{"id":"e2604a41-96a9-4f1b-ab3d-1fac7aa46144","arxiv_id":"2505.01669","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A banded high-dimensional Hettmansperger-Randles estimator is introduced and used to build robust location tests and a quadratic discriminant classifier for elliptical data.","lead":"A high-dimensional version of the Hettmansperger-Randles robust estimator is proposed, built from a banded spatial-sign covariance and a sparse precision initializer. It is applied to one-sample location tests and to quadratic discriminant analysis, with simulations and a gene expression example.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theory proves one-step SGLASSO plug-in estimator, not the iterated banded Algorithm 2 that is proposed and simulated; the gap is unaddressed.","rationale":"After reading the full text, the reader's weakest-assumption is precisely right. The proof of Lemma 1 treats Ωhat from (2) as fixed and never accounts for the subsequent scatter updates; Lemma 5 and Lemma 7 similarly bound quantities involving Ωhat only. Algorithm 2's trace-normalized, banded scatter is a different object, and no fixed-point or convergence analysis is supplied. The simulations in Section 5 (Table 1, Figure 1-2, Table 2) report no iteration count or estimator variant, making it impossible to attribute the empirical results to the theoretical theorems. The real-data section additionally selects genes by two-sample t-tests on the full dataset before splitting, which leaks label information and inflates reported accuracy; though secondary, it strengthens the case that the empirical validation is not aligned with the central claims. Because the distributional limits, size control, and QDA consistency all depend on the unproven carry-over, the rejection is warranted. I found no reason to change the reader's verdict, and I agree with the identified weakest assumption.","tokens_in":37502,"tokens_out":8845,"duration_ms":77530,"concrete_test":"Replicate Table 1 and Table 2 twice: once using the one-step SGLASSO-based estimator exactly as in Lemma 1 (Ωhat from (2), µhat as its minimizer) and once using Algorithm 2 iterated to convergence, with the same tuning. Compare empirical sizes of TSUM/TMAX and classification accuracy of HRQDA. If the iterated version differs materially (e.g., size > 0.10 when nominal 0.05) or if the original paper's code used the one-step version, the concern that the theory does not cover the implemented method is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3's theoretical results (Lemma 1, Lemma 2, Corollary 1, Theorems 1-6) all rely on a fixed precision matrix Ωhat obtained from SGLASSO in (2). Lemma 1's proof defines the location estimator as the minimizer of (13), equivalently the solution of ∑ U{Ωhat^{1/2}(Xi−µhat)}=0, and Lemma 4, Lemma 5, and Lemma 7 bound the error of that particular Ωhat. In contrast, Algorithm 2 iteratively updates both location and scatter: Step 3 applies banding B_h with h=3 and trace renormalization, so the precision matrix used in the final estimator is not the SGLASSO solution (2). No theorem analyzes the iterated estimator, no convergence proof for the Algorithm 2 fixed point is given, and Section 5 never states whether simulations use the one-step estimator or the iterated algorithm. Consequently, the size control of TSUM/TMAX and the Bayes-rate consistency of HRQDA are not established for the method actually proposed. This is the load-bearing gap in the paper's central claim, and the direct contradiction between the title's 'Hettmansperger-Randles' and the one-step plug-in analyzed only compounds it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a high-dimensional version of the Hettmansperger-Randles (HR) estimator for elliptical distributions, aiming to simultaneously estimate location and scatter in an affine-invariant manner. It then applies this estimator to one-sample location testing (sum-type, max-type, and Cauchy combination tests) and to quadratic discriminant analysis (HRQDA). The theoretical part derives a Bahadur representation and a Gaussian approximation for the standardized estimator, yielding Gumbel and normal limits for the test statistics and a Bayes-rate consistency result for the classifier. Simulations and a gene-expression data application are used to support the methods.","tokens_in":37760,"tokens_out":11858,"duration_ms":108799,"significance":"The paper's goals are valuable: robust high-dimensional inference for elliptical distributions is an active area, and a joint location-scatter procedure with a QDA application would be a useful contribution if the theorems were