{"id":"00999540-de54-4821-b668-ea0031631291","arxiv_id":"2507.00889","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Under covariate shift with target data near a smooth d-dimensional manifold in D dimensions, the minimax regression rate switches between a manifold-dominated and a noise-dominated regime at a threshold set by source size, target size, smoothness, and d versus D.","lead":"This paper derives minimax-optimal rates for transfer learning when target data lie near a low-dimensional manifold inside a larger source domain. It shows a phase transition driven by the distance of the target data to the manifold, and it proposes a local polynomial estimator that achieves the optimal rates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Small-ρ achievability by LPR is unsupported when target samples dominate: Lemma A.2's uniform eigenvalue lower bound fails at ρ_n=o(h_n) because the ambient polynomial design restricted to a d-manifold is nearly rank-deficient.","rationale":"The reader's weakest_assumption (single global chart) is less central: pointwise estimation only needs a local chart around x⋆, and the claimed 'minor bookkeeping' extension is plausible. The L1/L2 mismatch in Theorem 2.3 is real but concerns an auxiliary adaptive result. The design degeneracy in Lemma A.2 attacks the main upper-bound theorem in the small-ρ regime. It is not an external-consensus disagreement; it is an internal proof failure: the asserted uniform lower bound on λ_min(S_n) is false when target samples dominate and ρ_n/h_n → 0. This is exactly the regime where the rate becomes the intrinsic manifold rate, so the paper's claim that one LPR estimator attains the rate in both phases is at risk. The minimax rates themselves could survive with an intrinsic-dimensional estimator or with a source-dominance condition, so a conditional verdict remains appropriate, but the revision should address this gap.","tokens_in":42154,"tokens_out":22986,"duration_ms":281351,"concrete_test":"Analytic/simulation check: set nP=0, D=2, d=1, ϕ(v)=(v,v^2), ρ_n=0, x⋆=0, β=1.5 (ℓ=1), box kernel, and h_n = n_Q^{-1/(2β+1)}. Compute λ_min(S_n) and the MSE of the LPR estimator (2.4)-(2.6). Lemma A.2 requires λ_min(S_n) ≥ m; the calculation gives λ_min(S_n) ≍ h_n^2, and the simulated MSE scales like n_Q^{-1} h_n^{-5}, which is slower than h_n^{2β} ≍ n_Q^{-2β/(2β+1)}. Repeating with nP large enough that nP ≥ C n_Q h_n^{d-D} should restore the claimed rate, isolating the target-dominated regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma A.2 is the keystone of the Theorem 2.1 upper bound; it asserts λ_min(S_n) ≥ m uniformly in n. In case (ii) of its proof (ρ_n ≤ h_n), the argument reduces to a uniform lower bound on λ_min(B_n), where B_n is constructed from z(∇ϕ(0)s + t_n(u−u0)) with t_n = ρ_n/h_n. In the small-ρ phase, h_n is chosen so that ρ_n ≪ h_n, hence t_n → 0. For d < D and ℓ ≥ 1, the polynomial vector z(∇ϕ(0)s) spans at most N_d = binom(d+ℓ,ℓ) directions, strictly fewer than dim z = binom(D+ℓ,ℓ); the normal directions enter only through t_n(u−u0). Consequently λ_min(B_n) → 0 (for ρ_n=0 it is zero in the linearized calculation; with curvature restored the actual design eigenvalue is O(h_n^2)). When target samples dominate—in particular when nP=0—S_n converges to B_n, so the LPR design is nearly singular. The variance bound 1/(nψ_n h_n^D) used in Theorem 2.1 therefore does not follow; the true variance carries extra factors such as (ρ_n/h_n)^{-4} or h_n^{-4}. Thus the claimed attainment of the small-ρ rate by LPR, including the exact-manifold target-only rate in Remark 2.1, is unproven. For d=1, D=2, ϕ(v)=(v,v^2), ρ_n=0, β=1.5, the estimator does not achieve n_Q^{-2β/(2β+1)}. The minimax rates may still be correct, but the upper-bound theorem as stated has a real proof gap.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies pointwise estimation of a beta-Holder regression function at a target point under covariate shift, where the target covariates are generated as X = phi(V) + rho_n U with phi a d-dimensional chart in R^D and the source covariates admit a D-dimensional density. The main results, Theorems 2.1 and 2.2, assert that the minimax squared error is (n_P + n_Q rho_n^{d-D})^{-2 beta/(2 beta + D)} when rho_n is above a threshold and (n_P^{(2 beta + d)/(2 beta + D)} + n_Q)^{-2 beta/(2 beta + d)} when rho_n is below it, and that a single D-variate local polynomial regression estimator attains both rates. Theorem 2.3 and Corollary 2.1 propose a fully