{"id":"7248048d-1617-4cfc-a540-c2d3307fddc5","arxiv_id":"2607.06367","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"NOMAD unifies shrinkage and thresholding estimation by minimizing a data-driven approximate risk criterion derived via Stein's identity and Tweedie's formula, recovering James-Stein and lasso as special cases.","lead":"This paper introduces NOMAD, a method that automatically picks the best combination of shrinkage and thresholding for estimating noisy data by minimizing an approximate risk criterion. A smart generalist might read it because it unifies James-Stein shrinkage and lasso thresholding under one data-adaptive framework, potentially simplifying how we choose regularization methods across statistics and machine learning.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The oracle Tweedie accuracy condition (eq. 18) is assumed, not verified; its plausibility across signal regimes is the load-bearing gap for the consistency theory.","rationale":"The reader's verdict of CONDITIONAL with MODERATE confidence is appropriate. The reader correctly identified the single most load-bearing concern: the oracle Tweedie accuracy condition (eq. 18) is an unverified assumption whose plausibility across signal regimes — particularly sparse ones — is not discussed in the paper. I agree with this identification. The concern is not merely that the assumption is load-bearing (which the reader already notes), but that it may be specifically fragile in the sparse regime where the method claims its largest advantage. In sparse settings (small pi in spike-and-slab), the marginal density f_d is close to N(0,1), the score l'_d is near zero, and the Tweedie posterior mean z + l'_d(z) approximates z itself, which has risk d. This means condition (18) may fail exactly where NOMAD is supposed to reduce to lasso-type behavior. Three factors support keeping the verdict at CONDITIONAL rather than moving to REJECT: (1) The assumption is stated explicitly as an assumption, not hidden. The paper is transparent that (18) is required. (2) The supplement is not available, so we cannot verify whether the proof handles this regime-dependence or imposes additional conditions. It is possible the proof of Theorem 6.1 only requires (18) to hold along the specific sequence G_d for the given problem, not uniformly over all sparse regimes, and that the sub-exponential tail condition (17) combined with G_d -> G provides enough structure. (3) The simulations show NOMAD performing well in sparse regimes, which suggests either (18) holds approximately in finite samples or the method is robust to mild violations. However, simulation success does not verify the asymptotic condition. The reader also correctly notes two additional barriers to ACCEPT: deferred proofs and no reproducible code. These are real but secondary to the conceptual concern about (18). The reader's weakest_assumption field precisely identifies the same concern I identify, so agreement is 'agree.' The verdict should remain CONDITIONAL: the framework is legitimate and the unification is valuable, but the consistency theory's applicability to the sparse regime — where the method claims its primary advantage — depends on an assumption whose verification is deferred and whose plausibility in that regime is questionable. If the supplement shows that (18) holds under the stated conditions for sparse G_d, or that the theory applies in a dense-only regime with the sparse behavior being a finite-sample heuristic, the verdict could move to ACCEPT. If (18) provably fails for standard sparse sequences, the consistency claim would need to be restricted, weakening the paper's scope.","tokens_in":23723,"tokens_out":1171,"duration_ms":315358,"concrete_test":"Construct a sequence of sparse signal distributions G_d where pi_d -> 0 (e.g., pi_d = d^{-1/2}) with nonzero signals of magnitude sqrt(2 log d). Compute (1/d) sum_i E[(theta_i - z_i - l'_d(z_i))^2] numerically for the resulting marginal f_d. If this quantity does not converge to 0, Assumption C fails in the sparse regime and Theorem 6.1 does not apply where NOMAD claims to behave like lasso. Additionally, check the Supplement proof of Theorem 6.1 to see whether (18) is used with a uniform constant or whether the proof implicitly requires a denser signal regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies Assumption C, specifically the oracle Tweedie accuracy condition (eq. 18), as the weakest premise. The concern is more specific than 'is the assumption plausible': it is that the assumption is potentially circular in certain regimes and its verification is entirely deferred. Equation (18) requires (1/d) sum_i E[(theta_i - z_i - l'_d(z_i))^2] -> 0. This says the Tweedie posterior mean z + l'_d(z) must recover theta on average across coordinates. But l'_d is the score of the marginal mixture f_d(z) = integral phi(z-theta) dG_d(theta). For this posterior mean to be accurate, the empirical signal distribution G_d must be recoverable from the marginal f_d via the Tweedie deconvolution. When G_d is sparse (e.g., spike-and-slab with small pi), the marginal f_d is close to N(0,1) and the score l'_d(z) is near zero for most z, making the Tweedie estimate z + l'_d(z) approximately z itself — which has risk d, not o(d). The condition (18) may fail precisely in the sparse regime where the method claims to adapt to lasso-type behavior. The paper states that proofs are deferred to Supplement [22], which is not available in the provided text. Without the supplement, we cannot verify whether the proof of Theorem 6.1 actually uses (18) in a way that circumvents this regime-dependence, or whether additional unstated conditions on G_d (beyond sub-exponential tails) are needed. The simulations use spike-and-slab with pi ranging to 0, which is exactly the regime where (18) is most questionable. If (18) fails for sparse G_d, the approximate risk in Proposition 2 does not track the oracle risk, and Theorem 6.1 becomes vacuous in the regime where NOMAD claims its largest advantage over James-Stein.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This manuscript develops a unified approximate-risk-minimization framework for shrinkage-thresholding estimation in normal mean problems. The author introduces the GEST (general shrinkage-thresholding) class, which nests the MLE, James-Stein, ridge, and lasso as special cases through functional parameters t(|z|) and c(|z|). The oracle quadratic risk is expressed as a functional over this class via Stein's identity, and a feasible approximate risk criterion is constructed by replacing the unknown theta with the Tweedie posterior mean z + l'(z), where l'(z) is the marginal score estimated nonparametrically. The resulting estimator, NOMAD, is defined as the minimizer of this approximate risk over a spline sieve. The paper establishes oracle-risk consistency (Theorems 6.1, 6.2, 8.1, 8.2) under regularity conditions, shows that classical estimators arise as special cases under specific score structures (Section 5), extends the framework to correlated normal means via MLE-based and conditional-MLE-based constructions (Section 8), derives an equivalent penalized regression representation with a data-adaptive penalty (Section 9), and provides simulation studies across canonical, correlated, and regression settings (Section 10).","tokens_in":24671,"tokens_out":1635,"duration_ms":475962,"significance":"The paper's ambition is commendable: it provides a single risk-based framework that unifies shrinkage (James-Stein/ridge) and thresholding (lasso/soft-thresholding) as special cases of a two-parameter functional family, with the balance selected data-adaptively via the empirical marginal score. The recovery of classical estimators from specific score shapes (Section 5) is an elegant conceptual contribution. The penalized regression representation (Theorem 9.1) with a data-adaptive penalty shape is a potentially interesting bridge between normal mean theory and regression. The simulation studies are reasonably comprehensive. However, the significance is substantially diminished by the fact that all proofs are deferred to an unavailable supplement, making independent verification of the central theoretical claims impossible at this stage. The framework also introduces several named entities (GEST, NOMAD, power thresholding operator) whose novelty relative to existing shrinkage-thresholding families in the literature (e.g., Antoniadis-Fan, Johnstone-Silverman) is not clearly articulated.","major_comments":[{"comment":"§6.1, Assumption C, Eq. (18): The oracle Tweedie accuracy condition requires (1/d) sum_i E[(theta_i - z_i - l'_d(z_i))^2] -> 0. This is the load-bearing assumption connecting the approximate risk to the oracle risk. The manuscript does not verify this condition in any signal regime, nor does it discuss when it may fail. In sparse regimes (e.g., spike-and-slab with small pi), the marginal f_d is close to N(0,1), the score l'_d(z) is near zero for most z, and z + l'_d(z) approximates z itself, which has risk d. The condition (18) may fail precisely in the sparse regime where the method claims to adapt to lasso-type behavior. The paper should either (a) provide sufficient conditions on G_d under which (18) holds, (b) verify it for specific signal classes (e.g., sparse sequences with bounded sparsity), or (c) discuss the regime where it fails and the implications for the consistency theory.","section":null},{"comment":"All proofs of the main results (Propositions 1-2, Theorems 4.1-4.2, 6.1-6.2, 8.1-8.4, 9.1) are deferred to Supplement [22], which is not available in this submission. Without the supplement, the correctness of the risk functional in Proposition 1, the optimality conditions in Theorems 4.1-4.2, and the consistency arguments in Theorems 6.1 and 8.1-8.2 cannot be verified. This is a fundamental barrier to assessment. The supplement must be provided for the review to be complete.","section":null},{"comment":"§3.1, Proposition 1: The risk functional H involves terms with |z_i|^{c(|z|)} in the denominator and factors of |z_i|^{2-2c(|z|)}. When c(|z|) > 0 and |z_i| -> 0, these terms may diverge. The integrability conditions are mentioned ('the integrability conditions in Proposition 1 hold') but never stated explicitly. Since the GEST class includes c=1 (lasso), the behavior near |z_i|=0 needs careful treatment. The paper should state the integrability conditions explicitly and verify that they are satisfied for the special cases in Section 5.","section":null}],"minor_comments":[{"comment":"§2.2, Eq. (6): The notation |ZZZ| for the vector of absolute values is introduced but the Hadamard product symbol is used without explicit definition until later. Clarification would help.","section":null},{"comment":"§4.2, Eqs. (12)-(13): The optimality conditions involve sums over i of terms depending on the full vector through c(|z|), but c is treated as a function of the summary statistics s_c(z) in Section 6.1. The relationship between the functional form in Section 4 and the sieve parametrization in Section 6 should be clarified.","section":null},{"comment":"§5.2, Proposition 4: The claim that the NOMAD estimator reduces to James-Stein under Gaussian marginals requires the empirical-Bayes substitution of tau^{-2} by (d-2)/||z||^2. This substitution is a separate step from the approximate-risk minimization. The proposition should clarify that the reduction is to the shrinkage form, with the JS-specific constant arising from a post-processing step.","section":null},{"comment":"§8.3: The SURE formula for the CMLE-based construction involves the matrix Psi whose definition uses indicator functions for nonzero components. The differentiability of the positive-part operator at zero is a standard subtlety; the paper should note how this is handled.","section":null},{"comment":"§9.3, Theorem 9.1, Eq. (31): The penalty involves an integral with upper limit (X'X)^{1/2}_{ii} beta_i / sigma, which couples the penalty to the design through the diagonal of the precision matrix. This differs from standard coordinate-separable penalties. The remark should clarify whether this design-dependence is essential or an artifact of the reduction.","section":null},{"comment":"§10: No code or reproducible computational details are provided. Given the complexity of the NOMAD procedure (score estimation, Newton iteration for c, fixed-point iteration for CMLE), code availability would strengthen the submission.","section":null},{"comment":"The MSC2020 subject classifications (00X00) and keywords are placeholder codes. These should be filled in with appropriate classifications.","section":null},{"comment":"§7.3, Figure 2: The y-axis scales differ across the four panels, making cross-signal comparison difficult. Consider using a common scale or clearly labeling the different scales.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is ambitious in scope but the theoretical core is unverifiable without the supplement. The reader's concern about Assumption C (Eq. 18) is well-placed: in sparse regimes, the Tweedie plug-in may not approximate theta well, and this is exactly the regime where the method claims lasso-type adaptation. I recommend major revision with the supplement provided and the condition (18) either verified or discussed with respect to its failure regimes. The novelty of the GEST class relative to existing shrinkage-thresholding families (e.g., Antoniadis and Fan 2001, Johnstone and Silverman 2005) should also be clarified."