{"id":"d972b0b2-db44-4c01-a626-42bf8ac43061","arxiv_id":"2505.08639","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"The proposed VB-UKF plus Huber plus conformal outlier detection framework improves geomagnetic indoor localization accuracy in the authors' experiments, from 81.25% to 93.75% matching accuracy and from 0.62-6.87 m to 0.03-0.35 m error.","lead":"This paper combines variational Bayesian filtering, Huber robust weighting, and conformal prediction to make indoor geomagnetic localization more resistant to noise and outliers. It reports large accuracy gains on a building test, but its claimed statistical guarantees are not actually proved.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's coverage guarantee is unproven: the sliding-window conformal proof uses an unjustified block-exchangeability assumption and a non-martingale difference sequence, so the advertised user-controlled false alarm rate lacks support.","rationale":"I agree with the reader's REJECT verdict; my stress-test does not change it. The most load-bearing weakness is the conformal outlier detector's statistical guarantee, because the paper's headline claim of 'statistically rigorous confidence intervals' and a user-controlled false alarm rate depends on it. The reader identified the Modified Exchangeability Assumption as the weakest point; I go a step further and note that the proof in Appendix B is invalid even under that assumption, since D_k is not a martingale difference sequence and the normalization step cannot yield the stated pointwise bound. The Huber covariance theorem is also weak—it essentially proves that a sum of PSD matrices is PSD and does not match Algorithm 2's update exactly—but it is less load-bearing than the conformal guarantee. The UNGM simulation and Table I are genuine positive evidence for an empirical MSE improvement, and I credit them; however, they do not measure the false-alarm or coverage property that the paper advertises. The field experiment is too under-specified to verify the dramatic accuracy gains, but the theoretical gap alone is sufficient to support the REJECT verdict.","tokens_in":14767,"tokens_out":9035,"duration_ms":92716,"concrete_test":"Run the proposed COD detector on the geomagnetic field data from Section IV-B with a held-out set of clean (non-outlier) samples, fixing α = 0.05 and testing window sizes w = 50, 100, 500. Compute the empirical false-alarm rate as the fraction of clean samples with s_k > qhat_k, alongside Monte Carlo standard errors. If the observed rate exceeds α by more than twice the standard error for any window size, the claimed user-controlled false-alarm control fails in exactly the regime the paper targets; if it does not, the empirical claim survives despite the proof gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim needing support is the COD false-alarm guarantee in Theorem 2 / Eq. (31). That guarantee rests entirely on the 'Modified Exchangeability Assumption' in Appendix B, which asserts approximate exchangeability of the non-conformity scores in C_k ∪ {s_k} within temporal blocks of size w+1. No empirical validation or literature support is given for geomagnetic/inertial measurement streams, which are autocorrelated and nonstationary; the cited [31] is a jackknife+ paper, not a weak-dependence tool. More seriously, the proof in Appendix B does not establish the bound even if the assumption holds. The sequence D_k = 1{s_k ≤ qhat_k} − (1−α) is treated as a martingale difference, but qhat_k depends on C_k, so E[D_k | past] = P(s_k ≤ qhat_k | past) − (1−α), which is generally nonzero and random; Azuma–Hoeffding does not apply to this pointwise event. The normalization step also does not produce the stated O(1/√w): a concentration bound on (1/(K−w))∑D_k cannot be converted into the pointwise coverage of Eq. (31). Thus the advertised 'user-controlled false alarm rate' and 'statistically rigorous confidence intervals' are not demonstrated. The UNGM simulation gives some evidence of MSE improvement but never measures coverage or false-alarm rate against labeled outliers, so it does not substitute for the missing guarantee.