adaptive version using dimension estimation and Lepski's method. Numerical experiments illustrate the benefit of transfer when the target lies on a low-dimensional manifold inside the source support.","tokens_in":42530,"tokens_out":12160,"duration_ms":142747,"significance":"The claimed phase transition is a natural and potentially important refinement of existing transfer-learning and manifold-regression results: it extends Pathak et al. (2022) from beta in (0,1] and rho_n=0 to general beta>0 and rho_n>0, and Remark 2.1's target-only noisy-manifold rates appear new. The paper is largely self-contained and the lower-bound strategy via Le Cam alternatives is appropriately connected to the phases. However, the upper-bound proof has a rank-deficiency gap in the small-rho regime, the adaptivity theorem's statement and proof are mismatched, and the lower-bound argument invokes an unstated Gaussian likelihood assumption. These are substantial but localized correctness issues; if repaired, the contribution would be significant for the nonparametric transfer-learning literature.","major_comments":[{"comment":"The uniform lower bound lambda_min(B_n) >= m > 0 is not true when rho_n/h_n -> 0. For d < D and l >= 1, the polynomial vector z(nabla phi(0) s) spans only binom(d+l,l) directions, strictly fewer than dim z = binom(D+l,l); the normal directions enter only through t_n(u-u_0) with t_n = rho_n/h_n -> 0, so lambda_min(B_n) -> 0. In the exactly singular case rho_n=0 the linearized design is rank-deficient, and with curvature restored the design eigenvalue is O(h_n^2). Consequently, when target samples dominate (notably n_P=0 in Remark 2.1), the variance bound 1/(n_P h_n^D + n_Q (rho_n vee h_n)^{d-D} h_n^D) used in Theorem 2.1 does not follow, and the claimed attainment of the small-rho rate by the ambient LPR estimator is unproven. This is load-bearing for the upper half of the central claim.","section":"Appendix A, Lemma A.2, Case (ii)"},{"comment":"Theorem 2.3 states E[|f_adp(x_star)-f_star(x_star)|^2 / psi_n(beta)] <= C with psi_n(beta) equal to the squared rate (up to log factors). The proof in Appendix C, however, proves only E[|f_adp(x_star)-f_star(x_star)| / delta_{n,i_star}] <= C, where delta_{n,i_star} = h_{n,i_star}^{beta}; an L1 bound does not imply the displayed L2 bound. Additionally, the Lepski threshold is written as C_l h_{n,eta} in the theorem statement and Algorithm 1, but as C_l delta_{n,i} = C_l h_{n,i}^{beta_i} in the proof; these two thresholds are not equivalent for eta != 1. Please correct the statement or give an L2 proof.","section":"Section 2.4, Theorem 2.3 and Appendix C"},{"comment":"The proof uses KL(P_0|P_1) = E_P[(f_0 - f_1)^2] and the analogous identity for Q. This identity is specific to Gaussian conditional likelihoods with equal variance; Assumption 3 only bounds fourth moments and does not specify the noise distribution. As written, the minimax lower bound in Theorem 2.2 is not derived for the stated model. Please add the noise model used for the hard instances (e.g., Y = f(X) + epsilon with epsilon ~ N(0, sigma^2)) or replace the KL calculation by a valid total-variation/Hellinger bound under the stated moment assumptions.","section":"Appendix B, Eqs. (B.3)-(B.4)"},{"comment":"The model assumes a single global chart phi:[-1,1]^d -> [-1,1]^D satisfying (2.2), and the text after (2.2) asserts without proof that all results extend to a finite local-chart setting 'with minor bookkeeping'. Since the target point x_star need not lie in the range of a single global chart, and since the local-chart setting changes constants and boundary terms in Lemmas A.1 and A.2, this is not an entirely formal matter. Please either state and prove a local-chart version or explicitly restrict the claims to the global-chart case.","section":"Section 2.1, Assumption 2(2)"}],"minor_comments":[{"comment":"In the final display, the bound h_n^{2 beta} <~ (n_P + n_Q rho_n^{d-D})^{-2 beta/(2 beta + d)} should have denominator 2 beta + D to be consistent with Eq. (2.8); as printed it is dimensionally inconsistent with the chosen h_n.","section":"Appendix A, proof of Theorem 2.1"},{"comment":"The candidate set B is written as {0 ≡ beta_0 < beta_min ≡ beta_1 < …}, but beta_0 = 0 is never used in the bandwidth formula (C.1) and would be undefined there; remove beta_0 or define its treatment.","section":"Section 2.4 and Appendix C"},{"comment":"The descriptions of Figures 2-4 are informal ('performs on par', 'beneficial'); adding error bars or confidence intervals and stating the number of Monte Carlo repetitions in the captions would improve reproducibility.