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. The referee raises three major comments: (1) the oracle Tweedie accuracy condition (Assumption C, Eq. 18) is not verified in any signal regime and may fail in sparse settings; (2) all proofs are deferred to an unavailable supplement; and (3) the integrability conditions in Proposition 1 are not stated explicitly and the behavior near |z_i|=0 needs careful treatment. We address each point below.","responses":[{"response":"The referee raises an important point about the load-bearing nature of Assumption C, Eq. (18). We agree that the manuscript does not currently provide sufficient conditions or verification for specific signal classes, and that the sparse regime warrants careful discussion. We address the referee's three suggested remedies in turn. (a) Sufficient conditions: In the revised manuscript, we will add a new subsection (or a dedicated remark in Section 6.1) providing sufficient conditions on G_d under which (18) holds. Specifically, if G_d converges to G in Wasserstein distance and the marginal density f(z) = integral phi(z-theta) dG(theta) has a score l'(z) satisfying a Lipschitz condition and moderate growth, then (18) follows from standard empirical Bayes posterior mean consistency results (cf. Efron, 2011; Brown and Greenshtein, 2009). The key requirement is that G_d is not degenerate at zero in the limit, i.e., the limiting signal distribution G has positive variance or at least is not a point mass at zero. (b) Verification for specific signal classes: We will verify (18) explicitly for two signal classes. First, for dense signals where G_d converges to a non-degenerate distribution (e.g., N(0, tau^2) with tau > 0), the Tweedie posterior mean z + l'_d(z) is a consistent estimator of theta_i in the sense of (18) under standard density estimation rates. Second, for sparse signals with bounded sparsity pi_d -> pi in (0,1), the marginal is a two-component mixture and the score is well-defined and estimable; (18) holds provided the non-null component has bounded second moment and the density estimator achieves the uniform consistency rate stated in Assumption C. (c) The degenerate sparse regime: The referee correctly identifies a genuine limitation. If pi_d -> 0 (vanishing sp)","revision_made":"no","referee_comment":"§6.1, Assumption C, Eq. (18): The oracle Tweedie accuracy condition requires (1/d) sum_i E[(theta_i - z_i - l'_d(z_i))^2] -> 0. This is the load-bearing assumption connecting the approximate risk to the oracle risk. The manuscript does not verify this condition in any signal regime, nor does it discuss when it may fail. In sparse regimes (e.g., spike-and-slab with small pi), the marginal f_d is close to N(0,1), the score l'_d(z) is near zero for most z, and z + l'_d(z) approximates z itself, which has risk d. The condition (18) may fail precisely in the sparse regime where the method claims to adapt to lasso-type behavior. The paper should either (a) provide sufficient conditions on G_d under which (18) holds, (b) verify it for specific signal classes (e.g., sparse sequences with bounded sparsity), or (c) discuss the regime where it fails and the implications for the consistency theory."}],"tokens_in":23755,"tokens_out":747,"duration_ms":217362,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper introduces the GEST class — a two-functional-parameter family (t, c) that nests James-Stein shrinkage and lasso soft-thresholding as special cases — and builds a data-driven selection rule (NOMAD) by minimizing an approximate risk criterion constructed from Stein's identity plus Tweedie's formula. The unification is genuine and well-motivated. The reductions in Section 5 are clean: Gaussian marginals give James-Stein, Laplace-like tails give lasso, vanishing score gives MLE. The penalized regression representation in Section 9, where the power-thresholding operator induces a data-adaptive penalty interpolating ridge and lasso, is a nice structural observation. The wavelet and regression extensions are natural, and the simulations are favorable across regimes. Credit is earned for the framework design and the calculus-of-variations optimality conditions in Theorems 4.1–4.2, which are explicit and checkable. The CMLE-based correlated extension with its SURE formula (Section 8.3) is a practical contribution — the restriction to constant (t, c) there is honest and well-justified. The soft spots are real but concentrated. The load-bearing one is Assumption C, specifically equation (18): the oracle Tweedie accuracy condition requiring (1/d) sum E[(theta_i - z_i - l'_d(z_i))^2] -> 0. The stress-test concern here lands. In sparse spike-and-slab regimes with small pi, the marginal f_d is close to N(0,1), the score l'_d is near zero, and the Tweedie plug-in z + l'_d(z) is approximately z itself — which has risk d, not o(d). So (18) may fail precisely where NOMAD claims its largest advantage over James-Stein. The paper does not discuss this regime dependence. This is not fatal — the assumption could hold under conditions on G_d that exclude extreme sparsity, or the theory could be qualified — but it needs to be addressed. Second, all proofs are