{"id":"3bc455a4-08c7-4571-8b19-e449c0400e66","arxiv_id":"2506.23069","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Mapped-sieve estimators and bootstrap-based simultaneous confidence regions for time-varying nonlinear time series regression achieve uniform consistency and asymptotic coverage on unbounded support.","lead":"Statistical theory for estimation and simultaneous inference in time-varying nonlinear regression with locally stationary covariates. It introduces mapped sieve estimators, simultaneous confidence regions, and a multiplier bootstrap, with proofs and simulations; an R package is announced.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3.1 and Lemma G.2 are false as stated: the mapped orthonormal bases (3.8)/(G.1–G.3) vanish at |x|=∞, so functions like m(t,x)=sin(2πt) satisfy Assumption 3.1 yet the claimed sup-norm approximation error O(c^{-m1j}+d^{-m2j}) cannot hold.","rationale":"The reader's weakest_assumption correctly points to the decay/tail issue, but frames it as 'Assumption 3.1 requires the untransformed function to lie in a generalized Schwartz space.' In fact Assumption 3.1 is weaker: it only imposes smoothness of the mapped function on [0,1]^2. That weaker condition is insufficient because the mapped orthonormal basis functions all vanish at infinity, so the sieve space cannot uniformly approximate functions with non-vanishing tails. This is not merely a scope limitation; Proposition 3.1 and Lemma G.2 are false as stated, with an explicit counterexample. Since uniform consistency and SCR coverage rest on this approximation result, the central claim's proof fails under the stated assumptions. The authors can likely repair the paper by strengthening Assumption 3.1 to require m_j(t,·)∈C_0(R) and appropriate smoothness of m_j(t,x)/√(u'(x)) in the mapped coordinate, but this is a substantive correction to the main theorems. I therefore recommend REJECT of the current version, with the understanding that a corrected manuscript could be resubmitted. The reader's CONDITIONAL verdict is upgraded because the identified issue is a false theorem, not a missing code artifact or unverified technical condition.","tokens_in":61378,"tokens_out":20770,"duration_ms":236048,"concrete_test":"Take m(t,x)=sin(2πt), the algebraic mapping x = y/√(1−y^2) with s=1 (Example J.3), and the mapped Legendre basis (Example G.2). For fixed c (e.g., c=2) and d=10,50,100, compute the least-squares/projection coefficients and evaluate the approximation error at x=±M for large M, where all φ_j(±M)→0. The uniform error sup_t sup_{|x|≤M}|sin(2πt)−m_{c,d}(t,x)| will approach sup_t|sin(2πt)|=1 as M→∞ and will not decay with d, directly falsifying Lemma G.2 and Proposition 3.1 under Assumption 3.1.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's mapped orthonormal basis functions are φ_j(x) = √(u'(x))·J_j(u(x)) (Examples G.1–G.3). For every mapping in Definition 3.1 with unbounded image, u'(x)→0 as |x|→∞, so every φ_j(x)→0 and hence every finite linear combination m_{c,d}(t,x) tends to 0 in x. Thus the sieve space is contained in C_0(R) (functions vanishing at infinity). Assumption 3.1, however, only requires the mapped function em_j(t,y)=m_j(t,g(2y−1)) to be smooth with bounded derivatives on [0,1]^2; it does not require m_j(t,x) to vanish at infinity. The constant-in-x function m_j(t,x)=sin(2πt) satisfies Assumption 3.1 (em_j(t,y)=sin(2πt) is C^∞ on the square), but sup_x |sin(2πt) − m_{c,d}(t,x)| ≥ |sin(2πt)|, since the approximant is 0 at x=∞. The proof of Lemma G.2 incorrectly equates sup_{x∈R}|m_j(t,x)−Σ b_j φ_j(x)| with sup_{y∈[0,1]}|em_j(t,y)−Σ b_j J_j(y)|, dropping the √(u') factor. Consequently Proposition 3.1, the bias term in Theorem 3.2, and the coverage claims of Theorems 4.1–4.3 are not justified under the stated Assumption 3.1. Remark 3.1 states a Schwartz-space sufficient condition, but that condition is not part of the assumption, so the theorem is internally inconsistent as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes sieve estimators for the time-varying nonlinear regression model Y_i = m_0(t_i) + Σ_j m_j(t_i, X_{j,i}) + ε_i under a locally stationary physical-representation framework. It constructs pilot sieve estimators using mapped orthogonal bases on the unbounded covariate domain, corrects them for the identifiability condition E(m_j) = 0, and proves uniform consistency, simultaneous confidence regions over (t,x), a volume-of-tubes expansion for critical values, and a multiplier bootstrap procedure. Numerical simulations and two data applications are included, together with an R package.","tokens_in":61803,"tokens_out":7227,"duration_ms":84211,"significance":"If