{"id":"ceb94081-909a-4e95-8bd0-df2637ea43e1","arxiv_id":"1909.01675","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Khmaladze-transformed partial sum process of B-spline residuals yields a pivotal test for a wide class of nonparametric shape restrictions.","lead":"This paper proposes a general test for shape restrictions such as monotonicity, convexity, U-shape, and symmetry in nonparametric regression, using B-splines and a martingale transformation. A smart generalist might read it because the same test procedure covers many qualitative hypotheses and has a bootstrap implementation that works in finite samples.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pivotalness relies on an unproven active-set step: the Khmaladze transformation uses data-dependent binding constraints, but the Lemma 1 cancellation in Theorems 2 and 3 only holds if the estimated active set equals the true one.","rationale":"The reader's weakest assumption already locates the active-set and nonlinear-surface gaps; my read agrees and sharpens the mechanism. The proofs of Theorems 2 and 3 obtain ~M_n - M_n = op(1) by applying Lemma 1 to cancel the estimation error \\hat b - beta. This cancellation is exact only when \\hat b - beta lies in the span of the basis used in the transformation. In the algorithm, that basis is chosen after seeing the data, from the estimated binding constraints, so the cancellation requires the estimated active set to coincide with the true active set. No such consistency is proved; the paper explicitly declines to distinguish cases of binding constraints. The gap is not merely cosmetic: for a constant regression function under monotonicity, the true active set is all equality constraints, while the isotonic fit produces a random block partition, so the reduced basis changes with the sample. The same issue appears in log-convexity and r-convexity, where several quadratic constraints can be active at once and Section 3.2's single-surface implicit-function assumption is acknowledged as a simplification. Because the central claim is asymptotic pivotalness for a general class including boundary nulls, this unproved active-set step is load-bearing. The paper may still be correct for the linear interior case, but the general statement and the bootstrap for nonlinear constraints are not established as written. A conditional verdict is therefore appropriate, matching the reader's assessment.","tokens_in":48538,"tokens_out":16474,"duration_ms":169611,"concrete_test":"Run the Section 3.1 procedure under H0 with m(x)=1 (constant, so all monotonicity constraints are active) using n=1000, L'=9, cubic B-splines, and 10,000 Monte Carlo draws; compare the empirical 95% quantile of the KS statistic to the sup_{x in (0,1)} |B(F_X(x))| critical value. A deviation exceeding the Monte Carlo standard error (about 0.014) would show that the data-dependent active-set modification does not preserve the claimed pivotal limit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 instructs that \"the constraints that proved to be binding ... have to be incorporated\" by redefining P_L, and Section 3.2 does the same with ~P_i(\\hat b). The proofs of Theorems 2 and 3 kill the \\hat b-\\beta term by invoking Lemma 1, which requires \\hat b-\\beta to lie in the column space of the reduced basis. With a data-dependent active set this holds only if the estimated binding pattern equals the true one. No consistency of the active set is established, and in boundary null cases such as m(x)=1 under monotonicity or log-convexity the true active set has positive dimension while the finite-sample active set is random and typically different; mismatched active sets are not covered by the op(1) arguments. Section 3.2 compounds this by assuming a single implicit-function representation beta_l0 = h(beta_{-l0}) and deferring multiple active surfaces with \"the rest ... would remain the same\"; the quadratic constraints in Examples 3-4 naturally bind at several knots simultaneously. Since the bootstrap (Theorem 4, Section 4) is proved only for the fixed P_i object and Lemma 1, bootstrap validity for the nonlinear cases is also unsupported. Without active-set consistency, or a proof that mismatches are uniformly negligible, the claimed pivotal limit sigma_u B(F_X(x)) is not established for boundary nulls.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a test for a general class of nonparametric shape restrictions on a regression function, including monotonicity, convexity, U-shape, S-shape, symmetry, quasi-convexity, log-convexity, r-convexity, and MN-convexity. The method approximates the regression function with B-splines, expresses the null hypothesis as constraints on the spline coefficients, and applies a Khmaladze-type transformation to the partial sum empirical process. The authors claim that, under H0, the transformed process converges weakly to sigma_u B(F_X(x)), so that Kolmogorov-Smirnov, Cramer-von Mises, and Anderson-Darling tests become asymptotically pivotal. A residual bootstrap is also proposed and claimed to be asymptotically valid. The paper includes Monte Carlo simulations and two empirical applications.","tokens_in":48812,"tokens_out":6159,"duration_ms":63436,"significance":"If the main theorem were fully established, this would be a valuable unified framework for testing a wide range of shape restrictions, with a pivotal asymptotic distribution that does not depend on the particular constraint class. The paper