{"id":"d21bb439-068b-4674-a17a-5b6618b3a465","arxiv_id":"2508.10513","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"New product-of-exponentials spline algorithms on Lie groups, including a global spline that exactly reconstructs polynomial motions, derived from Poisson-equation approximations.","lead":"This paper builds smooth curves on Lie groups such as rotations and rigid-body motions by solving a Poisson equation in the Lie algebra, instead of relying on geodesics. It shows that existing product-of-exponentials splines cannot reproduce polynomial motions and introduces global splines that can, with applications to robot-arm shape reconstruction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global-spline exact-reconstruction claim is unproven: Eq. (46) uses v0 where the nonzero-initial Taylor coefficient is dexp^{-1}_{-ξ0}v0, and Algorithms 7–8 are non-runnable as printed.","rationale":"The negative result, Proposition 2, is well supported by the BCH argument and is not the issue. The load-bearing positive claim—that the global spline exactly reconstructs general kth-order curves—lacks a proof and, as printed, is contradicted by the nonzero-initial interpolation formulas. Eq. (46) is not a valid consequence of Theorem 1 because it substitutes the left-trivialized velocity v0 for the true Taylor coefficient Φ1=dexp^{-1}_{-ξ0}v0. The numerical example in §7 (Fig. 8) suggests the intended algorithm can work, so rejection is not warranted; however, the manuscript must supply a corrected derivation, define all variables in Algorithms 7–8, and prove or numerically verify the exact-reconstruction property for nonzero ξ0. This is consistent with the reader's conditional verdict, so no verdict change is needed, but the stated weakest assumption should be updated from 'series convergence for large ξ0' to 'the nonzero-initial interpolation formula itself is not derived correctly as printed.'","tokens_in":31540,"tokens_out":14520,"duration_ms":164782,"concrete_test":"Independently re-derive Eq. (46) from Theorem 1: compute Φ1=dexp^{-1}_{-ξ0}v0 and Φ2 via (23)–(24) and compare with the coefficients used in (46). Then on SO(3), take ξ0=(0.5,0,0), a=(0,0.7,0.2), c=(0.1,0,0.3), h(τ)=exp(ξ0+aτ+cτ^3), Δξ=a+c, v0=dexp^{-1}_{-ξ0}a, and v0' from the exact left-trivialized derivative. Substitute into Eq. (46) and evaluate ∥log(h(τ)^{-1} exp(ξ[3](τ)))∥ at τ=0.25,0.5,0.75. Rounding-level error would validate (46); any O(‖ξ0‖) error confirms the printed formula is wrong and the exact-reconstruction proof is missing. Re-run with Algorithm 5's d_i,e_i to verify the corrected form.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's positive headline claim—that the global spline (Def. 3, Eq. 49) can exactly reconstruct a kth-order curve—is not supported by the printed derivation. Section 6 obtains the ξ(0)≠0 formulas (45)–(48) by formally translating the ξ(0)=0 formula (16). But (16) relies on the identity v0^{(s)} = ξ^{(s+1)}(0), which holds only when ξ(0)=0 because dexp^{-1}_0=I and derivative terms of dexp^{-1} vanish. For ξ0≠0, Theorem 1(ii) and Eqs. (23)–(24) give Φ1 = dexp^{-1}_{-ξ0} v0 and Φ2 = dexp^{-1}_{-ξ0} v0' − (D dexp^{-1})(Φ1) v0, not v0 and v0'. Thus Eq. (46) cannot reproduce even a simple cubic ξ(t)=ξ0+a t+c t^3 on SO(3) when [ξ0,a]≠0. The algorithms in §7 partially correct this by introducing d_i,e_i, but the text never reconciles them with (46). Algorithms 7–8 are not reproducible: they set Δξ_i = log(h_{i-1}^{-1}h_i), the POE increment, instead of the global coordinate difference ξ_i−ξ_{i-1}, and their evaluation steps reference undefined β_{i-1}, d_i, e_i. Consequently, the claimed exact reconstruction rests on an unstated, corrected derivation; the printed equations do not establish it. The reader's smallness concern is real but secondary: the formulas already fail at arbitrarily small nonzero ξ0 if read literally.