the main results were correct, the paper would be a substantial contribution: it targets simultaneous inference for a general time-varying nonlinear regression model, provides two Gaussian approximation results for high-dimensional locally stationary affine forms, and offers a practical multiplier bootstrap implementation with an R package. The proof architecture is ambitious and the empirical comparisons against kernel estimators are informative. However, the central sieve approximation result is false under the stated Assumption 3.1, and this defect propagates into the consistency, under-smoothing, and coverage theorems. The contribution is therefore conditional on a substantial correction of the approximation theory and of the assumptions under which it is stated.","major_comments":[{"comment":"The stated approximation result is false under Assumption 3.1. The mapped orthonormal bases in (3.8) and Examples G.1–G.3 have the form φ_j(x) = √(u'(x)) J_j(u(x)); since u'(x) → 0 as |x| → ∞, every φ_j(x) tends to 0 at infinity. Any finite linear combination therefore belongs to C_0(R), the space of functions vanishing at infinity. Assumption 3.1 only requires smoothness of em_j(t, y) = m_j(t, g(2y−1)) on [0,1]^2, which is satisfied, for example, by m_1(t,x) = sin(2πt). For this function, sup_{t,x} |m_1(t,x) − m_{1,c,d}(t,x)| ≥ sup_t |sin(2πt)|, because the approximant tends to 0 as x → ∞. Thus the claimed O(c^{-m_1} + d^{-m_2}) bound in Proposition 3.1 cannot hold under Assumption 3.1. The proof of Lemma G.2 in the supplement equates sup_x |m(t,x) − Σ b_j φ_j(x)| with sup_y |em(t,y) − Σ b_j J_j(y)|; this is not an identity, since the mapped expansion is √(u'(x)) Σ b_j J_j(u(x)), not Σ b_j J_j(u(x)). The proof in Section L.5 merely cites standard compact-domain approximation results and does not repair this gap.","section":"§3.1.1, Proposition 3.1, Lemma G.2 (Supplement)"},{"comment":"Because Proposition 3.1 is false under the stated assumptions, the approximation-bias terms c_j^{-m_{1j}} + d_j^{-m_{2j}} in (3.35) and (4.1) are not justified. These terms are load-bearing: they are used to claim uniform consistency in Theorem 3.2 and to impose the under-smoothing condition in Assumption 4.1, and Theorems 4.1–4.3 then rely on this for the coverage of the simultaneous confidence regions (4.7) and (4.28). Remark 3.1 states a generalized Schwartz-space sufficient condition for the method to work, but that condition is not part of Assumption 3.1, so the theorem statements are internally inconsistent as written.","section":"Theorem 3.2, Assumption 4.1, Theorems 4.1–4.3"},{"comment":"The counterexample is not a peripheral technicality: the paper's advertised features are that the assumptions are 'mild' and that the method allows 'unbounded domain support.' Under the currently stated assumption, a constant-in-x smooth function such as sin(2πt) is admissible, yet it cannot be uniformly approximated by the proposed mapped sieve space. To repair the manuscript, the authors must add an explicit decay or vanishing-at-infinity condition on m_j(t, ·) (for example, membership in the generalized Schwartz space of order m_{2j} as described in Remark 3.1 and Section J.3), prove Proposition 3.1 from first principles under that condition while accounting for the √(u') weight, and then revisit all downstream rates and coverage results. Merely citing [60] or [12] is not sufficient.","section":"§1.2, §3.1.1, Remark 3.1"}],"minor_comments":[{"comment":"The sentence 'Note that {eφ_i} is a sequence of orthogonal basis of the functional space defined on R' should specify the space L^2(R) and the measure involved, since the orthogonality of mapped bases is an L^2 property, not a sup-norm property.","section":"§3.1.1, after (3.7)"},{"comment":"The displayed equality in the proof uses em_j(t,x) on the right-hand side; the second argument should be y, so that the expression reads em_j(t,y) and em_{j,d}(t,y).","section":"Lemma G.2 (Supplement)"},{"comment":"There are typos: 'SMIle' should be 'SIMle', and 'generayed' should be 'generated'.","section":"§1.2 and Figure A caption"},{"comment":"The lower index g in the second sum of (3.11) is introduced and then set to 1 for simplicity; the relationship between this convention and the definition of W_1 in (H.2), which uses φ_{ℓ_2+1}, should be clarified in the main text.","section":"Equation (3.11) and Section H"},{"comment":"The claim that the method 'allow[s] for unbounded domain support' should be reworded after the necessary decay assumption is imposed: the support may be unbounded, but the functions must vanish at infinity with prescribed derivative