{"id":"b6baebc1-047c-4293-8611-071998115520","arxiv_id":"1908.05255","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For rank-correlation M-estimators with U-process objectives, estimation works at the sqrt(p/n) rate when p/n converges to zero, but normal approximation needs the much stronger condition log(n/p^2) p^{3/2}/n^{1/4} to converge to zero.","lead":"This paper derives asymptotic theory for rank estimators when the number of covariates grows with the sample size. It shows that consistent estimation is possible at the sqrt(p/n) rate, but normal inference requires p to grow much more slowly.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.4's Bahadur bound loses a square root: the displayed rate is the additive objective error φε, while the proof's inequality only yields O_P(√φε).","rationale":"The reader's weakest assumption was Assumption 3(v), the exponential moment condition on the smoothed Hessian. That is a legitimate scope condition, but it is explicitly stated and, for Han's MRC, verified under primitive Conditions 1–3; a failure of the assumption simply means the theorem does not apply. The concern raised here is more load-bearing because it is an internal inconsistency: the proof of Theorem 2.4 does not imply the displayed Bahadur rate. The missing square root arises directly from (A.28) and the positive-definiteness of −V, so it is not a matter of omitted lemmas or overly strong assumptions. It affects the central advertised contribution, namely the explicit Bahadur-type bounds and the quantitative statement about normal approximation accuracy. The good news is that the main qualitative message—estimation rate √(p_n/n) and the much stronger scaling requirement for normal inference—appears to survive: the corrected bound still yields the same condition log(n/p_n^2)p_n^{3/2}/n^{1/4} = o(1) for asymptotic normality, up to constants and log factors. For this reason I do not recommend rejection; the paper needs a major correction to the rates in Theorem 2.4 and the four corollaries, plus a re-check of Theorem 2.6 if it uses the same type of argument. This strengthens the reader's CONDITIONAL verdict rather than overturning it.","tokens_in":45918,"tokens_out":28396,"duration_ms":266073,"concrete_test":"Independently re-derive the last two displays in the proof of Theorem 2.4: from (A.28), Assumption 3(ii), and the definition φε, solve first for ||t̂_n − t*_n|| and then for the unnormalized remainder ||θ̂_n − θ0 + V^{-1} P_n ∇_1 τ(·;θ0)||. If the result is √(nφε)/√n = √(φε), then the stated rate in Theorem 2.4(i) and Corollary 3.1(iii) is smaller by a square root than the proof supports. As a second check, set p_n = 1 and compare the corrected rate (√log n)/n^{5/8} with the fixed-dimension Bahadur bound in Subbotin (2008); the claimed log n / n^{5/4} is o(1/n), which is not implied by the displayed inequalities.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the proof of Theorem 2.4, the authors obtain, after equation (A.28), 0 ≤ −(1/2)(t̂_n − t*_n)^T V(t̂_n − t*_n) ≤ 2nφε. Assumption 3(ii) gives λ_min(−V) ≥ cmin > 0, so this inequality implies ||t̂_n − t*_n||^2 ≤ C nφε, hence ||t̂_n − t*_n|| ≤ C√(nφε). Dividing by √n, the Bahadur remainder is O_P(√φε), not O_P(φε) as stated in Theorem 2.4(i). For Corollary 3.1(iii), where ν_n = p_n and η_n = p_n/√n, this means φε is of order log(n/p_n^2) p_n^{3/2}/n^{5/4}, and the proof supports only ||θ̂^H_n − θ0 + (V^H)^{-1} P_n ∇_1 τ^H(·;θ0)|| = O_P( (log(n/p_n^2) p_n^{3/2}/n^{5/4})^{1/2} ), not the displayed rate without the square root. The same missing square root propagates to Corollaries 3.2–3.4. The asymptotic-normality scaling condition log(n/p_n^2) p_n^{3/2}/n^{1/4} = o(1) happens to be unchanged, because it is the condition √φε = o(1/√n) written explicitly. However, the quantitative Bahadur bounds advertised in the abstract and corollaries are not consequences of the preceding algebra. A fixed-dimension sanity check: for p_n = 1 the displayed rate is log n / n^{5/4} = o(1/n), whereas the inequality above gives √(log n)/n^{5/8}; the former is also smaller than standard n^{-3/4}-type Bahadur rates for nonsmooth M-estimators, making the displayed rate implausible.