{"id":"da881a7d-25d0-4049-bc30-90218c3c190a","arxiv_id":"2501.04973","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A factorial linear dynamical system with a sticky Indian buffet process prior is paired with a finite moving average stopping time to detect transient signals in multi-source backgrounds.","lead":"This paper builds a Bayesian nonparametric model of an unknown number of background signal sources and combines it with a finite moving average detector to find short radar-style pulses buried in interference. The method is aimed at electromagnetic environments where several emitters are active at once, a case that simpler single-source models handle poorly.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FKFF recursion assumes a block-diagonal posterior covariance that exact Gaussian filtering does not preserve, so the likelihood and the Gaussian performance bounds in Eqs. (28)-(30) and (43) are approximate rather than exact.","rationale":"The reader's weakest assumption correctly identifies the load-bearing issue. The paper claims to derive an exact recursive likelihood computation and exact Gaussian performance bounds for the FMA stopping time, but the block-diagonal covariance ansatz in Eq. (27) is not invariant under exact Gaussian conditioning for a factorial observation model. This makes the likelihood, LLR, and the theoretical bounds in Eqs. (35)-(43) approximations without a stated error bound. The additional typo in the false-alarm formula (Eq. 43, using μ_H1 where μ_H0 is required) and the absence of the supplement and code reinforce that the theoretical performance prediction is not established as stated. The empirical comparisons are suggestive and the proposed modeling direction has merit, but they do not rescue the central exactness claim. I therefore agree with the reader's REJECT verdict and would not change it. The proposed exact-filter comparison would settle whether the approximation is numerically negligible in the tested regime; absent such evidence, rejection remains appropriate.","tokens_in":25327,"tokens_out":5550,"duration_ms":55047,"concrete_test":"Implement exact Kalman filtering with full joint covariance over the M=2 FLDS states using the Section V-A parameters (Eqs. 44-45) and compare with the FKFF recursion on the same T=2000-sample CBS. Check (i) the maximum norm of the off-diagonal posterior covariance blocks produced by the exact filter relative to the diagonal blocks, and (ii) the per-step predictive log-likelihood log P∞(p_t|p_{1:t-1}) from Eq. (30) versus the exact marginal predictive likelihood. If the off-diagonal blocks are non-negligible (e.g., >1% of diagonal norm) or the cumulative log-likelihood difference is material (e.g., >0.1 nats/step), the FKFF is not exact and the Eq. (43) bounds do not apply to the true FLDS model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A parametrizes the forward distribution with ˆΣ_t = diag(...) (Eq. 27). For the FLDS observation model p_t = Σ_m C^m x^m_t + noise, conditioning on p_t couples the sources: the exact Kalman update produces nonzero off-diagonal posterior blocks Σ_t^{mn} = -P^{mm}(C^m)^T S^{-1} C^n P^{nn}, so the block-diagonal form is not preserved even when the prior is block diagonal. Consequently Eqs. (28)-(30) do not deliver the exact predictive likelihood, and the LLR in Eqs. (32)-(34) is computed for an approximate FKFF, not for the true FLDS model. The performance bounds in Theorem 2, Corollary 1, and Eq. (43) therefore bound the approximate detection statistic, not the actual FLDS-FMA procedure; no error analysis quantifies the gap. The paper does not label FKFF as an approximation, and the deferred supplement is not available for verification. This is load-bearing because the paper's central theoretical claim is exact closed-form false-alarm/missed-detection bounds. In addition, the PF_A formula in Eq. (43) uses μ_H1 inside the Φ term where the false-alarm bound requires μ_H0, an internal inconsistency that independently blocks use of the stated formulas.