{"id":"84ddcd3e-450b-4819-8263-10399480fb50","arxiv_id":"2506.09358","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new regression framework models scalar responses against functional tensor covariates using low Tucker rank and spline smoothness, with a provably quadratically convergent Riemannian Gauss-Newton algorithm.","lead":"This paper introduces functional tensor regression, a method for predicting a scalar outcome from a tensor covariate that changes smoothly over time or another continuous dimension. It combines low-rank tensor structure with a smoothness penalty and a fast Riemannian Gauss-Newton algorithm, and shows it can recover time-varying brain effects in ADHD data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Initialization guarantee is missing and the recommended H_r(Z^*y) is not covered by the basin condition; without it, Eq. (25) is not established.","rationale":"The paper's headline statistical claim is the local convergence guarantee: once inside the basin, iterations converge quadratically to an O(eta) ball, and the final estimator satisfies (25). The reader identified the initialization/basin condition as the weakest assumption; I agree, and the concern is concrete: H_r(Z^*y) is not a normalized estimate, so its Frobenius norm grows with n, whereas the basin in Corollary 1 is an absolute distance that also shrinks with lambda_min. The proof in Theorem 1 therefore does not cover the tested procedure as specified. This does not make the method useless: the exact tangent-space Gauss-Newton solve in (14) can rescale and take a large first step, and the simulations suggest the subspace information in H_r(Z^*y) is sufficient. But that is a different, unproved mechanism. The unavailable supplementary proofs add to the uncertainty, though the reader already flagged this. My stress test confirms and sharpens the reader's weakest-assumption analysis: the authors should either prove an initialization guarantee for a properly normalized spectral start, add an explicit lower bound on lambda_min and the relevant design constants, or clearly qualify the theorem as applying only after a separate warm-start phase. No change to the reader's CONDITIONAL verdict is needed.","tokens_in":18566,"tokens_out":18396,"duration_ms":217021,"concrete_test":"Run Algorithm 1 in the Section 4.1 setting (n=500, p0=12, p1=p2=8, ranks (2,3,3)) from three starts: (i) Theta_0 = H_r(Z^*y) as recommended; (ii) Theta_0 = H_r(n^{-1} Z^*y); (iii) an oracle point known to lie inside the Corollary 1 basin. For each start, record ||Theta_k - Theta||_F per iteration and estimate the basin radius empirically (compute lambda_min, R_u, R_l, eta from the data). If start (i) shows no quadratic phase (log ||Delta_k|| does not roughly halve each step) or its distance to Theta exceeds the basin radius by orders of magnitude, then Eq. (25) is not justified for the recommended initialization; if start (i) still converges quadratically, the paper should prove a one-step reachability or tangent-space guarantee instead of relying on the Corollary 1 basin condition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Corollary 1 and the final bound (25) apply only when the initial iterate lies in the basin described in Remark 1, whose radius is of order min{zeta_m, D^{-3/2} R_u^{-1/2} R_l^{1/2} lambda_min}. No theorem in the paper proves this for the recommended initialization Theta_0 = H_r(Z^*y). In fact, H_r is homogeneous and Z^*y is an unnormalized sum over n subjects, so ||H_r(Z^*y)||_F generically grows with n (at least like sqrt(n), and typically like n when the signal mean is nonzero), while ||Theta||_F and the basin radius are fixed O(1) quantities; moreover the radius contains lambda_min = min_d sigma_{r_d}(M_d(Theta)), which can be arbitrarily small for a tensor of exact Tucker rank r. Thus the actual initialization used in Section 4 can be far outside the local neighborhood to which Theorem 1 applies. The empirical success may come from the exact tangent-space solve in (14) taking a large first step, but that mechanism is not analyzed. Since neither the quadratic convergence phase nor the O_pr(.) error bound in (25) follows without the basin condition, the central claim is not yet supported for the algorithm as recommended.