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes spline interpolation on a connected Lie group G via the Poisson equation g' = gv. It derives kth-order two-point interpolation formulas from a truncated series expansion of the local coordinates ξ, then builds piecewise product-of-exponentials (POE) splines (Definition 2, Eq. (33)) and variants that prescribe the left-trivialized vector field at interpolation points. A central negative result (Proposition 2) states that standard zero-initial POE splines can exactly reconstruct only kth-order curves lying on a 1-parameter subgroup. To overcome this, the paper introduces 'global' splines with nonzero initial coordinate (Definition 3, Eq. (49)), based on a claimed Taylor expansion for nonzero ξ(0) (Theorem 1, Eq. (41)) and interpolation formulas (45)-(48). Numerical examples concern SO(3), SE(3), and Cosserat rod shape reconstruction.","tokens_in":31949,"tokens_out":6061,"duration_ms":66592,"significance":"If the global-spline construction were correct, this would be a useful contribution: a Poisson-equation-based derivation of Lie-group splines that does not depend on a Riemannian metric, a clear statement of the limitation of local POE splines, and velocity-prescribed interpolation algorithms with explicit SO(3)/SE(3) formulas and a realistic application. The paper does not provide machine-checked proofs or code, so the claims rest on the printed derivations. However, the global-spline part has load-bearing inconsistencies (see major comments): the nonzero-initial formulas (45)-(48) contradict Theorem 1, and Algorithms 5-8 are not reproducible as printed. Thus the headline exact-reconstruction claim is not established. The positive value of Sections 2-5 and the application examples does not compensate without repair.","major_comments":[{"comment":"The nonzero-initial interpolation formulas do not enforce the vector-field initial condition. From Eq. (5), at τ=0 one has ξ'(0) = dexp^{-1}_{-ξ0} v0, and the second derivative involves derivatives of dexp^{-1}. This is precisely what Theorem 1(ii) states via Φ1 and Φ2. Yet Eq. (46) gives ξ'(0)=v0, and Eq. (45) has the same structure. Therefore (46) cannot meet the stated kth-order initial-value conditions unless ξ0=0 or the relevant Lie brackets vanish. Consequently Eqs. (47)-(48) are not valid nonzero-initial boundary-value interpolations. Since Section 7 states that the global spline rests on these formulas, the exact-reconstruction claim in the introduction and Section 7 is unsupported by the printed equations.","section":"Section 6, Eqs. (45)-(48)"},{"comment":"The algorithms are not reproducible as printed. Algorithms 5 and 6 evaluate the segment using d_i, e_i (and f_i), which are not defined or derived in Section 6; the stated interpolation formulas (46)-(48) use v0 and v0' directly. Algorithm 5 also computes a_i and b_i that do not enter the evaluation. Algorithms 7 and 8 additionally use undefined β_{i-1}, d_i, e_i; set Δξ_i = log(h_{i-1}^{-1}h_i), which is the local POE increment rather than the global coordinate difference ξ_i - ξ_{i-1}; and Algorithm 8's evaluation combines ξ_{i-1} with Δξ_i in a way that is inconsistent at τ=1. Therefore the global spline algorithms cannot be executed from the manuscript, and Example 4's exact-reconstruction demonstration is not verifiable.","section":"Section 7, Algorithms 5-8"},{"comment":"The expansion (41) is asserted only for 'ξ(0) ≠ 0 small', and Remark 7 lifts this requirement in practice without a proof or quantitative bound. In a multi-segment global spline, ξ_{i-1} is the cumulative canonical coordinate log(h0^{-1} h_{i-1}) and can be large; on noncompact groups such as SE(3) or for points far from h0, the series domain and the dexp singularities are not controlled. No convergence radius or error estimate is supplied. This is secondary to the two issues above, but it is still load-bearing for the global spline's validity domain.","section":"Theorem 1 and Remark 7"}],"minor_comments":[{"comment":"The claim that the global spline 'can exactly reconstruct a kth-order curve in G' should be stated as a result of a corrected derivation; as written it is not proven by the current equations.","section":"Abstract / Introduction"},{"comment":"In Eq. (28), the expression for v0' appears algebraically inconsistent with Eq. (25); please check the sign and coefficient of the bracketed term. The final formulas (27) and (29) may be correct, but the intermediate equation should be verified.","section":"Eq. (27)-(29)"},{"comment":"The evaluation step says 't ∈ [ti – ti–1]'; it should read 't ∈ [ti–1, ti]'. The same convention appears in Algorithms 3-8.","section":"Algorithm 1 and 2"},{"comment":"The update for γ*_i has unmatched parentheses: 'γ∗ i := dexp– ¯ξi ci – 2(D– ¯ξidexp (ai) bi – ...'