{"id":"9aa8d570-e96b-4b76-aad3-b0fd9b085a31","arxiv_id":"2607.13170","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In random-design nonparametric regression with bounded density and constant conditional variance, the conjectured n^{-4s/(d+4s)} rate is unattainable for s>1, d>4s: the minimax root-mean-square risk is at least n^{-β} with β = [d(3s+1)+8s]/[(d+2s)(d+4)].","lead":"An open problem of James Robins asked whether a conjectured estimation rate for variance in random-design nonparametric regression holds for smooth regression functions. This paper proves it does not: the true minimax rate is strictly worse for every smoothness s>1 in dimension d>4s.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; main residual risk is unverified fixed-point construction in Proposition 3.8, but no concrete error found.","rationale":"The reader's weakest assumption correctly flags Proposition 3.8's fixed-point estimates as the most technically delicate load-bearing step. I did not find a concrete error; the key algebraic identities, scaling choices, and counting bounds are internally consistent. The disagreement is that the reader also emphasizes the varying-error-law limitation, which I regard as an explicit scope condition rather than a flaw in the stated theorem. My read therefore does not change the conditional verdict: the theorem is plausible but not fully verified, and independent verification of the Wiener-norm estimates would settle the residual risk.","tokens_in":39481,"tokens_out":33666,"duration_ms":297348,"concrete_test":"Independently verify Lemma 3.4 by implementing the localized kernel δJj,N for a representative setting (e.g., d=5, s=1.1, h=2^{-j}, N=2^j) and computing its truncated Fourier Wiener norm; confirm the bound scales as C c0 with C independent of h,N. If the scaling shows any dependence on h or an additional N factor, the contraction argument in Proposition 3.8 fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's Theorem 1.1 is a lower bound that relies on Le Cam's method with two priors. The construction hinges on Proposition 3.8, a contraction fixed point in a Wiener-norm ball that realizes the pair-density identity R(x,y)=E[p_T(x)p_T(y)]. I have checked the local moment identities (1.12)–(1.13), the two-point cancellation via Γ_N, the four-moment Wick identity in Lemma 3.10, the scaling choices in §2, and the component-counting bounds; all are internally consistent. The weakest point is Proposition 3.8: the O(c0)-Lipschitz estimates (3.81)–(3.82) and Lemma 3.4's N²-bound are stated with constants that depend only on fixed geometric data, but no independent or machine-checked verification is provided. If those estimates had a hidden dependence on h or N, the fixed point could fail and the lower-bound construction would collapse. This is a verification risk, not a demonstrated flaw.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the minimax rate for estimating a constant conditional variance σ² in nonparametric regression under random design. The class Θ allows an unknown design density bounded above and away from zero, an s-Hölder regression function, and conditional error laws that may depend on x but have mean zero, common variance σ², and uniformly bounded fourth moments. The main result, Theorem 1.1, states that for every s>1 and integer d>4s, the minimax root-mean-square risk satisfies R_n(Θ)^{1/2} ≥ c n^{-β} with β = [d(3s+1)+8s]/[(d+2s)(d+4)]. Since 4s/(d+4s) − β > 0, the conjectured rate n^{-4s/(d+4s)} is not uniformly attainable, and the minimax rate lies strictly between the regular-grid benchmark n^{-2s/d} and the conjectured rate. The proof uses Le Cam's two-prior method, constructing priors with variance gap Δ and nearly indistinguishable n-observation predictive laws. The construction combines periodic feature systems, a localized covariance dual, random design perturbations realized through a contraction fixed point in a Wiener-norm ball, and bounded coefficient vectors satisfying Gaussian fourth-moment identities. The full-sample Hellinger bound is obtained via a geometric-component decomposition and spanning-tree counting.","tokens_in":39511,"tokens_out":11525,"duration_ms":111423,"significance":"If the proof is correct, this is a substantial contribution: it resolves an open problem of Robins and shows that the random-design minimax rate is strictly