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an infinite factorial linear dynamical system (IFLDS) with a sticky Indian buffet process prior to model an unknown number of background sources, a parameter-learning method based on slice sampling and particle Gibbs with ancestor sampling, and a finite moving average (FMA) stopping time for transient signal detection. The detection statistic uses a factorial Kalman forward filtering (FKFF) recursion to compute the null likelihood and a reformulated log-likelihood ratio. The paper claims closed-form Gaussian bounds on the worst-case false-alarm and missed-detection probabilities in Eq. (43), and validates the method on simulated data and on a pulse-detection experiment under communication interference.","tokens_in":25625,"tokens_out":10014,"duration_ms":94832,"significance":"The problem of detecting a finite-duration signal in a multi-source, time-correlated background with an unknown number of sources is well motivated, and the proposed pipeline is coherent at a high level: a Bayesian nonparametric prior over LDS sources, a PGAS/slice-sampling inference scheme, and an online FMA detector. The paper's recursive formulation of the detection statistic is a useful practical feature, and the experimental section covers both synthetic and communication-interference scenarios. However, the central advertised result is a closed-form performance guarantee for the FMA test, and that result is not supported as written: the FKFF likelihood is not exact under the stated model, the false-alarm bound relies on an unjustified independence assumption, and Eq. (43) contains an internal inconsistency. Because these issues affect the main theoretical claims, the contribution cannot be accepted in its current form.","major_comments":[{"comment":"The block-diagonal covariance parametrization is not preserved by exact Gaussian filtering for the FLDS observation model. Conditioning the joint state (x_t^1,...,x_t^M) on p_t couples the sources through the v-structure; with a block-diagonal prior, the exact posterior covariance has off-diagonal blocks of the form -P^{mm}(C^m)^T S^{-1} C^n P^{nn}, where S = sum_m C^m P^{mm}(C^m)^T + M R. Equations (28)-(30) therefore compute an approximate predictive likelihood, not the exact P_∞(p_t | p_1,...,p_{t-1}) used in the LLR (34). Lemma 1 establishes only Gaussianity, not the factorized covariance form, and the paper does not label FKFF as an approximation. Since Theorem 2, Corollary 1, and Eq. (43) are derived from this likelihood, the stated closed-form performance bounds are unsupported without an error analysis of the approximation.","section":"IV-A, Eq. (27)"},{"comment":"The observation noise covariance used in the FKFF gain and predictive variance is inconsistent with model (2). In (2), n_t = sum_{m=1}^M (C^m x_t^m + v) with v ~ N(0,R); under the natural reading that each source contributes an independent noise term, the total observation noise covariance is M R, while Eq. (29) places M^2 R in the denominator and Eq. (30) adds M R inside the sum over m, again yielding M^2 R. If instead the same v is reused for all M sources, the covariance is M^2 R and this should be stated explicitly. Because the gain and predictive covariance determine the LLR numerically, this ambiguity changes the detection statistic and all subsequent bounds. The authors should specify the intended observation model and correct Eqs. (29)-(30) accordingly.","section":"IV-A, Eqs. (29)-(30)"},{"comment":"The false-alarm bound α(h,w_α) = 1 - (P_∞(Ŵ_w < h))^{w_α} treats the overlapping window sums Ŵ_t as independent. The event in (5) is the existence of t in an interval of length w_α with Ŵ_t ≥ h, and consecutive Ŵ_t share w-1 observations, so they are strongly dependent. Without a mixing or independence argument, the product form is not a valid upper bound; a union bound would instead require summing marginal tail probabilities over t. Because Corollary 1 and the first line of Eq. (43) use this formula to set the threshold h, the false-alarm guarantee P_FA ≤ ᾱ is not established.","section":"IV-C, Eq. (36)"},{"comment":"The false-alarm expression uses μ_H1 in the Gaussian CDF, but under H0 the accumulated statistic Ŵ_t has mean w μ_H0, not w μ_H1. Thus the displayed formula is not a false-alarm bound. If the intended correction is to use μ_H0 and σ²_H0, the threshold h produced by Eq. (43) would change, and the numerical comparisons in Section V would need to be recomputed. This internal inconsistency independently prevents the stated formulas from being used for detector calibration.","section":"IV-C, Eq. (43), third line"},{"comment":"The convergence of e_t in (42) is asserted only on the basis of Fig. 4, which uses the same parameter settings as the evaluation in Section V-A, and the paper explicitly states that a theoretical investigation of the convergence is beyond its scope. The closed-form missed-detection bound therefore depends on a numerically calibrated steady-state value of e_t and μ_H1 rather than