{"id":"28efd220-2fb0-4a70-9af2-e5ee1561630e","arxiv_id":"2411.15326","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new functional regression model predicts scalar responses from the phase-invariant shape of predictor functions using Fisher-Rao geometry.","lead":"This paper introduces scalar-on-shape (ScoSh) regression models that predict a scalar from the shape of a function while discarding its timing or phase. It shows how to fit these models with an index function and reports strong prediction in simulations and COVID and weather datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's centering step uses ordinary composition instead of the SRVF action, so the final estimated β is not the invariant regression mean the paper defines.","rationale":"The reader's weakest-assumption concern was that the single-index form in Eq. (4) is a strong structural restriction not tested against more flexible shape regression. That is a legitimate scope concern, but it concerns model flexibility rather than the paper's internal correctness. The more load-bearing technical issue is that the algorithm as written does not carry out the centering that the invariance theory requires. Section 2.2 item 3 defines β's equivalence class using the SRVF action β ⋆ γ0, and the phase-invariance property of the supremum term holds exactly for that action. Algorithm 1 Step 9 instead performs ordinary composition β ∘ ¯γ^{-1}, which is not the same operation. Because the fitted h and aligned q_i are tied to the uncentered β during coordinate descent, replacing β with a differently transformed function at the end can alter the fitted cost and predictions. The paper's claimed 'regression mean' is therefore not a well-defined output of the procedure as stated. This is a concrete, checkable inconsistency, and it is fixable: if Step 9 is corrected to the SRVF action, the centered β preserves the model's predictions and the regression mean becomes well-defined, assuming the average warping is computed consistently. The reader's verdict of CONDITIONAL remains appropriate, because the paper needs a correction and a verification step, but the central idea is not fundamentally invalidated; hence UNCHANGED relative to the reader's verdict.","tokens_in":16858,"tokens_out":9248,"duration_ms":93868,"concrete_test":"On the simulation behind Figure 1, rerun Algorithm 1 to convergence before Step 9, recording the fitted h and the uncentered β_unc. On the held-out test set, compare predictions from (a) β_unc, (b) β_unc ⋆ ¯γ^{-1} (the SRVF action), and (c) β_unc ∘ ¯γ^{-1} (the composition written in Step 9). If (c) changes test R2 relative to (a) while (b) leaves it identical, the published centering step is inconsistent with the fitted model and with the paper's invariance argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The model's invariance claim in Section 2.2 item 3 depends on the SRVF action: sup_γ <β ⋆ γ0, q_i ⋆ γ> = sup_γ <β, q_i ⋆ γ>. Therefore β is identified only up to its orbit under ⋆, and the paper proposes to center it by requiring the average estimated warping to be the identity. However, Algorithm 1 Step 9 says bβ = bβ ∘ ¯γ^{-1}, ordinary function composition, not bβ ⋆ ¯γ^{-1}. These two transformations differ. The SRVF action is needed to preserve the supremum scores and the fitted h; ordinary composition does not. As written, after Step 9 the reported β is not the one used in the coordinate descent or the model whose h was fitted, so the final model for prediction and the 'regression mean' are not well-defined. This is load-bearing because the paper's central contribution includes the new concepts of regression phase and regression mean. Even if this is only a typographical error, the manuscript must state the correct SRVF centering and verify that all reported estimates and bootstrap intervals use it.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a scalar-on-shape (ScoSh) regression model in which a scalar response is regressed on the shape of a functional predictor, with phase variability removed inside the regression model. The model replaces the usual L2 inner product with the phase-invariant supremum term sup_γ <β, q_i ⋆ γ>, where q_i is the square-root velocity function (SRVF) of the predictor and ⋆ is the SRVF warping action. A single-index extension (SI-ScoSh) introduces a nonlinear link h and an offset g, and the paper introduces the concepts of regression phase and regression mean, defined through the optimizing warpings and the estimated coefficient β. Estimation is performed by a coordinate-descent algorithm with dynamic programming for the