worse than the conjectured close-pair rate. The construction is novel and technically deep, combining harmonic analysis, random design perturbations, and high-dimensional probability. The paper is careful and honest about its scope: it explicitly states in §1.2 that the lower-bound construction uses error laws Q_x that vary with x, and therefore does not settle the common-error-law submodel. This transparency is a strength. The exponent arithmetic and the cancellation structure in the proof are internally consistent; my own checks of (1.3), (1.4), (2.3)–(2.4), (2.8), the variance-gap matching Δ = κ²δ, the V₀-term cancellation in Lemma 3.12, and the dyadic-shell bound in Lemma 3.4 did not reveal an error. The paper does not ship machine-checked proofs or code, but the argument is self-contained and the constants are tracked with explicit dependence on c₀, h, and N.","major_comments":[{"comment":"The contraction fixed point is the linchpin of the whole construction: it alone realizes the pair-density identity R(x,y) = E[p_T(x)p_T(y)] (Eq. (3.74)). The proof of Proposition 3.8 derives the O(c₀)-Lipschitz estimates (3.81)–(3.82) by saying that summing (3.79) and (3.80) over indices j, taking the supremum in k, and applying (3.70)–(3.71) yields the result. This step is compressed: it implicitly uses the bounded-overlap constant D₀ from (3.28) and the fact that the number of j whose U_j¹ meets a fixed U_k² is uniformly bounded, but the constants and the summation are not shown. A hidden dependence on h, N, or c₀ in these estimates would invalidate the fixed point and hence the lower bound. I have not found a concrete error, but this is a load-bearing point and the proof should be expanded to make the Lipschitz constant explicit and independent of the chart index and scale.","section":"§3.3.5, Proposition 3.8 and Lemma 3.4"}],"minor_comments":[{"comment":"The sentence that the construction uses error laws Q_x that vary with x is important. I recommend adding a remark in the introduction after Theorem 1.1 stating explicitly that the common-error-law submodel remains open; currently this is only in the related-literature section.","section":"§1.2"},{"comment":"The dyadic-shell estimate (3.46) is stated with a brief justification. In particular, the behavior of the cutoff ϑ(N/t ‖R‖) in the intermediate range N/2 < t < 4N is only discussed verbally. Expanding this derivation would help the reader verify the uniform derivative bounds.","section":"§3.3.2, Lemma 3.4"},{"comment":"In (2.2), the exponent (4s−d)/(2(d+2s)) is negative exactly because d>4s. It would be helpful to point out this connection when introducing d>4s, since it also makes N_n → ∞ in (1.23).","section":"§2, proof of Theorem 1.1"},{"comment":"The notation p and p̄ for the lower and upper density bounds is easy to confuse, especially in displayed inequalities. Consider using p_min and p_max or adding a one-line reminder after (1.5).","section":"Notation"},{"comment":"The completeness argument for the ball B_{M*} is written in a long paragraph. It is correct but could be shortened or made into a separate lemma for readability.","section":"§3.3.5, Proposition 3.8"}],"recommendation":"major_revision","confidential_remarks":"The paper is very long and technically demanding. The main risk is the fixed-point construction in Proposition 3.8; I did not find a specific error, but the proof is compressed at the critical point where the Wloc-Lipschitz estimates are assembled. I recommend asking the authors to expand that part. The common-error-law limitation is honestly stated and should not be counted against the paper, but it should be highlighted in the introduction. The reference list appears appropriate. No concerns about authorship or citation patterns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proves a real negative answer: for s>1 and d>4s, the rate n^{-4s/(d+4s)} is not uniformly attainable under random design with a density bounded above and away from zero. The lower-bound exponent beta is strictly larger, and the gap formula checks out. This is new — Shen et al. needed vanishing design densities, and Robins et al.'s upper bound only covered 0<s<1. The variance-hiding mechanism, with conditional error laws varying in x, is the right wedge: it lets them match one-observation laws while shifting the variance, and the three-point channel makes the algebra explicit.