decay.","section":"Abstract, §1.2, §3.1.1"}],"recommendation":"major_revision","confidential_remarks":"The counterexample given in the major comments is decisive: Proposition 3.1 is false as stated, and the downstream theorems inherit the problem. I recommend major revision rather than rejection because the framework could plausibly be repaired by explicitly assuming generalized Schwartz decay for m_j(t, ·) and by redoing the mapped-basis approximation proof, including the effect of the √(u') factor. After such a repair, please re-check that the volume-of-tubes expansion in Theorem 4.2 and the bootstrap proof in Theorem 4.3 remain valid under the revised assumptions. Note that the simulations use rapidly decaying regression functions, so they do not expose the counterexample."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a serious and unusually complete contribution: it claims the first simultaneous inferential method for the general model (1.1), and the claim holds up against the cited literature. The proof architecture is coherent — uniform consistency, pointwise and uniform Gaussian approximation, volume-of-tubes critical values, and a multiplier bootstrap that actually works in simulations. The two-step bias correction for identifiability is sensible, the simulations are extensive, and an R package ships. Second, there is a real, load-bearing gap in the approximation theory. The mapped sieve basis functions in Examples G.1–G.3 all carry a factor sqrt(u'(x)), which decays to zero as |x| -> infinity for the paper's own mappings. So every finite linear combination of these basis functions vanishes at infinity, and so does every sieve approximant. But Assumption 3.1 only requires the mapped function to be smooth on the unit square; it does not require the untransformed m_j(t,x) to decay. A function like m_j(t,x) = sin(2πt) satisfies Assumption 3.1, yet its sup-norm distance to any sieve approximant is at least |sin(2πt)|, because the approximant is zero at infinity. Proposition 3.1, which gives the O(c^{-m1j} + d^{-m2j}) bias rate, is therefore false as stated. Since that bias term is what Theorem 3.2's uniform consistency and Assumption 4.1's under-smoothing condition both rely on, the coverage claims in Theorems 4.1–4.3 are not justified under the stated assumptions. The fix is straightforward: either strengthen Assumption 3.1 to require generalized Schwartz-space decay, or restrict all uniform statements to compact sets and adjust the claims accordingly. Note that the authors do mention decay in Remark 3.1 as a sufficient condition, but it is not part of the assumption, so the paper is internally inconsistent as written. Two smaller soft spots: the condition in Theorem 4.1 that |r| is uniformly bounded below is asserted but not verified for the recommended bases, and several lemmas in the supplement are proved only 'by a discussion similar to,' which is a bit casual given how delicate the uniform Gaussian approximation is. My verdict: this deserves peer review — the gap is identifiable and fixable, and the framework is valuable — but the referee should insist on the corrected assumption or a compact-domain version before the results can be trusted as stated. I'd bring it to a reading group, not to cite it in its current form.","headline":"A serious, well-built framework for simultaneous inference in time-varying nonlinear regression, but the unbounded-domain approximation result has a genuine gap: the mapped sieve space only contains functions vanishing at infinity, and Assumption 3.1 doesn't require that.","tokens_in":62325,"tokens_out":2832,"would_cite":false,"duration_ms":30825,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62G08","62G05","62G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Mapped sieve bases let time-varying nonlinear regression be estimated and covered uniformly over [0,1]×R, with bootstrap-based simultaneous confidence regions.","keywords":["nonstationary time series","time-inhomogeneous nonlinear regression","sieve method","simultaneous inference","Gaussian approximation","multiplier bootstrap","locally stationary processes","mapped basis functions"],"falsifier":"Take a model in which one regression function does not decay as |x|→∞—for example m_j(t,x)=cos(x) or m_j(t,x)=2+sin(x)—so the mapped function is not smooth at the boundary of [0,1]. Then the claimed uniform approximation rate O($c^{{−m1j}}$+$d^{{−m2j}}$) should fail; a simulation could check whether the sup-norm error of the sieve estimator over a growing covariate interval [−L,L] fails to shrink as L grows, or whether the nominal 95% simultaneous coverage drops well below 95%.","tokens_in":61103,"feed_emoji":"📈","tokens_out":6299,"duration_ms":70138,"temperature":0.7,"pith_summary":"Time-varying