warpings, and the paper reports simulation studies and several real-data applications (Spanish weather, COVID hospitalization, COVID infection, Australian rainfall). The paper also sketches a multiple-index extension and proposes bootstrap-based uncertainty evaluation.","tokens_in":17135,"tokens_out":5094,"duration_ms":50891,"significance":"If the method is technically sound, the paper makes a useful contribution to functional data analysis: it offers a principled way to regress scalars on function shapes without a separate registration step, and the phase-invariant Fisher-Rao/SRVF construction is appropriate. The paper contains substantial experimental work, including comparisons with functional linear models and single-index models, and it explicitly studies the important distinction between informative and uninformative phases. The bootstrap analysis in Section 2.4 is a good empirical check, and the real-data examples address practically relevant problems. However, the paper's central inferential claims depend on identifiability and estimation details that are not fully established, and at least one step of the proposed algorithm appears inconsistent with the model's own invariance argument.","major_comments":[{"comment":"The centering step in Algorithm 1, Step 9 uses ordinary composition, bβ = bβ ∘ γ̄^{-1}, but the invariance property used to define the equivalence class of β is the SRVF action: sup_γ <β ⋆ γ0, q_i ⋆ γ> = sup_γ <β, q_i ⋆ γ>. These two transformations differ in general, and ordinary composition does not preserve the supremum scores used in the fitted model. As written, after Step 9 the reported β is not the function whose sup scores were used to fit h and to run the coordinate descent, so the final model and the claimed 'regression mean' are not well-defined. This is load-bearing because the regression mean and regression phase are stated as central contributions. Please correct Step 9 to use the SRVF action (or state explicitly that β is an SRVF and that centering is performed by β ⋆ γ̄^{-1}), and verify that all reported estimates, bootstrap intervals, and real-data results use the corrected definition.","section":"Section 2.2, Items 3-5"},{"comment":"The identifiability constraints are imposed but not justified. The condition (1/n) Σ_i γ_i = γ_id is a nonlinear equation in β because each γ_i is defined as argmax_γ <β, q_i ⋆ γ>; no existence or uniqueness proof is given, and it is not shown that every orbit under ⋆ contains exactly one β satisfying this condition. Similarly, the normalization of h by fixing its highest coefficient is only meaningful for a polynomial parameterization, and the paper does not state how the general continuous h is identified, nor does it discuss the joint additive ambiguity between h and g except for the g(0)=0 constraint. These are not merely technicalities: Section 2.2 claims that 'with these constraints, the model is fully specified, and the parameters are well-defined.' Please either provide a rigorous identifiability argument or explicitly weaken the claim to say that the algorithm selects a representative of an equivalence class.","section":"Section 2.2, Items 3-5"},{"comment":"The estimation procedure is heuristic: Algorithm 1 steps from 3 to 7 use dynamic programming, continuous optimization, and a stopping rule based only on whether H is 'large,' with no proof that the alternating updates decrease the objective or converge to a stationary point. The bootstrap analysis in Section 2.4 is purely empirical and covers only one simulation setting; it does not establish consistency or asymptotic validity of the bootstrap. The real-data sections report point predictions (R2 values) without error bars or intervals, so it is unclear whether the reported advantages over competing models are stable. This is a load-bearing issue because the paper's practical claims rest on the estimated parameters and their uncertainty. Please add convergence diagnostics, a clearly specified stopping criterion, and uncertainty quantification for the real-data results, or explicitly state that the method is heuristic and the real-data comparisons are exploratory.","section":"Algorithms 1-2 and Section 2.4"}],"minor_comments":[{"comment":"The initialization line reads 'Compute initial bβ(t) = Σ_{j=1}^J c_j,' which omits the basis functions b_j(t); it should be corrected to include them.","section":"Algorithm 1, Step 2"},{"comment":"There is a typo: 'haevn' should be 'have' in the first sentence of Section 2.4.","section":"Section 2.4"},{"comment":"The inner product notation in the displayed objective is broken: the expression should