{"id":"83c15277-b8bd-4d51-be5d-70f99dd4894b","arxiv_id":"2607.04734","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A recursive adaptive risk-sensitive Kalman filter is derived for discrete-time singular systems under round-robin measurement scheduling, with stability conditions and numerical gains over standard KF.","lead":"The paper builds a risk-sensitive Kalman filter for singular (descriptor) systems whose sensors take turns sending data under a round-robin schedule. It gives recursive equations, an adaptive risk parameter, and stability conditions that improve robustness over ordinary Kalman filtering when the model is uncertain.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Gaussian closure of the information state is assumed, not proved, for the non-causal RRIPS under adaptive risk and periodic measurements.","rationale":"The reader correctly isolates the Gaussian-closure assumption as the weakest link supporting the strongest claim. The paper never proves that Φk stays Gaussian once the non-causal Weierstrass dynamics and the round-robin measurement schedule are present; it simply assumes the property for “sufficiently small” µ1 and proceeds. All subsequent algorithmic and stability statements inherit this gap. The concrete numerical check proposed above would either confirm that the classical closure still holds in the transformed coordinates or expose a quantitative breakdown, thereby deciding whether the CONDITIONAL verdict can be upgraded or must remain. No stronger internal inconsistency appears, so the reader’s overall assessment stands.","tokens_in":18286,"tokens_out":568,"duration_ms":5863,"concrete_test":"Specialize to a nilpotency-index-1 system (N=0) with constant full measurement matrix (σ=1) and fixed µ1=0; the recursion must recover the classical Kalman filter for the Weierstrass-transformed pair. Then re-introduce a small fixed µ1>0 and a two-period schedule; compute the true information-state density by numerical quadrature (or particle approximation) over a short horizon and compare its first two moments with those produced by (18)+(26). If the relative discrepancy exceeds a few percent, the Gaussian-closure assumption fails for the non-causal periodic case.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The recursive RSKF (Theorems 1 and 4) and the adaptive rule of §3.3 rest on the claim that the information state Φk remains an unnormalized Gaussian whenever µ1,k−1 is small enough that 2µ1,k−1P̄k−1|k−1<I. The paper states this after (13) and invokes it to obtain the closed-form prediction (18) and update (26). For ordinary causal linear-Gaussian systems the claim is classical, but here the Weierstrass transformation produces a non-causal process (5b)–(6) whose process noise is correlated with the initial state (Remark 2) and whose measurement matrix is periodically time-varying. No inductive argument is supplied showing that the product of the non-causal transition density, the exponential risk factor, and the periodic likelihood stays Gaussian (or even that the resulting covariance remains finite) under the online choice of µ1. The adaptive rule only enforces the algebraic inequality after the fact; it does not restore the missing closure proof. If the Gaussianity fails, the claimed recursion is no longer exact and the subsequent stability claim (Theorem 6) loses its foundation.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper develops a risk-sensitive Kalman filter for discrete-time linear stochastic singular systems under a round-robin communication protocol. Using the Weierstrass canonical form, the singular system is converted into an equivalent non-causal augmented state-space model with periodically scheduled measurements (RRIPS). A recursive RSKF is derived via a Bayesian information-state formulation that minimizes an exponential quadratic cost; an adaptive rule for the risk parameter µ1 is introduced to keep the predicted covariance positive definite; and uniform boundedness of the posterior covariance is claimed under uniform complete observability/controllability of the uncertain periodic system (Theorems 1, 4, 6). The filter recovers ordinary KF when the risk parameter vanishes and ordinary RSKF when E = I. Numerical Monte-Carlo comparisons with the standard KF under parametric uncertainty are provided.","tokens_in":18554,"tokens_out":1111,"duration_ms":8820,"significance":"If the Gaussian-closure and stability arguments hold, the work fills a genuine gap: risk-sensitive filtering for networked singular systems under explicit scheduling has not been treated, and the combination of WCF, periodic measurements, adaptive risk, and a stability claim is a natural and useful extension of both the authors’ earlier KF-for-RRP result and classical RSKF. The adaptive mechanism that enforces positive-definiteness online is a practical contribution. The reductions to known filters are clean. The main technical value therefore hinges on whether the information-state recursion remains exact for the non-causal RRIPS; that point is currently the load-bearing open issue.","major_comments":[{"comment":"After (13) and throughout Theorems 1 and 4 the information state Φk is assumed to remain an unnormalized Gaussian for sufficiently small µ1,k-1 satisfying 2µ1P̄k-1|k-1 < I. For ordinary causal linear-Gaussian systems this is classical, but the WCF produces a non-causal process (5b)–(6) whose process noise is correlated with the initial state (Remark 2) and whose measurement matrix is periodically time-varying. No inductive argument is supplied