{"id":"951beee0-dc63-4c88-9e3f-851882454c87","arxiv_id":"1908.05875","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A framework for Hermite interpolation on Riemannian manifolds that transports derivative data through the differential of the logarithm map, plus an asymptotic error estimate governed by sectional curvature.","lead":"This paper introduces a general method for Hermite interpolation on Riemannian manifolds, using only exponential and logarithm maps. It also relates interpolation errors to the manifold's curvature, and demonstrates the approach on matrices with orthonormal columns.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Hermite construction in Theorem 2.1 is sound, but the error-bound contribution is overclaimed: Theorem 3.1 gives only an asymptotic expansion with unquantified remainders, not the 'general error bounds' promised in the abstract.","rationale":"The reader's weakest_assumption already mentions the δ, δ_tilde < 1 hypothesis in Theorem 3.1, so there is partial overlap. My analysis sharpens that point: the real issue is not merely smallness of δ but the absence of explicit, curvature-dependent control on the remainder terms, which means Theorem 3.1 is an asymptotic estimate rather than a bound. The Hermite interpolation theorem itself appears mathematically correct, and the numerical experiments support it as a practical method. The overstatement in the abstract and conclusions about 'general error bounds' is real but does not invalidate the central Hermite construction; it makes the paper conditional on a corrected and properly quantified statement of Theorem 3.1. The reader's conditional verdict therefore remains appropriate, and no stronger adjustment is needed.","tokens_in":22268,"tokens_out":15694,"duration_ms":165233,"concrete_test":"Specialize Theorem 3.1 to the sphere of radius R with constant curvature K = 1/R^2 and derive the exact distance d(s0,δ) = R arccos(cos^2(δ/R) + sin^2(δ/R) cos s0) between Exp_q(δw) and Exp_q(δw(s0)). Fix R so that Kδ^2/6 is not small, e.g., R = 1 and δ = 2.5, still below the injectivity radius π, and compute the remainder d(s0,δ) − δs0(1 − Kδ^2/6) for a sequence of s0 values. Check whether the remainder is O(s0^2) with a constant that remains bounded as δ approaches the injectivity radius; if no such uniform constant can be established, Theorem 3.1 does not provide a general error bound as claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2.1 is internally consistent: with the exact choice v_hat_p = d(Log_q)_p(v_p), the inverse-function identity d(Exp_q)_{Delta_p} ∘ d(Log_q)_p = id yields the prescribed endpoint velocities, and the piecewise assembly is C^1 by construction. The load-bearing weakness is the paper's second stated contribution, the 'general error bounds' of Section 3. Theorem 3.1 concludes dist_M(p,p_tilde) ≤ |δ−δ_tilde| + s0δ(1 − K_q(σ)δ^2/6 + o(δ^2)) + O(s0^2), but this is an asymptotic expansion, not a computable bound: the o(δ^2) and O(s0^2) terms carry no explicit constants, and no curvature-dependent smallness condition is given. The hypothesis δ, δ_tilde < 1 alone does not control the size of the remainder; for a fixed manifold with large sectional curvature, the stated bound cannot be evaluated or certified, and Section 5.3 uses only the sign of the curvature, not the estimate itself. Thus the abstract's claim that the paper 'provides general error bounds' is not established; the theorem supports a qualitative curvature heuristic, not a bound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Riemannian generalization of local cubic Hermite interpolation. Given two sample points p, q and tangent vectors v_p, v_q on a Riemannian manifold, the method maps the data into the tangent space T_qM via the Riemannian logarithm, constructs the ordinary Euclidean cubic Hermite polynomial in that tangent space using the basis functions a0, b0, b1, and maps the result back to the manifold via the exponential map. The endpoint velocity at q is automatic, while the velocity at p is matched by transporting v_p through the differential d(Log_q)_p. The paper proves this construction in Theorem 2.1, discusses the practical finite-difference approximation of d(Log_q)_p, compares the computational effort with Bézier-type methods, and then derives an asymptotic relation between data-processing errors in normal coordinates and the sectional curvature of the manifold (Theorem 3.1). The numerical part specializes the framework