read 〈Σ_{j=1}^J c_j b_j, q_i ⋆ γ_i〉 with closing parentheses in the correct places.","section":"Equation (5)"},{"comment":"The text states that the shape metric d_s, rather than RMSE, is used to evaluate estimation error for β, but Tables 5 and 6 label the error as RMSE and do not define the metric. Please make the table labels and captions consistent with the text.","section":"Section 3.2 and Tables 5-6"},{"comment":"The text contains the typo 'ScoSch' instead of 'ScoSh' in the sentence 'The high performance of the ScoSch model underscores its invariance.'","section":"Section 3.3"},{"comment":"The real-data sections refer to detailed tables in the Supplementary Material, but this preprint does not include the supplementary material; without those tables, the real-data results cannot be fully checked. Please include the supplementary tables or summarize the additional results in the main text.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a reproducibility statement and, ideally, code or data for the real-data analyses. The most serious issue is the inconsistency between the SRVF-based invariance argument and the ordinary-composition centering in Algorithm 1; this should be fixed before the manuscript is considered further. The identifiability claims also need to be made precise or softened. I do not see a fundamental flaw in the core modeling idea, which is why I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea is solid: model a scalar response as a function of the phase-invariant Fisher-Rao inner product between the predictor's SRVF and a coefficient function, then add a nonlinear index function. That is a genuine extension of Ahn et al., and the concepts of regression phase and regression mean are clearly motivated and nicely illustrated. The simulations are extensive, the bootstrap confidence intervals in the synthetic experiment are a real step forward, and the COVID and weather applications are relevant. This is a paper a shape-FDA person should read.\n\nThe stress-test note holds. Algorithm 1, Step 9 centers the estimated coefficient by ordinary composition, bβ = bβ ∘ ¯γ^{-1}. The model's invariance property lives in the SRVF action, β ⋆ γ, not composition. After ordinary composition, the reported β is no longer the function whose supremum scores were used to fit h and g, so the fitted model and the 'regression mean' are not well-defined. This is load-bearing, not a cosmetic typo. Even if the intended operation was ⋆, the manuscript must state it explicitly and the code must implement it; the bootstrap intervals also need to be recomputed with the correct centering.\n\nOther soft spots, in descending order of importance. Identifiability of β is handled by imposing a centering condition on the average warping, but the paper does not prove that this condition uniquely selects an orbit representative, or that the argmax warping is unique. Estimation is coordinate descent on a non-convex objective with a supremum over an infinite-dimensional group; there are no convergence guarantees, only empirical bootstrap evidence. The real-data model selection appears to choose polynomial degrees and basis sizes using test-set R2, which inflates reported performance. And the single-index structural assumption is strong; the paper shows the model can fail when phases are informative, but never compares against a more flexible shape-based nonparametric alternative.\n\nAll that said, the framework is coherent, the experiments are honest in aggregate, and the flaws are repairable. I would send this to peer review, but with a firm request to fix the centering step, prove or carefully state the identifiability conditions, and move model selection to a proper validation scheme. The current version is not ready as is, but it deserves serious referee time.","headline":"A useful phase-invariant functional regression framework, but the estimation algorithm's final centering step uses ordinary composition instead of the SRVF action, so as written the reported β is not the fitted β.","tokens_in":17614,"tokens_out":1781,"would_cite":false,"duration_ms":18232,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62R10","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"Scalar responses can be predicted from functional shape alone—invariant to time warping—via a Fisher-Rao inner product maximized over warpings; the fit defines regression phases and a regression mean.","keywords":["functional data analysis","shape regression","scalar-on-shape model","Fisher-Rao metric","square-root velocity function","phase-amplitude separation","single-index model","regression