{"id":"edade217-692f-4a55-be95-2866ffb0d3c1","arxiv_id":"2506.05728","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A geometric extended Kalman filter using affine connections, parallel transport, and curvature to transform Gaussian covariances between coordinates improves state estimation on manifolds.","lead":"This paper derives a version of the extended Kalman filter that respects the curved geometry of the state space, using parallel transport and curvature to translate Gaussian uncertainty between different reference points. The authors show that adding these geometric corrections reduces estimation error compared to the standard filter on a simplified inertial navigation problem.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (37) appears to give the wrong sign for the posterior mean: in the Euclidean identity-output limit it moves the estimate away from the measurement, contradicting the objective in Eq. (33).","rationale":"The reader's concern about the O(tr(Sigma^2)) validity in the transient is a domain-of-validity issue, but it presupposes that the update formula is internally correct. The sign contradiction in Eq. (37) and Eq. (44) is more fundamental: if the paper's displayed update equations are taken literally, the filter corrects in the wrong direction in the simplest possible limit, so the reported improvements cannot be attributed to the published algorithm. Because the numerical section does not specify how the update mean is implemented and no code is shipped, the discrepancy cannot be resolved from the manuscript. A Euclidean specialization is the cleanest probe: all geometric Jacobians become identity and the formulas must reduce to the standard Kalman update, but Eq. (37) instead gives the negative of the standard correction. This warrants rejection of the current version, with the expectation that a sign correction, a derivation fix, and reproducible code would allow resubmission.","tokens_in":24542,"tokens_out":18568,"duration_ms":191571,"concrete_test":"Run the scalar Euclidean consistency check: set M = N = R, F(x) = x, h(x) = x, prior (xi_hat, Sigma), and measurement y with noise covariance R. Minimize the posterior objective in Eq. (33) using the linearized residual eta(epsilon) approximately r - epsilon from Eq. (35); the minimizer is epsilon* = Sigma (Sigma + R)^{-1} r. Compare with Eq. (37), which gives mu+ = -Sigma (Sigma + R)^{-1} r. If the discrepancy is confirmed, implement Algorithm 1 on this scalar example as written and observe that the estimate diverges, proving the update formula cannot be correct.","verdict_should_be":"REJECT","load_bearing_attack":"The central update step is internally inconsistent as written. Specialize Section 4.4 to M = N = R^m, h = identity, so that J1 = J2 = I, C = I, R-dagger = R, and K = Sigma (Sigma + R)^{-1}. With r = phi_yhat(y) = y - xi_hat, Eq. (35) gives eta(epsilon) approximately r - epsilon. Substituting into the quadratic objective (33) and minimizing over epsilon yields epsilon* = K r. Equation (37), however, states mu+ = -K r, i.e., the opposite correction; the reset step would move the state away from the measurement and the filter would diverge. The same sign appears in the iterated EKF mean formula (44). The simulations report good performance, so either the implementation uses a corrected formula not shown in the paper or the displayed formulas are wrong; with no code or details on how the update mean is computed, the reader cannot tell. This is a direct algebraic flaw in the paper's central algorithmic claim, not merely an asymptotic-regime limitation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a framework for extended Kalman filtering on smooth manifolds equipped with an affine connection. It defines concentrated Gaussian distributions in normal coordinates, derives a coordinate-change formula (Lemma 3.2) with Jacobian corrections involving the exponential map, and proposes geometric modifications to the propagation, update, and reset steps of the EKF. An iterated variant is also presented. The method is benchmarked on an SE2(3) inertial navigation example with direct pose measurements, reporting improved RMSE and ANEES relative to classical EKF and iterated EKF, with an ablation study separating the contributions of the update and reset modifications. The central formal contribution is the covariance transformation (14) and the associated Jacobian approximations in the appendix.","tokens_in":24743,"tokens_out":12761,"duration_ms":129879,"significance":"If the proposed update and reset formulas were correct, the paper would make a useful contribution to intrinsic state estimation on manifolds. The coordinate-change Lemma 3.2 is a clean and potentially reusable result, and the ablation study is a strength because it isolates the effect of each geometric correction. The paper is also transparent in admitting that the proposed propagation correction is not implementable in its current form. However, the central algorithmic formulas for the update mean appear internally inconsistent: the displayed equations would diverge in a simple Euclidean limit, and the simulations therefore cannot be trusted as evidence for the claimed