to the compact Stiefel manifold, including formulas for differentiating the Stiefel exponential via Mathias' theorem, and presents experiments for the QR decomposition, low-rank SVD interpolation, and a Black-Scholes option-pricing application.","tokens_in":22552,"tokens_out":8887,"duration_ms":89635,"significance":"The Hermite construction is elementary, geometrically natural, and potentially useful: it applies to any Riemannian manifold for which exp and log algorithms are available, and the explicit Stiefel derivative formula in Lemma 4.1 is a concrete computational contribution. The numerical experiments are relevant and cover nontrivial applications. The paper also openly identifies its main practical constraint, namely that consecutive samples must lie close enough for the Stiefel logarithm to converge. The curvature-error relation in Theorem 3.1 is conceptually attractive and gives a clean qualitative explanation of why the manifold errors in the Stiefel experiments are slightly smaller than the tangent-space errors. However, as written the theorem is an asymptotic expansion, not a computable error bound; this is the main gap between the abstract's claims and the actual result. The central Hermite derivation is sound, and the overclaim is local and fixable, so the appropriate outcome is a major revision rather than rejection.","major_comments":[{"comment":"The abstract states that the paper 'provides general error bounds for manifold data processing methods', but Theorem 3.1 is an asymptotic estimate rather than a bound in the usual numerical-analysis sense. The right-hand side of Eq. (3.1) contains the unquantified terms o(δ^2) and O(s0^2), and the hypotheses δ, δ~<1 do not by themselves control these remainders; no explicit smallness condition involving the curvature or the injectivity radius is given. Consequently, for a fixed manifold with large sectional curvature, the displayed inequality cannot be evaluated or certified as a worst-case bound. Section 5.3 uses only the sign of the Stiefel curvature, not the quantitative content of the theorem. I recommend replacing 'general error bounds' by 'asymptotic error estimates' throughout, or alternatively adding explicit constants and a sufficient smallness condition under which the estimate is a true inequality with a computable remainder.","section":"Abstract; Theorem 3.1, Eq. (3.1)"},{"comment":"Theorem 2.1 is proved for the exact choice v_hat_p = d(Log_q)_p(v_p), but the general implementation described in Section 2.2 replaces this differential by the central finite-difference approximation in Eq. (2.11). The paper does not provide an error analysis showing how the O(h^2) truncation error and the convergence threshold of the iterative Stiefel logarithm affect the satisfaction of the Hermite conditions (1.2) or the C^1 assembly in Remark 1. The table in Section 5.1 gives empirical reconstruction errors for h = 10^-2 through 10^-7, but no theoretical statement links h to the interpolation error. Since the claimed generality is 'only exp and log algorithms are needed', the manuscript should either include a clear statement that the implemented curve satisfies the Hermite conditions only up to O(h^2) plus solver tolerance, or provide an explicit error estimate for the finite-difference step.","section":"Section 2.2, Eq. (2.11); Section 5.1"}],"minor_comments":[{"comment":"The figure numbering is inconsistent: the error curves in Section 5.2 are first introduced as Fig. 5.1 and then referenced as Fig. 5.2, and Section 5.3 again labels a plot as Fig. 5.2. Please renumber the figures and check all cross-references.","section":"Section 5.2 and Section 5.3"},{"comment":"The section title refers to 'left singular values', but the experiment interpolates the left singular vectors U(μ); please adjust the title or the wording.","section":"Section 5.4"},{"comment":"The notation σ = span(Δ, Δ~) assumes that Δ and Δ~ span a two-plane; when the two vectors are collinear, the sectional curvature K_q(σ) is not defined. Please add a convention for the collinear case or handle s0 = 0 separately.","section":"Theorem 3.1"},{"comment":"The relation ε ≥ δ‖w(0)-w(s0)‖ and the subsequent bound δs0 ≤ 2δ arcsin(ε/(2δ)) are used to pass to Eq. (3.5), but Eq. (3.5) is only an asymptotic replacement for the exact angle-dependent expression; please state explicitly that the equality there is valid up to O(ε^2).","section":"Remark 2(i)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central