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The rank-deficiency concern in Case (ii) of Lemma A.2 is the main correctness risk; it affects the proposed estimator's claimed attainment of the small-rho rate, not necessarily the minimax rates themselves. The KL identity issue in the lower bound is a similarly essential but localized correction. I would not recommend rejection because the phase-transition formulation is promising and the lower-bound framework could survive with a specified noise model or a different divergence bound. The revision should be substantial, however, because the upper-bound proof needs a different design or regularization analysis for the small-rho regime."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious read: it identifies a genuinely new phase transition in minimax rates for covariate shift when the target distribution sits within rho_n of a d-manifold in R^D, for general Holder beta > 0. The lower bound construction is standard but clean, and the rates in Theorems 2.1 and 2.2 match. The threshold kappa_n* is the right object. Credit also for situating the result against Pathak et al. and Kpotufe-Martinet.\n\nThe soft spots are real but not evenly distributed.\n\nThe biggest one: the upper bound proof appears to have a load-bearing gap in the small-rho regime when target samples dominate. Lemma A.2 claims a uniform lower bound on lambda_min(S_n). In case (ii) (rho_n <= h_n), the argument rests on the polynomial z(grad phi(0) s + (rho_n/h_n)(u-u0)) being nonzero on a set of positive measure. That only gives positivity for fixed n. As rho_n/h_n -> 0, the normal-direction components of the design are of order O((rho_n/h_n)^2) or O(h_n^2) (curvature terms), so the minimum eigenvalue tends to 0. For nP = 0, the claimed target-only rate n_Q^{-2 beta/(2 beta + d)} is not attained by the plain ambient LPR estimator; the variance picks up a factor like h_n^{-4}. I think this is not a minor technicality: it affects the achievability of the very rates the paper advertises when the target dominates. The lower bound may still be correct, and the rates might be attained by an estimator that does local PCA or otherwise adapts to the tangent plane, but that is not in the paper.\n\nA smaller issue: Theorem 2.3 is stated as an L2 bound normalized by the squared rate, but the proof in Appendix C delivers an L1 bound. The squared statement does not follow from the displayed argument. This is likely fixable, but as written it is a gap.\n\nAlso: the global-chart assumption in Assumption 2(2) is asserted to be relaxable to local charts without proof. That is a fair caveat, and I would want the bookkeeping or a citation. The experiments are not reproducible from the preprint (no code, seeds, or data).\n\nOverall: the minimax characterization is important enough that I would not desk reject. The right outcome is a major revision: fix the design-matrix argument (or change the estimator), prove the adaptive result at the stated L2 risk, and address the chart issue. If the achievability can be repaired, this will be a strong paper in nonparametrics and transfer learning.\n\nMy recommendation: send to a serious referee with the expectation of heavy revision.","headline":"New minimax phase transition with a solid lower bound, but the LPR upper bound has a real gap in the small-rho target-dominated regime.","tokens_in":43069,"tokens_out":11468,"would_cite":false,"duration_ms":135479,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G08","62C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The pointwise minimax squared error in covariate-shift regression switches between two rates at a single manifold-noise threshold.","keywords":["covariate shift","approximate manifold","minimax rates","phase transition","local polynomial regression","transfer learning","Hölder smoothness","adaptive estimation"],"falsifier":"On a synthetic $d=1$ circle (or $d=2$ surface) embedded in $\\mathbb{R}^D$, generate target covariates with noise level $\\rho_n$ scaling as $n_Q^{-a}$ for a grid of exponents $a$ around $1/(2\\beta+d)$, run the local polynomial estimator with the two bandwidth rules, and compare the empirical MSE slopes to the two predicted rates; a mismatch in the exponent or in the exponent at which the slopes switch would refute the phase-transition claim.","tokens_in":41936,"feed_emoji":"📉","tokens_out":9159,"duration_ms":92394,"temperature":0.7,"pith_summary":"Under