{"id":"a61990a2-db99-49da-aa4f-7813ebe3fb86","arxiv_id":"2501.06881","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Rauch-Tung-Striebel smoother built on exact Gaussian integration of polynomial functions is shown to outperform existing Gaussian approximation smoothers on a Van der Pol oscillator.","lead":"This paper proposes a new smoother for nonlinear systems that combines Bayesian smoothing with exact Gaussian integrals for polynomial nonlinearities. It reports improved accuracy over existing smoothers on a Van der Pol oscillator test case.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 (Eq. 12) misstates the Gamma-function argument; as printed the central exact-integral formula computes wrong moments.","rationale":"The reader's weakest assumption pointed at Eq. (12) as unproved and typo-laden; our stress-test confirms this is the load-bearing concern and identifies a concrete mathematical error, not just a notation issue. The Gamma argument in Eq. (12) is missing a factor of 1/2: the correct Gaussian moment formula requires Γ((E+1)/2), where E = Σ_j a_{j,i+1}, but the paper writes Γ(E + 1/2). This is easily verified in the scalar limit: for N(0,1), E[x²] should be 1, but the printed formula gives 1.5; for E[x⁴] it gives 26.25 instead of 3. Therefore the central claim 'exact solution for polynomial integrals' is not backed by the printed mathematics. The proof is cited to [19], so the paper is not self-contained. The secondary sign error in Algorithm 1 line 10 reinforces that the pseudocode is unreliable as a specification. However, the simulation results are plausible and the corrected formula is standard, so the underlying method may be sound; this warrants a conditional acceptance pending correction of Eq. (12), a derivation or verified implementation, and code release. Our recommendation therefore does not change the reader's CONDITIONAL verdict.","tokens_in":8892,"tokens_out":25751,"duration_ms":211152,"concrete_test":"Evaluate Theorem 1's Eq. (12) for the scalar case n=1, x̂=0, P=1, with polynomials x^2 and x^4. The printed formula gives 2Γ(2.5)/√π = 1.5 and 4Γ(4.5)/√π = 26.25, instead of the true Gaussian moments 1 and 3. If the formula with Γ((Σa+1)/2) gives 1 and 3, the paper's Eq. (12) is misprinted and must be corrected or the theorem re-derived from [19].","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (12) in Theorem 1 contains a mathematical error in the Gamma-function argument. It reads Γ(Σ_j a_{j,i+1} + 1/2), but the correct argument for the Gaussian moment is (1/2)(Σ_j a_{j,i+1}) + 1/2, equivalently (Σ_j a_{j,i+1} + 1)/2. As written, the scalar case n=1, x̂=0, c=1, m=2 yields (1/√π)·2·Γ(2.5) = 1.5, while the true E[x²] = 1; for m=4 it yields 26.25 instead of 3. Because Theorem 2 and Algorithm 1 both invoke Eq. (12), the central 'exact integration' claim rests on a formula that is false as printed. The proof is only cited to [19], so the reader cannot tell whether the implementation used a corrected version. Separately, Algorithm 1 line 10 writes the Kalman mean update with a minus sign (should be plus), suggesting the pseudocode does not match the simulations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Gaussian-integral RTS smoother (GIRTSS) for discrete-time nonlinear state-space models with Gaussian noise. The key idea is to evaluate the Gaussian integrals appearing in the forward-filtering and backward-smoothing equations exactly when the nonlinear functions are polynomials, using a multivariate Gaussian integral formula from the authors' earlier work [19]. The resulting smoother is then inserted into the standard RTS smoothing recursions. The algorithm is tested on a Van der Pol oscillator model and compared with ERTSS, URTSS, and CRTSS in terms of RMSE and relative execution time.","tokens_in":9108,"tokens_out":5758,"duration_ms":52298,"significance":"If the central formulas were correct, the contribution would be a clean and potentially useful specialization: for polynomial nonlinearities, the filtering and smoothing integrals are computed exactly rather than by linearization or cubature, which should improve accuracy when the Gaussian assumption is