{"id":"0cefb74d-891a-4bdd-998d-0208020fbeef","arxiv_id":"1908.09431","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper proposes two selective detectors and one tunable detector for multichannel signals in subspace interference with signal mismatch, and gives analytical detection and false-alarm probabilities verified by simulation.","lead":"This paper designs new detectors that decide whether a weak signal is present when interfering signals and noise obscure it, and when the signal is not exactly what was expected. The detectors offer a tuning knob that lets an operator choose how strictly to reject imperfectly matched signals, with formulas for detection and false-alarm probabilities verified by simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (37) is not the ρ=0 limit of Eq. (34): it omits all but the k=0 term, so the printed PFA expressions are incorrect; the PD integrals also drop β-regions where the conditional PD equals 1.","rationale":"The central claim is that Eqs. (22)-(27) give the PD/PFA of ABORT-I, W-ABORT-I, and T-W-ABORT-I. The derivation routes everything through Eq. (34), an imported finite-sum CDF. Eq. (37) is presented as the H0 specialization of Eq. (34), but it is not: all IG_{k+1}=1 terms must be summed. This makes every printed PFA expression incorrect unless Eq. (37) is only a typographical stand-in for the sum. Since Section 5 states that thresholds are generated by 10^5 Monte Carlo trials, the TH-vs-MC PD agreement in the figures does not exercise the PFA formulas, so the paper gives no numerical support for Eq. (37). Independently, the averaging domains in Eqs. (22), (23), (25), and (26) are too small: wherever the CDF argument would be negative, the conditional probability is exactly 1, not 0 or undefined, so those β-regions must be included with integrand 1 (equivalently, write PD=1-∫P1 f1 over the admissible set). For selective settings such as κ=2.5, the omitted β>η_t^{1/(κ-1)} mass can be comparable to the target PFA. Both defects are concrete and checkable, and both lie in the load-bearing analytical formulas, so the conditional verdict is justified. The reader's identified assumption about deterministic mismatch parameters is a scope limitation rather than the most immediate correctness risk; the printed equations can be tested without leaving the paper's own Gaussian model. The Monte Carlo results strongly suggest the underlying statistical machinery and the tunable-detector concept are sound, so a revision rather than a rejection is the appropriate outcome.","tokens_in":12160,"tokens_out":34302,"duration_ms":330259,"concrete_test":"Recompute P0(η) from Eq. (34) with ρ_eff=0 for the paper's N=12, L=24, p=1, q=2 and compare with Eq. (37). For η=1, Eq. (37) gives 15·1/2^15≈4.6e-4, whereas the full sum gives 1-2^{-15}=0.99997. Then insert both forms into the PFA integral for T-W-ABORT-I with κ=2.5 at PFA=10^-3 and compare the resulting thresholds with the Monte Carlo threshold; the single-term version will be off by orders of magnitude. Separately, evaluate Eq. (26) at η_t=0.8 and compare with the Monte Carlo PD/PFA: the printed formula misses the mass Pr(β>0.8^{2/3})≈Pr(β>0.86), where the conditional PD is 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 builds every PFA and PD formula on P1(η) in Eq. (34), the finite-sum noncentral F CDF. Setting ρ_eff=0 in Eq. (34) gives P0(η)=Σ_{k=0}^{L-N+q} C_{k+p}^{L-N+p+q} η^{k+p}/(1+η)^{L-N+p+q}, because IG_{k+1}(0)=1 for every k. The printed Eq. (37) keeps only the k=0 term. Consequently, the PFA integrals obtained by replacing P1 with P0 in Eqs. (22)-(27) are not the false-alarm probabilities of the proposed detectors; any threshold computed from Eq. (37) will not achieve the designed PFA. In addition, Eqs. (22), (23), (25), and (26) restrict the β-integration to the domain where the CDF argument is non-negative and discard the complementary domain, but on that discarded domain the conditional PD (or PFA) is identically 1. For example, for κ>1 and η_t≤1 (case iii), any β>η_t^{1/(κ-1)} guarantees β^{κ-1}(1+t_GLRT)>η_t, so the correct PD is 1-∫_0^{η_t^{1/(κ-1)}} P1(η_t β^{1-κ}-1) f1(β)dβ, not Eq. (26). Unless the authors supply a corrected derivation, the claimed analytical expressions are not the expressions actually verified in the plots.