{"id":"498beece-779f-4d1b-beb7-5eda07d62040","arxiv_id":"2608.02539","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Under local-to-target asymptotics, √n times the ridge estimation error is approximately Gaussian with mean −λ(Σ+λI)^{-1}b and variance (Σ+λI)^{-1}Ω(Σ+λI)^{-1}, yielding closed-form tuning rules for isotropic features.","lead":"This paper derives a simple Gaussian approximation to the distribution of the ridge regression estimator when the penalty grows with the sample size and the true coefficients lie near the shrinkage target, allowing heteroskedasticity and autocorrelation. It uses the approximation to propose new rules for choosing the ridge penalty by minimizing worst-case or average excess prediction risk.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Local-to-reference assumption (Eq. 14) is load-bearing: if it fails, Theorem 1's bias term diverges and Theorem 3's tuning rules, which depend on unidentifiable b, lose justification; simulations calibrate B oracularly.","rationale":"The paper proves Theorem 1 correctly: the algebra in (30) plus Assumption 1 gives the stated normal limit. Theorems 2 and 3 are valid derivations from that limit. The paper is candid that b is not identified (Section 3.3). The central weakness is not a proof error but a gap between the asymptotics and the advertised 'data-driven' tuning rules: the rules depend on B or π, which are user-specified and not estimable. In the simulations B is set to the true norm or calibrated from the same DGP that generates the coefficients, so the comparison with LOO-CV is favorable to the proposed rule. A sensitivity analysis varying B would show the extent of the dependence. This does not invalidate Theorem 1, but it tempers the practical claim. The reader's CONDITIONAL verdict is appropriate; I recommend no change.","tokens_in":24048,"tokens_out":12074,"duration_ms":112468,"concrete_test":"Re-run the DGP-1 simulation (Section 4.1) with the minimax radius B set to 0.5×, 1×, and 2× the true value ||b||=√10, keeping all other settings fixed, and report the excess risk of minimax ridge relative to LOO-CV at n=500 and n=3000. If the minimax advantage disappears or reverses for misspecified B, the Figure 2 comparison depends on oracle knowledge of b, confirming that the tuning-rule claim is not fully data-driven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core approximation is Theorem 1: √n(bβλn − β_n) → N(−λ(Σ+λI)^{-1}b, (Σ+λI)^{-1}Ω(Σ+λI)^{-1}). The bias term is finite only if b = lim √n(β_n−β_0) is finite. If √n(β_n−β_0) diverges, the second term in (30), (λ_n/n)√n(β_n−β_0), diverges and the Gaussian approximation (and the finite-sample version (4)) ceases to hold. This restricts the theory to coefficients within O(1/√n) of the shrinkage target β_0 — an untestable restriction on the unknown DGP. Moreover, b is not consistently estimable, yet Theorem 3's tuning rules require it: λ*_minimax = n·trace(Ω)/(σ_x²B²) and λ*_π = n·trace(Ω)/(σ_x²Eπ[b'b]). The simulations set B equal to the true ||b|| in DGP-1 (B=√10) and calibrate B from the coefficient-generating scheme in DGP-2/3, i.e., using oracle knowledge. Thus the reported ~20% risk improvement over LOO-CV in Figure 2 is not a fully data-driven result; it depends on a user-supplied bound/prior that the data cannot identify. The mathematical claim in Theorem 1 is correct under its assumptions, but the practical claim that the proposed selection rules are data-driven and preferable to CV is not established without a sensitivity analysis for B/π.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Gaussian approximation to the finite-sample distribution of the ridge regression estimator under high-level assumptions that allow heteroskedasticity and autocorrelation. The key asymptotic setup lets the regularization parameter grow with the sample size, λ_n/n →_p λ, and treats the true coefficient as local to the shrinkage reference vector, √n(β_n−β_0)→b. Theorem 1 derives the limiting distribution of √n(β̂_{λ_n}−β_n). Theorem 2 uses this to approximate the excess prediction risk, and Theorem 3 gives closed-form tuning rules for isotropic features: λ* = n·trace(Ω)/(σ_x²E_π[b'b]) under a prior and λ* = n·trace(Ω)/(σ_x²B²) under a norm bound. Simulation evidence compares minimax ridge with LOO-CV and OLS in three DGPs.","tokens_in":24460,"tokens_out":6180,"duration_ms":70125,"significance":"If the results hold, the paper provides a clean and useful generalization of Knight