adequate. The manuscript is honest about the restriction to polynomial nonlinearities and notes that non-polynomial functions require Taylor approximation. However, the paper does not prove its central Theorem 1 (only cites [19]), and the theorem as printed is false; moreover, the pseudocode contains a sign error in the Kalman update. The numerical validation is a single Monte Carlo study without statistical error bars or code release. With a corrected theorem, a self-contained derivation, and a corrected algorithm, the idea is defensible but incremental.","major_comments":[{"comment":"The Gamma-function argument in Eq. (12) is incorrect. For the univariate case n=1, xhat=0, c=1, f1(x)=x^2, the formula as printed returns (1/sqrt(pi))*2*Gamma(2.5)=1.5, whereas E[x^2]=1; for x^4 it returns 26.25 instead of 3. The standard zero-mean Gaussian moment requires Gamma((m+1)/2), so the printed argument should be (Σ_j a_{j,i+1}+1)/2, not Σ_j a_{j,i+1}+1/2. Because Theorem 2 (Eq. (15)) and Algorithm 1 (lines 4-8) inherit this expression, the central exact-integration claim is not supported by the manuscript as written.","section":"Section III, Theorem 1, Eq. (12)"},{"comment":"Algorithm 1 line 10 writes xhat_{k|k} = xhat_{k|k-1} - K_k(y_k - yhat_{k|k-1}); the Kalman update requires a plus sign before the Kalman gain term. As printed, the filtering recursion is incorrect. If the simulations used the correct plus sign, the pseudocode must be corrected; if the simulations used the printed minus sign, the reported RMSE values are not produced by the algorithm described.","section":"Section IV, Algorithm 1, line 10"},{"comment":"The proof of Theorem 1 is only cited to [19], and Theorem 2 follows directly from Theorem 1. Because Eq. (12) is the load-bearing exact-integration identity, the manuscript should give a self-contained derivation or at least a complete, correct statement with all indices carefully defined. The current notation is ambiguous: the role of the matrix S in the factor ∏_{j=1}^n S_{i,j}^{a_{j,i+1}} and the summation convention in the Gamma argument are not clear enough for a reader to implement Algorithm 1 without the external source.","section":"Section III, Theorem 1 and Theorem 2"}],"minor_comments":[{"comment":"The joint density in Eq. (6) is written as p(x_k, x_{k+1}|y_{1:k}) = p(x_{k+1}|x_{1:k}) p(x_k|y_{1:k}); the first factor should be p(x_{k+1}|x_k), and the conditioning on y_{1:k} should be made explicit.","section":"Section II, Eq. (6)"},{"comment":"There are several typographical issues: Eq. (11) contains 'Pxk.xk+1' with a dot instead of a proper subscript, Note 1 writes 'I = [I1 I2 ... I3]' where the last entry should be I_n, and Table I reports the URTSS RET as 1.962 in one row and 1.96 in the neighboring block.","section":"Throughout"},{"comment":"The numerical comparison is based on a single Van der Pol scenario with 1000 Monte Carlo runs but no standard errors, confidence intervals, or statistical significance tests; the average RMSE differences are small for some states, so the superiority claim would be more convincing with additional scenarios and error bars.","section":"Section IV, Tables I and Figures 1-2"},{"comment":"No code or pseudocode-verification details are provided; given the discrepancy in Algorithm 1 line 10, the authors should state explicitly whether the simulations were run with the plus or minus sign and ideally release the implementation.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is incremental over the authors' earlier work [19], applying that paper's Gaussian integral formula to RTS smoothing; the novelty is limited but could be acceptable for a signal-processing venue if the formulas are correct. The lack of a proof for the central formula and the heavy reliance on a self-citation need careful handling. I recommend requiring a corrected, self-contained statement of the integral formula, a corrected Algorithm 1, and a reproducibility statement before reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible incremental contribution — adding an RTS backward pass to the authors' earlier Gaussian-integral filter — and the Van der Pol experiment shows the expected accuracy gain. But the paper has a load-bearing typo: Eq. (12), the master formula, misstates the Gamma-function argument. As printed it computes the wrong moments; the scalar formula given two paragraphs earlier has Gamma((m+1)/2), while Eq. (12) has Gamma(sum a + 1/2). That inconsistency makes the central 'exact integration' claim false as written. The likely fix is to replace the Gamma argument with (sum a + 1)/2, and I'd bet the simulations used the corrected version, but I can't verify without the cited source [19] or code.