clearly demonstrates how many shape restrictions can be encoded as linear or nonlinear constraints on B-spline coefficients, and the bootstrap algorithm is straightforward to implement. However, the central proof has a load-bearing gap: the Khmaladze transformation is defined using a data-dependent active set of binding constraints, and the proof that the estimation error term vanishes requires that the estimated active set coincide with the true one, which is never established. The nonlinear extension is also asserted rather than proved for multiple binding surfaces. These issues prevent the main result from being accepted as stated, despite the evident promise of the approach.","major_comments":[{"comment":"The proof of Theorem 2 eliminates the contribution of P_i'(hat b - beta) by invoking Lemma 1, but Lemma 1 only annihilates linear combinations of the B-spline basis actually used in the transformation. In Section 3.1 the basis is redefined after estimation according to the constraints that 'proved to be binding.' With a data-dependent active set, hat b - beta lies in the column space of the reduced basis only if the estimated binding pattern equals the true one. No consistency of the active set is established, and for boundary nulls such as m(x)=1 under monotonicity the true active set has positive dimension while the finite-sample active set is random and typically different. Consequently, the claimed convergence of ~M_n(x) to sigma_u B(F_X(x)) is not proven for boundary nulls.","section":"Section 3.1, Theorem 2 and its proof (Section 7.2); Lemma 1 (Appendix B)"},{"comment":"The generalization from a single binding surface to multiple active surfaces is asserted with the sentence 'The rest of the methodology and the asymptotic result would remain the same as above,' without a proof. This is not a cosmetic omission: the quadratic constraints in Examples 3 and 4, e.g. the system in (2.8), naturally bind at several knots simultaneously, so the single-surface case is not representative. The proof of Theorem 3 also relies on an implicit-function representation beta_l0 = h(beta_-l0) and assumes the Hessian of h has finite norm, but this bounded-Hessian condition is not stated among the maintained assumptions.","section":"Section 3.2, last paragraph"},{"comment":"The bootstrap validity proof uses Lemma 1 in exactly the same way as the proof of Theorem 2; in particular, the argument that sup_x |~M*_n(x) - M*_n(x)| = o_p*(1) rests on the identity hat u*_i - hat u_i = P_i'(hat b* - hat b) and the cancellation of that term by Lemma 1. Since the active-set issue already undermines Theorem 2, bootstrap validity for the nonlinear constrained case and for boundary nulls with estimated active sets is equally unsupported. A separate treatment covering data-dependent active sets is needed before Theorem 4 can be accepted.","section":"Section 4, Theorem 4 and its proof (Section 7.5)"}],"minor_comments":[{"comment":"The wording 'its asymptotic distribution is a functional of the standard Brownian motion' is imprecise; the transformed process itself converges to sigma_u B(F_X(x)), and it is the test statistics such as KS, CvM, and AD that are functionals of that process.","section":"Abstract and Section 1"},{"comment":"The 2000 Mathematics Subject Classification codes (05C38, 15A15, 05A15, 15A18) appear to be copied from another paper and are unrelated to the statistical content; appropriate codes from statistics or econometrics should be used.","section":"Title page"},{"comment":"All null scenarios used in the Monte Carlo experiment are strictly interior to the shape class (e.g., m(x)=x^(13/4) is strictly increasing, the log-convex function exp(x^2) is strictly log-convex). None of the simulations exercises a boundary null such as a constant function under monotonicity, which is precisely the case where the active-set issue arises. Adding such a scenario would help assess the practical impact of the theoretical gap.","section":"Section 5.1, Scenarios 1, 2, and 5"},{"comment":"The statement that the test 'requires no more than recursive least squares' may be overstated: implementation also requires recomputing the active set and updating the generalized inverse at each step, and for nonlinear constraints the optimization can be substantially more involved.","section":"Section 3.3"},{"comment":"The stated bound E||~P_i||^s = O(L^{s/2} + L n^{(s/2-1)ς}) does not match the proof: equation (8.3) gives O(n^{(s/2-1)ς} + L^{s/2-1}), and the subsequent lines also suggest the L^{s/2-1} order. Please check whether the stated order is correct or whether the proof is missing a factor.","section":"Lemma 2 (Appendix B)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising idea and a substantial amount of technical work, but the central theorem is currently not proven as stated because of the data-dependent active set issue. I would recommend that the editors ask for a major revision: the authors should either prove consistency of the active set under H0 (or a suitable variant), or modify the construction so that the transformation does not depend on the estimated active set in a way that invalidates Lemma 1. The nonlinear multi-surface case also needs a real proof, not an assertion. The Monte Carlo results are interesting but avoid the problematic boundary-null case, so they do not currently provide reassurance on the main theoretical gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for the linear-constraint machinery, but the nonlinear claims outrun the proofs. The stress-test concern is real and lands on both Theorem 2 and Theorem 3, not just the nonlinear part.