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical framework for robust indoor localization that combines variational Bayesian (VB) parameter learning, Huber M-estimation, and a conformal outlier detection (COD) layer. The central claims are: (i) VB-UKF adaptively estimates state and noise covariances; (ii) Huber weighting suppresses mild outliers while preserving positive semi-definiteness of the posterior covariance; and (iii) COD provides a distribution-free, user-controlled false-alarm guarantee through a sliding-window conformal threshold. The theoretical part includes Theorem 1 on Huber-weighted covariance and Theorem 2 on time-varying coverage for sliding-window conformal prediction. The experiments use a Univariate Nonstationary Growth Model (UNGM) with four noise scenarios and a geomagnetic fingerprint indoor localization testbed; the reported results show consistent MSE reductions when the COD layer is added and an improvement in fingerprint matching accuracy from 81.25% to 93.75% with positioning errors reduced from 0.62–6.87 m to 0.03–0.35 m. The paper frames these results as evidence that the framework provides statistically rigorous confidence intervals and robust outlier mitigation in non-Gaussian environments.","tokens_in":15028,"tokens_out":5131,"duration_ms":55462,"significance":"If the claimed coverage guarantee were valid, the integration of VB adaptation, Huber robustness, and conformal outlier detection would be a useful contribution to robust indoor localization under non-Gaussian noise. The UNGM Monte Carlo comparison is a genuine empirical study and the reported MSE gains are internally consistent with the paper's algorithmic claims. However, the load-bearing theoretical result—the sliding-window conformal coverage guarantee—is not established: the proof in Appendix B is mathematically invalid, and the 'Modified Exchangeability Assumption' on which it depends is asserted without empirical or theoretical support for autocorrelated, nonstationary geomagnetic/IMU streams. Moreover, the experiments do not measure the advertised coverage or false-alarm rate, so the central statistical claim remains unverified. The paper also does not provide code or data, which limits reproducibility.","major_comments":[{"comment":"The proof of Theorem 2 is invalid as written. The sequence D_k = I{s_k ≤ qhat_k} − (1−α) is not a martingale difference sequence with respect to the natural filtration, because qhat_k depends on the calibration window C_k and hence on past scores; E[I{s_k ≤ qhat_k} | F_{k−1}] = P(s_k ≤ qhat_k | F_{k−1}) is a random variable that is generally not equal to 1−α. Azuma–Hoeffding therefore cannot be applied to this sequence. In addition, the normalization step does not yield the claimed pointwise bound: a concentration inequality for (1/(K−w)) Σ D_k would control an average over time, not the marginal probability P(s_k ≤ qhat_k) in Eq. (31). The proof also cites reference [31] as the source of a 'weakly dependent sequences' inequality, but [31] is the jackknife+ paper, which does not contain such a result.","section":"§III-C2 and Appendix B, Eq. (31)"},{"comment":"The coverage guarantee rests entirely on the assumption that the non-conformity scores in C_k ∪ {s_k} are approximately exchangeable within temporal blocks of size w+1. This assumption is stated without any empirical validation or literature support for geomagnetic and inertial measurement streams, which are strongly autocorrelated and nonstationary. It is essentially the same exchangeability condition that conformal prediction requires, so the proof does not derive the guarantee; it assumes it. Consequently, the advertised 'user-controlled false alarm rate' and the claim of working 'without making assumptions about the underlying noise distribution' are not established. A concrete test would be to report empirical coverage and false-alarm rates on labeled outliers under Cases c/d and on the real geomagnetic data, which the paper does not provide.","section":"Appendix B, 'Modified Exchangeability Assumption'"},{"comment":"The empirical evaluation does not test the central statistical claim. The UNGM experiments report only MSE values; no empirical coverage of the conformal interval, no false-alarm rate, and no precision/recall against labeled outliers are given. The Introduction's claim of '95% outlier detection precision with controlled false alarms' is not supported by any table or figure in the paper. The real-world experiment in Section IV-B reports only single-point matching accuracy (81.25% to 93.75%) and an error range (0.62–6.87 m to 0.03–0.35 m), without specifying the number of test points, the experimental protocol, the number of independent runs, or comparisons against the baseline filters used in simulation. As a result, the empirical study cannot substitute for the missing coverage guarantee.","section":"§IV-A, Table I and §IV-B"},{"comment":"The non-conformity score in Eq. (27) is defined through a linearized innovation covariance H P H^T + R_k, but the filtering framework is an unscented filter with a nonlinear measurement function h(·). The paper never defines H for the UNGM or for the geomagnetic experiment, and it does not explain how Eq. (27) relates to the UKF innovation covariance P_zz computed in Eq. (15). If H is a Jacobian, its definition is missing; if H is intended as a linearization, its use in an unscented filter needs justification. This makes the actual conformal scores used in the experiments ambiguous and hinders reproducibility.","section":"Eq. (27) and §III-C2"}],"minor_comments":[{"comment":"The