on a derived property of the model. In addition, the real-data experiment in Section VI sets h empirically to maintain a 15% false-alarm rate, so the theoretically predicted threshold from (39)-(43) is not actually validated. The dependence of the claimed bounds on simulation-calibrated quantities should be stated clearly, and the bounds should not be presented as closed-form model-based guarantees.","section":"IV-C, Lemma 2 and Eq. (43)"},{"comment":"Several load-bearing proofs are deferred to an unavailable supplement: the derivation of the FKFF update (28), the proof of Theorem 1, the proof of Theorem 2, and the proof of Lemma 2 (Supplement Sections I-IV). The manuscript is not self-contained, and the central claims cannot be verified from the submitted text. In particular, Theorem 1 relies on Assumption 2, which states that the latent variable at arrival time ν is independent of the latent variable at ν-1 but depends on previous observations; this is not a consequence of the FLDS transition (2) or of Assumption 1, and no derivation is given in the main text. The omitted proofs and the unjustified assumption should be addressed before the theoretical claims can be evaluated.","section":"IV-B, IV-C; Supplemental Material"}],"minor_comments":[{"comment":"The update for μ̂_t^m uses x_{t-1}^m inside the innovation term where μ̂_{t-1}^m is evidently intended; as written, the recursion is not closed and cannot be implemented without additional definitions.","section":"IV-A, Eq. (28)"},{"comment":"The same matrix C^m is given for all m, so the simulated background sources differ only in their transition matrices G^m; the text says the output matrices are defined for each source, so the formula should be reconciled with the statement.","section":"V-A, Eq. (44)"},{"comment":"The x-axis is labeled SINR, but the conversion from the constant signal amplitude y_t (between 0.001 and 1.5) to SINR in dB is not defined; without this mapping the experimental curves cannot be reproduced.","section":"V-C, Fig. 6"},{"comment":"The text refers to 'MSE' while the defined metric is reconstruction error RE; please use consistent notation throughout the figure and the discussion.","section":"V-B, Fig. 5"},{"comment":"The threshold h is said to be adjusted to maintain a 15% false-alarm probability, but the corresponding theoretical threshold from Section IV-C is not reported; reporting both would clarify the discrepancy between the empirical calibration and the theoretical prediction.","section":"VI, Fig. 9"},{"comment":"Several figure captions and equations contain garbled unicode strings (e.g., the caption lines beginning '/uni00000013...' near Figs. 4-9), indicating a rendering or encoding problem; the final version should be regenerated with correct math and text.","section":"Figures 4-9"}],"recommendation":"reject","confidential_remarks":"The core theoretical contribution is the closed-form performance bound for the FMA detector, and that contribution is not valid as submitted: the FKFF likelihood is approximate but presented as exact, the false-alarm bound assumes independence of overlapping windows, and Eq. (43) contains a false-alarm formula that uses the wrong mean. Some of these issues could in principle be patched, but only by redefining the contribution as an approximate method with numerical calibration and by providing the missing supplement. As submitted, the central advertised guarantees do not hold, so I cannot recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper works on a real problem—transient detection in a background of multiple unknown LDS sources—and the IFLDS model plus the FMA stopping time form a plausible pipeline. But the central theoretical contribution, the closed-form performance bounds in Eq. (43), is not established. The FKFF recursion assumes a block-diagonal posterior covariance that exact Gaussian filtering would not preserve; the paper never labels this as an approximation or bounds the error. The false-alarm bound in Eq. (36) treats overlapping window statistics as independent without justification, and Eq. (43) appears to use μ_H1 in the false-alarm expression where μ_H0 is needed. These are not peripheral typos; they sit on the load-bearing claim that the detector's false-alarm and missed-detection probabilities can be computed in closed form.