\n\nWhat is new: the first unified Khmaladze-transformation treatment of shape constraints via B-spline coefficient restrictions, with the number of constraints growing with the sample size. The linear case (monotonicity, convexity, U-shape with known switch, symmetry) is developed in detail; the appendix proofs are serious, and the simulations are honest and reasonably extensive. The bootstrap is a sensible addition and the WARP implementation is standard. This is a genuine contribution for practitioners testing qualitative restrictions in nonparametric regression.\n\nThe soft spots are concentrated and important. The proof of Theorem 2 uses Lemma 1 to kill the (hat b - beta) term, but Lemma 1 only applies if the estimated active set equals the true one. The paper never establishes active-set consistency, and under boundary nulls like a constant function under monotonicity, the true active set is large while the finite-sample active set is random. Without an argument that mismatches are uniformly negligible, the pivotal limit for boundary nulls is not proven. The nonlinear case in Theorem 3 assumes a single implicit-function representation and defers multiple active surfaces with \"the rest ... would remain the same\"; for quadratic constraints like r-convexity or AH-convexity, multiple binding knots are natural, so this is not a minor technicality. The bootstrap (Theorem 4) is proved for the full P object, not the active-set-reduced transformation used in the test statistic, so its theoretical validity for the actual procedure is also unsupported. Heteroscedasticity is handled in practice but not in the theory; that is a minor issue because the authors explicitly say the conditions can be weakened.\n\nProportionate verdict: for the core linear cases with non-boundary nulls, the method probably works; the gap is concentrated at boundary nulls and nonlinear constraints. That is enough to make the current version conditional, not a rejection.\n\nRecommendation: send it to a serious referee. The referee should ask for a proof of active-set consistency or a modification that avoids the issue (for example, using a superset of constraints or a different transformation). If that gap is closed, this becomes a useful paper for empirical economists and nonparametric statisticians. It deserves referee time, but it is not ready as is.","headline":"Worth a serious referee for the unified linear-constraint framework, but the nonlinear extension and the active-set handling are promises, not proofs.","tokens_in":49331,"tokens_out":3088,"would_cite":false,"duration_ms":34469,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62G10","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single transformation makes shape-restriction tests distribution-free, so standard Brownian-motion critical values apply across many shape classes.","keywords":["shape restrictions","nonparametric regression","Khmaladze transformation","partial sum process","B-splines","bootstrap","monotonicity","convexity"],"falsifier":"Run a simulation with a null model where the constrained spline fit has two binding constraints at the same point (e.g., both monotonicity and convexity are tight), using the paper's bootstrap and the asymptotic critical values, and compare the empirical rejection rate to the nominal level; a systematic size distortion would contradict the claim that the generalized multiple-surface case is straightforward.","tokens_in":48299,"feed_emoji":"📈","tokens_out":4786,"duration_ms":44068,"temperature":0.7,"pith_summary":"The paper establishes a unified test for whether a nonparametric regression function satisfies a shape restriction such as monotonicity, convexity, U-shape, S-shape, symmetry, log-convexity, r-convexity, or MN-convexity. The test applies Khmaladze's martingale transformation to the partial sum process of residuals from a shape-constrained B-spline fit. The transformed process converges weakly to a functional of standard Brownian motion, making the test asymptotically pivotal: the same Kolmogorov–Smirnov, Cramér–von Mises, and Anderson–Darling critical values work for every shape in the class. Because asymptotic critical values can be poor in finite samples, the paper also proves that a simple bootstrap is asymptotically valid and provides Monte Carlo evidence that it restores nominal size.","feed_headline":"Khmaladze transform makes shape tests distribution-free","feed_subtitle":"One nonparametric test covers monotonicity, convexity, U-shapes and more, with standard Brownian-motion critical values.","key_machinery":"The central object is Khmaladze's martingale transformation $T$ applied to the partial sum empirical process $K_n(x) = n^{-1}\\sum_{i=1}^n \\hat{u}_i I(x_i < x)$. The transformation projects the process onto the orthogonal complement of the space spanned by the B-spline basis (with binding constraints incorporated), so that the estimated parameters drop out and the process becomes asymptotically a Brownian motion. The B-spline basis is chosen because shape constraints translate into simple inequalities on coefficients, and because the transformation requires a basis with local support.","core_discovery":"Under the null hypothesis that the regression function $m$ belongs to a shape class $\\mathcal{M}_0$ satisfying Condition C0, the Khmaladze-transformed