performance unless the implementation uses different formulas. Because the load-bearing algorithmic claim is affected, the manuscript needs a substantive technical revision before its conclusions can be accepted.","major_comments":[{"comment":"The posterior-mean formula has the wrong sign and is applied to the wrong innovation. In the Euclidean identity-output limit M = N = R^m, h = identity, one has J1 = J2 = C = I and R-dagger = R; Eq. (35) gives eta(epsilon) = r - epsilon with r = phi_haty(y). Minimizing the quadratic objective (33), (Sigma^{-1} + R^{-1}) epsilon = R^{-1} r, yields epsilon* = K r with K = Sigma (Sigma + R)^{-1}. Equation (37) instead states mu^+ = -K r, which moves the estimate away from the measurement and would cause the filter to diverge. The same sign error appears in Eq. (44) and in the pseudocode of Algorithms 1 and 2. In addition, Eq. (37) applies K to the raw innovation r, whereas consistency with the R-dagger covariance in Eq. (36) requires the transformed innovation J1^{-1}J2 r; the mean and covariance updates therefore do not solve the same quadratic minimization. Since the simulations report good performance, the implementation must differ from the displayed formulas; without code or a corrected derivation, the central algorithmic claim is not verifiable.","section":"Section 4.4, Eqs. (35)-(37), and Algorithm 1"},{"comment":"The claimed propagation modification is not actually part of the proposed filter. The corrected process-noise covariance Q-diamond in Eq. (30) depends on the virtual noise-free state xi_diamond_{k+1} = F(xi_k, u_{k+1}), which is unknown at the time of propagation. The text states that the authors do not have a way to overcome this difficulty and that the best approximation is to set Q_{k+1} to approximate Q-diamond_{k+1}; Eq. (32) then uses Q_{k+1} unmodified, and Algorithm 1 contains no propagation correction. The abstract's claim of \"novel geometric modifications to the propagation\" is therefore not supported by the presented algorithm. The paper should either restrict its contribution claims to the update and reset steps or provide an implementable, error-bounded approximation to Q-diamond rather than abandoning the modification.","section":"Section 4.3, Eqs. (30)-(32)"},{"comment":"The central coordinate-change formula (14) is proved only to second order in the covariance, O(tr(Sigma_1^2)), yet the reported performance gains are largest in the initial transient (Table 1 and Figures 4-5), where the covariance and linearization errors are large. The paper provides no bound or numerical evidence that the truncated formula remains accurate outside its asymptotic regime. This is a load-bearing concern because the update and reset steps rely on Eq. (14); the ablation study in Section 6.4 shows that the reset-only variant, which applies the coordinate change without the update correction, degrades performance dramatically. A discussion of the domain of validity of the approximation, or a simulation with a smaller initial covariance, would be needed to support the transient-phase claims.","section":"Lemma 3.2 and Section 6"}],"minor_comments":[{"comment":"Independently of the sign error, the information term in Eq. (44) appears to use J2^T where the derivation from Eq. (42) and the transformed innovation require J2^{-T}; please verify the matrix transposes and show the intermediate steps.","section":"Section 5, Eq. (44)"},{"comment":"The implementation of the Jacobians J1 and J2 in the simulation is not described: it is unclear whether exact algebraic formulas for SE(3)/SE2(3) are used or the o(|v|^3) approximations of Lemmas A.1 and A.3. Providing the exact formulas used would substantially improve reproducibility.","section":"Appendix A and Section 6.2"},{"comment":"The text says the output function (46) is \"linear in the state\" and hence the geometric ItEKF terminates on the first iteration, but the projection from SE2(3) to SE(3) is not linear in the exponential coordinates used for the error state; please clarify what is meant by linearity here.","section":"Section 6.3"},{"comment":"There are several typographical errors and minor inconsistencies, including \"Substitude\" (Eq. 28), \"consistancy\" (conclusion), \"implmentation\" (Section 6.4), and the use of both Q^+ and Q_diamond for the same quantity in Section 4.3; a careful proofreading pass is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eqs. (37) and (44) is likely a fixable algebraic slip, but it is central: the displayed formulas would diverge in the Euclidean limit, and the simulation results are presented without code. The revision should correct the update-mean formulas, verify the matrix transposes in Eq. (44), and either provide implementation details or re-run the simulations with the corrected formulas. I would not recommend rejection on novelty grounds; the coordinate-change framework is a reasonable contribution, but the technical consistency of the algorithmic core must be established first."