Hermite construction is mathematically sound, and the Stiefel-specific derivative formulas are a useful addition. The main concern is that the second advertised contribution, the 'general error bounds' of Section 3, is an asymptotic result with unquantified remainders; this should be repositioned in a revision. The finite-difference approximation of the velocity transport also deserves a clearer caveat or a small theoretical statement. With those fixes the paper would be acceptable for publication; I do not see a load-bearing error that would require rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the Hermite interpolation construction in Theorem 2.1 is mathematically sound and practically useful: it transfers derivative samples through the differential of the log map, then builds a cubic Hermite curve in a single tangent space. The cost count (one log plus two logs/two exps per subinterval) is genuinely lower than the Bezier-based alternatives. Second, the advertised 'general error bounds' are not bounds. Theorem 3.1 is an asymptotic expansion with o(delta^2) and O(s0^2) remainders and no explicit constants; you cannot evaluate or certify it for a given manifold. The abstract overstates this.\n\nWhat is new: the Log-differential transport in Theorem 2.1 and the explicit directional derivative of the Stiefel exponential in Lemma 4.1 (via the block matrix exponential trick) are real additions. The paper also works out the QR and SVD derivative bookkeeping cleanly. The numerical experiments on Stiefel show Hermite interpolation beating quasi-linear and RBF tangent-space interpolation by a wide margin, and the positive-curvature heuristic from Section 3 is consistent with the observed manifold errors being slightly smaller than tangent-space errors.\n\nSoft spots: (1) As said, Theorem 3.1 is a qualitative curvature heuristic, not a bound. That should be fixed in the abstract and conclusions. (2) The numerical section does not compare with existing Hermite manifold methods (e.g., the blended-arc approaches of Jakubiak et al. or Gousenbourger et al.); the 'lower cost' claim is argued analytically but not demonstrated experimentally. (3) No code or data is released, so the numbers are not independently reproducible. (4) The Stiefel log algorithm used is the author's own (Zimmermann 2017), but that is a legitimate computational tool and the paper does not reduce to it; the main framework is independent. (5) The non-conjugate assumption and injectivity-radius caveat are stated, and Section 5.4 honestly shows the log failing for distant samples.\n\nVerdict: this is a solid, useful paper for anyone doing manifold-valued interpolation with derivative data, especially on Stiefel/orthogonal-matrix problems. It deserves peer review; it needs a revised abstract and ideally a direct comparison with prior Hermite methods. I would send it to a knowledgeable referee.","headline":"Sound Hermite construction, but the 'error bounds' are asymptotic heuristics and the numerical comparison skips existing Hermite methods.","tokens_in":23035,"tokens_out":2120,"would_cite":true,"duration_ms":20578,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A16","15B10","33B30","33F05","53-04","65F60"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single tangent-space curve solves Hermite interpolation on any Riemannian manifold carrying exponential and logarithm maps.","keywords":["Hermite interpolation","Riemannian manifolds","Stiefel manifold","Riemannian logarithm","sectional curvature","matrix manifolds","SVD interpolation","QR decomposition"],"falsifier":"On a sphere, choose two tangent vectors $\\Delta,\\tilde\\Delta$ at a point $q$ with small angle $s_0$ and norms $\\delta,\\tilde\\delta<1$, compute the endpoints via the exponential map, and measure their geodesic distance; the inequality (3.1) predicts a specific shortening relative to the flat tangent-space distance, so a measured lengthening beyond the $O(s_0^2)$ term would refute the curvature relation.","tokens_in":1802,"feed_emoji":"📐","tokens_out":2114,"duration_ms":101302,"temperature":0.7,"pith_summary":"This paper claims that Hermite interpolation of manifold-valued data need not be approached as a variational or control-point problem: if the manifold has computable Riemannian exponential and logarithm maps, a single tangent-space formula built from the classical cubic Hermite polynomials yields a $C^1$ curve through sampled points with sampled