{"id":"e7c72652-cef7-42d3-88a0-ba23fb3a4aa6","arxiv_id":"1908.08760","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"M-type penalized spline estimators for functional linear regression achieve the same prediction-error rates as least-squares penalized splines while resisting outliers.","lead":"This paper designs a robust version of penalized spline regression for predicting a scalar outcome from a functional predictor, using M-estimators to limit the influence of outliers. It proves that the robust estimator can match least-squares convergence rates and demonstrates good performance in simulations and on archaeological glass spectra.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rate theorem assumes C^2 ψ, but the implemented default is the piecewise-linear Huber loss; no approximation argument is given, so the central rate guarantee does not currently cover the estimator used in all simulations.","rationale":"The paper's central theoretical result is a rate bound for penalized M-estimators whose score function ψ is twice continuously differentiable with bounded second derivative. The numerical default is the Huber loss, which is the most prominently discussed robust loss and is used with c=1.345 in Section 4.1. Because Huber's ψ is only piecewise linear, Assumption 4 fails, and the paper provides no smoothing or approximation theorem to close the gap. This makes the headline claim about the proposed estimator's rates broader than what the proof supports. I do not see this as an internal contradiction: the theorem can be read as applying to smooth losses such as bisquare and exponential squared, and the gap is plausibly repairable by a more delicate argument or by using a smoothed Huber loss in practice. It does, however, justify a conditional rather than unconditional acceptance, exactly as the reader concluded. I therefore keep the reader's verdict unchanged.","tokens_in":15072,"tokens_out":12291,"duration_ms":133738,"concrete_test":"Independently re-derive the proof of Theorem 1 for the Huber ψ instead of a C^2 ψ. The decisive step is to bound R_n(δ) = n^{-1} Σ_i [ψ_H(ε_i - z_i^T δ) - ψ_H(ε_i) + ψ_H'(ε_i) z_i^T δ] z_i^T δ on the event used in the supplement, using only the bounded variation of ψ_H'. If |R_n(δ)| ≤ C ||δ||^2 + o_P(K/n + K^{-2p} + λ) holds, add a lemma covering the default loss; if an extra factor such as K^{1/2} appears, the rate statement needs amendment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 4 requires ψ=ρ' to be in C^2(R) with sup|ψ''|<∞, and the proof outline in Section 3 uses this regularity to control the Taylor remainder of the estimating equations. The implementation in Section 4.1 fixes the Huber loss with c=1.345, whose ψ is continuous but only piecewise linear, with a jump in ψ' at ±c; it is not even C^1, hence not C^2. The paper explicitly notes that Assumption 4 is satisfied by bisquare or exponential squared losses, so the authors are aware of the restriction, but they never prove a separate theorem for Lipschitz or piecewise-smooth ψ, nor an approximation argument showing that the Huber estimator inherits the asymptotic behavior of smooth losses. Consequently, Theorem 1 and Corollary 1, as written, do not apply to the estimator that produces the simulation and real-data results. The central claim that the proposed robust penalized spline estimator attains least-squares rates is therefore broader than the verified statement. This is a gap rather than a contradiction: smooth redescending losses such as bisquare are covered by the theorem. The auxiliary-scale corollary inherits the same regularity gap and additionally assumes a root-n scale estimate without proving it for the recommended S-estimator in the increasing-K setting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a family of M-type penalized spline estimators for the functional linear model, replacing the quadratic loss with a bounded loss function and adding a roughness penalty on the spline basis expansion. The main theoretical results, stated as Theorem 1 and Corollary 1, assert that under certain design and smoothness conditions the estimator achieves the same prediction-error rates as least-squares penalized splines, both with and without an auxiliary scale estimate. The paper also describes an iterative reweighted least squares algorithm, a robust corrected AIC for smoothing parameter selection, and reports a simulation study and a real-data application on archaeological glass spectra. The manuscript is clearly written and the proposed methodology is a natural and practically