{"id":"9aee0ba3-028f-4f18-9013-b9323d74f645","arxiv_id":"2607.05350","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"Stochastic mirror ascent provably finds an approximately least-favorable distribution and minimax estimator for the Bounded Normal Mean problem, yielding 6–18% risk improvements over the minimax linear estimator.","lead":"The paper gives a computational algorithm—stochastic mirror ascent—that provably finds an approximately minimax estimator for the classical Bounded Normal Mean problem, with explicit error bounds. A smart generalist might read it because it makes statistical decision theory computationally implementable, including a macroeconometric application to combining VAR and local projection impulse response estimates.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Theoretical guarantees are sound; the main concern is that formal certification via Theorem 1 only applies when the full J(ε) iterations complete, which for m ≥ 2.2 exceeds practical time budgets—leaving reported risk improvements for moderate m empirically validated but not formally certified.","rationale":"The reader correctly identified the gap between theoretical iteration counts and practical performance as the main concern. My detailed check of the proofs confirms that the theoretical results (Theorems 1–4) are correctly derived from standard optimization and decision-theoretic tools. The mirror descent regret bound, the Hoeffding concentration lemma, the discretization error bound, and the Wasserstein convergence argument are all applied correctly. The concern is not about correctness but about practical certification: the Theorem 1 guarantee only applies when J(ε) iterations are completed, which for m ≥ 2.2 exceeds reasonable time budgets. The authors acknowledge this explicitly (Section 3, remark 2; Section 4.4). For the application (m ≈ 1.5–1.6), the algorithm completes in under a minute and the guarantee applies. The reader's CONDITIONAL verdict with MODERATE confidence is appropriate: the theoretical contribution is solid, but the lack of shipped code and the certification gap for moderate m prevent full acceptance. I add one observation the reader did not emphasize: the 'certified' upper bound in Appendix D.3 does not account for Monte Carlo noise in the risk evaluation (N = 1000 draws), which means the graphical certification in Figure 3 is not a strict upper bound. This is a minor issue for the theoretical claims but affects the empirical validation. The concrete test I propose—verifying the Theorem 1 bound against the known closed-form solution for m = 1 across many seeds—would directly test whether the high-probability guarantee is properly calibrated.","tokens_in":33517,"tokens_out":9382,"duration_ms":457060,"concrete_test":"For m = 1 (where the exact minimax value v*(1) is known from Casella and Strawderman [1981] and J(ε) ≈ 63K iterations is feasible in ~3 seconds), run Algorithm 1 with the full Theorem 1 parameter set across 500 independent random seeds. For each seed, compute inf_d r(d, π_ε) and check whether it exceeds v*(1) − ε(1 + √(ln(1/α)/ln(I(ε)))) for α = 0.05. If the fraction of seeds satisfying this inequality is below 95%, the high-probability guarantee in Theorem 1 is miscalibrated. Additionally, recompute the certified upper bound in Figure 3 (Appendix D.3) using a Hoeffding confidence interval for each grid-point risk estimate bR_N(d, θ_i) with N = 1000; if the corrected upper bound exceeds the linear minimax risk for any m ∈ {1, 1.6, 2}, the claimed risk improvements lack formal certification even in the feasible regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I traced the proof of Theorem 1 in Appendix A.3 step by step. The mirror descent regret bound (inequality 4.10 from Bubeck [2015]) is applied correctly: with the negative-entropy mirror map on Δ_{I-1}, the Bregman divergence term is ln(I), the dual norm is L∞, and ||g̃_j||_∞ ≤ 4m² = M holds almost surely since (d_{π_j}(Y)−θ_i)² ∈ [0, 4m²]. The Hoeffding-type concentration in Lemma 3 is also correct: each ⟨Δ_j, x_j⟩ is bounded in [−8m², 8m²] (since ||x_j||_1 ≤ 2 and ||Δ_j||_∞ ≤ 4m²), giving L = 8m² and the tail bound exp(−Jδ²/(128m⁴)). The parameter choices η = (ε/2)/M² and J = 2M²ln(I)/(ε/2)² exactly cancel the deterministic error terms to ε/2, and the stochastic correction δ(α,ε) = ε√(ln(1/α)/ln(I(ε))) follows algebraically. Lemmas 1–2 (Lipschitz risk, discretization error) and Theorems 2–4 (Wasserstein convergence, stochastic minimax sequence, uniform convergence over compacts) all follow from standard arguments applied correctly. The theory is internally consistent and correctly derived. The load-bearing concern is practical, not theoretical: the guarantee in Theorem 1 requires completing J(ε) = ⌈2M²ln(I(ε))/(ε/2)²⌉ iterations. For m = 3 with ε(m) ≈ 0.18, this is ~2.3 million iterations (~2 hours). Figures 6–8 show that wall-clock limits of 5–15 minutes bind for m ≥ 2.2, meaning the Theorem 1 guarantee does not apply to the time-limited results. The 6–18% risk improvements reported for m ≥ 2.2 are thus empirically observed but lack the formal ε-minimax certification that the paper's theoretical framework provides. For the application (m ≈ 1.5–1.6, ~32–54 seconds), the full iteration count is feasible and the guarantee applies. Additionally, the certified upper bound in Appendix D.3 accounts for grid discretization via Lemma 1 but does not account for Monte Carlo noise in the N = 1000 risk estimates (Eq. 17), so the 'certifiable' bound in Figure 3 is not a strict upper bound on the worst-case risk.