nonlinear regression with locally stationary covariates is hard because the regression functions live on [0,1]×R, so classical sieve approximations on a compact box fail near the tails. This paper claims a mapped, hierarchical sieve expansion approximates such functions uniformly over the whole domain, and that a two-step OLS correction makes the resulting estimators uniformly consistent while restoring the centering condition that identifies each additive component. On top of that, the paper builds simultaneous confidence regions covering every time point and covariate value at asymptotic level 1−α, with a multiplier bootstrap that produces the critical values. If correct, this turns sieve estimation into a tool for structural testing—exact form, time-homogeneity, separability—without kernel boundary corrections or compact-support restrictions.","feed_headline":"Sieve method puts uniform bounds on drifting nonlinear regressions","feed_subtitle":"A fast two-step sieve plus multiplier bootstrap covers every time and covariate at level 1−α.","key_machinery":"The paper's workhorse is the mapped hierarchical sieve basis: a monotone map g(y;s) sends the unbounded covariate domain (R or R+) to [−1,1], and tensor products of an orthonormal basis in t with the mapped basis in x form the approximation space (3.8). Because the mapping is smooth, smoothness of m_j(t,x) in x becomes smoothness of em_j(t,y) on a compact square, so classical sieve approximation rates transfer to unbounded domains. A second OLS step estimates the time-varying mean shift χ_j(t)=E[m_{j,c,d}(t,X_{j,i})] and subtracts it from the pilot estimator, which enforces the identifiability condition (1.2); Gaussian approximation and volume-of-tubes devices then control the sup-norm of the estimation error and the critical value.","core_discovery":"The central claim is that for model (1.1), where each m_j is smooth and decays sufficiently fast as |x|→∞, the bias-corrected sieve estimators (3.21)-(3.22) are uniformly consistent over [0,1]×R, and the simultaneous confidence regions (4.7)/(4.28) have asymptotic coverage exactly 1−α. The proof rests on three technical pillars: a uniform approximation theorem (Proposition 3.1) for 2-D functions on unbounded domains via mapped sieve bases; two Gaussian approximation results for affine forms of high-dimensional locally stationary time series (Theorems L.2 and L.3); and a volume-of-tubes expansion (Theorem 4.2) for the critical value of the maximum of the resulting Gaussian field. The multiplier bootstrap (Theorem 4.3) makes the construction operational by approximating both the variance function h_j(t,x) and the critical value from one realization.","pith_inferences":["The same mapped-sieve and Gaussian-approximation machinery likely transfers to locally stationary nonlinear autoregressions with lagged covariates, which the paper notes requires only minor changes; this is our inference, not a theorem of the paper.","Because the approximation rate depends on how fast the mapped functions approach the boundary, choosing the map's scale parameter data-adaptively, rather than fixing s=1 by convention, could noticeably improve finite-sample coverage.","The volume-of-tubes formula is tailored to a Gaussian field indexed by a 2-D manifold; for higher-dimensional covariate vectors the same reasoning would require a higher-dimensional manifold and a different critical-value expansion, an extension the paper does not pursue.","The identifiability correction is estimated rather than imposed by design, which suggests the method may also work when covariates are mutually dependent, a setting where standard additive-model centering is harder to justify."],"forward_implications":["If the central claim holds, practitioners can construct simultaneous 1−α confidence regions for each time-varying regression function over the entire unbounded covariate range from a single observed time series.","Structural tests for time-invariance, multiplicative separability, and exact parametric form are valid at asymptotic level α and have power tending to 1 for deviations larger than the order of the band width.","The estimator achieves the optimal uniform rate O(n^{−1/2} log^3 n) when the regression functions are infinitely smooth and the sieve dimensions grow logarithmically with n.","The multiplier bootstrap procedure is theoretically sound and is implemented in an accompanying R package, so the method is ready for routine use.","The two Gaussian approximation results for affine forms of high-dimensional locally stationary time series are stated as having independent interest beyond