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops asymptotic theory for M-estimators whose objective functions are U-processes, possibly discontinuous, in an increasing-dimension setting where both the data dimension m_n and parameter dimension p_n grow with n. The main results are: a maximal inequality for degenerate U-processes in increasing dimensions (Theorem 2.1); consistency under ν_n/n → 0 (Theorem 2.2); a (ν_n∨p_n)/n rate of convergence (Theorem 2.3); a Bahadur-type bound and normal approximation under stronger scaling (Theorem 2.4); and consistency of numerical-derivative covariance estimators with step-size calibration (Theorem 2.6). The general results are applied to four rank estimators: Han's maximum rank correlation, Cavanagh–Sherman, Khan–Tamer, and Abrevaya–Shin, with corollaries giving p_n/n^{1/2} estimation rates, Bahadur-type bounds of order log(n/p_n^2)p_n^{3/2}/n^{5/4}, and normal approximation under log(n/p_n^2)p_n^{3/2}/n^{1/4}=o(1). Simulations for Han's MRC illustrate coverage deterioration with p_n.","tokens_in":46371,"tokens_out":9783,"duration_ms":80038,"significance":"The paper addresses a genuine gap: existing increasing-dimension M-estimation theory excludes discontinuous U-process objectives, and rank estimators have only been analyzed for fixed p. The new maximal inequality for degenerate U-processes (Theorem 2.1) and the careful tracking of ν_n, p_n, and m_n are valuable technical contributions, and the conclusion that normal inference requires a much stronger scaling condition than estimation is economically important. The extensive proofs for the general M-estimator and for Han's MRC are detailed and represent serious work. However, the advertised Bahadur-type rates are not supported by the proof as written (missing square root; see major comments), and the minimax optimality claim is not backed by a lower bound. With those corrected, the framework would be a solid contribution to the nonparametric and semiparametric econometrics literature.","major_comments":[{"comment":"The displayed Bahadur bound is not implied by the proof. After (A.28) one has 0 ≤ −(1/2)(t̂_n − t*_n)^T V(t̂_n − t*_n) ≤ 2nφε, and with Assumption 3(ii) this yields ‖t̂_n − t*_n‖ = O_P(√(nφε)), so ‖θ̂_n − θ0 + V^{-1}P_n∇_1τ(·;θ0)‖_2 = O_P(√φε), not O_P(φε). The rate displayed in Theorem 2.4(i) and the rates in Corollaries 3.1(iii), 3.2(iii), 3.3(iii), and 3.4(iii) must be replaced by square-root versions; for Corollary 3.1(iii), for example, the proof supports O_P((log(n/p_n^2))^{1/2}p_n^{3/4}/n^{5/8} + p_n^{5/4}/n^{3/4}) under the stated scaling, not the displayed log(n/p_n^2)p_n^{3/2}/n^{5/4}. The asymptotic-normality scaling condition log(n/p_n^2)p_n^{3/2}/n^{1/4}=o(1) is unchanged because it is equivalent to √φε = o(1/√n), but the quantitative claims in the abstract and corollaries need revision.","section":"§A.3.5, Theorem 2.4(i)"},{"comment":"The abstract and Section 1.1 call (p_n/n)^{1/2} the 'minimax optimal' rate, but no minimax lower bound is proved for the increasing-dimension problems studied. The reference (Yu, 1997) supplies lower bounds only in fixed-dimensional settings; the triangular-array framework with changing parameter spaces requires a new lower-bound argument. As it stands, Theorems 2.3 and Corollaries 3.1–3.4 only establish upper bounds, so the optimality claim should either be proved or softened.","section":"Abstract and §1.1"},{"comment":"Lemmas A.7–A.9, which bound sup_{θ∈B(θ0,r)} E{h^C(·,·;θ)}^2, E{h^K(·,·;θ)}^2, and E{h^A(·,·;θ)}^2, are omitted with the statement that their proofs are similar to Lemma A.6. These lemmas are load-bearing for Corollaries 3.2–3.4, and the similarity is not immediate: the censored-duration objective includes R_i and V_i, and the Abrevaya–Shin objective includes a kernel K_b(W_i−W_j) with bandwidth, which changes the differentiation and moment arguments. The proofs should be supplied or at least the differences from Lemma A.6 detailed.","section":"Appendix A.4.1, Lemmas A.7–A.9"},{"comment":"Assumption 3(v), the exponential moment condition on the smoothed Hessian, is verified only for Han's MRC in Theorem 3.1 under Conditions 1–3. For the Cavanagh–Sherman, Khan–Tamer, and Abrevaya–Shin estimators, no primitive conditions are given under which Assumption 3(v) (or Assumption 3(iii)) holds; the corresponding corollaries therefore rely on an unverified high-level condition. The authors should state sufficient design and smoothness conditions for these estimators or explicitly flag Assumption 3 as a high-level condition that must be checked case by case.","section":"§3.2–§3.4, Assumption 3(v)"}],"minor_comments":[{"comment":"The quantity φε is used in (A.26) before it is defined in (A.27); reorder the display or define φε earlier.","section":"§A.3.5"},{"comment":"The notation 'P− →' in Section 1.4 appears garbled; standard notations for convergence in probability should be used.","section":"§1.4"},{"comment":"In Corollary 3.4(iii), the term n^{-δJ} should be made explicit as O_P(n^{-δJ}), and the dependence of the constant on J and on the kernel K(·) should be stated.","section":"Corollary 3.4(iii)"},{"comment":"Tables 1–3 report coverage probabilities for three projection directions, and Figures 1–3 are not explicitly cross-referenced to those directions in the text; a sentence stating which figure corresponds to which projection would improve readability.