relevant extension of robust estimation to penalized spline functional regression.","tokens_in":15364,"tokens_out":4026,"duration_ms":42235,"significance":"If the theorems are correct, the paper makes a useful contribution: it provides a computationally feasible robust estimator for functional linear regression that attains the same asymptotic rates as least-squares penalized splines under weak conditions on the design and error distribution. The authors also provide R code and a supplementary proof file, which is a strength for reproducibility. However, the central theoretical result currently does not cover the loss function actually used in the implementation (Huber's loss with c=1.345), and the auxiliary-scale result relies on a root-n scale condition that is not verified for the recommended S-estimator in the increasing-knot setting. These gaps must be addressed before the paper's claims can be fully accepted as stated.","major_comments":[{"comment":"Assumption 4 requires psi = rho' to be in C^2(R) with bounded second derivative, but the implementation in Section 4.1 fixes the Huber loss with tuning constant c=1.345. The Huber psi function is continuous but piecewise linear, with a jump in psi' at the points +/- c; it is not C^1, hence not C^2. Therefore Theorem 1 and Corollary 1, taken literally, do not apply to the estimator used to produce the simulation results and the real-data analysis. The paper notes that Assumption 4 is satisfied by the bisquare or exponential squared losses, but it does not provide an approximation argument showing that the Huber estimator inherits the same asymptotic rates. This is a load-bearing gap between the theory and the numerical evidence. I recommend either proving a version of the theorems for Lipschitz or piecewise-smooth psi, or changing the implementation to a loss satisfying Assumption 4 (e.g., bisquare) and rerunning the experiments.","section":"Section 3, Assumption 4 vs Section 4.1"},{"comment":"Assumption 8 postulates that sqrt(n)(sigma_hat - sigma) = O_P(1) for a preliminary scale estimate, and the paper recommends obtaining sigma_hat from an initial unpenalized S-estimator. However, in this setting the number of spline basis functions K grows with n, so the dimension of the regression parameter increases. It is not demonstrated that the recommended S-estimator achieves the required root-n consistency under the growing-dimension regime; also, the scale is that of the residuals after estimating a growing-dimensional parameter. Without verifying Assumption 8 for the proposed scale estimator, Corollary 1's practical relevance is not established. The authors should either prove the needed property for their proposed S-estimator or replace Assumption 8 with a condition that is verifiable in the described context.","section":"Section 2.1 and Corollary 1, Assumption 8"},{"comment":"Theorem 2 introduces the condition rho_n^{-1} K^{-1} C_n = O_P(1), where rho_n is the smallest eigenvalue of n^{-1} X^T X and C_n = K/n + lambda + K^{-2p}. This condition appears abrupt and is not discussed in the text; the authors do not explain whether it is plausible for the simulation designs, nor how it relates to the earlier Assumption 5. Since Theorem 2 is used to claim rates in the population prediction norm, the condition should be motivated and, ideally, checked in the simulation settings. Please clarify the interpretation of this condition and its relation to the design assumptions.","section":"Section 3, Theorem 2"}],"minor_comments":[{"comment":"The formula for the corrected AIC contains an expression of the form \"log(sigma_hat^2(lambda)) + 1 + 2 Tr(H(lambda)) + 1/(n - Tr(H(lambda)) - 2)\" that appears to lack parentheses; please verify that this matches the formula in Hurvich et al. (1998).","section":"Section 2.2"},{"comment":"In the description of the FPCRR estimator, some symbols are used without definition, such as V_A, A, and the norm ||·||_E; please provide explicit definitions so the reader can interpret the formula.","section":"Section 4.1"},{"comment":"For the closely-spaced case with beta3 and beta4, the reported mean errors for SSp and RKHS are very large even under Gaussian errors (e.g., 633.1, 686.5, 5823, 7737); this striking behavior is not commented on in the text and deserves an explanation or a caveat about the experimental setup.","section":"Table 1"},{"comment":"The term \"redesceding M-estimators\" appears to be a typo for \"redescending M-estimators\"; please correct it.","section":"Section 6"},{"comment":"The proof sketch in the main text is informative but the full proofs are deferred to a supplementary file; the authors should at least list the main lemmas or steps in the main text to allow readers to follow the argument without opening the supplement.