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a serious geometric extension of concentrated-Gaussian EKF to manifolds with affine connection. The genuinely new piece is Lemma 3.2: changing the reference point of a concentrated Gaussian requires the two Jacobians J1 and J2 of the exponential map, with the KL-optimal covariance at second order. The update-step covariance correction R† = J1^{-1} J2 R J2^T J1^{-T} is also new to me, and the Jacobi-field approximations in the appendix are useful.\n\nThat said, there is a load-bearing sign error in the displayed update formulas. In the Euclidean identity-output limit, Eq. (37) gives mu+ = -K r, which moves the estimate away from the innovation; the correct posterior mean is +K r. The same negative sign appears in the iterated formula (44), which also has J2^T where the derivation requires J2^{-T}. If the implemented filter used these formulas literally, it would diverge. The simulations report good performance, so the code must use corrected formulas, but with no code or explicit description of the implemented update mean, the paper as written is internally inconsistent with its own objective (33). This is not a small asymptotic-regime quibble; it is an algebraic contradiction in the central algorithm.\n\nThe other soft spots are less severe. The propagation modification is derived and then dropped because the virtual noise-free state is unavailable; the paper is honest about this, but the abstract promises a propagation modification that the actual filter does not use. Lemma 3.2 is only second-order optimal, and the reported gains are largest in the transient where errors are large, so the practical benefit rests on an unproven extrapolation. No code or data is shipped, and the simulation section does not say exactly how J1, J2, and parallel transport are computed, which matters because the approximations are the implementable content. The ablation study is good, and the observation that update and reset corrections are synergistic is a real finding.\n\nThe geometric core is coherent and the literature is cited fairly. With corrected signs and a small reproducibility appendix, this would be a solid paper. As it stands, the displayed algorithm contradicts the simulations, so I would not cite it until the authors fix the sign and clarify what was actually implemented. I would still send it to peer review: the theory is substantial and the flaw looks corrigible.\n\nRecommendation: engage, but require a corrected update/reset derivation and code or detailed implementation before acceptance.","headline":"Strong geometric filtering theory with a central sign error in the displayed update formulas that contradicts the paper's own simulations; the flaw looks corrigible and the paper deserves peer review.","tokens_in":25264,"tokens_out":7565,"would_cite":false,"duration_ms":69296,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["53B05","53C22","62M20","93E11"],"pacs":[],"model":"deepseek-v4-flash","headline":"The classical EKF on a curved state space is missing two Jacobian corrections, and adding them reduces position and velocity error by 40-60 percent in inertial navigation simulations.","keywords":["extended Kalman filter","manifold state space","affine connection","concentrated Gaussian distribution","normal coordinates","exponential map Jacobian","covariance reset","inertial navigation"],"falsifier":"Re-run the $\\mathrm{SE}_2(3)$ Monte-Carlo benchmark with initial covariance scaled up by one or two orders of magnitude and check whether the geometric EKF keeps its ANEES advantage over the classical EKF in the first seconds; if the advantage shrinks or reverses, the filter is operating outside the regime where Lemma 3.2 is proven. A complementary test is to replace the approximate $J_1,J_2$ expressions with Jacobians obtained by numerically integrating the Jacobi equation, and ask whether the reported gains persist.","tokens_in":24324,"feed_emoji":"🧭","tokens_out":12027,"duration_ms":116739,"temperature":0.7,"pith_summary":"The paper claims that the classical extended Kalman filter, applied to a system whose state space is a curved manifold, is missing two geometric corrections: one when a measurement noise covariance is fused in coordinates anchored at the predicted state, and one when the posterior is re-expressed around the updated estimate. The central object is a concentrated Gaussian distribution written in normal coordinates, and the main technical result (Lemma 3.2) gives the covariance transformation between two normal coordinate charts in terms of the Jacobians $J_1$ and $J_2$ of the exponential map. On the extended pose group $\\mathrm{SE}_2(3)$ with direct pose measurements, the geometrically corrected EKF is more accurate and more consistent than the classical EKF and the iterated EKF, cutting position and velocity RMSE by roughly 40 percent in the transient. The ablation study shows the two corrections act synergistically: the reset modification alone degrades performance, and the full benefit appears only when both update and reset are changed.","feed_headline":"Add two Jacobian corrections and EKF pose error drops by 40%","feed_subtitle":"On curved manifolds, moving a Gaussian between coordinates changes its covariance; the corrected filter is more accurate and consistent.","key_machinery":"The machinery is the normal-coordinate system generated by geodesics of