\n\nWhat is new and good: the smoothing extension itself is new, it follows naturally from the forward filter, and the VDP example is well suited to the method (polynomial nonlinearity). The comparison against ERTSS, URTSS, and CRTSS is standard and the reported RMSE improvements are credible. The paper is short, readable, and honest about the computational cost.\n\nSoft spots beyond the Gamma error: Algorithm 1's Kalman mean update has a minus sign where a plus belongs; again likely a typo, but it undermines confidence in the pseudocode. The theorem is only cited, not proved, so a referee cannot check the implementation. The validation is a single Monte Carlo run on one system, with no error bars, no different noise levels, and no code release — fine for a first report, but not a complete empirical case.\n\nBottom line: I'd send this to review. The idea is sound, the audience (people building Gaussian-moment smoothers for polynomial dynamics) exists, and the errors are fixable. But the referee will need the corrected formula, a derivation or a verified implementation, and ideally code before the paper is publishable as-is.","headline":"A sensible smoothing extension with a misprinted core formula: as printed, the exact-integration theorem gives the wrong moments, though the fix looks trivial.","tokens_in":9576,"tokens_out":5982,"would_cite":false,"duration_ms":52570,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a Gaussian-integral RTS smoother that evaluates polynomial moments exactly and reports lower RMSE than three standard smoothers on a Van der Pol oscillator.","keywords":["Gaussian integral","Rauch-Tung-Striebel smoother","Bayesian smoothing","nonlinear state estimation","polynomial nonlinearity","Van der Pol oscillator","Gaussian approximate filtering"],"falsifier":"For a scalar Gaussian $x \\sim N(\\mu,\\sigma^2)$, evaluate Eq. (12) for $f(x)=x^2$ and $f(x)=x^3+2x$ and compare with the closed forms $\\mu^2+\\sigma^2$ and $\\mu^3+3\\mu\\sigma^2$; any mismatch invalidates the formula, and the same check can be run on the multivariate example used in the Van der Pol simulation.","tokens_in":8724,"feed_emoji":"📉","tokens_out":9509,"duration_ms":82791,"temperature":0.7,"pith_summary":"The paper proposes a Gaussian-integral Rauch-Tung-Striebel smoother (GIRTSS) for nonlinear state-space models whose nonlinearities are polynomial. Its central claim is that exact evaluation of the integrals of polynomial functions against a Gaussian density removes the moment-approximation error that extended, unscented, and cubature smoothers incur, so the smoother should be more accurate for polynomial-type dynamics. The authors implement this by using a combinatorial Gaussian-integral identity for all means, covariances, and cross-covariances in the RTS recursion, and they demonstrate it on a Van der Pol oscillator, where the dynamics are polynomial. In their 1000-run Monte Carlo test, GIRTSS reports the lowest average RMSE on all three states compared with ERTSS, URTSS, and CRTSS, at the price of higher relative computation time. The reason a reader should care is that the same recipe targets the main source of error in Gaussian-approximation smoothing rather than layering another approximation on top.","feed_headline":"Exact Gaussian integrals beat three standard smoothers on Van der Pol","feed_subtitle":"A polynomial-exact RTS smoother reports the lowest RMSE across all three Van der Pol states.","key_machinery":"The machinery is the Gaussian-integral moment identity for multivariate polynomials (Theorem 1, Eq. (12)), together with its extension to $x f(x)^\\top$-type integrals (Theorem 2, Eq. (15)). The identity diagonalizes the covariance $P = S\\,\\mathrm{diag}(c_1,\\ldots,c_n)\\,S^{-1}$, expands each monomial about $\\hat{x}$ with multinomial coefficients, and uses the fact that a one-dimensional Gaussian integral of $x^m$ vanishes for odd $m$ and equals a Gamma-function expression for even $m$. This converts every integral in the forward filter and in the backward smoother cross-covariance (8) into a finite combinatorial sum, so the smoother's only Gaussian assumption is the moment-matched posterior; no quadrature points, $\\sigma$ points, or Taylor linearization are introduced. The RTS update equations (10)-(11) themselves remain unchanged.","core_discovery":"The central claim is that every integral the RTS smoother needs, both in the forward filtering pass and in the backward cross-covariance calculation, can be written as a finite sum when the process and measurement functions are polynomials. For a polynomial $f(x)$ and a Gaussian pdf $N(x;\\hat{x}, P)$, the integral $\\int f(x)N(x;\\hat{x},P)\\,dx$ is evaluated exactly by diagonalizing $P$ into an orthogonal matrix $S$ and eigenvalues $c_i$, expanding the shifted monomials multinomially, and applying the one-dimensional odd-even moment identity, so each term is a product of Gamma functions, powers of $\\hat{x}$, and entries of $S$. Theorem 2 extends the same identity to cross-integrals $\\int x_g f_{g'}(x)N(x;\\hat{x},P)\\,dx$, which supply the cross-covariance $P_{x_k,x_{k+1}}$ in the backward step. On the Van der Pol oscillator, whose dynamics contain $x_1^2x_2$ and $x_3(1-x_1^2)x_2$ terms, the resulting GIRTSS reports lower RMSE than ERTSS, URTSS, and CRTSS at essentially every backward time step, with the same estimate as the Gaussian-integral filter at the final time. The paper's own caveat is that for non-polynomial nonlinearities the exactness is lost once the function is replaced by a truncated Taylor expansion.","pith_inferences":["If the combinatorial identity is verified and numerically stable, the same exact-moment substitution could be dropped into other Gaussian-assumed estimators, including fixed-lag smoothers and Gaussian filters, wherever the integrand is a polynomial times a Gaussian density.","A natural stress test beyond the paper's example is a higher-degree polynomial or stiff Van der Pol dynamics with a smaller sampling interval; the index-combination sum in Eq. (15) grows quickly with degree and state dimension, so the computational cost may erode the accuracy gain in that regime.","The method's reliance on an eigendecomposition of the covariance every step suggests near-degenerate covariances, where eigenvalues $c_i$ are very small or close together, could be the practical limit on the formula's numerical reliability."],"forward_implications":["For any state-space model with polynomial process and measurement functions, GIRTSS evaluates the Gaussian moments appearing in the RTS recursion exactly, so the remaining error comes only from the Gaussian approximation of the posterior, not from numerical integration.","At the final time step the GIRTSS estimate coincides with the Gaussian-integral filter's, and moving backward the smoother error decreases, so the backward pass strictly adds information rather than re-approximating it.","In the Van der Pol simulation, GIRTSS achieves lower average RMSE than ERTSS, URTSS, and CRTSS on all three states; the margin over ERTSS is the largest, and the method's relative execution time is about 6.3 times that of the EKF-based smoother.","For non-polynomial nonlinearities the paper's method still applies only after a Taylor expansion, and then the exactness property no longer holds; the accuracy advantage will depend on how well the truncated polynomial represents the function."],"supporting_citations":[{"why":"Supplies the proof of Theorem 1, the Gaussian-integral formula for polynomial moments that the smoother's exactness rests on.","marker":"[19]"},{"why":"Provides the Bayesian filtering and RTS smoothing derivation, including the joint Gaussian lemma and the backward recursion the algorithm implements.","marker":"[7]"},{"why":"Supplies the forward-filtering