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a statistics journal and addresses a useful problem, but the gap between the theoretical assumptions and the implemented loss function is significant. The authors should either extend the theory to cover the Huber loss or adjust the numerical studies to use a loss that satisfies the regularity conditions. I would also recommend asking the authors to make the supplementary proofs available for review, as the main text's proof outline is concise and the details are essential for verifying the claims. The paper is not ready for acceptance in its current form, but the issues seem addressable with careful revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine methodological contribution: it introduces M-type penalized lower-rank spline estimators for scalar-on-function regression and proves prediction-error rates of order K/n + K^{-2p} + lambda under relatively weak conditions. The novelty is the combination of bounded-loss M-estimation with penalized B-splines, and the rates match the least-squares penalized spline rates, which is what you'd want from a robust method. The authors also provide a neat pseudo-observation representation, a thorough simulation study against FPCR, smoothing spline, RKHS, and robust FPCA competitors, and a real data example on archaeological glass spectra. The citation pattern is solid. I did not verify the supplement proofs, but the assumptions align with standard penalized spline arguments, and the rates are plausible.\n\nThe main soft spot is exactly what the stress-test note highlights: Assumption 4 requires psi = rho' to be twice continuously differentiable with bounded second derivative, but the implementation in Section 4.1 uses Huber's loss with c = 1.345, whose psi is only piecewise linear and not C^1 at the knots. The authors are aware of this—they note that bisquare or exponential-squared losses satisfy Assumption 4—but they never provide an approximation argument or a separate theorem for Lipschitz losses. Consequently, Theorem 1 and Corollary 1, as written, do not cover the estimator that produces the simulation and real-data results. This is a gap rather than a contradiction: if a smooth redescending loss were used, the rates would apply. Still, the default loss is the one advertised, so the mismatch should be fixed.\n\nA second, smaller gap is Assumption 8, which assumes a root-n consistent preliminary scale estimate without proving it for the recommended S-estimator in the increasing-K functional setting. That issue is worth a sentence in the paper. The corrected AIC for selecting lambda is heuristic and not part of the asymptotic rates; that's fine for a methods paper, but the authors should state that lambda is treated as deterministic in the theory.\n\nOverall, I'd accept this for peer review and would probably give a conditional accept after revision: prove the result for Huber-type losses (or switch the default to a smooth loss) and address the scale assumption. The paper is aimed at functional data analysts who want a robust, computationally feasible estimator and at methodologists interested in penalized spline asymptotics.","headline":"Robust functional regression paper with a genuine contribution but a central theory-implementation gap: the rates require C^2 losses, while the default Huber loss is not covered.","tokens_in":15842,"tokens_out":3700,"would_cite":true,"duration_ms":37730,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62G35","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that M-type penalized spline estimators for the functional linear model match least-squares convergence rates while resisting outliers.","keywords":["functional linear regression","M-estimators","penalized splines","prediction error rates","auxiliary scale","B-spline basis","heavy-tailed errors","outlier resistance"],"falsifier":"Run the default implementation at increasing sample sizes with $K\\asymp n^{1/(2p+1)}$ and $\\lambda\\to0$: if the prediction error does not shrink at roughly $n^{-2p/(2p+1)}$, the smoothness condition on the loss is doing real work. A direct check is whether the pseudo-observation representation still yields the promised leading term for residuals that land exactly on the kink.","tokens_in":14895,"feed_emoji":"📊","tokens_out":17490,"duration_ms":164234,"temperature":0.7,"pith_summary":"The