mean"],"falsifier":"Generate responses from two shape indices, $y_i = h_1(\\sup_{\\gamma} \\langle \\beta_1, q_i \\star \\gamma \\rangle) + h_2(\\sup_{\\gamma} \\langle \\beta_2, q_i \\star \\gamma \\rangle)$ with $\\beta_1, \\beta_2$ orthogonal, and fit the single-index SI-ScoSh model; if out-of-sample $R^2$ falls well below a two-index fit, the single-index structure is falsified. A quicker check already appears in the paper: data generated with informative phases give ScoSh test $R^2$ below zero, confirming the model is only sound when phases are genuine nuisances.","tokens_in":16652,"feed_emoji":"📈","tokens_out":12856,"duration_ms":100668,"temperature":0.7,"pith_summary":"Functional data such as daily COVID curves mixes two components: shape, the timing-normalized pattern of peaks and valleys, and phase, the time warping that moves those peaks left or right. This paper tries to establish that scalar responses can be predicted from shape alone by building time warping into the regression model instead of removing it beforehand. The key move is to replace the ordinary inner product with the Fisher-Rao inner product on square-root velocity functions, so the alignment term $\\sup_{\\gamma_i} \\langle \\beta, q_i \\star \\gamma_i \\rangle$ is provably invariant to replacing a predictor by any time-warped version of itself. When that works, the fitted coefficient $\\beta$ acts as a regression mean—a shape whose alignment with each predictor best explains the response—and the optimal warpings are regression phases. The method delivers high predictive accuracy on simulated and real data when phases are uninformative, and the paper is explicit that discarding informative phases hurts.","feed_headline":"Shape-only regression predicts scalars without pre-aligning curves","feed_subtitle":"Optimizing time warpings inside the fit turns the regression coefficient into a regression mean.","key_machinery":"The machinery is the phase-invariant supremum of a Fisher-Rao inner product, made computable by the square-root velocity function $q = \\text{sign}(f')\\sqrt{|f'|}$. Because the Fisher-Rao metric satisfies $\\langle q_1, q_2 \\rangle = \\langle q_1 \\star \\gamma, q_2 \\star \\gamma \\rangle$ for every warping $\\gamma$, the value $\\sup_{\\gamma \\in \\Gamma} \\langle \\beta, q_i \\star \\gamma \\rangle$ depends only on the orbit of $q_i$ under $\\Gamma$—that is, only on the shape of $f_i$. The supremum itself is the registration step: it searches over warpings to align each predictor to the current $\\beta$, and dynamic programming makes that search feasible. The index function $h$ then wraps the aligned inner product into a nonlinear predictor-response map, and the offset $g(f_i(0))$ absorbs baseline differences.","core_discovery":"The paper's central object is the single-index scalar-on-shape model $y_i = g(f_i(0)) + h(\\sup_{\\gamma_i \\in \\Gamma} \\langle \\beta, q_i \\star \\gamma_i \\rangle) + \\epsilon_i$, with $q_i$ the square-root velocity function of the predictor $f_i$ and $\\Gamma$ the group of boundary-preserving warpings. The Fisher-Rao invariance $\\langle q_1, q_2 \\rangle = \\langle q_1 \\star \\gamma, q_2 \\star \\gamma \\rangle$ makes the supremum independent of the phase of $f_i$, so the response is a function of the shape class $[f_i]$ alone. Identifiability is restored by constraining $h$ and $g$ and by centering $\\beta$ so the average of the optimal warpings is the identity. Estimation alternates dynamic-programming alignment of each $q_i$ to the current $\\beta$ with least-squares updates of $\\beta$, $h$, and $g$, and bootstrap experiments show the true parameters fall inside 95% intervals. The authors call the solution $\\beta$ the regression mean and the optimizing warpings the regression phases, and the multiple-index version fits residuals iteratively.","pith_inferences":["A natural extension the paper leaves implicit: the same supremum machinery could feed the fitted regression phases back as predictors, turning phase information into a covariate when it is informative rather than discarding it.","The single-index restriction is a strong structural assumption; comparing SI-ScoSh against a nonparametric shape regression would test whether one aligned index is sufficient or merely convenient.","The centering condition on the average of optimal warpings suggests a connection to means on the warping group; one could test estimator stability by replacing the Fourier basis with data-adaptive principal components and checking whether the shape estimate of $\\beta$ changes.","Since the supremum is computed by dynamic programming at each