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The missing square root in Theorem 2.4 is a genuine error that affects the paper's headline quantitative claims. I do not think this warrants rejection: the conceptual framework and the asymptotic-normality scaling condition survive, and the error is local to the algebra after (A.28). I recommend asking for a careful revision that corrects the rates, supplies the omitted lemmas, and removes or proves the unsupported minimax claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper’s main qualitative findings are right and worth having. It is the first to handle increasing dimension for M-estimators with discontinuous U-process objectives, including Han’s MRC and the other rank estimators. The new maximal inequality for degenerate U-processes is a real contribution, and the general theorems are proved carefully. The applied message — that rank estimators achieve the sqrt(p/n) estimation rate but need much stronger scaling for normal inference, and that the numerical-derivative step size must shrink with p — is well supported and practically useful.\n\nThe soft spot is real and load-bearing. In the proof of Theorem 2.4, the inequality at (A.28) gives ||t̂_n − t*_n||^2 ≤ C n φε, hence ||t̂_n − t*_n|| = O(√(nφε)). Dividing by √n, the Bahadur remainder is O_P(√φε), not O_P(φε) as stated. The displayed rates in Theorem 2.4(i) and Corollaries 3.1–3.4 are therefore not consequences of the proof; the correct bound is the square root of what is advertised. The asymptotic-normality scaling condition survives, because it is equivalent to √φε = o(1/√n), but the quantitative Bahadur bounds in the abstract are overstated.\n\nOther soft spots are minor. The proofs of Lemmas A.7–A.9 are omitted as “similar” to Lemma A.6; that is acceptable if the similarity is genuine, but it leaves three of the four estimators partially unverified. The minimax optimality claim is an overreach without a lower bound. The simulations initialize the optimizer at the truth, which limits what they show about practical implementation, though they still illustrate the deterioration of coverage quite clearly.\n\nThe right audience is econometricians and statisticians working on rank estimation and high-dimensional inference. The framework and the scaling warnings are valuable enough that the paper deserves a serious referee, but the quantitative results need a major correction before they can be used as stated.","headline":"The qualitative message is solid, but the advertised Bahadur-type bounds are off by a square root — the proof only supports the square root of the stated rate.","tokens_in":46908,"tokens_out":1592,"would_cite":true,"duration_ms":16915,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","60F17","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes the first high-dimensional theory for rank estimators such as Han’s maximum rank correlation estimator, showing the estimator stays accurate at the minimax rate but its normal approximation requires a much stronger…","keywords":["rank estimators","maximum rank correlation","increasing dimension","U-processes","Bahadur-type bounds","normal approximation","asymptotic covariance estimation","semiparametric index models"],"falsifier":"Simulate Han's MRC estimator with heavy-tailed covariates (for example, $t$-distributed with few degrees of freedom) while keeping $p_n$ and $n$ within the paper's scaling regime, and examine whether the coverage probability of the normal confidence interval for a fixed projection deteriorates substantially faster than in the Gaussian-design simulations, or whether the Bahadur expansion's error term grows at a rate larger than the paper's bound.","tokens_in":45704,"feed_emoji":"📈","tokens_out":4071,"duration_ms":30397,"temperature":0.7,"pith_summary":"We now know what happens to rank estimators (like Han's maximum rank correlation estimator) when the number of coefficients grows with the sample size. The paper proves that estimation remains attractive: as long as the parameter count $p_n$ grows more slowly than the sample size $n$, the estimator converges at the minimax-optimal $(p_n/n)^{1/2}$ rate. Inference is another story, because the paper shows that a normal approximation for these estimators requires a much stronger scaling than $p_n^2/n \\to 0$; the Bahadur-type bound and the consistency of the covariance estimator both demand stricter