paper proposes replacing the squared-error loss in penalized B-spline estimation of the functional linear model with a general bounded loss, producing M-type penalized spline estimators that are outlier-resistant and computationally feasible. Its main theorem states that these estimators attain the same prediction-error rate as least-squares penalized splines, $\\|\\hat\\beta-\\beta\\|_{\\Gamma_n}^2 = O_P(K/n + K^{-2p} + \\lambda)$, under mild design conditions. This matters because least-squares functional regression degrades under contamination, smoothing splines are computationally heavy, and principal-component methods need smooth trajectories or eigenvalue decay. The paper also shows the rate survives auxiliary scale estimation, and its Monte Carlo and real-data analyses place the new estimator at or above the level of competitors under Gaussian, heavy-tailed, and contaminated errors.","feed_headline":"Bounded-loss splines match least-squares rates in functional regression","feed_subtitle":"The estimator keeps the same prediction error even when errors are heavy-tailed or contain outliers.","key_machinery":"The central object is the penalized M-estimator defined by minimizing the paper's criterion (3): a B-spline expansion of the slope function, a symmetric bounded loss $\\rho$ applied to residuals, and a quadratic roughness penalty on the $q$th derivative. The theorem's proof is carried by an asymptotic equivalence: after a Taylor expansion of the score equations, the M-estimator behaves like the minimizer of a least-squares problem on pseudo-observations, so existing penalized-spline rate theory applies unchanged. The lower-rank B-spline basis keeps the design at dimension $K+p$ instead of $n$, and the banded roughness matrix keeps computation feasible through iteratively reweighted least squares.","core_discovery":"The central claim is that robustifying the penalized B-spline estimator for the functional linear model does not degrade its asymptotic prediction accuracy. Under bounded predictors, a sufficiently smooth regression function, nonsingular design behavior, and a score function satisfying standard moment conditions, there is a sequence of minimizers of the penalized M-criterion for which $\\|\\hat\\beta_{\\mathrm{MPS}_p}-\\beta\\|_{\\Gamma_n}^2 = O_P(K/n + K^{-2p} + \\lambda)$. The same rate holds when the loss uses a root-$n$-consistent auxiliary scale, and the population prediction norm adds only an $n^{-1/2}$ term from estimating the covariance operator. The proof represents the M-estimator asymptotically as a least-squares estimator on pseudo-observations $\\tilde Y_i = \\int_{[0,1]} X_i(t)\\beta(t)\\,dt + \\psi(\\epsilon_i)/E\\psi'(\\epsilon)$, with the penalty rescaled by $1/E\\psi'(\\epsilon)$, which is why a bounded loss protects against outliers without changing the leading error terms.","pith_inferences":["A smoothing approximation argument, not given in the paper, would close the gap between the stated $C^2$ loss condition and the piecewise-linear default loss used in the implementation; the theorems as written cover the smooth bounded losses they name.","The pseudo-observation representation suggests an outlier diagnostic: observations with large transformed residuals $\\psi(\\hat\\epsilon_i)/\\widehat{E\\psi'(\\epsilon)}$ are down-weighted rather than deleted, which the paper only explores informally through residual QQ-plots.","Since the rates need only boundedness of the predictor in $L^2$, the same argument should transfer to tensor-product B-splines for functional predictors over higher-dimensional domains, a generalization the abstract advertises without carrying out."],"forward_implications":["Choosing $K\\asymp n^{1/(2p+1)}$ and $\\lambda=o(n^{-2p/(2p+1)})$ gives the optimal nonparametric rate $O_P(n^{-2p/(2p+1)})$, so robustness costs no first-order accuracy.","For a new observation, the same rate holds up to an extra $n^{-1/2}$ term when the smallest eigenvalue of the design matrix does not decay too fast.","A root-$n$-consistent preliminary scale estimate leaves the rate unchanged, requiring only a finite first error moment.","The assumptions require no second error moments, no trajectory smoothness, and no eigenvalue decay rate, so the estimator is usable for heavy-tailed and rough-predictor settings where established competitors are harder to justify.","The proof is not specific to the B-spline basis, so the same rate argument extends to other sieved estimators with a quadratic penalty."],"supporting_citations":[{"why":"Defines the