisions with ground truth; if the empirical misclassification rate exceeds $p$ for any target contrast, the bound does not transfer to practice.","tokens_in":17056,"feed_emoji":"📡","tokens_out":10065,"duration_ms":89101,"temperature":0.7,"pith_summary":"The paper tries to make the discrimination between instantaneous and delayed radar scatterers in coordinate-delay SAR images statistically reliable. It proposes a classifier with two thresholds on a log-likelihood score: above $l_+$ call the target delayed, below $l_-$ call it instantaneous, and in between declare uncertainty. The thresholds are quantiles of cumulative distribution functions built from simulated speckle ensembles; taking worst-case minima and maxima over target contrasts makes the thresholds contrast-independent. The payoff would be that a single SAR acquisition can carry a confidence label on the target type, which is currently missing because speckle makes deterministic discrimination unreliable.","feed_headline":"Two-threshold radar test caps errors without target contrast","feed_subtitle":"Speckle-model thresholds hold delayed-versus-instantaneous radar classification errors at or below a preset rate, with no contrast…","key_machinery":"The load-bearing object is the coordinate-delay point spread function $W(t_y-t_z,\\mathbf{y}-\\mathbf{z})$, whose slow decay along the ambiguity lines $T_0=\\mathrm{const}$ creates the range-delay ambiguity that makes instantaneous and delayed scatterers hard to separate. Around it, the paper places Gaussian white-noise models for the background, the delayed $t$-scatterer, and the instantaneous $s$-scatterer, from which the image second-order statistics (31)-(32) follow. Classification runs on the maximum-likelihood score $l=\\log\\hat{p}_t-\\log\\hat{p}_s$, and the confidence levels come from the cdf quantile equations (58), generalized to all contrasts by (61).","core_discovery":"The paper's central claim is that a maximum-likelihood score $l=\\log\\hat{p}_t-\\log\\hat{p}_s$, the difference of the fitted likelihoods for the delayed and instantaneous models, can be turned into a calibrated three-outcome classifier. For known contrast, choosing $l_-$ and $l_+$ from $\\mathrm{cdf}_t(l_-)=p$ and $\\mathrm{cdf}_s(l_+)=1-p$ makes the misclassification rates exactly $p$. For unknown contrast, replacing these with the contrast-independent extremes of (61) yields $r'_s\\le p$ and $r'_t\\le p$ for every contrast distribution, and Monte Carlo simulations show the dependence of these rates on $\\kappa$ and $\\zeta_{\\max}$.","pith_inferences":["Extension: the same cdf-quantile construction could be reused for any binary SAR classification task whose two hypotheses have known covariance structure, not only delay detection.","Extension: the contrast-independent envelope of (61) is conservative for a known contrast distribution; averaging the cdfs over that distribution would shrink the uncertainty band while sacrificing the worst-case guarantee.","Extension: a calibrated delay-line corner reflector in a real SAR scene would provide a direct test of whether the simulated cdfs match reality; until then the $p$ bound is a property of the model world.","Extension: feeding the full eight-dimensional maximum-likelihood solution into a learned classifier is mentioned in the paper but not systematically explored; that parameter space remains an open testable extension."],"forward_implications":["A single SAR acquisition can be assigned a confidence label for 'delayed response present', because the thresholds need only system parameters and a simulation ensemble, not repeated imaging.","Operators can preset the maximum tolerable misclassification probability $p$ and get thresholds that keep $r'_s$ and $r'_t$ at or below $p$ for any unknown contrast distribution.","The required amount of delay is quantified: separation improves as $\\zeta_{\\max}$ grows, and the condition $\\kappa\\zeta_{\\max}\\gtrsim 20$ indicates when the ambiguity becomes resolvable.","Wider uncertainty intervals lower the error rates but raise the fraction of uncertain outcomes, an explicit trade-off measured by $r''_s$ and $r''_t$.","Saturation of reliable classification for $\\kappa\\gtrsim 0.4$ suggests a regime where increasing aperture width no longer improves discrimination quality."],"supporting_citations":[{"why":"Supplies the earlier maximum-likelihood discrimination algorithm, the Monte Carlo ensemble simulation method, and the baseline results that this paper extends with confidence levels.","marker":"[7]"},{"why":"Introduces coordinate-delay reflectivity and the range-delay ambiguity lines that define the discrimination problem.","marker":"[8]"},{"why":"Provides the speckle and Gaussian random-field model of distributed SAR targets used for the statistical scattering models.","marker":"[6]"},{"why":"Supplies the maximum likelihood and confidence-interval/quantile formalism underlying equations (49), (54), and (58).","marker":"[18]"},{"why":"Gives the Fresnel-integral identities used to derive the PSF asymptotics that set the resolvability threshold (35).","marker":"[11]"},{"why":"Supports the physical speckle-interference picture behind the circular Gaussian statistics of the image.","marker":"[9]"}],"fun_headline_variants":["Radar delay test bounds errors without contrast assumptions","Three-outcome SAR classifier with guaranteed error caps","ML radar test sets confidence levels for scatterer delay","Contrast-free radar discrimination with certified error rates","SAR imaging: calibrated test for delayed vs instant scatterers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The error bound $p$ transfers to real radar images only if the Gaussian white-noise scattering models reproduce actual speckle statistics and the assumed independence between ambiguity lines holds; the paper never tests the thresholds against measured SAR data.","fun_headline_variants_meta":{"raw":{"variants":["Radar delay test bounds errors without contrast assumptions","Three-outcome SAR classifier with guaranteed error caps","ML radar test sets confidence levels for scatterer delay","Contrast-free radar discrimination with certified error rates","SAR imaging: calibrated test for delayed vs instant scatterers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000647,"raw_usage":{"total_tokens":2933,"prompt_tokens":870,"completion_tokens":2063,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1989}},"tokens_in":486,"tokens_out":2063,"duration_ms":15255,"temperature":1.0,"reasoning_tokens":1989,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:48:50.126960+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run algorithm (52) with the contrast-independent thresholds from (61) on real coordinate-delay SAR images of a scene with independently calibrated instantaneous and delayed scatterers, and compare the decisions with ground truth; if the empirical misclassification rate exceeds $p$ for any target contrast, the bound does not transfer to practice.","supporting_citations":[{"cited_title":"Detection of delayed target re- sponse in SAR","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier maximum-likelihood discrimination algorithm, the Monte Carlo ensemble simulation method, and the baseline results that this paper extends with confidence levels."},{"cited_title":"Hyperspectral SAR","cited_arxiv_id":null,"evidence_quote":"Introduces coordinate-delay reflectivity and the range-delay ambiguity lines that define the discrimination problem."},{"cited_title":"Understanding Synthetic Aperture Radar Images","cited_arxiv_id":null,"evidence_quote":"Provides the speckle and Gaussian random-field model of distributed SAR targets used for the statistical scattering models."},{"cited_title":"Scheaffer","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum likelihood and confidence-interval/quantile formalism underlying equations (49), (54), and (58)."},{"cited_title":"http://dlmf.nist.gov/, Release 1.0.23 of 2019-06-15","cited_arxiv_id":null,"evidence_quote":"Gives the Fresnel-integral identities used to derive the PSF asymptotics that set the resolvability threshold (35)."},{"cited_title":"Statistical properties of laser speckle patterns","cited_arxiv_id":null,"evidence_quote":"Supports the physical speckle-interference picture behind the circular Gaussian statistics of the image."}],"review_version":1}