showing that the product of the non-causal transition density, the exponential risk factor and the periodic likelihood stays Gaussian (or that the resulting covariance remains finite) under the online choice of µ1. The adaptive rule of §3.3 only enforces the algebraic inequality after the fact; it does not restore the missing closure. If Gaussianity fails, the claimed closed-form recursion (18),(26) is no longer exact.","section":"§3.1–3.2, Theorems 1 & 4"},{"comment":"Theorem 6 asserts uniform boundedness of P̄k|k under uniform complete observability/controllability of the uncertain RRIPS, but the proof is only a one-sentence reference to “Appendix B of [27]”. That appendix treats a different (causal, delayed-measurement) setting. The non-causal structure, the periodic measurement schedule, and the adaptive risk parameter all alter the Gramian and Riccati arguments; a self-contained sketch (or an explicit verification that the cited appendix applies verbatim) is required before the stability claim can be accepted.","section":"§4, Theorem 6"}],"minor_comments":[{"comment":"The process-noise covariance Q̄k is written as a double sum involving future noises and then reduced to a single sum (Remark 1). A short explicit verification that the cross terms vanish under the white-noise assumption would remove any ambiguity.","section":"Remark 1"},{"comment":"Figures 1–3 report RMSE/Avg-MSE for only two of the three state components and a single uncertainty level range; adding the third component and a brief discussion of the algebraic subsystem would strengthen the numerical section.","section":"§5"},{"comment":"Notation for the partitioned measurement matrices H̄mk and Rmk is introduced twice with slightly different indexing; a single consistent definition would improve readability.","section":"§2.2 and Lemmas 2–3"},{"comment":"Several self-citations ([38],[39]) are appropriate background, but the relation of the present adaptive-risk construction to the fixed-risk RSKF of Zhang et al. could be stated more sharply in the introduction.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a natural incremental step from the authors’ own recent KF-for-RRP paper and from classical RSKF. The technical gap (Gaussian closure for the non-causal RRIPS) is real but fixable; if the authors supply a short inductive argument or a precise reduction to a known causal case, the paper becomes a solid contribution for Automatica or a comparable journal. Scope fit is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The paper gives a recursive risk-sensitive Kalman filter for regular discrete singular systems under round-robin measurement scheduling. They reduce via Weierstrass form to a non-causal but non-singular periodic system (their RRIPS), derive the usual information-state prediction/update under an exponential quadratic cost, add an online adaptive µ1 that keeps the predicted covariance positive definite, and recover the ordinary KF when µ vanishes. That combination, plus the adaptive rule, is new relative to Zhang et al. (2003) and the authors’ own earlier KF-for-RRP paper.\n\nWhat works: the Bayesian derivation is clean once you grant the Gaussian information state; the special-case reductions are correct; the adaptive rule is practical and well-motivated; the numerical example (modest uncertainty, 500 MC runs) shows the expected RMSE improvement over KF. Observability of the periodic system is characterized properly. Citations are appropriate and not circular.\n\nThe soft spot is real but not fatal. After (13) they simply assume Φ k stays unnormalized Gaussian for small enough µ1. For ordinary causal LQG this is classical; here the Weierstrass step produces a non-causal process whose noise is correlated with the initial state and whose measurement matrix is periodic. No inductive argument is given that the product of the non-causal transition, the exponential risk factor, and the periodic likelihood remains Gaussian under the online µ choice. The adaptive rule only enforces the algebraic inequality after the fact. Theorem 6 on uniform boundedness is likewise only sketched by reference to an appendix of a prior paper. These gaps keep the soundness mid-range; they do not invalidate the recursion for practical use, but a referee will want them tightened.\n\nThis is for people already working on networked descriptor filtering or risk-sensitive estimators under scheduling. It is not a broad advance, but it is a usable, carefully assembled algorithm with honest special-case checks. I would send it to peer review; the contribution is real enough to deserve a serious referee who can push on the Gaussian-closure and stability arguments.","headline":"Solid incremental RSKF for singular systems under round-robin: the recursion and adaptive µ rule are usable, but Gaussian closure of the information state is assumed rather than proved for the non-causal RRIPS.","tokens_in":19177,"tokens_out":538,"would_cite":false,"duration_ms":6689,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E11","93C55","93B07"],"pacs":[],"model":"grok-4.5","headline":"A recursive risk-sensitive Kalman filter for singular systems under round-robin measurement scheduling improves robustness by adapting the risk parameter online from covariance information.","keywords":["singular systems","networked control systems","round-robin protocol","Weierstrass canonical form","Kalman filtering","risk-sensitive filtering","periodic systems","adaptive risk parameter"],"falsifier":"Run the adaptive risk-sensitive recursion on a regular singular plant with