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces functional tensor regression, in which a scalar response y is modeled as ∫_T ⟨X(t),B(t)⟩dt + ε with X(t) a functional tensor covariate and B(t) a smooth functional tensor coefficient. The coefficient is discretized on a grid through natural splines, B(t)=Θ×_0{ψ(t)^⊤(Ψ^⊤Ψ)^{-1}Ψ^⊤}, and Θ is assumed to have Tucker rank r. The estimator minimizes the penalized least squares objective (9), L(Θ)+ρJ(Θ), where J enforces smoothness along the functional mode. Estimation is performed by a functional Riemannian Gauss–Newton algorithm (Algorithm 1) that combines the exact tangent-space solve (14) with T-HOSVD retraction (16). The main theoretical results are Theorem 1 (one-step contraction), Corollary 1 (quadratic convergence to a neighborhood of the truth), Theorem 2 (the size of the statistical error η), and Theorem 3 (a minimax lower bound). The paper reports simulation evidence and an ADHD fMRI data analysis. Technical proofs are deferred to an online supplement.","tokens_in":18772,"tokens_out":7412,"duration_ms":82102,"significance":"The proposed estimator and algorithm are potentially significant: if the local convergence theory holds, this is the first functional tensor regression method with a second-order convergence guarantee, and the error bound (25) exhibits a plausible trade-off between the nonparametric smoothing error p0^{-2κ} and the tensor statistical error n^{-1/2}p0^{1/2}(Σ p_d r_d + ∏ r_d). The paper is also usefully accompanied by code and a real-data illustration. My main reservation is not about the overall strategy but about the gap between the theory and the algorithm as actually initialized: the convergence theorems require an initialization basin that the suggested warm start H_r(Z^*y) is not shown to satisfy, and a key lower isometry condition in Lemma 1 is not verified for the iterates. Because these conditions enter the contraction factor and the final bound, the central claim of provable quadratic convergence is not yet fully supported for the recommended implementation.","major_comments":[{"comment":"The initialization condition is not verified for the recommended Θ0 = H_r(Z^*y). Corollary 1 and hence the final bound (25) require ||Θ0 − Θ||_F to be of order min{ζ_m, D^{-3/2} R_u^{-1/2} R_l^{1/2} λ_min}, where λ_min = min_d σ_{r_d}(M_d(Θ)). The map H_r is homogeneous, while Z^*y = Σ_i y_i Z_i has Frobenius norm of order at least √n (and typically of order n when the signal part of y_i has nonzero mean), with Θ and λ_min fixed; moreover λ_min can be arbitrarily small for a tensor of exact Tucker rank r. Thus H_r(Z^*y) can lie far outside the basin required by Theorem 1, and neither the quadratic convergence phase nor the O_pr(·) bound in (25) is established for the algorithm as used in Section 4. The authors should either prove that this warm start lies in the basin under Assumptions 1–3, or analyze a normalized or projected initialization, or restate the guarantees as conditional on an initialization that is verifiably in the basin.","section":"Section 3, Remark 1; Corollary 1; Eq. (25)"},{"comment":"The lower isometry condition (19) is required at Υ = \\checkΘ_{k+1} − Θ, but this difference does not lie in the low-rank manifold M_r: \\checkΘ_{k+1} is in the tangent space T_{Θ_k}M_r, so its matricizations can have rank larger than r. Condition (18) is therefore not automatic, and Lemma 2 only gives a sufficient condition involving the singular values of Φ^⊤U_0 and the coefficient α for the leading r_0 columns of M_0(Υ). No argument is supplied that the iterates of Algorithm 1 satisfy (18) with a uniform constant c. Since R_l appears in the contraction factor and in the basin radius, this is a second gap in the induction used to derive Theorem 1 and Corollary 1. The paper should either prove (18) for the relevant Υ under Assumptions 1–3, or state an explicit high-level condition and verify it for the proposed algorithm.","section":"Section 3, Lemma 1 and Theorem 1"},{"comment":"The minimax lower-bound statement is not in standard form. It asserts a bound on sup_{n≤∏ p_d} n^{1/2} p0^{-1/2} sup_{(y,Z,Θ)∈M} ||\\tildeΘ−Θ||_F, with the supremum taken over varying sample sizes n. A minimax lower bound should fix n and give a lower bound on the minimax risk as a function of n (or on the risk multiplied by a known rate). With the supremum over n, the statement could be satisfied by a single favorable n and does not establish that the rate in (25) is unavoidable for every n in the range. Please restate Theorem 3 for fixed n and make explicit how it compares with the upper bound (25).","section":"Section 3, Theorem 3"}],"minor_comments":[{"comment":"The pseudocode says 'Initialize Θ0 = S0 × ...' but does not specify how S0 and U_d^0 are obtained, while Remark 1 recommends Θ0 = H_r(Z^*y). Please align the pseudocode with the practical initialization used in Section 4.","section":"Algorithm 1 and Remark 1"},{"comment":"The caption of the middle panel states 'RISE ∝ n^{-1/2}', but the theoretical relative error in (26) is O_pr(p0^{-2κ} + n^{-1/2} p0^{1/2}); the dashed line may not match the displayed setting unless p0 and κ are specified. Please clarify how the theoretical curve is computed.","section":"Section 4.1, Figure 2"},{"comment":"The acknowledgment contains a likely typo: 'Newcorner Stone science foundation' should probably read 'New Corner Stone Science Foundation'.","section":"Acknowledgment"}],"recommendation":"major_revision","confidential_remarks":"I could not verify the proofs of Theorems 1–3 because the supplementary PDF is not included in the submitted manuscript; please ensure the supplement is available to referees. The theoretical framework is largely imported from Luo & Zhang (2023), co-authored by one of the present authors, and the paper should make the boundary between imported and new results explicit. The main substantive issue is the initialization gap: the recommended H_r(Z^*y) is not covered by the basin condition, so the central convergence claim is not yet supported for the algorithm as implemented. The lower isometry verification for the iterates in Lemma 1 is a related issue that also needs attention."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper extends tensor regression to covariates with one smooth functional mode, and it does so in a way that is genuinely new. The model is natural, the roughness penalty is handled inside a Riemannian Gauss–Newton update, and the per-iteration cost is acceptable. The central claims—quadratic convergence and a near-minimax error bound—are structurally plausible because they follow the Luo & Zhang (2023) machinery, which one of the authors knows well. Proposition 1, giving the explicit normal equations for the penalty term, is real work and not a trivial tweak. The theory also separates approximation error, regularization bias, and noise, producing the rate (25) that sensibly trades off smoothing against tensor dimensionality. The simulation study is honest; the functional method beats tabular tensor regression by about 1% in RISE, a small gain that the authors explicitly acknowledge. That is acceptable for a methodological contribution.\n\nThe soft spots are real but not fatal. The largest one is initialization. Theorem 1 and Corollary 1 require the initial iterate to lie in a basin whose radius contains λ_min = min_d σ_{r_d}(M_d(Θ)), and for an exact-rank tensor this quantity can be arbitrarily small. The suggested initialization Θ_0 = H_r(Z^* y) is not analyzed, and the stress-test concern is valid: ||H_r(Z^* y)|| generically grows with n while the basin radius, as stated, does not. The paper needs to either prove the basin condition for a reasonable data-generating process or visibly qualify the theorem. Second, the proofs are deferred to a supplementary file that is not included in the arXiv submission, so the theorems cannot currently be verified. That is an editorial problem and should be fixed before publication. Third, the real-data R^2 of 0.991 on n=50 subjects is likely overfit; the cross-validation comparison with ρ=10^-2 helps, but the language describing the fit should be toned down.\n\nThis paper deserves a serious referee. The core model and algorithm are new, the rates are plausible, and the initialization gap is a revision issue rather than a desk-reject issue. I would send it to someone with expertise in both tensor optimization and functional data analysis, and would ask the authors to provide the supplementary proofs and to address the initialization condition directly.","headline":"A genuinely new functional tensor regression model with a fast algorithm and plausible rates, but the advertised guarantees rely on an initialization condition the paper never verifies.","tokens_in":19332,"tokens_out":1723,"would_cite":true,"duration_ms":17483,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62J05","62R30","15A69"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that functional tensor regression—estimating a scalar response from a tensor-valued smooth curve—can be solved