this regression setting."],"supporting_citations":[{"why":"Supplies the sieve least-squares framework and compact-domain approximation rates that Proposition 3.1 extends to unbounded covariate domains.","marker":"[12]"},{"why":"Provides the mapped spectral basis construction for unbounded domains that defines the hierarchical sieve basis (3.8).","marker":"[60]"},{"why":"Presents the kernel Nadaraya-Watson estimator for locally stationary regression that serves as the main baseline and pointwise-inference target.","marker":"[70]"},{"why":"Develops kernel estimation and model selection for time-varying nonlinear regression under physical dependence, the model setting generalized here.","marker":"[80]"},{"why":"Gives the high-dimensional multiplier bootstrap for maxima of sums that Algorithm 1 adapts to dependent locally stationary data.","marker":"[16]"},{"why":"Introduces the blockwise multiplier bootstrap and minimum-volatility block-length selection used for the simultaneous bands.","marker":"[83]"},{"why":"Supplies the convex-set Gaussian approximation bound that underlies Theorems L.2 and L.3.","marker":"[30]"},{"why":"Gives the volume-of-tubes tail expansion for simultaneous confidence bands used to solve for cα in Theorem 4.2.","marker":"[63]"}],"fun_headline_variants":["Sieve uniform bounds for drifting nonlinear regressions","Uniform sieve inference for time-varying nonlinear regression","Bootstrap sieve confidence regions for nonstationary series","Time-varying regression sieves with exact asymptotic coverage","Sieve estimation and simultaneous inference for local stationarity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The foundation is Assumption 3.1: after mapping the covariate domain to [0,1], each regression function must be smooth with uniformly bounded derivatives, which in practice means the original functions decay rapidly as |x|→∞; if a true function does not decay at infinity, the approximation rate in Proposition 3.1 and everything built on it collapses.","fun_headline_variants_meta":{"raw":{"variants":["Sieve uniform bounds for drifting nonlinear regressions","Uniform sieve inference for time-varying nonlinear regression","Bootstrap sieve confidence regions for nonstationary series","Time-varying regression sieves with exact asymptotic coverage","Sieve estimation and simultaneous inference for local stationarity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1217,"prompt_tokens":910,"completion_tokens":307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":235}},"tokens_in":526,"tokens_out":307,"duration_ms":3670,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:51:01.043833+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a model in which one regression function does not decay as |x|→∞—for example m_j(t,x)=cos(x) or m_j(t,x)=2+sin(x)—so the mapped function is not smooth at the boundary of [0,1]. Then the claimed uniform approximation rate O($c^{{−m1j}}$+$d^{{−m2j}}$) should fail; a simulation could check whether the sup-norm error of the sieve estimator over a growing covariate interval [−L,L] fails to shrink as L grows, or whether the nominal 95% simultaneous coverage drops well below 95%.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sieve least-squares framework and compact-domain approximation rates that Proposition 3.1 extends to unbounded covariate domains."},{"cited_title":"and W ANG , L.-L","cited_arxiv_id":null,"evidence_quote":"Provides the mapped spectral basis construction for unbounded domains that defines the hierarchical sieve basis (3.8)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents the kernel Nadaraya-Watson estimator for locally stationary regression that serves as the main baseline and pointwise-inference target."},{"cited_title":"and W U, W","cited_arxiv_id":null,"evidence_quote":"Develops kernel estimation and model selection for time-varying nonlinear regression under physical dependence, the model setting generalized here."},{"cited_title":"and K ATO, K","cited_arxiv_id":null,"evidence_quote":"Gives the high-dimensional multiplier bootstrap for maxima of sums that Algorithm 1 adapts to dependent locally stationary data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the blockwise multiplier bootstrap and minimum-volatility block-length selection used for the simultaneous bands."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the convex-set Gaussian approximation bound that underlies Theorems L.2 and L.3."},{"cited_title":"and L OADER , C","cited_arxiv_id":null,"evidence_quote":"Gives the volume-of-tubes tail expansion for simultaneous confidence bands used to solve for cα in Theorem 4.2."}],"review_version":1}