. Please correct the bracket structure.","section":"Algorithm 2"},{"comment":"Typos: 'repsectiely' (Example 1), 'initial an terminal' (Example 5), 'pints' (Example 6 figure caption), 'rations' (Example 3), 'algotithm' (Example 6 caption), 'minium' (Remark 4). Minor but should be cleaned.","section":"Examples"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the POE-spline limitation (Section 5) is a genuine and useful result. The problem is internal consistency of Sections 6-7: the printed nonzero-initial formulas contradict Theorem 1, and the algorithms are not reproducible. This is fixable, but it is a substantial rewrite rather than a local edit. I would not reject outright because the underlying idea is sound and Sections 2-5 appear coherent; however, the exact-reconstruction claim must be supported by a correct derivation and runnable algorithms before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the two things you should know. (1) The limitation result is real: a kth-order POE spline built from identity-starting local coordinates cannot exactly reconstruct a general kth-order curve on a non-abelian Lie group; only 1-parameter subgroups are exactly recovered. That is new and worth taking seriously on its own. (2) The paper's positive solution, the global spline with xi(0) != 0, is not established by the printed derivation. Equations (45)-(48) are obtained by formally translating the xi(0)=0 formulas, but the series coefficients in Theorem 1 are Phi_1 = dexp^{-1}_{-xi0} v0 and Phi_2 = dexp^{-1}_{-xi0} v0' - (D dexp^{-1})(Phi_1) v0, not v0 and v0'. So the exact-reconstruction claim rests on unstated corrections. The algorithms in Section 7 (d_i, e_i) point in the right direction but are never reconciled with the formulas, and Algorithms 7 and 8 are not runnable as printed: undefined variables, and Delta xi_i is set to the local increment log(h_{i-1}^{-1} h_i) rather than the global difference xi_i - xi_{i-1}.\n\nWhat is genuinely good: the derivation of the 3rd- and 4th-order 2-point interpolations from a Poisson-equation/Magnus-type expansion is clean and reproduces known minimum-energy curves as a by-product; the velocity-prescribed splines (Algorithms 3-4) are a useful addition for trajectory planning; and the numerical experiments on SO(3) and on Cosserat rod reconstruction are extensive and demonstrate the practical motivation. Proposition 2 is a solid contribution even standing alone.\n\nThe other soft spots: the smallness assumption in Theorem 1 (xi(0) small) is lifted in Remark 7 without any bound or proof; if the series diverges for large initial coordinates, the global spline loses its foundation. Also, equation (28) has a typo—it shows v0' on both sides, presumably should be v0''. These are fixable but need addressing. I concur with the reader's conditional verdict; the stress-test concern about Eq. (46) is real and actually more basic than the smallness issue.