covariate shift, the paper derives the exact pointwise minimax rate for estimating a $\\beta$-Hölder regression function at a target point when the target covariates lie near, but not exactly on, a $d$-dimensional manifold inside the source support in $\\mathbb{R}^D$. Because the target measure is supported on a thin neighborhood of the manifold, a density ratio with the source need not exist, and the analysis replaces the density ratio with the geometric distance $\\rho_n$. The minimax rate changes sharply at the threshold $\\kappa_n^\\star = (n_P^{(2\\beta+d)/(2\\beta+D)} + n_Q)^{-1/(2\\beta+d)}$: for $\\rho_n$ above it the ambient dimension $D$ controls the rate, while for $\\rho_n$ below it the intrinsic dimension $d$ controls the rate and the source sample contributes like $n_P^{(2\\beta+d)/(2\\beta+D)}$ extra target points. A local polynomial regression estimator attains both rates, and an adaptive version attains them up to logarithmic factors when $\\beta$ and $d$ are unknown.","feed_headline":"Noise threshold flips covariate-shift minimax rates","feed_subtitle":"When target data hug a manifold, source samples multiply the effective sample size; past a small noise level they don't.","key_machinery":"The load-bearing object is the $\\rho_n$-approximate manifold $M_{\\rho_n} = \\{x \\in [-1,1]^D : \\inf_z \\|x - \\phi(z)\\| \\le \\rho_n\\}$, generated as $X = \\phi(V) + \\rho_n U$ with $\\phi:[-1,1]^d \\to [-1,1]^D$ bi-Lipschitz and with densities of $V$ and $U$ bounded below. The proof runs through the local polynomial regression normalizer $\\psi_n = n_P/n + (n_Q/n)(\\rho_n \\vee h_n)^{d-D}$ and a family of integral bounds (Lemma A.1) that count how much kernel mass the source and target place in a bandwidth-$h_n$ cube around $x^\\star$. These bounds make the target contribution equal to a $d$-dimensional mass when $h_n \\ge \\rho_n$ and to $\\rho_n^{d-D}$ times a $D$-dimensional mass when $h_n \\le \\rho_n$. Balancing the bias term $h_n^{2\\beta}$ against the variance $[n_P h_n^D + n_Q(\\rho_n \\vee h_n)^{d-D} h_n^D]^{-1}$ yields exactly the two rate regimes and the threshold $\\kappa_n^\\star$.","core_discovery":"On its own terms, the paper's central claim is that the pointwise minimax squared error at a target point $x^\\star$ is $(n_P + n_Q \\rho_n^{d-D})^{-2\\beta/(2\\beta+D)}$ when $\\rho_n$ is above $\\kappa_n^\\star$, and $(n_P^{(2\\beta+d)/(2\\beta+D)} + n_Q)^{-2\\beta/(2\\beta+d)}$ when $\\rho_n$ is below it. In the first regime the target samples behave like $n_Q\\rho_n^{d-D}$ equivalent points in the ambient space, so they join the source samples in a $D$-dimensional rate. In the second regime the target covariates behave as if they lay exactly on the manifold, so the problem is $d$-dimensional and the source sample inflates $n_Q$ to an effective size $n_P^{(2\\beta+d)/(2\\beta+D)} + n_Q$. Theorems 2.1 and 2.2 prove, respectively, that a local polynomial regression estimator with the bandwidth in (2.8) attains these rates and that no estimator can do better. A lower bound is obtained by a two-point Le Cam construction using localized Hölder bumps at scale $h_n$.","pith_inferences":["A natural next test is whether multi-chart manifolds shift the threshold by chart-overlap constants; the paper asserts but does not prove this extension.","The mechanism suggests that a direct estimator of the local noise level $\\rho_n$, obtained from target points' distances to a fitted manifold, could be plugged into the bandwidth rule and make the phase transition visible in practice.","The same bias-variance split should apply to other local estimators such as kernel regression, $k$-nearest neighbors, and neural networks, giving the same threshold but possibly different constants.","For target distributions that are mixtures of approximate manifolds at different distances, one would expect each component to have its own $\\rho$ and its own threshold, an extension not considered in the paper."],"forward_implications":["Source samples provably increase the effective target sample size to $n_P^{(2\\beta+d)/(2\\beta+D)} + n_Q$ when $\\rho_n$ is below the threshold, which is how transfer helps without a density ratio.","Above the threshold, transfer still helps, but through the combination $n_P + n_Q\\rho_n^{d-D}$ rather than through the raw target count.","Setting $n_P=0$ gives new minimax rates for noisy-manifold regression, including a phase transition at $\\rho_n \\asymp