\n\nWhat I verified by hand: the exponent arithmetic, the variance gap Delta = kappa^2 delta, the cancellation of V0 terms in R4, the Fourier estimate in Lemma 3.3, and the dyadic-shell bound ||L_k(delta Jj,N)||_W <= C c0. All consistent. The paper is honest about scope: it does not establish the lower bound for a common error law independent of X, and it says so in §1.2. The density band containing 1, the integer-s interpretation, and the finiteness of constants are all stated.\n\nThe soft spots are in proportion. The load-bearing piece is Proposition 3.8, the contraction fixed point realizing the pair-density identity. I have not machine-checked it, and I cannot certify the O(c0)-Lipschitz estimates (3.81)-(3.82) from reading alone. But there is no concrete error, and the structure is plausible: the Wiener-norm machinery is standard, and Lemma 3.4 is exactly the estimate needed to absorb the N^2 growth. The disclosure that LLMs suggested arguments and did exploratory checks does not, by itself, undermine confidence; the identities are checkable and the proof is written densely but honestly.\n\nFor a reader: this is a serious paper for anyone working on semiparametric minimax rates or quadratic functionals. The reader's conditional verdict is right. The residual risk is verification depth on a roughly 54-page proof; that is not a reason to desk-reject. If I were an editor, I would send it to a careful referee (ideally someone who knows Wiener norms) and expect heavy but bounded revision. It deserves referee time.\n\nRecommendation: engage.","headline":"A convincing negative answer to Robins' question, with an intricate construction whose main risk is verification depth rather than a found flaw.","tokens_in":40310,"tokens_out":1674,"would_cite":true,"duration_ms":123553,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that in random-design nonparametric regression with s-Hölder regression functions and dimension d>4s, the minimax root-mean-square risk for estimating a constant conditional variance is at least n^{-β} with β=[d(3s+1)+8s]/","keywords":["variance estimation","nonparametric regression","random design","minimax lower bounds","Hölder smoothness","quadratic functional","two-prior testing","semiparametric functional estimation"],"falsifier":"Check the contraction proof at the heart of the construction: on a discretized torus with small h and N=2, compute the map T in Proposition 3.8 at a small c0; if it fails to have a fixed point satisfying R(x,y)=E[p_T(x)p_T(y)], the cancellation identities fail and the lower bound collapses. Alternatively, any explicit estimator whose worst-case root-mean-square error is o(n^{-β}) on the stated class would refute the theorem.","tokens_in":39090,"feed_emoji":"📉","tokens_out":7224,"duration_ms":73207,"temperature":0.7,"pith_summary":"The paper settles a question about estimating a constant conditional variance when the covariate distribution is random and unknown. It proves that for regression functions with s Hölder smoothness and dimension d>4s, no estimator can achieve the previously conjectured root-mean-square rate n^{-4s/(d+4s)} uniformly over the model class. Instead the minimax risk is bounded below by n^{-β} with β=[d(3s+1)+8s]/[(d+2s)(d+4)], which lies strictly between the fixed-grid benchmark n^{-2s/d} and the conjecture. The proof builds two nearly indistinguishable data-generating processes with different error variances, hiding the variance change through specially tuned conditional error distributions and random design perturbations. The result matters because it closes a gap in the theory of semiparametric functional estimation: extra smoothness beyond one derivative is less useful under random design than was hoped.","feed_headline":"New lower bound blocks conjectured variance-estimation rate","feed_subtitle":"Specifies the gap: minimax risk scales as n^{-β}, strictly between the old conjecture and the grid rate.","key_machinery":"The proof uses a standard two-prior testing argument with two priors whose error variances differ by Δ but whose n-observation predictive laws are nearly indistinguishable. The central object is a three-point response distribution on {−a,0,a} whose probabilities are linear in q=f(x) and q²+σ², enabling exact cancellation of all one-observation differences through three moment identities. Random