by a Riemannian Gauss–Newton algorithm with quadratic convergence and near-minimax error in the tensor…","keywords":["functional tensor regression","Tucker decomposition","quadratic convergence","Riemannian Gauss–Newton","smoothing spline penalty","tensor restricted isometry property","neuroimaging","minimax estimation"],"falsifier":"Run the algorithm on a simulated low-rank functional tensor problem where the true coefficient has a very small smallest nonzero singular value in one mode, start from $H_r(Z^*y)$, and record the per-iteration contraction ratio; a ratio near $1/2$ instead of the squared-error halving predicted by Corollary 1 would show the basin condition is not satisfied by that initialization.","tokens_in":18285,"feed_emoji":"🧠","tokens_out":7382,"duration_ms":69428,"temperature":0.7,"pith_summary":"This paper proposes a regression model for a scalar response and a covariate that is a tensor-valued function of time: $y = \\int_T \\langle X(t), B(t)\\\\rangle\\, dt + \\varepsilon$. The coefficient $B$ is recovered through a discrete tensor $\\Theta$ with low Tucker rank, with a roughness penalty enforcing smooth evolution along the time mode. The paper's central claim is that the proposed functional Riemannian Gauss–Newton algorithm converges quadratically to a neighborhood of the true coefficient and that the final estimator's squared error scales with the tensor dimension rather than the full parameter count. A minimax lower bound is given to show this tensor-dimension rate cannot be improved in general. A reader should care because time-varying tensor covariates arise in neuroimaging and longitudinal studies, where simply tabular tensor regression ignores smooth temporal structure.","feed_headline":"Quadratic convergence for time-varying tensor regression","feed_subtitle":"A Riemannian Gauss-Newton algorithm learns smooth coefficient tensors with near-optimal error.","key_machinery":"The load-bearing object is the functional Riemannian Gauss–Newton iteration on the manifold of Tucker-rank-$r$ tensors, where Tucker rank means that the tensor can be written as a small core tensor multiplied by orthonormal factor matrices along each mode. At each step, the tangent space is parameterized by a core increment $C$ and mode-wise corrections $D_d$; the penalized quadratic loss is minimized exactly in those coordinates, then the resulting tangent vector is retracted to the manifold by truncated higher-order singular value decomposition. The second-order convergence argument relies on two quantitative ingredients: a functional analog of the tensor restricted isometry property bounding how the covariate map scales tensor norms, and a spline penalty matrix $A_m$ whose smallest eigenvalue is at least $O(p_0^{-1})$, so the roughness penalty is positive definite and well conditioned.","core_discovery":"The central claim is that, under a functional tensor restricted-isometry condition and a suitable start, the functional Riemannian Gauss–Newton iterates satisfy $\\|\\Theta_{k+1}-\\Theta\\|_F \\le C_1 \\|\\Theta_k-\\Theta\\|_F^2 + C_2\\eta$, so the error decreases quadratically until it reaches a ball of radius $O(\\eta)$ around the truth. The saturation term has the form $\\eta^2 = O_{\\\\mathrm{pr}}(p_0^{1-2\\\\kappa}C_m + \\rho p_0 C_m + (n\\\\rho)^{-1}p_0(\\sum_d p_d r_d + \\prod_d r_d))$, and after tuning $\\rho$ the final estimator obeys $\\|\\\\hat\\\\Theta-\\\\Theta\\|_F^2 = O_{\\\\mathrm{pr}}((p_0^{-2\\\\kappa}+n^{-1/2}p_0^{1/2})(\\sum_d p_d r_d + \\prod_d r_d))$, where $O_{\\\\mathrm{pr}}$ means the bound holds with probability tending to one. A minimax lower bound of the same order in the tensor dimension shows this part of the rate cannot be improved.","pith_inferences":["The theory leaves open whether the recommended warm start $H_r(Z^*y)$ actually lies in the basin; one testable extension is to prove a high-probability basin-entry guarantee for a spectral initialization, or to add a safeguarded initial phase.","Because the functional mode enters only through a fixed spline basis and quasi-uniform grid, the same proof structure should extend to irregular grids or to reproducing-kernel penalties, with the constants reworked.","The minimax lower bound is stated for tensor dimensionality only, so a full minimax rate over a Sobolev class of $B(t)$ is still open; the paper's rate suggests such a bound would mix $p_0^{-2\\kappa}$ with the tensor-dimension term.","One can check