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers adaptive detection of a multichannel signal in subspace interference and Gaussian noise with unknown covariance, when the actual signal may be mismatched with respect to the nominal signal subspace. It proposes two selective detectors, ABORT-I and W-ABORT-I, and a tunable detector T-W-ABORT-I whose tunable parameter kappa is intended to interpolate continuously between robustness and selectivity. The statistical analysis expresses the detectors as functions of the GLRT-I statistic and a loss factor beta, and Section 4 derives analytical PD and PFA expressions from the conditional distributions of these quantities. The paper claims that these expressions are verified by Monte Carlo simulations and that the tunable detector can smoothly trade off rejection of mismatched signals against detection of matched signals.","tokens_in":12469,"tokens_out":13321,"duration_ms":129716,"significance":"The detector construction is a natural and potentially useful extension of ABORT/W-ABORT to the subspace-interference setting, and the idea of a single tunable parameter that continuously trades selectivity for robustness is attractive for practical operators. The conditional-distribution route is standard, and the Monte Carlo results in the tested regime are broadly consistent with the underlying distributional model. However, two load-bearing defects in the printed analytical expressions mean that the paper's central claim of verified PD/PFA formulas is not currently established: the printed H0 CDF is not the rho_eff=0 limit of the stated noncentral CDF, and several PD/PFA integrals discard regions where the conditional probability is exactly one. These errors are correctable but require re-derivation and re-verification.","major_comments":[{"comment":"The stated P0(eta) is not the rho_eff=0 limit of Eq. (34). Since IG_{k+1}(0)=1 for every k, setting rho_eff=0 in Eq. (34) gives P0(eta)= sum_{k=0}^{L-N+q} C_{k+p}^{L-N+p+q} eta^{k+p}/(1+eta)^{L-N+p+q}; the printed expression keeps only the k=0 term. The printed P0 tends to 0 as eta tends to infinity rather than to 1, so it cannot be a valid CDF. Because the PFA formulas are obtained by substituting P0 into Eqs. (22)-(27), all printed PFA expressions are incorrect, and thresholds computed from Eq. (37) would not achieve the designed false-alarm probability.","section":"Section 4, Eq. (37)"},{"comment":"The integrals restrict beta to the region where the CDF argument is non-negative and omit the complementary region, but on that complementary region the conditional detection probability is exactly 1. For example, for ABORT-I the correct expression is PD_ABORT-I = integral_0^{min(1,eta_a)} [1-P1(eta_a-beta)] f1(beta) dbeta + integral_{min(1,eta_a)}^1 f1(beta) dbeta, because t_GLRT-I is non-negative. The same omission affects W-ABORT-I and T-W-ABORT-I cases ii and iii; in particular, Eq. (26) should be 1 - integral_0^{eta_t^{1/(kappa-1)}} P1(eta_t beta^{1-kappa}-1) f1(beta) dbeta, not the two-part expression that drops the region where beta>eta_t^{1/(kappa-1)}. The PFA integrals formed by replacing P1 with P0 have the same defect. Unless the thresholds in the plotted examples always fall in cases where the omitted mass is negligible, the displayed theoretical curves do not correspond to the printed formulas.","section":"Section 4, Eqs. (22), (23), (25), (26)"},{"comment":"The numerical section verifies the PD curves against Monte Carlo detection probabilities, but it does not verify the analytical PFA expressions; the thresholds themselves are generated by 10^5 Monte Carlo trials. Given the error in Eq. (37), the claim that the analytical PFA expressions are verified is unsupported. The authors should either display simulated-versus-theoretical PFA curves for the proposed detectors or explicitly state that the analytical PFA is not used and is not verified by the figures shown.","section":"Section 5, numerical verification"}],"minor_comments":[{"comment":"The sentence containing 'Equation (10) as be recast as' should read 'Equation (10) can be recast as'.","section":"Section 3, text before Eq. (11)"},{"comment":"The condition written as 'beta^{1-kappa} > eta_t^{-1}' is hard to parse; it would be clearer to state the equivalent inequalities for each case, such as beta < eta_t^{1/(kappa-1)} for kappa>1 and eta_t<=1.","section":"Eq. (21) and surrounding text"},{"comment":"The phrase 'a special case of the T-W-ABORT-I with <= kappa = 1.0' contains a stray '<=' symbol and should be corrected.","section":"Section 5, Fig. 3 discussion"},{"comment":"The paper does not report the actual threshold values used for the PD curves in Figs. 1-4; reporting these values would allow readers to identify which case of