deferred to a supplement not available in the submission. The main theorems (6.1, 6.2, 8.1, 8.2) cannot be verified without it. Third, no code or data is provided, so the simulations are not independently reproducible. These are standard issues for a first submission. The circularity concern raised by the reader is overstated — the approximate risk is a plug-in approximation, not a circular definition. The optimality conditions are derived by standard calculus of variations on the approximate functional. The real question is whether the approximation tracks the oracle risk, which is exactly the (18) issue. This paper is for researchers in high-dimensional estimation and regularization who want a principled bridge between shrinkage and thresholding. It deserves a serious referee who can check the supplement and press on Assumption C. I'd recommend peer review with a request to the author to (1) provide the supplement, (2) explicitly discuss when (18) holds and when it fails, and (3) provide code.","headline":"Unified shrinkage-thresholding framework via approximate risk minimization; consistency theory rests on an unverified Tweedie accuracy assumption that may fail in sparse regimes","tokens_in":24781,"tokens_out":723,"would_cite":true,"duration_ms":126905,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F10","62F12","62J07"],"pacs":[],"model":"glm-5.2","headline":"One risk criterion unifies shrinkage and thresholding","keywords":["NOMAD","GEST","approximate risk minimization","shrinkage","thresholding","James-Stein","lasso","Tweedie's formula"],"falsifier":"If the Tweedie plug-in z + l'(z) fails to approximate theta on average — for instance, if the empirical signal distribution G_d does not converge, or if the score estimator is not uniformly consistent — then the approximate risk criterion diverges from the oracle risk, and Theorem 6.1's consistency claim becomes vacuous. One could construct a counterexample with an oscillating or non-convergent signal distribution where the score estimate is systematically biased.","tokens_in":23853,"feed_emoji":"⚖️","tokens_out":1055,"duration_ms":219713,"temperature":0.7,"pith_summary":"The paper introduces a single family of estimators, called GEST, that contains both classical shrinkage rules (James-Stein, ridge) and classical thresholding rules (lasso soft-thresholding) as special cases. The family is parameterized by two functional components: a threshold function t and a rate function c, which together determine how aggressively each coordinate is shrunk toward zero and whether small coordinates are set exactly to zero. The central construction is a data-dependent approximate risk criterion, built by combining Stein's identity with a nonparametric empirical Bayes plug-in (Tweedie's formula), that replaces the unknown true mean in the oracle risk with an estimate derived solely from the observed data. Minimizing this approximate risk over the GEST class yields an estimator called NOMAD. The paper proves that, under regularity conditions, NOMAD's risk converges to the oracle-optimal risk within the class at rate o_P(d). The framework extends to correlated normal means (via MLE-based and conditional-MLE-based constructions with SURE criteria) and to linear regression, where the shrinkage-thresholding map induces an equivalent penalized regression with a data-adaptive penalty that recovers ridge (c=0) and lasso (c=1) as special cases.","feed_headline":"One risk criterion unifies shrinkage and thresholding","feed_subtitle":"A data-driven approximate risk picks between James-Stein and lasso automatically, recovering each as a special case.","key_machinery":"The GEST class (equation 6) parameterized by (t, c); the approximate risk criterion (Proposition 2) replacing unknown theta with Tweedie plug-in z+l'(z); the optimality conditions (Theorem 4.2) yielding implicit equations for t and c; the sieve-based consistency proof (Theorem 6.1); the CMLE fixed-point construction for correlated settings (equation 25); the inverse power-thresholding operator inducing the penalized regression representation (Theorem 9.1).","core_discovery":"The key structural finding is that the shape of the empirical marginal score function l'(z) determines where on the shrinkage-to-thresholding spectrum the optimal estimator lies. When the marginal is Gaussian, the score is linear, the optimality conditions force c=0, and NOMAD reduces to James-Stein-type pure shrinkage. When the marginal has Laplace-like tails, the score magnitude is approximately constant, the optimality conditions force c=1, and NOMAD reduces to lasso soft-thresholding. For intermediate marginal shapes, the optimizer selects 0