the basin empirically by comparing runs from random starts with runs from $H_r(Z^*y)$: if quadratic convergence is rare from random starts, the local nature of the guarantee is the practical bottleneck."],"forward_implications":["Each iteration at most doubles the number of correct digits once the iterate is in the basin, so the algorithm reaches the statistical error floor in $O(\\log\\log(1/\\eta))$ steps.","The squared estimation error is proportional to $\\sum_d p_d r_d + \\prod_d r_d$, the effective number of parameters in a Tucker decomposition, not to the ambient tensor size.","Penalizing the functional mode reduces the nonparametric approximation error $p_0^{-2\\kappa}$ and controls the noise term, while the minimax lower bound shows the tensor-dimensional part of the rate is optimal.","The plug-in estimate of $B(t)$ achieves integrated squared error of order $p_0^{-2\\kappa} + n^{-1/2}p_0^{1/2}$ relative to the norm of $B$.","The roughness penalty gives numerical gains over plain tabular tensor regression in simulations and makes the ADHD neuroimaging analysis computationally feasible with only $n=50$ subjects."],"supporting_citations":[{"why":"Supplies the Riemannian Gauss–Newton template, the tangent-space projection, and the quadratic-convergence analysis that this paper adapts to the functional setting.","marker":"Luo & Zhang (2023)"},{"why":"Provides the smoothing-spline representation of the functional coefficient and the roughness-penalty construction used to estimate $B(t)$.","marker":"Crambes et al. (2009)"},{"why":"Gives the tangent-space characterization of the Tucker-rank manifold that parameterizes each Gauss–Newton step.","marker":"Koch & Lubich (2010)"},{"why":"Provides the truncated HOSVD quasi-projection property used for the retraction step and for controlling the approximation error.","marker":"Hackbusch (2019)"},{"why":"Defines the tabular tensor-regression baseline for neuroimaging data that the functional model extends and outperforms in the ADHD analysis.","marker":"Zhou et al. (2013)"},{"why":"Supplies the functional linear-regression rate framework that motivates the smoothness conditions and the nonparametric part of the error bound.","marker":"Hall & Horowitz (2007)"},{"why":"Supports the sample-size discussion for the restricted-isometry-type condition on low-rank tensor covariates.","marker":"Rauhut et al. (2017)"},{"why":"Defines the higher-order singular value decomposition and its truncated version used in the retraction and initialization.","marker":"De Lathauwer et al. (2000)"},{"why":"Establishes the eigenvalue behavior of the natural-spline penalty matrix that yields the positive-definiteness bound $A_m \\ge c_m p_0^{-1}$.","marker":"Utreras (1983)"}],"fun_headline_variants":["Quadratic convergence for functional tensor regression","Riemannian Gauss-Newton learns tensors at quadratic rate","Smooth tensor regression with near-optimal bounds","Tensor regression reaches quadratic error decay","Functional tensors: quadratic convergence proven"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quadratic convergence and the stated error bound hold only when the initial coefficient tensor already lies in a small neighborhood of the truth, and the paper does not prove that its recommended warm start is inside that neighborhood.","fun_headline_variants_meta":{"raw":{"variants":["Quadratic convergence for functional tensor regression","Riemannian Gauss-Newton learns tensors at quadratic rate","Smooth tensor regression with near-optimal bounds","Tensor regression reaches quadratic error decay","Functional tensors: quadratic convergence proven"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000472,"raw_usage":{"total_tokens":2312,"prompt_tokens":878,"completion_tokens":1434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":1367}},"tokens_in":494,"tokens_out":1434,"duration_ms":12214,"temperature":1.0,"reasoning_tokens":1367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:50:52.384817+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a simulated low-rank functional tensor problem where the true coefficient has a very small smallest nonzero singular value in one mode, start from $H_r(Z^*y)$, and record the per-iteration contraction ratio; a ratio near $1/2$ instead of the squared-error halving predicted by Corollary 1 would show the basin condition is not satisfied by that initialization.","supporting_citations":[],"review_version":1}