the affine connection, together with the two partial pushforwards of the exponential map: $J_2$ differentiates the exponential with respect to the tangent vector and $J_1$ differentiates it with respect to the base point. $J_2$ carries the covariance transformation in Lemma 3.2 and in the reset step, while $J_1$ and $J_2$ together convert the measurement noise covariance into the state chart in the update step. The appendix supplies computable approximations of these transcendental maps using Jacobi field theory: $J_2(\\xi_1,\\xi_2)[w] = P^{0,1}_\\gamma(w + \\tfrac{1}{6}R(v,w)v) + o(|v|^3)$ and $J_1(\\xi_1,\\xi_2)[w] = P^{0,1}_\\gamma(w - \\tfrac{1}{2}R(v,w)v) + o(|v|^3)$, where $P$ is parallel transport along the connecting geodesic and $R$ is the curvature tensor. This turns the geometric corrections into implementable linear-algebra operations.","core_discovery":"The paper's central claim is that a concentrated Gaussian information state is not invariant under a change of normal coordinate chart in the way the classical EKF implicitly assumes. Moving the reference point from $\\xi_1$ to $\\xi_2$ while keeping the same underlying distribution over the manifold changes the covariance parameter to $\\Sigma_2 = J_2^{-1}(\\xi_2,\\xi_\\diamond)\\, J_2(\\xi_1,\\xi_\\diamond)\\, \\Sigma_1\\, J_2^\\top(\\xi_1,\\xi_\\diamond)\\, J_2^{-\\top}(\\xi_2,\\xi_\\diamond)$, and Lemma 3.2 proves this is the concentrated Gaussian closest to the original in Kullback-Leibler divergence up to $\\mathcal{O}(\\mathrm{tr}(\\Sigma_1^2))$. The paper applies the same Jacobian reasoning to the update step, expressing the measurement noise covariance $R_k$ in the normal coordinates of the state estimate, and to the reset step, recentring the posterior covariance at the updated state estimate. In the $\\mathrm{SE}_2(3)$ inertial navigation benchmark, the resulting geometric EKF reduces position and velocity RMSE by roughly 40--60 percent relative to the classical EKF, improves the ANEES consistency measure by roughly halving transient overconfidence, and, per the ablation study, only delivers the full gain when update and reset corrections are combined.","pith_inferences":["An implication the paper leaves implicit is that any boxplus-style error-state filter on a Lie group or homogeneous space carries the same coordinate-dependence, so applying the $J_2$ reset and the $J_1,J_2$ measurement correction could cure known overconfidence of multiplicative and error-state filters even when the prediction model is unchanged.","The propagation step still approximates the process noise covariance by leaving $Q_{k+1}$ unchanged because the noiseless intermediate state $\\xi^\\diamond_{k+1}$ is unavailable; a testable extension is to estimate or marginalise over that state and use the full $J_2^{-1}(\\hat{\\xi}_{k+1|k},\\xi^\\diamond_{k+1}) Q_{k+1} J_2^{-\\top}(\\cdot)$ transform, which may give further gains under large process no","Since Lemma 3.2 is proved only to second order in the covariance, the theory predicts that at sufficiently large initial errors the Jacobian-only corrections should eventually stop helping; if that happens, a second-order or sigma-point variant would be the natural next step rather than a contradiction of the paper's asymptotic claim.","The coordinate-change formula is a statement about concentrated Gaussian distributions themselves, so the same $J_2$ transformation should transfer to smoothing and information-form estimators on manifolds, not only to filtering."],"forward_implications":["Any error-state or boxplus EKF that ignores the covariance change under recentring will be overconfident; the geometric corrections directly address that, so the consistency gain is a general consequence of the theory rather than a tuning artifact.","In the benchmark problem the full geometric EKF reduces transient position RMSE from $0.7226\\,\\mathrm{m}$ to $0.4168\\,\\mathrm{m}$ and velocity RMSE from $0.6271\\,\\mathrm{m/s}$ to $0.3602\\,\\mathrm{m/s}$ relative to the classical EKF, with smaller but persistent gains in the asymptotic phase.","The update-step modification alone is beneficial, but the reset-step modification alone is harmful; the two corrections work together, which the paper suggests explains why reset-only schemes in the existing literature have not been adopted by the filter community.","For the iterated filter, the geometric update correction supplies nearly all the benefit; the full geometric iterated EKF and an update-only geometric iterated EKF perform almost identically, because the iterated update already ends at the best linearisation point and the reset then does little.","The measurement correction matters even when the output function is linear, because the measurement noise is defined in coordinates around the true output rather than the predicted output; this is why the position and velocity errors, emphasised by the inhomogeneous measurement noise, improve substantially."],"supporting_citations":[{"why":"provides the affine connection, geodesic, exponential map, normal coordinates, and Jacobi field theory that the whole construction rests on.","marker":"[36]"},{"why":"introduces concentrated Gaussian distributions on