\n\nWhat's genuinely new: the sticky-MIBP prior for the factorial LDS, the slice/PGAS inference for automatically determining the number of sources, and the particular way the FMA statistic is computed using the FKFF predictive likelihood. The simulations are reasonably thorough—they show clear gains over single-source LDS, Gaussian, CUSUM, and Shewhart baselines, and the radar pulse-detection experiment gives practical credibility. The FKFF itself, even as an approximation, is a useful algorithmic contribution.\n\nThe soft spots are proportional. The biggest one is the gap between the exact-sounding theory and the approximate recursion, and I don't see an error analysis anywhere. The dependence of Lemma 2 on a numerically verified convergence of e_t, rather than a proof, weakens Eq. (43) further. The supplement is deferred and no code/data are provided, so verification is currently impossible. These issues cut to the core of the theoretical contribution, but they do not kill the empirical method. If the paper were reframed as an approximate detector with empirical validation and a corrected false-alarm formula, it could be a solid engineering result.\n\nWho it's for: people working in sequential detection, especially with non-i.i.d. backgrounds, and researchers interested in Bayesian nonparametric models for multi-source signal processing. It deserves a serious referee, but I would not accept it in its current form. The referee should ask for either a proof of the approximation error or a clear statement that FKFF is an approximation, a corrected Eq. (43), and a supplement with the proofs.","headline":"The model and experiments are worth a look, but the central claim of exact closed-form performance bounds is not established: FKFF is an unlabeled approximation and Eq. (43) has an internal inconsistency.","tokens_in":26151,"tokens_out":3377,"would_cite":false,"duration_ms":32493,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Bayesian nonparametric model of an unknown number of background emitters, plus a recursive factorial Kalman filter, makes finite moving average detection of transient signals computable online with bounded missed-detection and…","keywords":["infinite factorial linear dynamical system","sticky Indian buffet process","transient signal detection","finite moving average test","factorial Kalman forward filtering","Bayesian nonparametric","particle Gibbs with ancestor sampling","slice sampling"],"falsifier":"For a single time step with $M=2$, compute the exact Gaussian posterior over the two latent sources given the observation and compare the off-diagonal covariance block to zero; the FKFF equations assume that block is zero. If it is nonzero and large enough to shift the predictive likelihood or the FMA threshold by a non-negligible amount, the exactness claim of the recursive likelihood computation fails. Concretely, simulate from the FLDS, run both the FKFF and a full-covariance or particle-filter approximation, and check whether the empirical false alarm and missed detection probabilities violate the Eq. (43) bounds for a fixed threshold.","tokens_in":25035,"feed_emoji":"📡","tokens_out":6974,"duration_ms":63923,"temperature":0.7,"pith_summary":"This paper addresses transient signal detection when the background is not a single source but a superposition of signals from multiple emitters, with the number of emitters unknown. It proposes to learn the background offline with the infinite factorial linear dynamical system (IFLDS), a Bayesian nonparametric model built on a sticky Indian buffet process, and then detect the signal online with a finite moving average (FMA) stopping time. To make the stopping time computable, it derives a factorial Kalman forward filtering recursion and a dependence structure that turns the log-likelihood ratio into a recursive statistic; it also gives closed-form Gaussian bounds for the worst-case false alarm and missed detection probabilities. The significance is that a detector can handle time-correlated multi-source backgrounds without knowing how many sources are present, with calibrated performance guarantees.","feed_headline":"Unknown emitter count no longer blocks transient signal detection","feed_subtitle":"A Bayesian filter plus finite moving average test detects finite signals in multi-source backgrounds.","key_machinery":"The central machinery has three pieces. First, the IFLDS, a conjugate Bayesian nonparametric model in which a sticky Indian buffet process (a prior over infinite binary activity matrices with persistence) gives each emitter a Markov activity state and an LDS, allowing unbounded source number and temporally persistent sources. Second, the factorial Kalman forward filtering (FKFF), a