proven for the implemented method. The paper leverages a substantial body of existing work (Lu-Feng SGLASSO bounds, Gaussian approximation results, and spatial-sign methods) and provides explicit asymptotic relative efficiency comparisons. However, the central theoretical statements apply to a one-step plug-in estimator that is not the algorithm proposed and simulated; as a result, the headline claims about Algorithm 2, including its affine invariance and the validity of the test sizes and classification consistency, are not supported.","major_comments":[{"comment":"The theoretical core of the paper (Lemmas 1-2, Corollary 1, and Theorems 1-6) is established for a one-step location estimator that takes the SGLASSO precision matrix from equation (2) as a fixed plug-in: the proof of Lemma 1 in §8.2.6 defines μhat as the minimizer of L(θ) in (13) and solves Σ_i U{Ωhat^{1/2}(X_i−μhat)}=0 with Ωhat from (2). The method proposed in the paper, however, is Algorithm 2, whose Step 3 iteratively updates scatter via banding with h=3 followed by trace renormalization, so that the final precision matrix is not the SGLASSO solution (2). The paper contains no convergence analysis, fixed-point characterization, or asymptotic result for this iterated banded estimator. Consequently, the claimed null distributions for TSUM and TMAX (Theorems 1 and 3) and the Bayes-rate consistency of HRQDA (Theorem 6) are not established for the estimator used in the simulations.","section":"§2 (Algorithm 2), §3 (Lemma 1, Eq. (13))"},{"comment":"The simulation section does not specify whether the reported empirical sizes and powers are computed with the one-step estimator analyzed in the theorems or with the iterated Algorithm 2. This omission is consequential: if Algorithm 2 is used, the simulations evaluate a procedure for which no asymptotic theory is provided; if the one-step estimator is used, the simulations do not evaluate the headline HR estimator of Section 2. In addition, the SGLASSO tuning parameter λ_n used in the simulations is not reported, even though Lemma 1 and Lemma 4 state specific choices of λ_n that are needed for the theoretical rates to hold.","section":"§5 (Table 1, Figures 1-2)"},{"comment":"The claim that the proposed high-dimensional HR estimator estimates location and scatter 'in an affine-invariant manner' is not supported by the algorithm as written. Step 3 of Algorithm 2 applies the coordinate-dependent banding operator B_h with h=3 and trace normalization, and the initial spatial median μhat_0 in (1) is not affine equivariant, so the estimator does not satisfy the usual affine equivariance property under transformations X ↦ AX+b. The further statement in §3 that TMAX is 'obviously' affine invariant is likewise not justified: even for an affine-equivariant estimator pair, the max-norm ‖Ωhat^{1/2} μhat‖_∞ is not invariant under a general nonsingular transformation A. Because affine invariance is advertised as a key advantage over prior scalar-invariant estimators, this claim must be either proved for the actual algorithm or withdrawn.","section":"§1 Introduction; §2 Algorithm 2; §3, TMAX definition"}],"minor_comments":[{"comment":"The title page contains a typo: 'and i ts Applications' should read 'and its Applications'.","section":"Title page"},{"comment":"In the description of the mixture normal distribution, 'Xi2∼MN (µ2, Σ2, 10.0.8)' should read 'MN (µ2, Σ2, 10, 0.8)'.","section":"§5.2"},{"comment":"The sentence 'We found that all the test could control the empirical sizes very well' should read 'all the tests'.","section":"§5.1"},{"comment":"The bootstrap bias correction is introduced informally; a few sentences explaining why the bootstrap mean and variance correct the asymptotic bias would improve readability, and the statement that 'M = 50 is always enough' should be supported or softened.","section":"§3, bootstrap paragraph"}],"recommendation":"reject","confidential_remarks":"The central objection is the theory-implementation mismatch: all theorems concern a one-step SGLASSO plug-in estimator, while the proposed and simulated method is the iterated banded Algorithm 2. A revision that merely adds proofs of the existing claims for Algorithm 2 would require a substantial new convergence and asymptotic analysis; alternatively, the paper could be repositioned around the one-step estimator, but then the title and motivation would need to change. Please also ask the authors, in any future version, to explicitly state which estimator is used in the simulations and to report the SGLASSO tuning parameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. The first is that there is a real, substantive theory here: a Bahadur representation for a spatially standardized median, Gaussian approximation over simple convex sets, asymptotic independence of L2- and L∞-based tests, a Cauchy combination test, and a QDA classifier with a consistency bound. The proofs are long and non-trivially assembled from the spatial-sign literature, and the simulation study is fairly extensive.