Lie groups, the stochastic model generalised here to geometric manifolds.","marker":"[23]"},{"why":"supplies the robotics notion of the Jacobian of the exponential map $J_2$ and the transcendental-function issue that the appendix approximates.","marker":"[37]"},{"why":"introduces the boxplus/boxminus operators that the paper uses as its local-coordinate perturbation formalism.","marker":"[14]"},{"why":"first identified covariance reset for the multiplicative EKF, the reset-step problem the paper revisits with Jacobian corrections.","marker":"[21]"},{"why":"generalised the attitude covariance reset step that the paper's manifold reset step extends.","marker":"[33]"},{"why":"gives the full-order attitude reset solution that motivates treating the reset as a coordinate change.","marker":"[34]"},{"why":"prior work on fusing concentrated Gaussians on Lie groups via parallel transport and curvature that the present Jacobian approximation builds on.","marker":"[32]"},{"why":"provides the baseline formulation of the extended Kalman filter on a manifold that the geometric EKF is compared against.","marker":"[45]"},{"why":"defines the iterated Kalman update whose geometric version and simulation baseline the paper implements.","marker":"[44]"}],"fun_headline_variants":["Two Jacobian corrections in EKF cut pose error by 40%","Jacobian corrections make EKF geometry-aware, cutting error 40%","Curved-space EKF: two Jacobian fixes improve accuracy","Manifold EKF gains from Jacobian corrections in update and reset","Intrinsic Gaussian updates: Jacobian fixes improve EKF consistency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the truncated, second-order coordinate-change formulas of Lemma 3.2 remain accurate when errors and covariances are large, because the paper's largest performance gains occur during the initial transient rather than in the asymptotic regime where the $\\mathcal{O}(\\mathrm{tr}(\\Sigma_1^2))$ error term is small.","fun_headline_variants_meta":{"raw":{"variants":["Two Jacobian corrections in EKF cut pose error by 40%","Jacobian corrections make EKF geometry-aware, cutting error 40%","Curved-space EKF: two Jacobian fixes improve accuracy","Manifold EKF gains from Jacobian corrections in update and reset","Intrinsic Gaussian updates: Jacobian fixes improve EKF consistency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3309,"prompt_tokens":1034,"completion_tokens":2275,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":2197}},"tokens_in":650,"tokens_out":2275,"duration_ms":16157,"temperature":1.0,"reasoning_tokens":2197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:14:04.282542+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the $\\mathrm{SE}_2(3)$ Monte-Carlo benchmark with initial covariance scaled up by one or two orders of magnitude and check whether the geometric EKF keeps its ANEES advantage over the classical EKF in the first seconds; if the advantage shrinks or reverses, the filter is operating outside the regime where Lemma 3.2 is proven. A complementary test is to replace the approximate $J_1,J_2$ expressions with Jacobians obtained by numerically integrating the Jacobi equation, and ask whether the reported gains persist.","supporting_citations":[{"cited_title":"Error propagation on the euclidean group with applications to manipulator kinematics,","cited_arxiv_id":null,"evidence_quote":"introduces concentrated Gaussian distributions on Lie groups, the stochastic model generalised here to geometric manifolds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the robotics notion of the Jacobian of the exponential map $J_2$ and the transcendental-function issue that the appendix approximates."},{"cited_title":"Integrating generic sensor fusion algorithms with sound state representations through encapsulation of manifolds,","cited_arxiv_id":null,"evidence_quote":"introduces the boxplus/boxminus operators that the paper uses as its local-coordinate perturbation formalism."},{"cited_title":"Attitude error representations for kalman filtering,","cited_arxiv_id":null,"evidence_quote":"first identified covariance reset for the multiplicative EKF, the reset-step problem the paper revisits with Jacobian corrections."},{"cited_title":"Covariance correction step for kalman filtering with an attitude,","cited_arxiv_id":null,"evidence_quote":"generalised the attitude covariance reset step that the paper's manifold reset step extends."},{"cited_title":"Full-order solution to the attitude reset problem for kalman filtering of atti- tudes,","cited_arxiv_id":null,"evidence_quote":"gives the full-order attitude reset solution that motivates treating the reset as a coordinate change."},{"cited_title":"A geometric perspective on fusing gaussian distributions on lie groups,","cited_arxiv_id":null,"evidence_quote":"prior work on fusing concentrated Gaussians on Lie groups via parallel transport and curvature that the present Jacobian approximation builds on."},{"cited_title":"A note on the extended kalman filter on a manifold,","cited_arxiv_id":null,"evidence_quote":"provides the baseline formulation of the extended Kalman filter on a manifold that the geometric EKF is compared against."}],"review_version":1}