main text labels the coverage result as Theorem 2, while Appendix B labels the same result as Theorem 3; the numbering should be reconciled.","section":"Section III-C2 vs. Appendix B"},{"comment":"The quantile definition in Eq. (29) is not the standard conformal quantile. Standard split conformal uses the ⌈(w+1)(1−α)⌉-th order statistic of the calibration scores; writing it as an empirical quantile at level ⌈(w+1)(1−α)⌉/w is inaccurate for finite w and should be corrected or justified.","section":"Eq. (29)"},{"comment":"The text says the variational framework 'maximizes the Kullback-Leibler divergence' between the variational distribution and the posterior; this should be 'minimizes the KL divergence' (equivalently, maximizes the ELBO), otherwise the exposition is technically wrong.","section":"Section III-B, paragraph before Eq. (12)"},{"comment":"Several figures lack full axis labels or explicit descriptions of the plotted quantities (e.g., Figure 2's 'Value' axis and Figure 6's 'Error' axis are not defined in the text); the caption of Figure 2 mentions 20 test samples but does not describe how the conformal interval is constructed from those samples, which limits reproducibility.","section":"Figures 2, 4, 5, and 6"},{"comment":"The paper states that 'Sections IV–V describe the experimental setup and results, followed by conclusions in Section V,' which conflates Section V with both results and conclusions; the structure should be clarified.","section":"Section V and Abstract"}],"recommendation":"reject","confidential_remarks":"The central advertised contribution is a statistical coverage guarantee for the conformal outlier detector, but the proof is invalid and the assumption on which it depends is unverified. The empirical section, while showing consistent MSE improvements, does not measure coverage or false-alarm rates, so the main claim is unsupported. The misapplication of reference [31] and the ambiguous definition of the non-conformity score further suggest that the theoretical framing needs substantial rework. I recommend rejection, though the underlying idea of combining VB adaptation, Huber robustness, and conformal outlier detection may be salvageable with a correct proof and appropriate experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe system-level combination is genuinely new: VB-AUKF plus Huber downweighting plus sliding-window conformal outlier detection for geomagnetic fingerprinting. The Monte Carlo study is a real comparison—50 runs, four noise scenarios, five filters—and the MSE gains are consistent and sometimes large (UKF drops 53% in Case c). That empirical core deserves credit.\n\nThe theoretical core does not hold up. Theorem 1 proves positive semi-definiteness of a Joseph-form covariance update, but Algorithm 2 implements a different update: (I - W K H) P + W K R K^T, without the outer products. PSD of one does not imply PSD of the other, so the proof does not cover the implemented filter. Theorem 2's coverage guarantee is unproven. The 'Modified Exchangeability Assumption' in Appendix B is doing all the work and is simply asserted for geomagnetic/inertial streams. Even granting it, the proof treats D_k = 1{s_k <= qhat_k} - (1-alpha) as a martingale difference, but qhat_k depends on C_k, so the conditional expectation is not (1-alpha); Azuma-Hoeffding does not apply. The normalization step also converts a bound on averaged coverage error into pointwise coverage in Eq. (31) without justification. So the advertised 'user-controlled false alarm rate' and 'statistically rigorous confidence intervals' are not demonstrated, and the simulation never measures coverage or false-alarm rate against labeled outliers.\n\nThe field experiment is too under-specified to verify the dramatic accuracy claims: 81.25% to 93.75% matching, 0.62-6.87 m to 0.03-0.35 m error, but no trial count, path length, number of test points, or error bars. No code or data are provided, so reproducibility is low.\n\nThat said, this is not a throwaway. The integration idea could be useful if the theory were either corrected or honestly demoted to a heuristic. My verdict is reject, but I would send it to peer review rather than desk-reject: the empirical study and the novel combination merit referee time, and referees should push for either fixed proofs or weakened claims.","headline":"A genuinely new engineering combination with a solid Monte Carlo study, but both theoretical guarantees are broken and the field experiment is too vague to verify; reject after review, not desk-reject.","tokens_in":15614,"tokens_out":2124,"would_cite":false,"duration_ms":22072,"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":"Variational Bayesian learning, Huber M-estimation, and sliding-window conformal outlier detection can keep indoor localization accurate in non-Gaussian noise, reducing worst-case positioning error