n_Q^{-1/(2\\beta+d)}$.","The pooled local polynomial estimator does not need to know which observations are source or target, because one bandwidth is used for both samples.","Combining a nearest-neighbor intrinsic-dimension estimator with Lepski's method yields nearly optimal rates when $\\beta$ and $d$ are unknown."],"supporting_citations":[{"why":"supplies the target-only optimal manifold rate that the $\\rho_n \\ll \\kappa_n^\\star$ regime matches and extends.","marker":"Bickel and Li (2007)"},{"why":"establishes the $\\rho_n=0$ minimax rate for $\\beta \\le 1$ that this paper generalizes to $\\beta>0$ and $\\rho_n>0$.","marker":"Pathak et al. (2022)"},{"why":"introduces the transfer-exponent framework and a $k$-NN estimator whose limited smoothness regime is extended here.","marker":"Kpotufe and Martinet (2021)"},{"why":"provides the local polynomial regression machinery whose bias-variance tradeoff attains the upper bounds.","marker":"Fan et al. (1995)"},{"why":"supplies the manifold-adaptive intrinsic dimension estimator used in the adaptive procedure.","marker":"Farahmand et al. (2007)"},{"why":"supplies the Lepski method for adaptive smoothness selection.","marker":"Lepski and Spokoiny (1997)"},{"why":"uses the same approximate-manifold model and motivates the $\\rho$-neighborhood definition.","marker":"Jiao et al. (2023)"},{"why":"defines the Hölder classes and the Le Cam/Pinsker lower-bound toolkit used for minimaxity.","marker":"Tsybakov (2009)"}],"fun_headline_variants":["Manifold gap flips minimax rate in covariate shift","When target hugs manifold, source samples multiply","Phase transition in transfer learning: distance matters","Source data help past a threshold, not below it","Covariate shift: minimax rate jumps at a critical distance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorems assume a single global coordinate chart $\\phi$ covering the whole target manifold with uniform bounds on $\\nabla\\phi^\\top\\nabla\\phi$ and with target covariates generated as $X=\\phi(V)+\\rho_n U$; if real target manifolds require several charts or have strongly varying curvature, the stated rates and threshold are not proven and could fail.","fun_headline_variants_meta":{"raw":{"variants":["Manifold gap flips minimax rate in covariate shift","When target hugs manifold, source samples multiply","Phase transition in transfer learning: distance matters","Source data help past a threshold, not below it","Covariate shift: minimax rate jumps at a critical distance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1728,"prompt_tokens":1077,"completion_tokens":651,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":573}},"tokens_in":693,"tokens_out":651,"duration_ms":7521,"temperature":1.0,"reasoning_tokens":573,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:06:37.713980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic $d=1$ circle (or $d=2$ surface) embedded in $\\mathbb{R}^D$, generate target covariates with noise level $\\rho_n$ scaling as $n_Q^{-a}$ for a grid of exponents $a$ around $1/(2\\beta+d)$, run the local polynomial estimator with the two bandwidth rules, and compare the empirical MSE slopes to the two predicted rates; a mismatch in the exponent or in the exponent at which the slopes switch would refute the phase-transition claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the target-only optimal manifold rate that the $\\rho_n \\ll \\kappa_n^\\star$ regime matches and extends."},{"cited_title":"Ma, and M","cited_arxiv_id":null,"evidence_quote":"establishes the $\\rho_n=0$ minimax rate for $\\beta \\le 1$ that this paper generalizes to $\\beta>0$ and $\\rho_n>0$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the transfer-exponent framework and a $k$-NN estimator whose limited smoothness regime is extended here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the local polynomial regression machinery whose bias-variance tradeoff attains the upper bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the manifold-adaptive intrinsic dimension estimator used in the adaptive procedure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Lepski method for adaptive smoothness selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"uses the same approximate-manifold model and motivates the $\\rho$-neighborhood definition."},{"cited_title":"(2009, 01)","cited_arxiv_id":null,"evidence_quote":"defines the Hölder classes and the Le Cam/Pinsker lower-bound toolkit used for minimaxity."}],"review_version":1}