{"id":"104a3efd-39f3-4d18-a53c-02c353c6532f","arxiv_id":"1909.02426","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proposes EFRM, a scalar-on-function regression that optimizes a norm-preserving time warping of predictors inside the model and shows improved prediction RMSE on simulated and real datasets.","lead":"This paper introduces a regression model for functional data that aligns predictor curves inside the model, using a norm-preserving time warping, so predictions are robust to timing differences. The method reports lower prediction error than standard functional regression on gait, NMR, and stock price datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-data superiority of EFRM is not established: Table 4 reports the best of three h degrees selected on the same test split, so EFRM's advantage over NP-shape/NP-L2 may be selection bias.","rationale":"The reader's conditional verdict already flags the post hoc h selection in the rationale, so the current recommendation need not change. I depart from the reader's stated weakest assumption: the norm-preserving phase model is a modeling-choice concern, but the test-set selection of h is more directly load-bearing for the paper's headline prediction claim because it undermines the only real-data evidence of superiority. A single nested-CV check can settle whether the advantage survives an honest evaluation protocol. The paper does have independent support: the MATLAB code is publicly available, the simulations are internally consistent, and the invariance property within the norm-preserving model is correctly stated. Those strengths do not overcome the evaluation weakness in Table 4, but they justify a conditional rather than a reject verdict.","tokens_in":15692,"tokens_out":10684,"duration_ms":126400,"concrete_test":"Re-run the three real-data analyses with a nested cross-validation: for each training split, select h in {linear, quadratic, cubic} by 5-fold cross-validation inside the training set only, then evaluate the selected EFRM on the held-out test split; repeat over 50 random splits and report mean (sd) RMSE for EFRM, FLM, PAFLM, NP-L2, and NP-shape. If the selected-h EFRM RMSE is not lower than NP-shape/NP-L2 in all three datasets, or if the difference is within the split-to-split error, Table 4 does not support the headline claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim, that EFRM improves prediction over FLM, PAFLM, and kernel methods, rests on the real-data RMSE comparisons in Table 4. In that table, EFRM's reported value is the minimum over three fitted index functions (linear, quadratic, cubic) evaluated on the same test split: Section 3.2 reports that 1H-NMR does best with a cubic h, while Gait and Stock do best with lower-order h. Because the degree of h is selected after inspecting test-set RMSE, the EFRM row is not an honest prediction comparison: the test set has been used for model selection. The reported margins are small (Gait: 2.466 vs 2.625 for NP-shape; Stock: 9.007 vs 9.540; 1H-NMR: 4.025 vs 4.251), the real datasets are small (n=61, 31, 200), and each experiment uses a single random split with no repeated-split error bars. Selection bias plus split variability could plausibly account for the observed advantage. The model formulation itself is not invalidated, but the sole real-data support for the headline claim needs a valid evaluation protocol in which h is chosen without access to test responses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an elastic functional regression model (EFRM) for scalar-on-function regression when predictor functions contain phase variability. The model replaces the L2 inner product in functional linear regression by sup_{γ∈Γ} ⟨β, (f_i*γ)⟩, where f_i*γ = (f_i∘γ)√(γ̇) is a norm-preserving warping, and applies an unknown link function h. Estimation proceeds by alternating estimation of β (via a basis expansion and dynamic programming over γ) and h (a low-order polynomial), with an average-warping constraint for identifiability. The method is compared with FLM, pre-aligned FLM, and kernel nonparametric regression on two simulations and three real datasets (gait, 1H-NMR, stock).","tokens_in":15986,"tokens_out":8249,"duration_ms":79606,"significance":"If properly validated, the idea of folding alignment into the regression model via a norm-preserving group action is a useful contribution to functional regression: it avoids the derivative-based noise amplification of SRVF preprocessing and provides a unified objective for prediction and phase removal. The algorithm is clearly described and reproducible code is provided. However, the current empirical support is not sufficient: the real-data results are based on a single test split with model selection on the test set, baselines are not regularized for small samples, and the simulations only exercise the model's own data-generating mechanism. The model also lacks a rigorous identifiability analysis. These gaps currently prevent acceptance.","major_comments":[{"comment":"The EFRM row reports the minimum RMSE among three polynomial degrees for h, and the narrative selects the degree by looking at test-set performance (e.g., 'cubic index function does the best'). This