from 6.87 m to 0.35 m.","keywords":["indoor localization","conformal prediction","variational Bayesian inference","Huber M-estimation","outlier detection","geomagnetic fingerprinting","Kalman filtering","uncertainty quantification"],"falsifier":"Run the conformal outlier detection layer on a long recorded geomagnetic walk with known outliers injected at random positions, fixing $\\alpha$ in advance, and compare the empirical rate at which clean samples are flagged with $\\alpha$. If the empirical false-alarm rate systematically exceeds the nominal level, or if the empirical coverage of the sliding-window prediction sets falls below $1 - \\alpha - O(1/\\sqrt{w})$, the modified exchangeability assumption is violated and the central statistical claim fails.","tokens_in":14494,"feed_emoji":"📍","tokens_out":12978,"duration_ms":109235,"temperature":0.7,"pith_summary":"This paper targets indoor localization in environments where measurement noise is non-Gaussian, time-varying, or contaminated by outliers. It proposes a hierarchical filter that combines variational Bayesian learning of the noise covariance, Huber M-estimation to downweight mild outliers, and a sliding-window conformal outlier detector that recalibrates its threshold online. The central claim is that the conformal layer gives the user control over the false-alarm rate, with a coverage guarantee that does not assume a Gaussian or other parametric noise distribution. On a geomagnetic fingerprint test site, the framework raises single-point matching accuracy from 81.25% to 93.75% and reduces positioning error from a 0.62–6.87 m range to 0.03–0.35 m. A sympathetic reader would care because the same outlier layer improves every baseline filter in the paper's Monte Carlo runs, with the biggest gains in heavy-tailed mixture noise.","feed_headline":"Conformal outlier detection cuts indoor positioning error to 0.35 m","feed_subtitle":"Outlier rejection becomes statistically controlled; match accuracy rises to 93.75 percent.","key_machinery":"The load-bearing mechanism is the filter-aware conformal outlier detector. At each time step it forms a non-conformity score from the normalized innovation, $s_k = \\|z_k - H\\hat{x}_{k|k-1}\\|_{S_k^{-1}}$ with $S_k = HP_{k|k-1}H^\\top + R_k$, and compares it with the sliding-window quantile $\\hat{q}_k$ over the past $w$ scores. A score above $\\hat{q}_k$ triggers covariance inflation, $R_k \\leftarrow \\gamma R_k$ for $\\gamma > 1$, which dampens the suspect measurement while preserving filter stability. Variational Bayesian inference, using an inverse-Wishart prior on $R_k$, supplies the adaptive noise covariance; the Huber weight function handles mild outliers before the conformal layer fires. This stack is what converts a distribution-free prediction-set method into a real-time robust filtering update.","core_discovery":"The paper demonstrates, through Monte Carlo simulations and a real geomagnetic/inertial experiment, that wrapping a robust filter with a sliding-window conformal outlier detector improves estimation under non-Gaussian and outlier-contaminated noise. Its theoretical contributions are a proof that the Huber-modified posterior covariance remains positive semi-definite and a proof that the sliding-window conformal threshold has coverage $\\mathbb{P}(s_k \\le \\hat{q}_k) \\ge 1 - \\alpha - O(1/\\sqrt{w})$ under a modified exchangeability assumption. In the reported experiments the full VB-Huber-AUKF plus conformal detector achieves the lowest mean squared error in every noise scenario, and the largest relative gains appear in heavy-tailed mixture noise cases.","pith_inferences":["The conformal layer is not Kalman-specific; any estimator that can produce a predictive mean and covariance could be wrapped with the same normalized-innovation threshold, including particle filters or learned positioning models.","The modified exchangeability assumption is the untested hinge. A direct empirical coverage audit on recorded geomagnetic walks with injected outliers would show whether the nominal $1-\\alpha$ false-alarm rate survives real nonstationarity.","A soft alternative to the binary covariance-inflation rule would weight measurements continuously by the conformal score, potentially reducing variance near the decision boundary.","Persistent threshold violations could double as a sensor-health signal, flagging a drifting magnetometer or a changed environment rather than a single bad measurement."],"forward_implications":["A user can set a target false-alarm rate $\\alpha$ for outlier rejection and expect the detection threshold to adapt to changing signal statistics without retraining.","The conformal layer is portable: the paper's simulations show it reduces mean squared error for PF, UKF, HUKF, VB-AUKF, and VB-HAUKF, so existing filters can be hardened by wrapping them with the same detector.","The largest