conditions. This matters because rank estimators are popular for semiparametric models with many covariates, and the results say the usual $p_n^2/n$ rule used for smooth M-estimators is not enough when the objective function is discontinuous. The paper's simulations show normal confidence intervals losing coverage quickly as $p_n$ grows, even for modest $p_n$.","feed_headline":"Rank estimators keep their speed but lose normal inference in high dimension","feed_subtitle":"Estimation stays optimal at $(p_n/n)^{1/2}$, but valid confidence intervals demand a much stronger scaling condition than $p_n^2/n\\to0$.","key_machinery":"The central object is the U-process objective $\\Gamma_n(\\theta) = \\frac{1}{n(n-1)}\\sum_{i\\neq j} f(Z_i,Z_j;\\theta)$ and its Hoeffding decomposition $\\Gamma_n(\\theta) = \\Gamma(\\theta) + P_n g(\\cdot;\\theta) + U_n h(\\cdot,\\cdot;\\theta)$. The key technical contribution is a new maximal inequality for degenerate U-processes in increasing dimensions, which controls the uniform decay of the remainder $\\sup_{\\theta\\in B(\\theta_0,r_n)}|U_n h(\\cdot,\\cdot;\\theta)|$ in terms of the VC dimension $\\nu_n$, the radius $r_n$, and a variance proxy $\\tilde\\epsilon_n$. This inequality makes possible a Bahadur-type representation for $\\hat\\theta_n$ by transferring the analysis to the smoothed objective $\\tilde\\Gamma_n(\\theta) = \\Gamma(\\theta) + P_n \\tau(\\cdot;\\theta)$, whose theoretical properties are handled via Assumption 3 and exponential-moment bounds. The machinery is what turns the discontinuous loss into a tractable smooth one while keeping track of how $p_n$ affects all rates.","core_discovery":"The paper's central claim is that for M-estimators whose objective functions are U-processes and possibly discontinuous, in the increasing-dimension regime, estimation still achieves the minimax-optimal $(p_n/n)^{1/2}$ rate, but normal approximation demands far more stringent scaling. Specifically, for Han's maximum rank correlation estimator under Assumptions 4–6, the paper establishes $\\|\\hat\\theta_n^H - \\theta_0 + (V^H)^{-1}P_n\\nabla_1\\tau^H(\\cdot;\\theta_0)\\|_2 = O_P\\bigl(\\log(n/p_n^2)p_n^{3/2}/n^{5/4}\\bigr)$ whenever $p_n^2/n = o(1)$ and $\\log(n/p_n^2)p_n^{3/2}/n^{5/4} = o(1)$. It further shows that $\\sqrt{n}\\gamma^T(\\hat\\theta_n^H - \\theta_0) / (\\gamma^T (V^H)^{-1}\\Delta^H (V^H)^{-1} \\gamma)^{1/2} \\Rightarrow N(0,1)$ under the stronger condition $\\log(n/p_n^2)p_n^{3/2}/n^{1/4} = o(1)$. The same pattern holds for the other three rank estimators: a minimax estimation rate, followed by a much more demanding condition for valid normal inference. The paper also proves that the numerical-derivative covariance estimator is consistent only if the step size is tuned with respect to $p_n$, not just $n$.","pith_inferences":["A likely consequence is that practitioners who want valid confidence intervals with many regressors should shift to alternative inferential methods for rank estimators, such as resampling or bootstrap calibrations, though the paper does not analyze those procedures here.","The maximal inequality for degenerate U-processes is a general tool that could be applied to other non-smooth econometric estimators beyond rank correlations, such as maximum score or other pairwise-comparison estimators in increasing dimensions.","A natural testable extension is to check whether the conditions imply that the bootstrap, or subsampling, can restore valid coverage under weaker scaling than the normal approximation; the paper leaves this unexplored.","Because the Assumption 3(v) exponential-moment bound is verified for Han's estimator only under subgaussian designs and smooth conditional densities, one might expect the normal approximation to break down even earlier for heavy-tailed designs; this is not tested in the paper."],"forward_implications":["For semiparametric index models with many covariates, rank estimators remain usable for point estimation under only $p_n/n \\to 0$, matching the minimax-optimal $(p_n/n)^{1/2}$ rate.","The usual scaling condition $p_n^2/n \\to 0$ used for smooth M-estimators is generally insufficient for normal approximation of rank estimators; inference requires the stronger $\\log(n/p_n^2)p_n^{3/2}/n^{1/4} = o(1)$ in the Han example.","To obtain consistent covariance matrices by numerical differentiation, the step size must shrink with $p_n$: the paper shows consistency under $\\varepsilon_n\\sqrt{p_n} = o(1)$ and $\\varepsilon_n^{-2}p_n/\\sqrt{n} = o(1)$, so the recommended $\\varepsilon_n \\asymp (p_n/n)^{1/6}$ depends on the dimension.","Confidence intervals based on normal approximation deteriorate quickly as $p_n$ increases for fixed $n$, as confirmed by the paper's simulations where even $p_n=3$ or $4$ shows severe coverage distortion at moderate sample sizes."],"supporting_citations":[{"why":"Defines the maximum rank correlation estimator whose increasing-dimension properties are the paper's leading example.","marker":"Han (1987)"},{"why":"Provides the fixed-dimension limiting distribution and identification conditions for Han's MRC that the paper extends.","marker":"Sherman (1993)"},{"why":"Gives the finite-dimensional maximal inequality for degenerate U-processes that the paper generalizes to increasing dimensions.","marker":"Sherman (1994)"},{"why":"Supplies the VC-class and U-process techniques and symmetrization arguments used to derive the new increasing-dimensional maximal inequality.","marker":"Nolan and Pollard (1987)"},{"why":"Provides the finite-sample analysis of M-estimators with differentiable objective functions in increasing dimensions, used to handle the smoothed component $\\tilde\\theta_n$.","marker":"Spokoiny (2012a)"},{"why":"Provides empirical-process tools such as Theorem 9.3 for controlling covering numbers of VC classes, used in the proofs.","marker":"Kosorok (2007)"},{"why":"Establishes the $p_n^2/n \\to 0$ benchmark for normal approximation of GLM estimators, which the paper contrasts with the stronger rank-estimator condition.","marker":"Portnoy (1988)"}],"fun_headline_variants":["Optimal estimation, but inference needs stricter scaling for rank estimators","Estimation hits optimal rate, but inference demands stricter conditions","High-dimensional rank estimators: efficient estimates, but normal inference requires more","Rank estimation stays sharp in high dimension, but inference lags behind","Rank estimators: optimal rates, but normal approximation needs tighter scaling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results for Han's estimator require an exponential moment bound (Assumption 3(v) under Conditions 1–3), which is verified only for subgaussian designs with smooth conditional densities and bounded second derivatives of the link function; if the covariates are heavy-tailed or the conditional density is not smooth, the Bahadur bound and normal approximation may fail.","fun_headline_variants_meta":{"raw":{"variants":["Optimal estimation, but inference needs stricter scaling for rank estimators","Estimation hits optimal rate, but inference demands stricter conditions","High-dimensional rank estimators: efficient estimates, but normal inference requires more","Rank estimation stays sharp in high dimension, but inference lags behind","Rank estimators: optimal rates, but normal approximation needs tighter scaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000806,"raw_usage":{"total_tokens":3621,"prompt_tokens":1105,"completion_tokens":2516,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":721,"completion_tokens_details":{"reasoning_tokens":2429}},"tokens_in":721,"tokens_out":2516,"duration_ms":15844,"temperature":1.0,"reasoning_tokens":2429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:19:18.210410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate Han's MRC estimator with heavy-tailed covariates (for example, $t$-distributed with few degrees of freedom) while keeping $p_n$ and $n$ within the paper's scaling regime, and examine whether the coverage probability of the normal confidence interval for a fixed projection deteriorates substantially faster than in the Gaussian-design simulations, or whether the Bahadur expansion's error term grows at a rate larger than the paper's bound.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the maximum rank correlation estimator whose increasing-dimension properties are the paper's leading example."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fixed-dimension limiting distribution and identification conditions for Han's MRC that the paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the finite-dimensional maximal inequality for degenerate U-processes that the paper generalizes to increasing dimensions."},{"cited_title":"and Pollard, D","cited_arxiv_id":null,"evidence_quote":"Supplies the VC-class and U-process techniques and symmetrization arguments used to derive the new increasing-dimensional maximal inequality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides empirical-process tools such as Theorem 9.3 for controlling covering numbers of VC classes, used in the proofs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the $p_n^2/n \\to 0$ benchmark for normal approximation of GLM estimators, which the paper contrasts with the stronger rank-estimator condition."}],"review_version":1}