and Fu (2000) to heteroskedastic and autocorrelated environments and to an arbitrary reference vector. The elementary proof of Theorem 1, the transparent use of local asymptotics, and the explicit admission that b is not identified are strengths. The risk approximation and the closed-form tuning rules are new and potentially useful in low-dimensional problems. The paper is also honest about the limitation that the tuning rules require user-specified prior/bound information. However, the empirical support for the proposed tuning rules is weakened by the oracle-based calibration of B in the simulations, so the practical advantage over LOO-CV is not yet established.","major_comments":[{"comment":"The tuning rules depend critically on the unidentifiable local parameter b, through either the bound B or the prior π. The paper is candid that b cannot be estimated, but it nevertheless describes the strategies as 'data-driven recommendations.' As stated, the data enter only through Ω̂ and σ̂_x²; the key tuning input B or E_π[b'b] is user-specified. More importantly, if the true b violates the bound or lies in a low-probability region of π, the selected λ can be far from the risk minimizer even though Theorem 1 still holds. The authors should add a sensitivity analysis for B and π and provide practical guidance on how to elicit these inputs, otherwise the prescriptive claim is incomplete.","section":"§3.3, Eqs. (20)–(25); Theorem 3"},{"comment":"The simulation evaluation of the minimax rule uses oracle knowledge of the true DGP to calibrate B. In DGP-1, B is set to the true ||b||=√10; in DGP-2 and DGP-3, B is the 90th percentile of √n||β_n|| computed from the same coefficient-generating scheme. LOO-CV receives no such oracle input. The reported ~20% improvement over LOO-CV in Figure 2 and the competitiveness in Figures 5 and 8 therefore reflect an oracle advantage, not a fully data-driven comparison. Please report risk as a function of B (e.g., multiples of a data-based estimate) and clarify what a practitioner must supply. This does not affect the validity of Theorem 1, but it is essential for the paper's practical claims.","section":"§4.1–4.3, Figures 2, 5, and 8"}],"minor_comments":[{"comment":"The error term is scalar, so 'ϵ_i ∼ N_k(0, σ²)' should be 'ϵ_i ∼ N(0, σ²)'.","section":"§4.1, first paragraph"},{"comment":"The text notes that DGP-2 and DGP-3 depart from Assumption 2 (strict stationarity) by allowing heterogeneous variances. Since Theorem 2 uses Assumption 2, this means the simulations do not exactly satisfy the theorem's conditions. The paper should state more clearly whether these simulations are intended as a robustness check outside the formal assumptions.","section":"§4.2–4.3"},{"comment":"The horizontal-axis notation is garbled: 'n( 1 n, 1)' should be something like '√n(β̂₁−β_{n,1})'. Please fix the typesetting.","section":"Figures 1, 4, and 7"},{"comment":"The uniform integrability condition is high-level. Providing primitive sufficient conditions, such as moment bounds on the design and errors and boundedness of λ_n/n, would make the theorem easier to verify in applications.","section":"Theorem 2, condition (iii)"},{"comment":"The estimator is denoted with λ_n in (3) and with λ in (13), and the text later says 'with λ replaced by λ_n.' Please unify the notation to avoid confusion.","section":"Eq. (3) vs. Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The central theoretical contribution is sound: Theorem 1 is correct under its stated assumptions and the paper is transparent about the non-identification of b. My recommendation is major revision rather than reject because the main gap is in the simulation-based evaluation of the tuning rules and the lack of sensitivity analysis, both of which can be fixed within the manuscript's scope. The paper fits the journal well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read for Theorem 1. Under λ_n/n → λ and √n(β_n − β_0) → b, they get a Gaussian approximation for ridge with general HAC dependence and an arbitrary shrinkage target. That genuinely generalizes Knight–Fu, and the proof is elementary. The risk approximation and the closed-form tuning rules for isotropic features are also new, though they reduce to known formulas in the homoskedastic case. The paper is honest about scope: low-dimensional only, and they don't claim the high-dimensional simulation (DGP-3) as a success—LOO-CV does slightly better there.