integral equations that GIRTSS evaluates exactly and defines the cubature filter used as comparison context.","marker":"[13]"},{"why":"Defines the unscented RTS smoother, one of the three baselines whose RMSE is compared with GIRTSS.","marker":"[12]"},{"why":"Defines the cubature RTS smoother, a baseline that GIRTSS reports as less accurate on the Van der Pol test.","marker":"[16]"},{"why":"Defines the extended RTS smoother, the Taylor-approximation baseline with the highest reported RMSE in the comparison.","marker":"[15]"},{"why":"Supplies the Van der Pol forcing amplitude and frequency used in the simulation, the data on which the RMSE comparison is made.","marker":"[22]"}],"fun_headline_variants":["Exact Gaussian integrals beat three standard smoothers on Van der Pol","Polynomial-exact RTS smoother wins on Van der Pol over three rivals","Gaussian integrals make smoothing exact for polynomial systems","Odd-even moment identity powers exact RTS smoothing for polynomials","New smoother: exact for polynomial nonlinearities, best on Van der Pol"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the combinatorial Gaussian-integral formula of Theorem 1, whose proof is cited to earlier work rather than given here and whose transcription into Theorem 2 contains notation that is hard to verify as printed, correctly computes polynomial moments under a Gaussian density; the smoother's exactness claim collapses if that identity is wrong.","fun_headline_variants_meta":{"raw":{"variants":["Exact Gaussian integrals beat three standard smoothers on Van der Pol","Polynomial-exact RTS smoother wins on Van der Pol over three rivals","Gaussian integrals make smoothing exact for polynomial systems","Odd-even moment identity powers exact RTS smoothing for polynomials","New smoother: exact for polynomial nonlinearities, best on Van der Pol"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001138,"raw_usage":{"total_tokens":4756,"prompt_tokens":1004,"completion_tokens":3752,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":3664}},"tokens_in":620,"tokens_out":3752,"duration_ms":25308,"temperature":1.0,"reasoning_tokens":3664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:51:38.626601+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a scalar Gaussian $x \\sim N(\\mu,\\sigma^2)$, evaluate Eq. (12) for $f(x)=x^2$ and $f(x)=x^3+2x$ and compare with the closed forms $\\mu^2+\\sigma^2$ and $\\mu^3+3\\mu\\sigma^2$; any mismatch invalidates the formula, and the same check can be run on the multivariate example used in the Van der Pol simulation.","supporting_citations":[{"cited_title":"A new method for nonlin ear state estimation problem,","cited_arxiv_id":null,"evidence_quote":"Supplies the proof of Theorem 1, the Gaussian-integral formula for polynomial moments that the smoother's exactness rests on."},{"cited_title":"S¨ arkk¨ a and L","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian filtering and RTS smoothing derivation, including the joint Gaussian lemma and the backward recursion the algorithm implements."},{"cited_title":"Cubature Kalman filters,","cited_arxiv_id":null,"evidence_quote":"Supplies the forward-filtering integral equations that GIRTSS evaluates exactly and defines the cubature filter used as comparison context."},{"cited_title":"Unscented Rauch–Tung–Striebel smoothe r,","cited_arxiv_id":null,"evidence_quote":"Defines the unscented RTS smoother, one of the three baselines whose RMSE is compared with GIRTSS."},{"cited_title":"Cubature Kalman smoothe rs,","cited_arxiv_id":null,"evidence_quote":"Defines the cubature RTS smoother, a baseline that GIRTSS reports as less accurate on the Van der Pol test."},{"cited_title":"Nonlinear s moothing theory,","cited_arxiv_id":null,"evidence_quote":"Defines the extended RTS smoother, the Taylor-approximation baseline with the highest reported RMSE in the comparison."},{"cited_title":"Oscillatory Kalman filtering for Duffing, Coulomb, and Van d er Pol oscillators,","cited_arxiv_id":null,"evidence_quote":"Supplies the Van der Pol forcing amplitude and frequency used in the simulation, the data on which the RMSE comparison is made."}],"review_version":1}