Eqs. (24)-(27) is active and to check whether the omitted-region issue affects the displayed results.","section":"Figures and thresholds"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope, and the detector concept is worth publishing if the probability analysis is corrected. The errors are serious but appear fixable: the H0 CDF must be the full sum, the PD/PFA integrals must include the regions where the conditional probability equals one, and the numerical section should verify the PFA expressions directly. I would encourage the editor to request a revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zac,\n\nThe detector constructions are the useful part of this paper. ABORT-I, W-ABORT-I, and the tunable T-W-ABORT-I are natural extensions of known selective detectors to the subspace-interference setting, and the reduction at kappa = 0, 1, 2 to AED, GLRT-I, and W-ABORT-I is clean. The tunable parameter does what the authors claim: the Monte Carlo curves show a smooth trade-off between robustness and selectivity, which is practically valuable for radar/sonar operators who face sidelobe targets or jamming. The analysis strategy is standard, and the paper is honest about importing distributional results from prior work rather than rederiving them.\n\nThe problem is that the analytical expressions are not right as printed. Equation (37) is not the rho_eff = 0 limit of (34). Setting rho_eff = 0 makes every incomplete gamma term equal to 1, so the central CDF is a sum over all k, not just the k = 0 term. The printed P0(eta) is a single binomial term. Since all PFA expressions are built from this P0, any threshold computed from them will not achieve the designed false-alarm rate. Separately, the PD integrals in Eqs. (22), (23), (25), and (26) restrict the beta-integration to the region where the CDF argument is non-negative, but on the complementary region the conditional PD is identically 1. That missing probability mass is not negligible when the threshold is below 1, which can easily happen at low PFA in this setting. The Monte Carlo \"verification\" in Section 5 appears to use empirically generated thresholds (100,000 simulations per threshold), so the agreement in the figures does not actually validate the analytic PFA/PD formulas.\n\nThese are fixable errors rather than a broken idea. The detector definitions, the statistical model, and the general approach are sound; the derivations just need to be redone carefully. The citation pattern is fine, and the deterministic mismatch model is a reasonable engineering assumption. No code or data is included, which is a minor inconvenience but not a flaw for this kind of paper.\n\nWho is this for? Engineers who want a tunable detector with a monotone selectivity knob. The novelty is moderate, but the practical value is real. I would not cite the analytic expressions in their current form, and the authors should be asked to correct the derivations. If they do, the paper has a legitimate place in the literature. It deserves a serious referee, but the review should request actual correction of the PFA/PD formulas rather than cosmetic revision.","headline":"The tunable detector idea is genuinely useful, but the printed analytical PFA and PD formulas are not correct, so the paper needs major revision before the performance analysis can be trusted.","tokens_in":13042,"tokens_out":4921,"would_cite":false,"duration_ms":45835,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H15","62F03","62E15","94A13"],"pacs":[],"model":"deepseek-v4-flash","headline":"A tunable detector smoothly trades robustness against selectivity for mismatched signals in subspace interference.","keywords":["adaptive detection","constant false alarm rate","multichannel signal","signal mismatch","subspace interference","tunable detector","ABORT","selective detection"],"falsifier":"Run the same Monte Carlo setup but draw the actual signal vector randomly around the nominal one on every trial, keeping the SNR fixed; the empirical detection probability will depart from Eqs. (22)-(27), because those integrals average only over a fixed mismatch realization rather than over a distribution of mismatch.","tokens_in":11958,"feed_emoji":"📡","tokens_out":10453,"duration_ms":86771,"temperature":0.7,"pith_summary":"This paper addresses the practical situation where a radar or communication receiver must detect a signal that is known to lie in a subspace but is imperfectly aligned with it, while strong interference occupies a separate known subspace. The authors