design perturbations are realized through a contraction fixed point: a provisional pair-density kernel R(x,y) is mapped to densities p_T satisfying E[p_T(x)p_T(y)]=R(x,y), via a projective decomposition into rank-one kernels and independent categorical variables. The load-bearing matrix identity is a","core_discovery":"The central claim is that the minimax root-mean-square risk of estimating σ² in this random-design regression model is at least c n^{-β} for every n, with β=[d(3s+1)+8s]/[(d+2s)(d+4)]. Equivalently, the conjectured rate n^{-4s/(d+4s)} is uniformly unattainable: the gap between the two exponents is d(d−4s)(s−1)/[(d+4s)(d+2s)(d+4)], which is positive for every s>1 and d>4s. The paper does not determine the exact minimax rate, only rules out the conjectured one, leaving the true rate somewhere between the fixed-grid benchmark n^{-2s/d} and the conjecture.","pith_inferences":["If the x-dependent error-law freedom is essential, then datasets with homogeneous noise may still admit the conjectured rate; the hardness may be driven by heteroscedasticity rather than by random design alone.","The pair-density fixed-point device is general: any exact cancellation requiring E[p_T(x)p_T(y)]=R(x,y) can be produced by the same contraction, so the technique may transfer to other quadratic semiparametric functionals.","The exponent β suggests a transition as d approaches 4s; a natural next step is to test numerically whether the lower bound is tight in moderate dimensions or to look for matching upper bounds.","Because the gap between the conjectured rate and β shrinks as s approaches 1, the phenomenon may be specific to genuine Hölder smoothness with s>1 rather than to Lipschitz-scale smoothness."],"forward_implications":["The conjectured n^{-4s/(d+4s)} rate is false for this entire model class: no estimator, however adaptive, attains it uniformly.","The true minimax rate lies strictly between n^{-2s/d} and n^{-4s/(d+4s)}; the paper's β gives a new lower bound that improves on the fixed-grid benchmark.","The obstruction is not an artifact of difference-based estimators; it applies to all measurable estimators of the variance.","The lower-bound construction does not apply when conditional error laws are forced to be identical across x; under a common error law independent of X, the question remains open.","For s≤1 or d≤4s, the argument gives no information, so the conjectured rate may still be attainable in those regimes."],"fun_headline_variants":["Variance estimation conjecture blocked by new lower bound","Robins' variance-rate conjecture fails under random design","Minimax lower bound refutes Robins' conjecture on variance","Variance estimation rate: conjecture not uniformly attainable","New lower bound: Robins' conjectured rate is unattainable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole argument hinges on allowing the noise distribution to be chosen afresh at every covariate value; if the conditional error laws had to be identical across the design, the indistinguishability construction is not known to work, and the paper explicitly does not claim the lower bound in that case.","fun_headline_variants_meta":{"raw":{"variants":["Variance estimation conjecture blocked by new lower bound","Robins' variance-rate conjecture fails under random design","Minimax lower bound refutes Robins' conjecture on variance","Variance estimation rate: conjecture not uniformly attainable","New lower bound: Robins' conjectured rate is unattainable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000497,"raw_usage":{"total_tokens":2246,"prompt_tokens":688,"completion_tokens":1558,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":1476}},"tokens_in":432,"tokens_out":1558,"duration_ms":13097,"temperature":1.0,"reasoning_tokens":1476,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:02:56.336720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the contraction proof at the heart of the construction: on a discretized torus with small h and N=2, compute the map T in Proposition 3.8 at a small c0; if it fails to have a fixed point satisfying R(x,y)=E[p_T(x)p_T(y)], the cancellation identities fail and the lower bound collapses. Alternatively, any explicit estimator whose worst-case root-mean-square error is o(n^{-β}) on the stated class would refute the theorem.","supporting_citations":[],"review_version":1}