least-squares penalized spline estimator for the functional linear model that this paper robustifies and extends.","marker":"Cardot et al. (2003)"},{"why":"Supplies the asymptotic rate decomposition for penalized splines that Theorem 1 adapts to the M-type objective.","marker":"Claeskens et al. (2009)"},{"why":"Establishes the design consistency condition for increasing-dimension M-estimators that Assumption 5 parallels.","marker":"Huber (1973)"},{"why":"Provides asymptotic theory for M-estimators under increasing dimension and auxiliary scale, used by Corollary 1.","marker":"Yohai et al. (1979)"},{"why":"Supplies the iteratively reweighted least-squares algorithm and standard robust-regression facts behind the implementation.","marker":"Maronna et al. (2006)"},{"why":"Proposes the smoothing-spline functional estimator used as a comparison baseline and motivation for lower-rank smoothing.","marker":"Crambes et al. (2009)"},{"why":"Gives the reproducing-kernel functional regression estimator whose rates and computational demands motivate the proposed class.","marker":"Yuan et al. (2010)"},{"why":"Provides the corrected AIC that the paper robustifies for smoothing parameter selection.","marker":"Hurvich et al. (1998)"}],"fun_headline_variants":["Robust functional regression matches least-squares prediction rates","Outlier-proof B-spline estimators keep optimal functional rates","Penalized M-estimators for functional data: robust, fast, consistent","Bounded-loss functional regression achieves LS error without robustness cost","Robust lower-rank functional estimators reach high efficiency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorems assume the loss's slope changes smoothly at every point, but the default loss used in the numerical work has a sharp corner at its cutoff, and the paper does not prove the rates survive for that cornered loss.","fun_headline_variants_meta":{"raw":{"variants":["Robust functional regression matches least-squares prediction rates","Outlier-proof B-spline estimators keep optimal functional rates","Penalized M-estimators for functional data: robust, fast, consistent","Bounded-loss functional regression achieves LS error without robustness cost","Robust lower-rank functional estimators reach high efficiency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001241,"raw_usage":{"total_tokens":5063,"prompt_tokens":887,"completion_tokens":4176,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":4093}},"tokens_in":503,"tokens_out":4176,"duration_ms":29472,"temperature":1.0,"reasoning_tokens":4093,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:29:28.992960+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the default implementation at increasing sample sizes with $K\\asymp n^{1/(2p+1)}$ and $\\lambda\\to0$: if the prediction error does not shrink at roughly $n^{-2p/(2p+1)}$, the smoothness condition on the loss is doing real work. A direct check is whether the pseudo-observation representation still yields the promised leading term for residuals that land exactly on the kink.","supporting_citations":[{"cited_title":", Ferraty, F","cited_arxiv_id":null,"evidence_quote":"Defines the least-squares penalized spline estimator for the functional linear model that this paper robustifies and extends."},{"cited_title":"J.(1973) Robust Regression: Asymptotics, Conjectures and Monte Carlo","cited_arxiv_id":null,"evidence_quote":"Establishes the design consistency condition for increasing-dimension M-estimators that Assumption 5 parallels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides asymptotic theory for M-estimators under increasing dimension and auxiliary scale, used by Corollary 1."},{"cited_title":"A., Martin, D","cited_arxiv_id":null,"evidence_quote":"Supplies the iteratively reweighted least-squares algorithm and standard robust-regression facts behind the implementation."},{"cited_title":"& Pascal, S.(2009) Smoothing splines estimators for functional linear regression","cited_arxiv_id":null,"evidence_quote":"Proposes the smoothing-spline functional estimator used as a comparison baseline and motivation for lower-rank smoothing."},{"cited_title":"& Cai, T.(2010) A reproducing kernel Hilbert space approach to functional linear regression","cited_arxiv_id":null,"evidence_quote":"Gives the reproducing-kernel functional regression estimator whose rates and computational demands motivate the proposed class."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the corrected AIC that the paper robustifies for smoothing parameter selection."}],"review_version":1}