partial sum process $\\widetilde{M}_n(x)$ converges weakly to $\\sigma_u B(F_X(x))$, a standard Brownian motion functional. Consequently, critical values for Kolmogorov–Smirnov, Cramér–von Mises, and Anderson–Darling statistics are available without simulation, and the proposed bootstrap reproduces the null distribution asymptotically. The class includes monotonicity, convexity/concavity, U- and S-shapes with known or estimated switch points, symmetry, quasi-convexity, r-convexity, and MN-convexity; the transformation automatically removes the estimation effect that otherwise makes the limit non-Gaussian at the boundary of the shape class.","pith_inferences":["The pivotal limit suggests that the same transformation could yield distribution-free tests for shape constraints in other settings, such as quantile regression or partial linear models, where partial sum processes arise.","The dependence on the correct active set implies that data-driven selection of binding constraints could be a source of size distortion in small samples; a diagnostic that checks stability of the result across different active sets would be a useful practical check.","The rate condition $L^2/n + n/L^4 = o(1)$ leaves room for adaptive choice of knots; one could envision a data-driven $L$ that minimizes a bootstrap estimate of size or power."],"forward_implications":["A single table of critical values (KS, CvM, AD) applies to tests of monotonicity, convexity, log-convexity, U-shape, S-shape, symmetry, quasi-convexity, r-convexity, and MN-convexity, as well as simultaneous constraints.","The bootstrap is asymptotically valid and corrects finite-sample size distortions, as documented in Monte Carlo experiments for sample sizes around 1000.","The test has non-trivial power against alternatives converging to the null at the parametric $n^{-1/2}$ rate, provided the binding constraints are correctly embedded in the transformation.","Because the implementation reduces to recursive least squares with no bandwidth choice, the procedure is computationally simple and does not require tuning parameters beyond the spline knot placement."],"supporting_citations":[{"why":"Supplies the martingale transformation that removes the estimation effect and makes the process asymptotically pivotal.","marker":"[57]"},{"why":"Provides the partial sums empirical process framework for regression model checks that the paper builds on.","marker":"[84]"},{"why":"Introduces recursive residuals and the CUSUM idea underlying the construction of the sample transformation.","marker":"[12]"},{"why":"Gives the approximation error bounds for B-splines used to control the bias term in the proofs.","marker":"[2]"},{"why":"Shows how the asymptotic distribution depends on parameter estimation, motivating the need for the transformation.","marker":"[33]"},{"why":"Supplies the WARP-speed bootstrap implementation used in the Monte Carlo experiments.","marker":"[41]"}],"fun_headline_variants":["Khmaladze transform gives standard critical values for shape tests","One test for many shape constraints, with Brownian-motion critical values","Distribution-free shape testing via Khmaladze transform","Unified nonparametric test for shape constraints with Brownian critical values"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For nonlinear shape constraints, the proof assumes the constrained estimator lies on a single smooth boundary surface described by an implicit function, and that the set of binding constraints is correctly incorporated; if several constraints bind at once or the active set is misspecified, the pivotal result may fail.","fun_headline_variants_meta":{"raw":{"variants":["Khmaladze transform gives standard critical values for shape tests","One test for many shape constraints, with Brownian-motion critical values","Distribution-free shape testing via Khmaladze transform","Unified nonparametric test for shape constraints with Brownian critical values"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000876,"raw_usage":{"total_tokens":3715,"prompt_tokens":799,"completion_tokens":2916,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":415,"completion_tokens_details":{"reasoning_tokens":2844}},"tokens_in":415,"tokens_out":2916,"duration_ms":22041,"temperature":1.0,"reasoning_tokens":2844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:10:34.759147+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a simulation with a null model where the constrained spline fit has two binding constraints at the same point (e.g., both monotonicity and convexity are tight), using the paper's bootstrap and the asymptotic critical values, and compare the empirical rejection rate to the nominal level; a systematic size distortion would contradict the claim that the generalized multiple-surface case is straightforward.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the martingale transformation that removes the estimation effect and makes the process asymptotically pivotal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the partial sums empirical process framework for regression model checks that the paper builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows how the asymptotic distribution depends on parameter estimation, motivating the need for the transformation."},{"cited_title":"and White, H","cited_arxiv_id":null,"evidence_quote":"Supplies the WARP-speed bootstrap implementation used in the Monte Carlo experiments."}],"review_version":1}