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes a computational approach to constructing an approximately minimax estimator for the classical Bounded Normal Mean (BNM) problem. The authors use stochastic mirror ascent with a negative-entropy mirror map to approximately solve the concave maximization problem of finding a least-favorable distribution over a discretized parameter space. The paper provides four main theorems: Theorem 1 gives a high-probability bound on the approximation quality of the least-favorable distribution as a function of the step size, number of iterations, and grid size; Theorem 2 establishes convergence in 1-Wasserstein distance to the true least-favorable distribution; Theorem 3 shows the resulting Bayes estimators form a stochastic minimax sequence; and Theorem 4 proves uniform convergence over compact sets to the exact minimax estimator. Simulations demonstrate 6–18% risk improvements over the minimax linear estimator for m in {1,...,4}, and the method is applied to aggregating LP and VAR impulse response estimates in four macroeconomic applications.","tokens_in":34478,"tokens_out":1606,"duration_ms":280053,"significance":"The paper makes a genuine contribution by providing explicit, finite-sample theoretical guarantees for a computational approach to the BNM problem—a classical problem where exact minimax estimators are unavailable for general m. The combination of a new Lipschitz continuity result for the risk function (Lemma 1), a clean discretization bound (Lemma 2), and the stochastic mirror ascent convergence guarantee (Theorem 1) with explicit parameter recommendations is a strength. The extension to stochastic minimax sequences (Definition 1, Theorem 3) and uniform convergence over compacts (Theorem 4) are well-motivated and correctly derived. The application to LP–VAR aggregation via the Armstrong–Kline–Sun framework connects the theoretical results to a problem of active interest in macroeconometrics. The authors are transparent about the conservativeness of their bounds and the gap between theoretical iteration counts and practical performance, which is commendable.","major_comments":[{"comment":"Section 4.4, Figures 6–8: The paper reports risk improvements of 6–18% for m in {1,...,4}, but for m >= 2.2 the wall-clock time limits (5 or 15 minutes) bind before the Theorem 1 iteration count J(epsilon) is reached. This means the formal epsilon-minimax certification from Theorem 1 does not apply to a subset of the headline simulation results. The authors acknowledge this (Remark 2 in Section 3, and the discussion in Section 4.4), but the abstract and introduction state the 6–18% improvement range without qualifying that some of these results lack formal certification. I recommend that the abstract and the simulation summary explicitly distinguish between results where the Theorem 1 guarantee applies (m approximately <= 2.2) and results where only empirical performance is demonstrated. This is a presentation issue that affects how readers interpret the central empirical claim, but itis","section":null},{"comment":"Section 4.4 and Figures 7–8: When the wall-clock limit binds and the algorithm does not complete J(epsilon) iterations, the output distribution pi_epsilon is not guaranteed to be epsilon-least-favorable. The paper reports a 'certified upper bound' on the worst-case risk in Figure 3 (using the Lipschitz property from Lemma 1), but it is unclear whether this certification strategy is still valid when the algorithm has been stopped early. The Lipschitz-based certification in Appendix D.3 bounds the gap between grid-evaluated risk and the true supremum, but it does not address the optimization error from incomplete iterations. The authors should clarify that the certified upper bound in Figure 3 applies only to the discretization/Monte Carlo gap, not to the optimization error, and that for time-limited runs the optimization error is uncontrolled. This distinction is important for readers who","section":null}],"minor_comments":[{"comment":"The notation v*(m) is used for the maximin value (Equation 6) and v-bar*(m) for the minimax value (Equation 2). These are shown to be equal (Equation 7), but the dual use of v* and v-bar* throughout the paper can cause confusion. Consider unifying the notation after Equation 7 is established, or adding a remark that they coincide henceforth.","section":null},{"comment":"In Algorithm 1, Step 6, the Bayes decision rule d_pi_j(y) is written as a ratio involving theta_i and phi(y;theta_i). The formula is correct but could benefit from a brief inline explanation that this is the posterior mean under prior pi_j, to help readers less familiar with the BNM literature.","section":null},{"comment":"Figure 1 shows J(m) and I(m) as functions of m, but the y-axis scale for J(m) is not clearly labeled (it appears to be in units of 10^6 for larger m). Adding explicit axis labels would improve readability.","section":null},{"comment":"The reference to 'Proposition 4.2 in Ben-Tal, Margalit, and Nemirovski [2001]' on page 2 is cited to support the claim that mirror ascent is 'essentially the best first-order iterative algorithm' for this problem. The claim is strong; the authors should verify that the proposition's setting (ordered subsets mirror descent for tomography) directly supports this optimality