\n\nThe second thing is that the main theorems do not cover the method the paper actually proposes. Lemma 1 and the results that follow treat μhat as the minimizer of (13) with a fixed SGLASSO precision matrix from (2). Algorithm 2, however, iterates a banded, trace-renormalized scatter update with h=3. No convergence or distributional result is given for that iterated procedure, and Section 5 never states which estimator the simulations use. This is not a side issue: the size control of TSUM and TMAX, the asymptotic independence, and the QDA consistency all rest on the one-step estimator. If the implemented algorithm differs, those guarantees do not automatically transfer. The stress-test note gets this right.\n\nThere are smaller soft spots. The affine-invariance claim in the introduction is overstated—SGLASSO is coordinate-dependent, so the resulting estimator is not affine-invariant in the usual sense. The bandwidth h=3 is arbitrary, the bootstrap bias correction with M=50 is ad hoc, and the real-data preprocessing selects features with t-tests on the full dataset before splitting, which risks data leakage.\n\nTo be fair, I don't see circularity in the derivations; the paper leans on earlier work from the same group, but that's normal in this area. The theory is not obviously wrong on its own terms—it just doesn't match the algorithm.\n\nWho gets value from it? Readers working on robust high-dimensional location tests and discriminant analysis will find the machinery interesting, but they should be careful to distinguish the analyzed estimator from the proposed one. I would like to see a revision that either proves the results for Algorithm 2 or explicitly demotes it to a computational heuristic, with the one-step version as the formal proposal. As it stands, I would not publish it, but I would send it to a serious referee—the gap is fixable and the underlying work deserves a fair reading.","headline":"A substantial high-dimensional robust-testing theory, but the theorems analyze a one-step plug-in estimator while the proposed method is an iterated banded algorithm—that gap is load-bearing.","tokens_in":38286,"tokens_out":3625,"would_cite":false,"duration_ms":34392,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","62H15","62F35","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a high-dimensional version of the Hettmansperger-Randles estimator—a spatial-median location estimator jointly fitted with a banded, trace-normalized scatter matrix—is affine invariant and simultaneously yields a…","keywords":["Hettmansperger-Randles estimator","spatial median","high-dimensional location test","affine invariance","elliptical distributions","quadratic discriminant analysis","Cauchy combination test","robust scatter estimation"],"falsifier":"Run T_MAX under the null using the fully iterated Algorithm 2 with n=100, p=120, and t_3 data, and compare the empirical null distribution to the Gumbel limit used for the critical value. If the 5% empirical quantile of the iterated estimator differs from the Gumbel quantile by more than simulation error (say more than two standard errors over 10,000 replications) while the one-step version matches it, the asymptotic theory does not cover the implemented method.","tokens_in":37288,"feed_emoji":"📊","tokens_out":8506,"duration_ms":76106,"temperature":0.7,"pith_summary":"This paper claims that the classic Hettmansperger-Randles joint estimator of multivariate location and scatter, which breaks down when the dimension exceeds the sample size, can be rebuilt for high-dimensional elliptical data by starting from a spatial median and a sparse graphical-lasso precision matrix and replacing the singular spatial-sign covariance matrix with its banded version. The paper proves that this high-dimensional HR estimator admits a Bahadur representation and a Gaussian approximation, and it derives test statistics for the one-sample location problem whose null limits are Gaussian (sum type) and Gumbel (max type) and are asymptotically independent, so they can be combined by a Cauchy combination. If these claims are right, robust high-dimensional inference no longer has to choose between location and scatter: the same estimator also yields a quadratic