velocities. The author proves this as Theorem 2.1 and gives a concrete recipe for the one nontrivial ingredient, the differential of the logarithm map, which can be approximated by central finite differences. The same normal-coordinate viewpoint yields a quantitative relation between the error made by processing data in the tangent space and the manifold's sectional curvature, stated as Theorem 3.1. Numerical tests on the Stiefel manifold, including interpolation of QR factors, SVD factors, and a parametric option-pricing model, show the method producing lower interpolation errors than geodesic and tangent-space RBF baselines, with fewer exponential and logarithm evaluations than comparable Hermite manifold methods.","feed_headline":"One formula does Hermite interpolation on any Riemannian manifold","feed_subtitle":"Works for Stiefel, QR and SVD data, needing fewer exponential and log evaluations.","key_machinery":"The load-bearing object is the Riemannian normal-coordinate chart centered at the terminal point $q$: the interpolating curve is the exponential image of a cubic Hermite curve $\\gamma(t)=a_0(t)\\Delta_p+b_0(t)\\hat v_p+b_1(t)\\hat v_q$ in $T_qM$. The only non-elementary ingredient is the push-forward of the start velocity through the logarithm transition map, $\\hat v_p=d(\\operatorname{Log}_q)_p(v_p)$, which the paper evaluates by central differences as $\\big((\\operatorname{Log}_q\\circ\\operatorname{Exp}_p)(hv_p)-(\\operatorname{Log}_q\\circ\\operatorname{Exp}_p)(-hv_p)\\big)/(2h)$; for the Stiefel manifold the paper derives an explicit formula for the derivative of the exponential via a block-matrix identity for the derivative of the matrix exponential. The curvature error bound is carried by the Jacobi-field expansion of $d(\\operatorname{Exp}_q)_{\\delta w}(\\delta w^\\perp)$, which is exactly the mechanism that turns tangent-space errors into manifold distances.","core_discovery":"The central discovery is that a curve of the form $$c(t)=\\operatorname{Exp}_q\\big(a_0(t)\\Delta_p+b_0(t)\\hat v_p+b_1(t)\\hat v_q\\big)$$ with $\\Delta_p=\\operatorname{Log}_q(p)$, $\\hat v_p=d(\\operatorname{Log}_q)_p(v_p)$, and $\\hat v_q=v_q$ solves the Hermite interpolation problem $c(t_0)=p$, $c(t_1)=q$, $\\dot c(t_0)=v_p$, $\\dot c(t_1)=v_q$ on any Riemannian manifold, provided $p$ and $q$ are not conjugate and the logarithm is defined. The proof works by expressing the whole curve in normal coordinates centered at the endpoint $q$, where the endpoint condition is automatic and the start condition becomes a single application of the differential of the logarithm. A second result relates the manifold distance between two exponential images of nearby tangent vectors to their tangent-space distance, with a first-order correction $\\delta s_0\\big(1-\\tfrac{K_q(\\sigma)}6\\delta^2+o(\\delta^2)\\big)$ determined by the sectional curvature $K_q(\\sigma)$: positive curvature shortens the propagated error, negative curvature amplifies it.","pith_inferences":["Because the local curve is anchored in normal coordinates at the right endpoint, swapping the endpoints changes the interpolant; an adaptive choice of anchor, for instance the endpoint with the larger remaining step, could reduce the chance that the logarithm fails, an effect the paper only observes numerically.","The curvature bound suggests an operational rule not stated by the author: on manifolds of positive sectional curvature, small tangent-space processing errors shrink on the manifold, so cheaper approximations such as retractions in place of exponentials may be safe; this could be tested by repeating the Stiefel experiments with retraction-based derivative transport.","The same transition-map derivative can in principle be computed to any order in the step size $h$ by higher-order finite differences or by differentiating the logarithm algorithm itself, which would remove the $O(h^2)$ error in the endpoint velocity without changing the interpolation formula."],"forward_implications":["The method transfers to any Riemannian manifold with computable exponential and logarithm maps, including manifolds that carry no Lie-group, homogeneous-space, or symmetric-space structure.","Composite interpolation over many sample points is globally $C^1$ by construction, because each local arc ends with exactly the velocity that the next arc starts with.","Constructing the full curve costs $3k$ logarithm and $2k$ exponential evaluations for $k$ intervals, and