\n\nWho this is for: researchers working on interpolation in Lie groups, robot motion planning, multibody dynamics, and geometric rod models. They will get a useful perspective and some algorithms, but they must not copy the printed formulas in Sections 6-7 without checking the corrections. Recommendation: send to peer review, expecting major revision before acceptance. The core idea is sound and the limitation result is original, but the central new claim needs a rigorous re-derivation and runnable algorithms.","headline":"Solid limitation result and a promising global-spline idea, but the printed nonzero-initial derivation is not self-consistent and the new algorithms are not reproducible as written.","tokens_in":32398,"tokens_out":4135,"would_cite":true,"duration_ms":41622,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D07","65D05","22E15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A kth-order product-of-exponentials spline on a Lie group can exactly reconstruct only one-parameter-subgroup motions; the paper's global spline with non-zero initial coordinates handles the rest.","keywords":["Lie group splines","product of exponentials","Poisson equation","SO(3)","SE(3)","rigid body motion interpolation","Cosserat rods","interpolation on manifolds"],"falsifier":"Two checks settle the claims. (1) For Proposition 2: interpolate a cubic, non-1-parameter-subgroup motion in $SO(3)$, e.g. $\\xi(t)=t\\xi_1+t^3\\xi_2$ (Example 3), with a 3rd-order POE spline; the residual should not vanish. (2) For the global spline: space samples so the cumulative norm $\\|\\xi(t_{i-1})\\|$ grows large (multiple full turns), compare against machine-precision Poisson-equation integration, and check that the 3rd/4th-order error still obeys Lemma 1's order law. If it breaks away, the 'small $\\xi(0)$' lifting in Remark 7 is the boundary.","tokens_in":31429,"feed_emoji":"🔄","tokens_out":23984,"duration_ms":207317,"temperature":0.7,"pith_summary":"Splines on Lie groups—the curved spaces that encode rotations and rigid-body motions—are usually built by chaining short exponential segments, each starting at the identity of the group. This paper establishes that such product-of-exponentials splines have a structural ceiling: a spline of order $k$ can exactly reproduce a $k$th-order motion only when that motion runs along a 1-parameter subgroup, because chaining segments on the group injects non-commutative corrections (nested commutators) that no polynomial coordinate function can absorb. The paper then removes the ceiling with a global spline $\\bar h(t) = \\bar h_0 \\exp \\xi(t)$ that joins the coordinate function on the Lie algebra—a vector space—instead of joining group elements, and that permits a non-zero starting coordinate $\\xi(0)$. The enabling tool is a generalized series solution of the Poisson equation (the reconstruction equation that feeds the spline), truncated to order $k$; it yields closed-form 3rd- and 4th-order algorithms, including variants that prescribe the velocity or twist at the data points. If the construction is right, it matters wherever motions must be interpolated exactly: robotics trajectory planning, animation, and the shape reconstruction of slender elastic (Cosserat) rods.","feed_headline":"Exponential splines fail exact motion reconstruction on Lie groups","feed_subtitle":"Piecewise exponentials on a Lie group only rebuild one-parameter motions; a global spline fixes that.","key_machinery":"The load-bearing mechanism is the generalized series solution of the local reconstruction equation $\\xi' = \\mathrm{dexp}^{-1}_{-\\xi}\\, v$—the Poisson equation in left-trivialized coordinates—expanded at $\\tau=0$ with $\\xi(0)$ possibly non-zero (Theorem 1, equation (41)): $\\xi(\\tau)=\\sum_{k\\ge 0}\\frac{\\tau^k}{k!}\\,\\Phi_k(\\xi(0),v(0),\\ldots,v^{(k-1)}(0))$, a generalization of the zero-initial-value expansion (7)–(10). Truncating this series at order $k$ produces the 2-point interpolation formulas that seed both spline families. The central structural object is equation (40), the Baker–Campbell–Hausdorff formula written as $\\eta(t)=\\xi(t)-\\xi(t_{i-1})-\\tfrac{1}{2}[\\xi(t_{i-1}),\\xi(t)]+\\cdots$:","core_discovery":"The paper's central claim is two-part. First (Proposition 2), a $k$th-order POE spline with segments $h_{i-1}\\exp\\xi_i^{[k]}(\\tau_i)$ and zero initial coordinates can exactly reconstruct a curve $h(t)=h_0\\exp\\xi(t)$ only when the motion runs along a 1-parameter subgroup; otherwise the Baker–Campbell–Hausdorff formula injects non-polynomial bracket terms. Second, a