propose two selective detectors, ABORT-I and W-ABORT-I, that reject mismatched signals aggressively, and a tunable detector, T-W-ABORT-I, whose non-negative parameter $\\kappa$ continuously adjusts this behavior from robust detection to strict selectivity. They derive closed-form expressions for the detection and false-alarm probabilities of all three detectors and verify them by Monte Carlo simulation. If correct, this gives operators a principled way to choose a detector for the mismatch conditions they expect, rather than forcing a choice between two extremes.","feed_headline":"One knob tunes detection from selective to tolerant","feed_subtitle":"A single knob shifts a detector from rejecting mismatched signals to finding them, with exact false-alarm formulas","key_machinery":"The load-bearing object is the loss factor $\\beta$ of Eq. (13), a scalar in $(0,1)$ equal to the reciprocal of the remaining energy after projecting the quasi-whitened data away from the interference and signal-plus-interference subspaces. Together with the GLRT-I statistic it satisfies $t_{\\mathrm{ABORT\\text{-}I}} = t_{\\mathrm{GLRT\\text{-}I}}+\\beta$, $t_{\\mathrm{W\\text{-}ABORT\\text{-}I}} = (1+t_{\\mathrm{GLRT\\text{-}I}})\\beta$, and $t_{\\mathrm{T\\text{-}W\\text{-}ABORT\\text{-}I}} = \\beta^{\\kappa-1}(1+t_{\\mathrm{GLRT\\text{-}I}})$. These identities let the paper substitute the known conditional distribution of $t_{\\mathrm{GLRT\\text{-}I}}$ and the density of $\\beta$ into the integrals in Eqs. (22)-(27), producing the closed-form PD and PFA expressions. Interference is removed by projecting the whitened test data onto the orthogonal complement of the interference subspace, so the interference power itself never appears in the probabilities.","core_discovery":"On the paper's own terms, the central discovery is that the tunable statistic defined in Eq. (9), $t_{\\mathrm{T\\text{-}W\\text{-}ABORT\\text{-}I}} = (1+\\tilde{x}^H P_{\\tilde J}^\\perp \\tilde{x})/(1+\\tilde{x}^H P_{\\tilde J}^\\perp \\tilde{x} - \\tilde{x}^H P_{P_{\\tilde J}^\\perp \\tilde H}\\tilde{x})^{\\kappa}$, interpolates between the adaptive energy detector at $\\kappa=0$, the GLRT-I at $\\kappa=1$, and the W-ABORT-I at $\\kappa=2$. For $\\kappa<1$ it is more robust to mismatched signals than the existing GLRT-I and 2S-GLRT-I, while for $\\kappa>2$ it rejects mismatched signals more strongly than the two selective detectors. With the loss factor $\\beta = (1+\\tilde{x}^H P_{\\tilde J}^\\perp \\tilde{x} - \\tilde{x}^H P_{P_{\\tilde J}^\\perp \\tilde H}\\tilde{x})^{-1}$, each proposed detector becomes a simple function of $\\beta$ and the GLRT-I statistic, so the detection and false-alarm probabilities reduce to one-dimensional integrals over $\\beta$ using known distributions of $t_{\\mathrm{GLRT\\text{-}I}}$ and $\\beta$ under mismatch. The resulting analytical PD and PFA formulas, Eqs. (22)-(27), are confirmed by Monte Carlo simulations for the parameters shown.","pith_inferences":["Because the trade-off is governed by a single power $\\kappa$, the same construction could be applied to other ratio-type detectors to generate a full family of selective-to-robust detectors.","If actual mismatch is random, as with a distribution of pointing errors, the fixed-mismatch distributions of $\\beta$ and $t_{\\mathrm{GLRT\\text{-}I}}$ would need to be averaged over that distribution; the paper's formulas cover each fixed realization, not the random-mismatch average.","A natural extension is to estimate the mismatch metric $\\cos^2\\vartheta$ online and choose $\\kappa$ adaptively, though the paper proposes no such feedback rule.","The verification is limited to exponentially correlated Gaussian noise with the displayed parameters; whether the integrals track simulations for non-Gaussian or heterogeneous training data remains untested."],"forward_implications":["With a small $\\kappa$, the T-W-ABORT-I preserves detection of targets whose steering vector is uncertain, going beyond the robustness of 2S-GLRT-I.","With a large $\\kappa$, for example 2.5, it is more selective than ABORT-I and W-ABORT-I, so it can reject sidelobe targets and jamming-like signals even at high SNR.","Because closed-form false-alarm probabilities are available, thresholds can be set analytically to meet a false-alarm constraint without Monte Carlo threshold searches.","For matched signals, ABORT-I and T-W-ABORT-I with $0.6 \\le \\kappa \\le 1.0$ nearly match the GLRT-I detection probability, so the added robustness costs little in the matched