discriminant classifier whose misclassification rate converges to the Bayes rate under mild moment conditions.","feed_headline":"One affine-invariant estimator powers high-dim robust tests and QDA","feed_subtitle":"Spatial median plus scatter yields Gumbel-limit tests and QDA whose error rate converges to the Bayes rule.","key_machinery":"The carrying object is Algorithm 2, the high-dimensional HR estimator: an alternating iteration that standardizes residuals by the current scatter, updates location by the mean spatial sign of the residuals, and updates scatter from the spatial-sign covariance matrix after banding with bandwidth $h=3$ and trace renormalization, seeded by the spatial median and the SGLASSO precision matrix from equation (2). The theoretical engine is the Bahadur representation (Lemma 1) followed by Gaussian approximation over simple convex sets (Lemma 2), which transfers the null distribution of the standardized location estimate to a standard Gaussian vector; from there the Gumbel and normal limits of the max- and sum-type statistics, their asymptotic independence, and the QDA error-rate consistency all follow.","core_discovery":"The central discovery is an estimator that solves the two HR fixed-point equations in high dimensions: the location update uses spatial signs of residuals standardized by the current scatter estimate, and the scatter update replaces the spatial-sign covariance matrix with a banded version and then renormalizes by trace. With a spatial-median start and an SGLASSO precision-matrix start, the one-step standardized estimator has a Bahadur representation whose remainder is small enough for Gaussian approximation, yielding a Gumbel limit for the $\\ell_\\infty$ form of the location statistic and a standard normal limit for its $\\ell_2$ form under the null, plus a noncentral normal limit under local alternatives. Sum and max statistics are asymptotically independent, and the four-way Cauchy combination adapts to alternatives sparse in either $\\Omega^{1/2}\\mu$ or $\\mu$. For QDA, plugging the HR location and scatter estimates into the generalized QDA classifier for elliptically symmetric distributions makes the misclassification rate converge to that of the Bayes rule at the stated rate.","pith_inferences":["The theorems cover the one-step version of the estimator with the SGLASSO precision matrix held fixed; since Algorithm 2 iterates banded, trace-renormalized updates and the simulations do not state which variant was used, the claimed null distributions would need separate proof if the iterations change the asymptotic law.","Affine invariance plus density-free null limits suggests the same estimator could be dropped into other affine-equivariant procedures, such as two-sample location tests or robust PCA, with comparable gains under heavy tails.","The efficiency identity $\\{E(r^{-1})\\}^2 E(r^2) \\ge 1$ implies HR-based tests dominate least-squares-based tests as tail weight increases; a direct check would be a power comparison at small $n$ with the bootstrap bias correction, where the paper's asymptotic ARE argument is least reliable.","Because the max-type statistic is calibrated to sparsity in $\\Omega^{1/2}\\mu$ rather than $\\mu$, a simulation with a rotated sparse signal (e.g., $\\Omega$ a banded matrix and $\\mu$ dense in the rotated coordinates) would isolate when $T_{\\mathrm{CC3}}$ beats sparsity-in-$\\mu$ procedures such as $T_{\\mathrm{CC2}}$."],"forward_implications":["$T_{\\mathrm{MAX}}$ has a Gumbel null limit $\\exp(-\\pi^{-1/2}e^{-x/2})$ and rejects when the $\\ell_\\infty$ norm of the standardized HR location estimate is large, with power tending to one against alternatives with $\\|\\Omega^{1/2}\\mu\\|_\\infty \\ge C n^{-1/2}(\\log p + q_{1-\\alpha})^{1/2}$.","$T_{\\mathrm{SUM}}$ is asymptotically standard normal under $H_0$, has noncentral normal power under local alternatives, and its asymptotic relative efficiency versus the sample-covariance-based $T_{\\mathrm{FLY}}$ is $\\{E(r^{-1})\\}^2 E(r^2) \\ge 1$, equal to $2.54$ for a $t_3$ distribution.","$T_{\\mathrm{SUM}}$ and $T_{\\mathrm{MAX}}$ are asymptotically independent under local alternatives, so the Cauchy combination tests $T_{\\mathrm{CC1}}$ and $T_{\\mathrm{CC3}}$ retain size while gaining power under both sparse and dense alternatives.","The four-statistic Cauchy combination $T_{\\mathrm{CC3}}$ stays competitive across the paper's four covariance models and three heavy-tailed distributions, adapting to unknown sparsity in either $\\Omega^{1/2}\\mu$ or $\\mu$.","The HR-based QDA classifier has misclassification rate differing from the Bayes rate by $O_p(\\lambda_n^{1-q/2} s_0(p)^{1/2} + \\lambda_n^{1-q} s_0(p))$, so it tracks optimal classification under heavy-tailed elliptically symmetric data."],"supporting_citations":[{"why":"Defines the joint location-scatter estimator and fixed-point equations that the paper extends.","marker":"Hettmansperger and Randles (2002)"},{"why":"Supplies the banding operator on covariance matrices and the short-range dependence assumptions used in Algorithm 2.","marker":"Bickel and Levina (2008b)"},{"why":"Provides the SGLASSO precision estimator in equation (2), whose convergence rates (Lemma 4) drive the Bahadur representation.","marker":"Lu and Feng (2025)"},{"why":"Yields the Gumbel limit for the maximum of standardized Gaussian variables from which T_MAX's null distribution is obtained.","marker":"Cai et al. (2013)"},{"why":"Gives the Gaussian approximation over simple convex sets that transfers the Bahadur representation to the limiting distributions.","marker":"Chernozhukov et al. (2017)"},{"why":"Supplies the generalized QDA classifier for elliptical distributions whose Bayes consistency HRQDA inherits.","marker":"Bose et al. (2015)"},{"why":"Provides the max-sum spatial-sign test and its asymptotic-independence theorem used to justify the four-statistic combination T_CC3.","marker":"Liu et al. (2024)"},{"why":"Gives the sample-covariance-based max-type test T_CFL used for comparison and the independence machinery reused in Theorem 5.","marker":"Chen et al. (2024)"},{"why":"Provides the sparse QDA method SQDA used as the main comparator in the QDA simulations and real-data analysis.","marker":"Li and Shao (2015)"}],"fun_headline_variants":["High-dim HR estimator: Gumbel-limit tests and QDA matching Bayes error","Solving HR equations in high dim yields Gumbel tests and Bayes QDA","High-dim robust location and scatter: HR estimator powers Gumbel tests and Bayes QDA","Gumbel limits for high-dim location tests and Bayes-consistent QDA from HR estimator","One-step HR estimator in high dim: robust tests with Gumbel limits and QDA to Bayes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the estimator whose asymptotics are proven is the estimator that is used: all theorems concern the one-step version with the SGLASSO precision matrix held fixed, while Algorithm 2 iterates banded, trace-renormalized scatter updates, and the paper's own simulations never state which version produced the reported tests.","fun_headline_variants_meta":{"raw":{"variants":["High-dim HR estimator: Gumbel-limit tests and QDA matching Bayes error","Solving HR equations in high dim yields Gumbel tests and Bayes QDA","High-dim robust location and scatter: HR estimator powers Gumbel tests and Bayes QDA","Gumbel limits for high-dim location tests and Bayes-consistent QDA from HR estimator","One-step HR estimator in high dim: robust tests with Gumbel limits and QDA to Bayes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001278,"raw_usage":{"total_tokens":5181,"prompt_tokens":854,"completion_tokens":4327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":4213}},"tokens_in":470,"tokens_out":4327,"duration_ms":27463,"temperature":1.0,"reasoning_tokens":4213,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:12:56.304428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run T_MAX under the null using the fully iterated Algorithm 2 with n=100, p=120, and t_3 data, and compare the empirical null distribution to the Gumbel limit used for the critical value. If the 5% empirical quantile of the iterated estimator differs from the Gumbel quantile by more than simulation error (say more than two standard errors over 10,000 replications) while the one-step version matches it, the asymptotic theory does not cover the implemented method.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the joint location-scatter estimator and fixed-point equations that the paper extends."},{"cited_title":"Liu, and Y","cited_arxiv_id":null,"evidence_quote":"Yields the Gumbel limit for the maximum of standardized Gaussian variables from which T_MAX's null distribution is obtained."},{"cited_title":"Chetverikov, and K","cited_arxiv_id":null,"evidence_quote":"Gives the Gaussian approximation over simple convex sets that transfers the Bahadur representation to the limiting distributions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the generalized QDA classifier for elliptical distributions whose Bayes consistency HRQDA inherits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the sample-covariance-based max-type test T_CFL used for comparison and the independence machinery reused in Theorem 5."}],"review_version":1}