recursive Gaussian update for the latent-state distribution of the FLDS; the recursion is made tractable by assuming the filtered covariance of the latent state remains block diagonal (Eq. 27), and it supplies the predictive likelihoods (Eq. 30) used in the detection statistic. Third, the FMA stopping time with the reformulated log-likelihood ratio of Eq. (34), which uses two parallel FKFF runs (signal present versus signal absent) and whose threshold is set through the closed-form bounds of Eq. (43).","core_discovery":"The central claim is that a composite background can be represented as an unbounded factorial linear dynamical system: each emitter has its own linear-Gaussian latent state, and the observation is the sum of the active emitters' outputs plus noise, with a sticky Indian buffet process prior deciding which emitters are active. Given that representation, the paper claims the factorial Kalman forward filtering equations (Eqs. 28-30) provide a recursive way to compute predictive data likelihoods, and that under Assumptions 1 and 2 the finite moving average stopping time (Eq. 33) is a valid detector whose worst-case missed detection and false alarm probabilities are bounded by the closed-form Gaussian expressions in Eq. (43). If correct, the detector can run online after an offline parameter-learning stage, estimate the number of background sources automatically, and deliver lower missed detection than single-source LDS, Gaussian, CUSUM, and Shewhart baselines in multi-source scenarios.","pith_inferences":["A full-covariance correction to the FKFF would produce true predictive likelihoods; if the cross-source posterior covariance stays small in steady state the current bounds may be close, but this remains to be demonstrated.","The block-diagonal approximation may also explain the gap the authors report between the theoretical and empirical missed-detection curves under the signal-present hypothesis; an exact-filtering version could close that gap.","The same IFLDS representation could serve adjacent tasks such as emitter enumeration, source activity tracking, and change-point localization, not just binary transient detection.","A natural testable extension is to relax Assumption 2 and quantify how much violation of the independence assumption degrades the bounds; the paper states the assumption but does not probe its sensitivity."],"forward_implications":["The IFLDS parameter-learning method (slice sampling plus particle Gibbs with ancestor sampling) automatically estimates how many background sources are present, removing the need to specify the source count in advance.","FKFF gives an online recursive likelihood for a factorial state-space model, bypassing the NP-hard exact likelihood computation noted for factorial hidden Markov models.","The threshold-setting formula ties the window length (tolerable detection delay) to worst-case false alarm and missed detection bounds, so a system designer can set detection thresholds analytically rather than by simulation.","In multi-source simulations the proposed FLDS-FMA detector shows lower missed detection than LDS-FMA, Gaussian-FMA, CUSUM, and Shewhart baselines, with the gap widening as the number of background sources grows.","In the pulse-detection experiment under BPSK/QPSK communication interference, the method reports detection probabilities of 63% versus 41% for LDS-FMA and 33% for Gaussian-FMA at a fixed 15% false alarm rate."],"supporting_citations":[{"why":"Supplies the Indian buffet process prior over infinite binary feature matrices, the basis of the unbounded-source representation.","marker":"[10]"},{"why":"Provides the stick-breaking construction of the Indian buffet process used to sample an unbounded number of chains.","marker":"[11]"},{"why":"Defines factorial HMMs and establishes that exact likelihood computation for factorial state-space models is NP-hard, the obstacle FKFF addresses.","marker":"[12]"},{"why":"Introduces the infinite factorial HMM with Markov Indian buffet process, the direct template for the IFLDS activity structure.","marker":"[13]"},{"why":"Provides the infinite factorial dynamical model with continuous latent states, the representational predecessor extended to electromagnetic signals.","marker":"[14]"},{"why":"Formulates sequential transient detection and motivates the finite moving average as the detection statistic to use.","marker":"[15]"},{"why":"Supplies slice sampling, the auxiliary-variable method used to sample the number of parallel chains in parameter