\n\nThe soft spots are real, but they are exactly where the reader's stress-test lands. The local-to-reference assumption is load-bearing: if √n(β_n − β_0) diverges, the bias diverges and the Gaussian approximation breaks. That's an untestable restriction on the unknown DGP. More importantly, the tuning rules depend on b, which is not identified, through user-supplied B or a prior. The simulations calibrate B with oracle knowledge: B = ||b|| in DGP-1, and B is the 90th percentile of √n||β_n|| from the coefficient-generating scheme in DGP-2/3. So the reported ~20% risk improvement over LOO-CV in Figure 2 is not a fully data-driven result. That doesn't invalidate Theorem 1, but it does mean the paper's practical claim—that these selection rules beat CV—is not yet established. A sensitivity analysis for B/π and a data-driven calibration would go a long way.\n\nMinor: no code shipped, which makes the simulations harder to check, and the plug-in step in Theorem 2 is plausible but not uniformly quantified. These are fixable.\n\nSo: the distributional approximation is solid and useful for econometric theory. The tuning-rule evidence is overstated relative to what a data-driven user would get. I'd send it to peer review and ask for sensitivity analysis and code. It's also a fine reading-group paper—there's a clean theorem and a clear discussion of what local asymptotics can and cannot buy you.","headline":"Theorem 1 is a clean, correct generalization of Knight–Fu to HAC errors; the tuning-rule story has an unidentifiability problem that the simulations paper over.","tokens_in":24944,"tokens_out":1724,"would_cite":true,"duration_ms":19626,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07","62E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Ridge regression's finite-sample distribution is a simple Gaussian whose bias and variance follow from the penalty and the shrinkage target.","keywords":["ridge regression","distributional approximation","nonstandard asymptotics","local-to-reference","bias-variance tradeoff","prediction risk","regularization parameter selection","heteroskedasticity and autocorrelation"],"falsifier":"Run a Monte Carlo with k fixed, x_i and epsilon_i satisfying Assumption 1, beta_n = beta_0 + b/sqrt(n), and lambda_n/n -> lambda; compare the empirical distribution of sqrt(n)(beta_hat_lambda_n - beta_n) to N(-lambda(Sigma + lambda I)^{-1}b, (Sigma + lambda I)^{-1}Omega(Sigma + lambda I)^{-1}). If the first two moments deviate systematically beyond Monte Carlo error, Theorem 1 is refuted. Setting beta fixed away from beta_0 instead should make the bias grow with sqrt(n), confirming that the local assumption is load-bearing.","tokens_in":23933,"feed_emoji":"⚖️","tokens_out":6007,"duration_ms":64505,"temperature":0.7,"pith_summary":"The paper aims to show that, in low-dimensional regressions with heteroskedastic or autocorrelated errors, the ridge estimator's distribution is well approximated by a Gaussian with an explicit bias and a sandwich-form variance. The approximation comes from two nonstandard asymptotics: the penalty grows proportionally to the sample size, and the true coefficients are local to the shrinkage reference vector. On top of this distributional result, the paper derives an approximation to the excess prediction risk of ridge, and then two closed-form rules for choosing the penalty: one that minimizes worst-case risk given a bound on the local parameter, and one that minimizes average risk given a prior. The paper is candid that the local parameter b is not consistently estimable, so the tuning rules require user-supplied prior or bound. If the approximation holds, it gives practitioners a way to think about ridge's bias-variance tradeoff without cross-validation or Gaussian-error assumptions.","feed_headline":"One Gaussian formula describes ridge's finite-sample distribution","feed_subtitle":"Penalty growing with n and coefficients near the target yield closed-form tuning rules.","key_machinery":"The central device is the joint rescaling lambda_n/n -> lambda and beta_n = beta_0 + b/sqrt(n) + o(1/sqrt(n)), which keeps shrinkage bias and sampling noise at the same order. The Gaussian law in Theorem 1 is the engine: its mean term -lambda(Sigma + lambda I)^{-1}b captures misspecification of the shrinkage target, and its covariance term (Sigma + lambda I)^{-1}Omega(Sigma + lambda I)^{-1} carries heteroskedasticity and autocorrelation through Omega. This approximation then feeds directly into the excess-risk