improvements occur where Gaussian assumptions fail hardest: in the heavy-tailed mixture-noise case, UKF's mean MSE drops by 53.2%.","The $O(1/\\sqrt{w})$ coverage gap makes the adaptivity-versus-guarantee tradeoff explicit: a longer window tightens the statistical bound, while a shorter window tracks environmental changes faster.","If the field experiment is representative, this filter stack can keep geomagnetic fingerprint positioning errors below roughly one-third of a meter on a 3-meter grid."],"supporting_citations":[{"why":"supplies the variational Bayesian noise-adaptation approach the filter's first layer builds on.","marker":"[4]"},{"why":"grounds the Huber M-estimation used to suppress mild outliers before conformal detection.","marker":"[5]"},{"why":"provides the variational Bayesian unscented Kalman filter that this paper extends with conformal outlier detection.","marker":"[18]"},{"why":"introduces the conformal prediction coverage guarantee that the outlier detector's threshold is built on.","marker":"[23]"},{"why":"cited for sliding-window conformal calibration methods that relax the i.i.d. assumption.","marker":"[24]"},{"why":"supplies sequential conformal inference for time series, the basis of the adaptive thresholding.","marker":"[25]"},{"why":"establishes the temporal stability of geomagnetic fingerprints that makes path matching a time-series matching problem.","marker":"[30]"},{"why":"used in the proof of the sliding-window coverage bound for weakly dependent sequences.","marker":"[31]"}],"fun_headline_variants":["Cuts max indoor positioning error to 0.35 m","Conformal + VB adaptive filter trims indoor error to 0.35 m","Fingerprint match accuracy jumps to 93.75% indoors","Statistical outlier control + adaptive filtering: indoor error 0.35 m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the non-conformity scores in each sliding calibration window of size $w+1$ are approximately exchangeable; the paper assumes this without testing it, and both the coverage guarantee and the user-controlled false-alarm rate depend on it.","fun_headline_variants_meta":{"raw":{"variants":["Cuts max indoor positioning error to 0.35 m","Conformal + VB adaptive filter trims indoor error to 0.35 m","Fingerprint match accuracy jumps to 93.75% indoors","Statistical outlier control + adaptive filtering: indoor error 0.35 m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001438,"raw_usage":{"total_tokens":5780,"prompt_tokens":909,"completion_tokens":4871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":4802}},"tokens_in":525,"tokens_out":4871,"duration_ms":34410,"temperature":1.0,"reasoning_tokens":4802,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:49:44.873063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the conformal outlier detection layer on a long recorded geomagnetic walk with known outliers injected at random positions, fixing $\\alpha$ in advance, and compare the empirical rate at which clean samples are flagged with $\\alpha$. If the empirical false-alarm rate systematically exceeds the nominal level, or if the empirical coverage of the sliding-window prediction sets falls below $1 - \\alpha - O(1/\\sqrt{w})$, the modified exchangeability assumption is violated and the central statistical claim fails.","supporting_citations":[{"cited_title":"V ariational inference based distribut ed noise adaptive bayesian filter,","cited_arxiv_id":null,"evidence_quote":"supplies the variational Bayesian noise-adaptation approach the filter's first layer builds on."},{"cited_title":"Minimax robust dete ction: Classic results and recent advances,","cited_arxiv_id":null,"evidence_quote":"grounds the Huber M-estimation used to suppress mild outliers before conformal detection."},{"cited_title":"A variational bayesian-base d unscented kalman filter with both adaptivity and robustness,","cited_arxiv_id":null,"evidence_quote":"provides the variational Bayesian unscented Kalman filter that this paper extends with conformal outlier detection."},{"cited_title":"Conformal prediction for natural language processing: A s urvey,","cited_arxiv_id":null,"evidence_quote":"cited for sliding-window conformal calibration methods that relax the i.i.d. assumption."},{"cited_title":"Conformal prediction for time series,","cited_arxiv_id":null,"evidence_quote":"supplies sequential conformal inference for time series, the basis of the adaptive thresholding."},{"cited_title":"Geomagnetis m-based indoor navigation by offloading strategy in nb-iot,","cited_arxiv_id":null,"evidence_quote":"establishes the temporal stability of geomagnetic fingerprints that makes path matching a time-series matching problem."},{"cited_title":"Pre dictive inference with the jackknife+,","cited_arxiv_id":null,"evidence_quote":"used in the proof of the sliding-window coverage bound for weakly dependent sequences."}],"review_version":1}