known nonzero modelling error and a fixed round-robin schedule; if the empirical RMSE never falls below that of the ordinary Kalman filter, or if the predicted covariance loses positive-definiteness despite the adaptive µ rule, the central claim fails.","tokens_in":19163,"feed_emoji":"📡","tokens_out":929,"duration_ms":12254,"temperature":0.7,"pith_summary":"The paper shows how to estimate the state of a linear singular system when only a rotating subset of sensors can talk at each time step. It first rewrites the singular plant, via the Weierstrass form, as an ordinary but non-causal periodic system induced by the round-robin schedule. On that model it derives a recursive risk-sensitive Kalman filter that minimises an exponential quadratic cost, then recovers the original singular-state estimate by a simple linear transformation. An adaptive rule chooses the risk weight from the current covariance so that the predicted covariance stays positive definite while the filter becomes more or less risk-averse as uncertainty grows. Under uniform complete observability and controllability of the uncertain periodic system the posterior covariance remains bounded, and the scheme collapses to ordinary Kalman filtering when the risk weight vanishes. Numerical trials indicate lower RMSE than the classical Kalman filter once model mismatch appears.","feed_headline":"Risk-sensitive filter beats Kalman under round-robin limits","feed_subtitle":"Adaptive risk weight keeps covariance valid and cuts error when only rotating sensor subsets arrive","key_machinery":"The Weierstrass-to-RRIPS transformation together with the information-state recursion that replaces the ordinary prediction covariance by (P^{-1}-2µI)^{-1}; the adaptive rule that selects µ so that this matrix remains positive definite is what carries both the recursion and the robustness claim.","core_discovery":"For a regular discrete-time linear singular system whose measurements are transmitted under a round-robin protocol, the Weierstrass canonical form yields an equivalent round-robin-induced periodic system on which a Bayesian risk-sensitive Kalman filter can be written in closed recursive form; an online covariance-dependent choice of the risk parameter keeps the predicted covariance positive definite and, under uniform complete observability and controllability, guarantees uniform boundedness of the posterior covariance even under bounded plant uncertainty.","pith_inferences":["The same adaptive-µ idea could be attached to other exponential-cost estimators (e.g., risk-sensitive H∞ hybrids) without re-deriving the entire singular-system theory.","Because the algebraic subsystem is non-causal, any future extension to multi-step packet loss will have to track finite-horizon noise correlations that ordinary causal filters never see.","If the nilpotency index of the Weierstrass block is larger than one, the effective process-noise covariance becomes a moving average of future noises; this may limit how aggressively µ can be increased before the Gaussian closure assumption breaks."],"forward_implications":["When the risk weight is set to zero the algorithm recovers the classical Kalman filter for singular systems under round-robin scheduling.","When the singular matrix is the identity the algorithm recovers the ordinary risk-sensitive Kalman filter for non-singular systems.","Uniform complete observability and controllability of the uncertain periodic system are sufficient for uniform boundedness of the filter covariance.","The same transformation-plus-adaptive-risk pattern can be applied once packet dropouts or random delays are added to the communication model."],"fun_headline_variants":["Adaptive RSKF for singular systems under round-robin measurements","WCF yields recursive Bayesian risk-sensitive filter for RRIPS","Online risk parameter keeps predicted covariance positive definite","RSKF stability under UCO/UCC for periodic singular systems","Risk-sensitive filter reduces to KF when risk weight vanishes"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The filter treats the information state as remaining an unnormalised Gaussian for every small enough risk weight, which is assumed rather than proved after the non-causal singular-to-state-space conversion and the periodic measurement selection.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive RSKF for singular systems under round-robin measurements","WCF yields recursive Bayesian risk-sensitive filter for RRIPS","Online risk parameter keeps predicted covariance positive definite","RSKF stability under UCO/UCC for periodic singular systems","Risk-sensitive filter reduces to KF when risk weight vanishes"]},"model":"grok-4.5","effort":"low","cost_usd":0.00554,"raw_usage":{"total_tokens":1532,"prompt_tokens":817,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":55400000,"prompt_tokens_details":{"text_tokens":817,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":651,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":817,"tokens_out":64,"duration_ms":5543,"temperature":1.0,"reasoning_tokens":651,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T14:21:12.654335+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the adaptive risk-sensitive recursion on a regular singular plant with known nonzero modelling error and a fixed round-robin schedule; if the empirical RMSE never falls below that of the ordinary Kalman filter, or if the predicted covariance loses positive-definiteness despite the adaptive µ rule, the central claim fails.","supporting_citations":[],"review_version":1}