learning.","marker":"[27]"},{"why":"Supplies particle Gibbs with ancestor sampling, the algorithm used to sample local latent variables in parameter learning.","marker":"[28]"},{"why":"Provides tighter performance bounds for finite moving average tests, the reason the paper selects FMA over window-limited CUSUM.","marker":"[32]"}],"fun_headline_variants":["Transient detection in multi-source backgrounds without knowing emitter count","Bayesian filtering and moving average test detect signals from many emitters","Unbounded source count: new transient detector adapts automatically","Factorial Kalman forward filtering for transient detection in unknown sources"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The filtering equations assume that the latent sources stay uncorrelated with each other after each update, even though conditioning a sum of Gaussian sources on a shared observation generally makes their posterior correlated; if that approximation is poor, the likelihoods and the detection bounds inherit the error.","fun_headline_variants_meta":{"raw":{"variants":["Transient detection in multi-source backgrounds without knowing emitter count","Bayesian filtering and moving average test detect signals from many emitters","Unbounded source count: new transient detector adapts automatically","Factorial Kalman forward filtering for transient detection in unknown sources"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2701,"prompt_tokens":993,"completion_tokens":1708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1637}},"tokens_in":609,"tokens_out":1708,"duration_ms":13630,"temperature":1.0,"reasoning_tokens":1637,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:22:08.244583+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a single time step with $M=2$, compute the exact Gaussian posterior over the two latent sources given the observation and compare the off-diagonal covariance block to zero; the FKFF equations assume that block is zero. If it is nonzero and large enough to shift the predictive likelihood or the FMA threshold by a non-negligible amount, the exactness claim of the recursive likelihood computation fails. Concretely, simulate from the FLDS, run both the FKFF and a full-covariance or particle-filter approximation, and check whether the empirical false alarm and missed detection probabilities violate the Eq. (43) bounds for a fixed threshold.","supporting_citations":[{"cited_title":"The indian buffet process: An introduction and review","cited_arxiv_id":null,"evidence_quote":"Supplies the Indian buffet process prior over infinite binary feature matrices, the basis of the unbounded-source representation."},{"cited_title":"Stick-breaking construction for the indian buffet process,","cited_arxiv_id":null,"evidence_quote":"Provides the stick-breaking construction of the Indian buffet process used to sample an unbounded number of chains."},{"cited_title":"Factorial hidden markov models,","cited_arxiv_id":null,"evidence_quote":"Defines factorial HMMs and establishes that exact likelihood computation for factorial state-space models is NP-hard, the obstacle FKFF addresses."},{"cited_title":"The infinite fac- torial hidden markov model,","cited_arxiv_id":null,"evidence_quote":"Introduces the infinite factorial HMM with Markov Indian buffet process, the direct template for the IFLDS activity structure."},{"cited_title":"Infinite factorial dynamical model,","cited_arxiv_id":null,"evidence_quote":"Provides the infinite factorial dynamical model with continuous latent states, the representational predecessor extended to electromagnetic signals."},{"cited_title":"Sequential detection of transient changes,","cited_arxiv_id":null,"evidence_quote":"Formulates sequential transient detection and motivates the finite moving average as the detection statistic to use."},{"cited_title":"Slice sampling,","cited_arxiv_id":null,"evidence_quote":"Supplies slice sampling, the auxiliary-variable method used to sample the number of parallel chains in parameter learning."},{"cited_title":"Particle gibbs with ancestor sampling,","cited_arxiv_id":null,"evidence_quote":"Supplies particle Gibbs with ancestor sampling, the algorithm used to sample local latent variables in parameter learning."},{"cited_title":"Performance bounds for finite moving average tests in transient change detection,","cited_arxiv_id":null,"evidence_quote":"Provides tighter performance bounds for finite moving average tests, the reason the paper selects FMA over window-limited CUSUM."}],"review_version":1}