case.","These detectors are the first in this setting specifically designed for signal mismatch, extending the ABORT/W-ABORT selectivity mechanism to subspace interference."],"supporting_citations":[{"why":"Supplies the subspace-interference model and the GLRT-I and 2S-GLRT-I baseline detectors that the proposed detectors extend and are compared with.","marker":"[8]"},{"why":"Provides the distributional results for $t_{\\mathrm{GLRT\\text{-}I}}$ and $\\beta$ under fixed signal mismatch, including the parameters $\\rho_{\\mathrm{eff}}$ and $\\delta^2$ used in the PD/PFA formulas.","marker":"[12]"},{"why":"Introduces the ABORT orthogonal-rejection test whose form is adapted to the interference case as ABORT-I.","marker":"[13]"},{"why":"Supplies the whitened ABORT statistic whose form is adapted to W-ABORT-I.","marker":"[15]"},{"why":"Shows that the $\\kappa=0$ limit of the tunable detector is the adaptive energy detector.","marker":"[16]"},{"why":"Gives the CDF and PDF identities used to evaluate $P_1(\\eta)$ and $f_1(\\beta)$ in the final integrals.","marker":"[17]"}],"fun_headline_variants":["Single parameter tunes detection from robust to selective","Adjustable detector interpolates between robust and selective","Tunable detector gives exact probabilities for mismatch cases","One knob shifts detection from mismatch-rejecting to mismatch-tolerant","A scaling factor interpolates detector between robust and selective"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed formulas assume the mismatch is a fixed signal vector completely captured by the two scalars $\\rho_{\\mathrm{eff}}$ and $\\delta^2$, and that the test data and training data share exactly the same Gaussian covariance matrix; if either assumption fails, the derived distributions and probabilities do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Single parameter tunes detection from robust to selective","Adjustable detector interpolates between robust and selective","Tunable detector gives exact probabilities for mismatch cases","One knob shifts detection from mismatch-rejecting to mismatch-tolerant","A scaling factor interpolates detector between robust and selective"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001006,"raw_usage":{"total_tokens":4290,"prompt_tokens":1017,"completion_tokens":3273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":3198}},"tokens_in":633,"tokens_out":3273,"duration_ms":26141,"temperature":1.0,"reasoning_tokens":3198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:11:52.348894+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Monte Carlo setup but draw the actual signal vector randomly around the nominal one on every trial, keeping the SNR fixed; the empirical detection probability will depart from Eqs. (22)-(27), because those integrals average only over a fixed mismatch realization rather than over a distribution of mismatch.","supporting_citations":[{"cited_title":"Adapti ve radar de- tection of distributed targets in homogeneous and partiall y homogeneous 19 noise plus subspace interference,","cited_arxiv_id":null,"evidence_quote":"Supplies the subspace-interference model and the GLRT-I and 2S-GLRT-I baseline detectors that the proposed detectors extend and are compared with."},{"cited_title":"Performan ce analysis of adaptive detectors for point targets in subspace interfere nce and Gaussian noise,","cited_arxiv_id":null,"evidence_quote":"Provides the distributional results for $t_{\\mathrm{GLRT\\text{-}I}}$ and $\\beta$ under fixed signal mismatch, including the parameters $\\rho_{\\mathrm{eff}}$ and $\\delta^2$ used in the PD/PFA formulas."},{"cited_title":"Adaptive beamformer orth ogonal rejection test,","cited_arxiv_id":null,"evidence_quote":"Introduces the ABORT orthogonal-rejection test whose form is adapted to the interference case as ABORT-I."},{"cited_title":"An ABORT-like det ector with im- proved mismatched signals rejection capabilities,","cited_arxiv_id":null,"evidence_quote":"Supplies the whitened ABORT statistic whose form is adapted to W-ABORT-I."},{"cited_title":"CFAR det ection in clutter with unknown correlation properties,","cited_arxiv_id":null,"evidence_quote":"Shows that the $\\kappa=0$ limit of the tunable detector is the adaptive energy detector."},{"cited_title":"Adaptive detection and p arameter estima- tion for multidimensional signal models,","cited_arxiv_id":null,"evidence_quote":"Gives the CDF and PDF identities used to evaluate $P_1(\\eta)$ and $f_1(\\beta)$ in the final integrals."}],"review_version":1}