uses the test responses for model selection, so the comparison is biased in favor of EFRM; the small margins over NP-shape (Gait 2.466 vs 2.625; Stock 9.007 vs 9.540; NMR 4.025 vs 4.251) could plausibly disappear under an honest protocol. The analysis should use nested cross-validation or repeated splits with h selected on training data only, and should report standard errors or intervals.","section":"Section 3.2, Table 4"},{"comment":"The FLM and PAFLM baselines are fit with 20 B-spline basis functions on training samples of size 21-140 with no penalization or basis truncation. This is likely to overfit and produce inflated RMSE (e.g., NMR FLM RMSE 213), making the comparison unfair. A regularized FLM (e.g., penalized least squares or functional principal components regression with tuning by CV) should be used.","section":"Section 3.2 and Tables 2-4"},{"comment":"Both simulation experiments generate predictors by applying the same norm-preserving warping used in EFRM (Eq. 5), so the simulation demonstrates that the estimator recovers the model when the model is true, but it does not test the robustness claim to value-preserving phase variability that motivates the paper in Section 1.1. The authors should add simulations where contamination follows value-preserving warpings (or other misspecified actions) and evaluate how EFRM and the baselines degrade.","section":"Section 3.1"},{"comment":"The identifiability constraint (1/n)Σγ̂_i = γ_id is introduced without proof or discussion of whether it is sufficient. Since the sup-term is invariant to pre-warping of the predictors, β is not identifiable without additional structure; the paper should either prove that the constraint identifies β up to the intended equivalence or acknowledge the limitation and examine sensitivity of predictions to the constraint.","section":"Section 2.1, item 3"}],"minor_comments":[{"comment":"The word 'diffeomoprhism' is misspelled; 'boundary preserving diffeomoprhisms' should be 'diffeomorphisms'.","section":"Section 2.1"},{"comment":"The 1H-NMR response is described as categorical with four values, but the model assumes Gaussian errors; the authors should discuss whether a linear/Gaussian model is appropriate.","section":"Section 3.2"},{"comment":"The instruction 'replace β by β∘γ̄^{-1}' is ambiguous; it should clarify whether this is the ordinary composition or the norm-preserving action (β*γ̄^{-1}).","section":"Algorithm 1, step 4"},{"comment":"The caption describes the second contamination as 'area preserving' but the displayed formula (f_i∘γ_i)√(γ̇_i) is the norm-preserving action; the caption conflates area- and norm-preserving warpings.","section":"Section 1.1, Fig. 2"},{"comment":"The text uses 'root squared error' for RSE; it should be 'root integrated squared error' or the acronym should be defined explicitly.","section":"Section 3.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of CSDA and the code availability is a strength. The central methodological idea is interesting, but the empirical evaluation needs major rework before the claims can be supported. The simulations being generated from the same model are acknowledged by the authors, which is honest, but the real-data analysis still has selection bias and unfair baselines. No concerns about citation patterns or novelty disclosure beyond the need to ensure proper credit to prior elastic FDA work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the Ahn–Tucker–Wu–Srivastava paper on elastic functional regression. Bottom line: the model is a sensible extension of elastic FDA into scalar-on-function regression, but the real-data evidence for its superiority is weaker than the tables suggest, because the polynomial degree of h is selected on the same test split. I'd send it to review, but with a request for a proper evaluation protocol.\n\nWhat's new: EFRM replaces the L2 inner product in FLM with a supremum over norm-preserving warpings, and adds a single-index link h. That gives a parametric model that is invariant to the group action f -> (f∘γ)√γ', avoiding the pinching effect that kills the naive sup under value-preserving warping. This is a real improvement over the authors' own workshop paper; the single-index h and the comparison set make it a self-contained methods paper. The exposition is clear and the code is on GitHub.\n\nWhat's done well: The invariance property is correctly stated, the estimation via iterative optimization and dynamic programming is practical, and the simulations show the method recovers β and h well when the data is generated from the same model. They are also honest that simulation success is partly a tautology.\n\nSoft spots: The real-data table (Table 4) is the main problem. EFRM is actually three models (linear, quadratic, cubic h), and the paper reports the minimum over the three per dataset as \"EFRM,\" while each competitor is a single model. That is test-set model selection, and it inflates the apparent advantage. The datasets are small (n=31, 61, 200), each uses a single random split, and there are no error bars. The margins over NP-shape are modest (e.g., 2.466 vs 2.625 on Gait). Selection bias and split variability could plausibly explain those gaps. Also, the norm-preserving warping is a strong assumption: it distorts peak heights, so \"phase\" here is not the classical one. They acknowledge this, but it limits the applicability and the interpretation of \"shape.