global spline $\\bar h(t)=\\bar h_0\\exp\\xi(t)$ that concatenates increments on the Lie algebra and allows $\\xi(0)\\neq 0$ exactly reconstructs general $k$th-order curves: Example 4 reproduces the cubic $\\xi(t)=t\\xi_1+t^3\\xi_2$ in $SO(3)$ while the same-order POE spline leaves a visible residual. The enabler is Theorem 1's series (41)","pith_inferences":["The Proposition 2 failure is structural, not an accuracy defect: the space of degree-$k$ POE splines does not contain the degree-$k$ polynomial curves, so POE splines can converge to general smooth motions only through shrinkage of bracket terms with segment length—a different regime from the exact reproduction the global spline offers.","The unproven step is Remark 7's lifting of the 'small $\\xi(0)$' condition: in a many-segment global spline, $\\xi(t_{i-1})$ accumulates and can leave the convergence domain of series (41). A natural testable extension is an explicit radius bound, or a segment-count limit, beyond which formulas (46)–(48) are only formal.","The conclusion concedes that exact polynomial reconstruction is 'a rather special situation'; the broadly useful features are the non-zero initial coordinate and the state-prescribing variants, which the rod examples exercise.","The algebra-level concatenation recipe should transfer to other interpolation problems on Lie groups or homogeneous spaces, and it suggests a route to the paper's noted open problem of computing control points so that Bézier-style curves pass through intermediate poses: the global spline's single coordinate function gives control data directly in the Lie algebra."],"forward_implications":["Closed-form 3rd- and 4th-order spline algorithms on $SO(3)$ and $SE(3)$ follow from the Poisson-series truncation with linear complexity in the number of points, including variants that prescribe the left-trivialized velocity (angular velocity or twist) at the knot points.","The 3rd-order POE spline recovers the established minimum-energy and minimum-acceleration interpolations as special cases, so the Poisson-equation derivation unifies that family of algorithms.","For Cosserat-rod shape reconstruction, prescribing the deformation/twist at the sample points markedly reduces error amplification as the number of segments grows, arguing for state-based rather than pose-only interpolation in soft-robotics applications.","The global spline treats the starting pose as an output rather than an input, which matches elastic-equilibrium problems; the uniqueness domain of the logarithm ($\\|\\xi\\|\\le\\pi$ on $SO(3)$) still bounds how far apart sampled poses may lie."],"supporting_citations":[{"why":"Supplies the kth-order series expansion and recursive coefficients (7)–(10) for zero-initial-value Poisson-equation solutions, which Theorem 1 generalizes to non-zero ξ(0).","marker":"[47]"},{"why":"The existing cubic orientation-spline algorithm built from identity-starting segments that Section 5 shows cannot reconstruct general cubic motions; the baseline the new splines must beat.","marker":"[35]"},{"why":"The smooth invariant rotation interpolation recovered as a special case of the 3rd-order POE spline; a main instance of the identity-starting construction Proposition 2 limits.","marker":"[56]"},{"why":"Source of the Bernoulli-number series (6) for the inverse of the dexp operator used to solve the local reconstruction equation.","marker":"[28]"},{"why":"Cited in the proof of Theorem 1 for convergence of the exponential-differential series underlying expansion (41).","marker":"[70]"},{"why":"Cited in the proof of Theorem 1 for convergence of the exponential solution of the linearized problem.","marker":"[44]"},{"why":"Establishes the minimum-acceleration SE(3) curve that the paper's 3rd-order boundary-value interpolation (27) recovers as a special case.","marker":"[79]"},{"why":"Prior global motion interpolation on SE(3) that the new globally parameterized splines extend to non-zero initial coordinates.","marker":"[27]"}],"fun_headline_variants":["Exponential splines