each evaluation of the final curve costs one exponential evaluation, which is below the cost of the composite Bezier alternatives the paper compares against.","The same normal-coordinate machinery derives explicit manifold error bounds for any data-processing scheme that maps data to a tangent space, processes it there, and maps back to the manifold.","On the Stiefel manifold, the method interpolates the factors of QR and SVD decompositions with smaller relative Frobenius errors than geodesic or tangent-space RBF interpolation in the reported experiments."],"supporting_citations":[{"why":"Supplies the iterative matrix-algebraic Stiefel logarithm used in the numerical experiments.","marker":"[37]"},{"why":"Provides the canonical metric and the sectional-curvature range for the Stiefel manifold.","marker":"[30]"},{"why":"Provides the Jacobi-field estimate of geodesic spreading used to prove Theorem 3.1.","marker":"[13]"},{"why":"Provides the differential-geometric facts about the exponential map and conjugate points used in Theorem 2.1.","marker":"[24]"},{"why":"Supplies the classical cubic Hermite basis and its error bound, which the manifold construction mirrors.","marker":"[20]"},{"why":"Defines the blended cubic-spline Hermite method whose computational effort is compared against the proposed method.","marker":"[16]"},{"why":"Supplies retraction alternatives and the basic Stiefel tangent-space decomposition used in the derivative-transport discussion.","marker":"[2]"},{"why":"Collects exponential and logarithm formulas for matrix manifolds and defines the quasi-linear geodesic interpolation baseline.","marker":"[38]"}],"fun_headline_variants":["One Hermite curve fits any Riemannian manifold","Curvature sets error bounds for manifold data processing","Universal Hermite formula, curvature-tied errors","One exponential-log formula solves Hermite on all Riemannian manifolds","Hermite interpolation made universal, errors tied to curvature"],"cache_read_input_tokens":25216,"weakest_assumption_plain":"The whole construction depends on consecutive sample points being close enough on the curved space that the map pulling one point back to the other's tangent space exists and that a small-step numerical derivative of that map is accurate.","fun_headline_variants_meta":{"raw":{"variants":["One Hermite curve fits any Riemannian manifold","Curvature sets error bounds for manifold data processing","Universal Hermite formula, curvature-tied errors","One exponential-log formula solves Hermite on all Riemannian manifolds","Hermite interpolation made universal, errors tied to curvature"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001474,"raw_usage":{"total_tokens":5927,"prompt_tokens":950,"completion_tokens":4977,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":4900}},"tokens_in":566,"tokens_out":4977,"duration_ms":31280,"temperature":1.0,"reasoning_tokens":4900,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:01:59.243282+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a sphere, choose two tangent vectors $\\Delta,\\tilde\\Delta$ at a point $q$ with small angle $s_0$ and norms $\\delta,\\tilde\\delta<1$, compute the endpoints via the exponential map, and measure their geodesic distance; the inequality (3.1) predicts a specific shortening relative to the flat tangent-space distance, so a measured lengthening beyond the $O(s_0^2)$ term would refute the curvature relation.","supporting_citations":[{"cited_title":"Rentmeesters","cited_arxiv_id":null,"evidence_quote":"Provides the canonical metric and the sectional-curvature range for the Stiefel manifold."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Jacobi-field estimate of geodesic spreading used to prove Theorem 3.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the differential-geometric facts about the exponential map and conjugate points used in Theorem 2.1."},{"cited_title":"Hohmann and P","cited_arxiv_id":null,"evidence_quote":"Supplies the classical cubic Hermite basis and its error bound, which the manifold construction mirrors."},{"cited_title":"Gousenbourger, E","cited_arxiv_id":null,"evidence_quote":"Defines the blended cubic-spline Hermite method whose computational effort is compared against the proposed method."},{"cited_title":"Absil, R","cited_arxiv_id":null,"evidence_quote":"Supplies retraction alternatives and the basic Stiefel tangent-space decomposition used in the derivative-transport discussion."}],"review_version":1}