function and the closed-form tuning rules.","core_discovery":"Under Assumption 1 (convergence of the design second-moment matrix Sigma and asymptotic normality of the score with covariance Omega), Theorem 1 states that if lambda_n/n converges in probability to lambda and sqrt(n)(beta_n - beta_0) converges to b, then sqrt(n)(beta_hat_lambda_n - beta_n) converges in distribution to N_k(-lambda(Sigma + lambda I_k)^{-1}b, (Sigma + lambda I_k)^{-1}Omega(Sigma + lambda I_k)^{-1}). This is the paper's central claim: in finite samples ridge is roughly Gaussian with shrinkage bias proportional to lambda and to the distance from the reference vector, and variance given by the usual sandwich form with the ridge-adjusted denominator. Theorem 2 converts this into a","pith_inferences":["Editorial inference: because b is unidentified, the practical content of the tuning rules is really a mapping from the user's bound or prior to a penalty; two researchers with different B or pi will get different lambda even from the same data, and nothing in the data can resolve the disagreement.","Editorial inference: the same local-to-reference rescaling could be applied to lasso-type estimators under Assumption 1, potentially producing analogous risk-based tuning rules; the paper suggests this extension but does not carry it out.","Editorial inference: the Gaussian approximation implies a simple shrinkage-bias-adjusted confidence ellipse for ridge predictions, but the paper does not develop this inferential use.","Editorial inference: in time-series settings Omega is the long-run variance, so the optimal lambda adapts to serial correlation; estimating Omega reliably in that setting may become the binding constraint."],"forward_implications":["If Theorem 1 is correct, ridge's finite-sample distribution in low-dimensional designs is a known Gaussian, so prediction intervals and risk comparisons can be made without assuming Gaussian errors or independence.","Heteroskedasticity and serial correlation enter only through Omega, so the approximation covers time-series and clustered data, not just i.i.d. samples.","The isotropic tuning formula lambda_n = n*tr(Omega_hat)/(sigma_hat_x^2 * B^2) gives an explicit ridge penalty that minimizes worst-case approximate excess risk, replacing cross-validation with a one-line calculation.","The paper's own DGP-3 results (k/n = 0.6) show the approximation and the minimax rule degrade away from the low-risk region and do not uniformly beat leave-one-out cross-validation, so the stated guarantees are low-dimensional."],"fun_headline_variants":["Ridge estimator is approximately Gaussian in finite samples","Gaussian approximation captures ridge's bias-variance tradeoff","A simple Gaussian for ridge's finite-sample distribution","Ridge estimator distribution: a Gaussian approximation","When ridge grows with n, distribution becomes Gaussian"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The true regression coefficients must lie within O(1/sqrt(n)) of the shrinkage reference vector beta_0; if they do not, the Gaussian approximation's bias term is wrong and the tuning rules that rely on the local parameter b lose their justification.","fun_headline_variants_meta":{"raw":{"variants":["Ridge estimator is approximately Gaussian in finite samples","Gaussian approximation captures ridge's bias-variance tradeoff","A simple Gaussian for ridge's finite-sample distribution","Ridge estimator distribution: a Gaussian approximation","When ridge grows with n, distribution becomes Gaussian"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1109,"prompt_tokens":713,"completion_tokens":396,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":324}},"tokens_in":457,"tokens_out":396,"duration_ms":4630,"temperature":1.0,"reasoning_tokens":324,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T05:08:05.949751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a Monte Carlo with k fixed, x_i and epsilon_i satisfying Assumption 1, beta_n = beta_0 + b/sqrt(n), and lambda_n/n -> lambda; compare the empirical distribution of sqrt(n)(beta_hat_lambda_n - beta_n) to N(-lambda(Sigma + lambda I)^{-1}b, (Sigma + lambda I)^{-1}Omega(Sigma + lambda I)^{-1}). If the first two moments deviate systematically beyond Monte Carlo error, Theorem 1 is refuted. Setting beta fixed away from beta_0 instead should make the bias grow with sqrt(n), confirming that the local assumption is load-bearing.","supporting_citations":[],"review_version":1}