\" The theory section is thin, but they explicitly defer to existing FDA asymptotics, which is acceptable for a methods paper.\n\nThe paper is for people doing scalar-on-function regression with misaligned curves who want a parametric alternative to pre-aligned FLM. It deserves serious peer review, but the current real-data analysis needs to be redone with repeated cross-validation, error bars, and model selection nested inside training folds. I'd ask for that before believing the claimed improvements.","headline":"A useful elastic functional regression model whose real-data advantage is likely inflated by test-set selection of the index polynomial degree.","tokens_in":16487,"tokens_out":5009,"would_cite":true,"duration_ms":47676,"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":"Aligning predictors inside the regression improves prediction.","keywords":["functional data analysis","scalar-on-function regression","elastic functional regression model","phase variability","function alignment","SRVF","single-index model","norm-preserving warping"],"falsifier":"Generate a test set from a value-preserving phase model, $f_i(t) = f_i^0(t + \\delta_i)$ with random shifts $\\delta_i$ and responses depending only on the unshifted shape, then fit EFRM and compare its prediction RMSE with PAFLM and FLM. If EFRM does not at least match the alternatives under this contamination, the phase-invariance claim holds only for norm-preserving warps, not general phase noise. A companion check is the model's own invariance: for any predictor $f$ and warp $\\gamma$, the fitted EFRM prediction should be unchanged between $f$ and $f * \\gamma$.","tokens_in":15511,"feed_emoji":"📈","tokens_out":9133,"duration_ms":82922,"temperature":0.7,"pith_summary":"The paper proposes elastic functional regression (EFRM), a scalar-on-function model in which the usual inner product $\\langle \\beta, f_i \\rangle$ is replaced by a supremum over time warpings of the predictor, wrapped in an unknown link function $h$. The warpings are norm-preserving, so the optimization avoids the pinching degeneracy of ordinary $L^2$ alignment. Because the supremum absorbs reparameterizations, the conditional mean of the response becomes invariant to this kind of predictor phase. The paper reports lower prediction RMSE than FLM, pre-aligned FLM, and kernel nonparametric baselines on simulated data and on gait, NMR, and stock-price datasets. The central point is that phase separation should be estimated jointly with the regression, using the response values, rather than as an unsupervised preprocessing step.","feed_headline":"Aligning predictors inside the regression improves prediction","feed_subtitle":"EFRM absorbs predictor time shifts with norm-preserving warps and reports lower RMSE on gait, NMR, and stock data.","key_machinery":"The central object is the norm-preserving warping action $f_i * \\gamma_i = (f_i \\circ \\gamma_i)\\sqrt{\\dot{\\gamma}_i}$ on $L^2$ functions, with $\\gamma_i$ ranging over boundary-preserving diffeomorphisms of $[0,1]$. This is the same action that the square-root velocity function representation induces on SRVFs, but applied directly to the observed functions, so it preserves the $L^2$ norm of each predictor and prevents the pinching effect that makes naive $L^2$ alignment degenerate. The model's key move is to put the supremum over this warping group inside the regression inner product, so that the response is a function of the best-aligned inner product rather than of the raw predictor. Estimation alternates between optimizing the coefficient function $\\beta$, with dynamic programming over $\\gamma_i$ for each candidate $\\beta$, and fitting a low-degree polynomial link $h$.","core_discovery":"The central claim is that a regression model of the form $y_i = h(\\sup_{\\gamma_i \\in \\Gamma} \\langle \\beta, (f_i \\circ \\gamma_i)\\sqrt{\\dot{\\gamma}_i} \\rangle) + \\epsilon_i$, with $\\Gamma$ the boundary-preserving diffeomorphisms of $[0,1]$, makes scalar responses depend only on the shape of the predictor function and not on its phase, and that fitting the model this way predicts better than regressing on unaligned or pre-aligned functions. The paper argues that the norm-preserving action $f_i * \\gamma_i = (f_i \\circ \\gamma_i)\\sqrt{\\dot{\\gamma}_i}$ inherits the good properties of the SRVF/Fisher-Rao approach while avoiding the pinching effect and without amplifying noise through derivatives. Under this model, the response mean is invariant to reparameterization of the predictor, and $\\beta$ is identifiable only up to composition with a warp; the paper resolves this by constraining the average estimated warp to be the identity. Empirically, EFRM gives the lowest root-mean-square prediction error among the models it compares on two simulated settings and three real datasets.","pith_inferences":["The same 'warp inside the loss' construction could extend to other group actions, such as area-preserving warping for density-valued predictors; the paper mentions other