iteration, the method's scalability hinges on that step; an obvious test is to benchmark runtime and accuracy on long curves against a pre-registered scalar-on-function baseline."],"forward_implications":["Registration moves inside the regression: estimated warpings are not a preprocessing choice but part of the fit, so results are invariant to how the input curves were initially parameterized.","The estimated $\\beta$ can be displayed as a curve, giving practitioners an interpretable regression mean shape that shows which timing-independent features drive the response.","In applications where phase is noise, such as pandemic waves with country-specific timings, shape-only prediction should outperform raw-function regression; the paper reports test $R^2$ above 0.92 on COVID hospitalization data and 0.89 on infection data.","With a nonlinear index $h$, the model captures curvature in the shape-response relation; simpler linear versions lose accuracy as the true $h$ becomes more complex.","Iterating the single-index fit on residuals yields a multiple-index ScoSh model, improving rainfall prediction from $R^2 = 0.39$ to $0.74$, so a small number of shape indices may suffice."],"supporting_citations":[{"why":"Supplies the square-root velocity function representation, the Fisher-Rao metric invariance, and the dynamic programming alignment used to define and estimate the phase-invariant inner product.","marker":"Srivastava and Klassen [2016]"},{"why":"Earlier scalar-on-shape regression that used a supremum over warpings of a non-invariant inner product; the paper positions its Fisher-Rao version as the fix for that invariance failure.","marker":"Ahn et al. [2020]"},{"why":"Defines the classical phase-amplitude separation and amplitude mean whose regression counterparts, regression phase and regression mean, the paper introduces.","marker":"Marron et al. [2015]"},{"why":"Provides the baseline functional linear model $y_i = \\alpha + \\langle \\beta, f_i \\rangle + \\epsilon_i$ that the scalar-on-shape model generalizes.","marker":"Ramsay and Silverman [2005]"},{"why":"Introduces the single-functional-index model with a link function $h$, the template for the index function in SI-ScoSh.","marker":"Ait-Saïdi et al. [2008]"},{"why":"Provides the multiple-index functional regression formulation that the paper's multi-index ScoSh extension follows.","marker":"Ferraty et al. [2013]"},{"why":"Earlier generative phase-amplitude separation that the paper contrasts with regression-based alignment when defining regression phases.","marker":"Tucker et al. [2013]"},{"why":"Supplies shape-matching regression analysis that motivates the paper's bootstrap approach to estimator evaluation in the absence of analytic distributions.","marker":"Du et al. [2015]"}],"fun_headline_variants":["Regression learns warpings and shapes to predict scalars","Scalar-on-shape: fit warpings, not pre-aligned curves","Fisher-Rao shape regression optimizes warpings during fit","Regression phases and means: new twist on functional data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every relevant shape feature of a predictor enters the response through one scalar index, $h(\\sup_{\\gamma} \\langle \\beta, q_i \\star \\gamma \\rangle)$, so shape information orthogonal to a single aligned coefficient $\\beta$ is assumed irrelevant; the paper's own experiments show the model loses predictive power when phase carries real signal.","fun_headline_variants_meta":{"raw":{"variants":["Regression learns warpings and shapes to predict scalars","Scalar-on-shape: fit warpings, not pre-aligned curves","Fisher-Rao shape regression optimizes warpings during fit","Regression phases and means: new twist on functional data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1473,"prompt_tokens":1029,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":375}},"tokens_in":645,"tokens_out":444,"duration_ms":4822,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:25:07.167653+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate responses from two shape indices, $y_i = h_1(\\sup_{\\gamma} \\langle \\beta_1, q_i \\star \\gamma \\rangle) + h_2(\\sup_{\\gamma} \\langle \\beta_2, q_i \\star \\gamma \\rangle)$ with $\\beta_1, \\beta_2$ orthogonal, and fit the single-index SI-ScoSh model; if out-of-sample $R^2$ falls well below a two-index fit, the single-index structure is falsified. A quicker check already appears in the paper: data generated with informative phases give ScoSh test $R^2$ below zero, confirming the model is only sound when phases are genuine nuisances.","supporting_citations":[],"review_version":1}