only exact along 1-parameter curves","Global Lie-group spline exactly rebuilds cubic motions","POE splines miss general curves; new method hits them","Identity-based splines fail; arbitrary-start splines win"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The global-spline algorithms rest on Theorem 1's series expansion (41), which is proved only for 'ξ(0) ≠ 0 small'; Remark 7 simply asserts that this smallness requirement can be lifted in practice when cumulative coordinates grow across segments, with no proof or bound given for the lifting.","fun_headline_variants_meta":{"raw":{"variants":["Exponential splines only exact along 1-parameter curves","Global Lie-group spline exactly rebuilds cubic motions","POE splines miss general curves; new method hits them","Identity-based splines fail; arbitrary-start splines win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1535,"prompt_tokens":883,"completion_tokens":652,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":584}},"tokens_in":627,"tokens_out":652,"duration_ms":7840,"temperature":1.0,"reasoning_tokens":584,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:23:17.768193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Two checks settle the claims. (1) For Proposition 2: interpolate a cubic, non-1-parameter-subgroup motion in $SO(3)$, e.g. $\\xi(t)=t\\xi_1+t^3\\xi_2$ (Example 3), with a 3rd-order POE spline; the residual should not vanish. (2) For the global spline: space samples so the cumulative norm $\\|\\xi(t_{i-1})\\|$ grows large (multiple full turns), compare against machine-precision Poisson-equation integration, and check that the 3rd/4th-order error still obeys Lemma 1's order law. If it breaks away, the 'small $\\xi(0)$' lifting in Remark 7 is the boundary.","supporting_citations":[{"cited_title":"ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik: Applied Mathematics and Mechanics 90(6), 514–521 (2010)","cited_arxiv_id":null,"evidence_quote":"Supplies the kth-order series expansion and recursive coefficients (7)–(10) for zero-initial-value Poisson-equation solutions, which Theorem 1 generalizes to non-zero ξ(0)."},{"cited_title":"International journal for numerical methods in engineering 46(1), 45–64 (1999)","cited_arxiv_id":null,"evidence_quote":"The existing cubic orientation-spline algorithm built from identity-starting segments that Section 5 shows cannot reconstruct general cubic motions; the baseline the new splines must beat."},{"cited_title":"ACM Transactions on Graphics (TOG)16(3), 277–295 (1997)","cited_arxiv_id":null,"evidence_quote":"The smooth invariant rotation interpolation recovered as a special case of the 3rd-order POE spline; a main instance of the identity-starting construction Proposition 2 limits."},{"cited_title":"Berichte der Königlich-Sächsischen Geselschaft der Wissenschaften zu Leipzig, Mathematisch-Physische Klasse 58, 19–48 (1906)","cited_arxiv_id":null,"evidence_quote":"Source of the Bernoulli-number series (6) for the inverse of the dexp operator used to solve the local reconstruction equation."},{"cited_title":"Springer Science & Business Media (2013)","cited_arxiv_id":null,"evidence_quote":"Cited in the proof of Theorem 1 for convergence of the exponential-differential series underlying expansion (41)."},{"cited_title":"Communications on pure and applied mathematics 7(4), 649–673 (1954)","cited_arxiv_id":null,"evidence_quote":"Cited in the proof of Theorem 1 for convergence of the exponential solution of the linearized problem."},{"cited_title":"IEEE Transactions on Robotics and Automation 14(4), 576–589 (1998)","cited_arxiv_id":null,"evidence_quote":"Establishes the minimum-acceleration SE(3) curve that the paper's 3rd-order boundary-value interpolation (27) recovers as a special case."},{"cited_title":"Computer Methods in Applied Mechanics and Engineering 337, 352–386 (2018)","cited_arxiv_id":null,"evidence_quote":"Prior global motion interpolation on SE(3) that the new globally parameterized splines extend to non-zero initial coordinates."}],"review_version":1}