warpings are possible but does not develop or test them.","Because EFRM discards phase entirely, data where timing carries signal would be better served by a two-index model that includes both the aligned inner product and a phase distance; the paper notes phase may be informative but does not build such a model.","The identifiability gauge chosen here, averaging the estimated warpings to the identity, is one of several possible constraints; other gauges could produce different $\\beta$ estimates, and the paper does not compare them.","The computational cost of running dynamic programming inside every coefficient update may limit EFRM to one-dimensional, short curves; scaling to images or multivariate functional predictors is an open extension."],"forward_implications":["If EFRM is right, scalar-on-function regression can be made phase-invariant without a separate alignment step, and the alignment is informed by the response values rather than by unsupervised matching.","The norm-preserving formulation avoids the pinching degeneracy, so no ad hoc restriction of the warping set is needed to make the supremum well behaved.","Prediction for a new function only requires aligning it to the estimated $\\beta$, so phase variability in future observations is absorbed at prediction time.","On the three real datasets studied, the reported RMSE of EFRM is below FLM, PAFLM, and kernel approaches, suggesting the gain is not limited to data generated by the model itself."],"supporting_citations":[{"why":"Supplies the SRVF representation whose norm-preserving action $f * \\gamma = (f \\circ \\gamma)\\sqrt{\\dot{\\gamma}}$ is the basis of EFRM.","marker":"[16]"},{"why":"Provides the shape-analysis framework and the argument that naive $L^2$ time-warping optimization is degenerate.","marker":"[22]"},{"why":"Documents the pinching effect in $L^2$ alignment that motivates replacing the raw inner product with a supremum over a norm-preserving action.","marker":"[15]"},{"why":"Contributes the phase-amplitude separation idea that EFRM makes response-dependent by moving alignment inside the regression.","marker":"[17]"},{"why":"Supplies the alternating estimation scheme of updating $\\beta$ and then updating the link function $h$ used in Algorithm 2.","marker":"[29]"},{"why":"Defines the Nadaraya-Watson kernel estimator used as a nonparametric baseline in the comparisons.","marker":"[24]"},{"why":"Provides the gait dataset used in the Parkinson's disease application.","marker":"[32]"},{"why":"Provides the 1H-NMR metabonomic dataset used as another real-data application.","marker":"[36]"}],"fun_headline_variants":["Warp predictors inside regression for better fit","Elastic regression: absorb phase shifts","Regression that uses function shapes, not phases","EFRM: phase-aligned regression, no preprocessing","Norm-preserving warps reduce prediction error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the only phase variability worth removing is a norm-preserving diffeomorphic reparameterization, so that a predictor's shape is the same before and after warping by $(f \\circ \\gamma)\\sqrt{\\dot{\\gamma}}$; if real phase acts by value-preserving shifts $f \\circ \\gamma$ or by non-diffeomorphic changes, the warping will change peak heights and the model may not recover the intended signal.","fun_headline_variants_meta":{"raw":{"variants":["Warp predictors inside regression for better fit","Elastic regression: absorb phase shifts","Regression that uses function shapes, not phases","EFRM: phase-aligned regression, no preprocessing","Norm-preserving warps reduce prediction error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1535,"prompt_tokens":968,"completion_tokens":567,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":500}},"tokens_in":584,"tokens_out":567,"duration_ms":6190,"temperature":1.0,"reasoning_tokens":500,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:51:03.556998+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a test set from a value-preserving phase model, $f_i(t) = f_i^0(t + \\delta_i)$ with random shifts $\\delta_i$ and responses depending only on the unshifted shape, then fit EFRM and compare its prediction RMSE with PAFLM and FLM. If EFRM does not at least match the alternatives under this contamination, the phase-invariance claim holds only for norm-preserving warps, not general phase noise. A companion check is the model's own invariance: for any predictor $f$ and warp $\\gamma$, the fitted EFRM prediction should be unchanged between $f$ and $f * \\gamma$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the pinching effect in $L^2$ alignment that motivates replacing the raw inner product with a supremum over a norm-preserving action."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the phase-amplitude separation idea that EFRM makes response-dependent by moving alignment inside the regression."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the alternating estimation scheme of updating $\\beta$ and then updating the link function $h$ used in Algorithm 2."},{"cited_title":"Winning, E","cited_arxiv_id":null,"evidence_quote":"Provides the 1H-NMR metabonomic dataset used as another real-data application."}],"review_version":1}