claim for the BNM simplex problem, or soften the language.","section":null},{"comment":"In Section 5.4, the values of m across applications (Table 1) range from 1.52 to 1.64. These are below 2.2, so the Theorem 1 guarantee should apply. It would be useful to explicitly state that the formal certification applies to all four applications, as this strengthens the applied results.","section":null},{"comment":"The paper mentions (footnote 3) that off-the-shelf methods like fmincon might perform comparably but lack theoretical guarantees. A brief comparison with such methods in the simulation section would strengthen the case for the proposed algorithm, even if only to confirm that the mirror ascent approach is competitive in practice.","section":null},{"comment":"Typo on page 28, Section 4.5: 'factors of approximately 2, 3.5, and 5' should clarify these are reduction factors in the number of iterations, not in runtime. Also, Figure 15 in Appendix D is referenced but the factors (2x, 3.5x, 5x) should be explicitly tied to the corresponding m values (1, 1.6, 2).","section":null}],"recommendation":"minor_revision","confidential_remarks":"The reader's report and stress-test note raise a valid concern about the gap between the theoretical iteration count and practical performance. Having traced the proofs, I agree that the theory is sound and correctly derived. The concern is primarily about presentation: the abstract and introduction conflate results with and without formal certification. This is fixable through clearer framing and does not undermine the theoretical contribution. The application to LP-VAR aggregation is well-motivated and the m values used (1.52-1.64) fall within the certified range, which is a point the authors should emphasize. I would not view the conservativeness of the bounds as a deficiency; it is inherent to providing a priori guarantees for a general concave maximization problem, and the authors are appropriately transparent about it."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. We are grateful for the positive assessment of the paper's theoretical contributions and the recommendation of minor revision. Both major comments raise legitimate and related points about the presentation of simulation results when the wall-clock time limit binds before the Theorem 1 iteration count is reached. We address each comment below.","responses":[{"response":"The referee is correct that the abstract and introduction do not currently distinguish between results with formal Theorem 1 certification and results where only empirical performance is demonstrated. We agree this is an important presentation issue and will revise accordingly. Specifically, we will modify the abstract to note that the 6-18% range includes both certified results (for smaller m, where the algorithm completes the theoretically required number of iterations) and empirically demonstrated improvements (for larger m, where wall-clock time limits bind before J(epsilon) is reached). We will add a similar clarification in the introduction and in the simulation summary in Section 4.4. This will make transparent which subset of results carries the formal epsilon-minimax guarantee from Theorem 1 and which subset reports only empirical performance.","revision_made":"yes","referee_comment":"Section 4.4, Figures 6-8: The paper reports risk improvements of 6-18% for m in {1,...,4}, but for m >= 2.2 the wall-clock time limits bind before the Theorem 1 iteration count J(epsilon) is reached. This means the formal epsilon-minimax certification from Theorem 1 does not apply to a subset of the headline simulation results. The authors acknowledge this, but the abstract and introduction state the 6-18% improvement range without qualifying that some of these results lack formal certification. The referee recommends that the abstract and simulation summary explicitly distinguish between results where the Theorem 1 guarantee applies (m approximately <= 2.2) and results where only empirical performance is demonstrated."},{"response":"The referee is correct on the substance. The Lipschitz-based certification in Appendix D.3 addresses only the gap between the worst-case risk evaluated on the finite grid and the true supremum over the continuous parameter space [−m, m]. It does not address the optimization error that arises when the algorithm is stopped before completing J(epsilon) iterations. We will revise the paper to make this distinction explicit. Specifically, we will add clarifying language in Section 4.4 and in Appendix D.3 stating that: (i) the certified upper bound in Figure 3 (and the corresponding bounds in Figures 7-8 for time-limited runs) controls only the discretization/Monte Carlo gap between grid-evaluated risk and the true supremum; (ii) for results where the algorithm completes J(epsilon) iterations, the optimization error is controlled by Theorem 1, so the overall certification is valid; and (iii) for time-limited runs where J(epsilon) is not reached, the optimization error is uncontrolled by our theory, and thus the reported upper bound should not be interpreted as a formal epsilon-minimax guarantee. We will also adjust the labeling in the figures and surrounding text to make this distinction clear to readers.","revision_made":"yes","referee_comment":"Section 4.4 and Figures 7-8: When the wall-clock limit binds and the algorithm does not complete J(epsilon) iterations, the output distribution pi_epsilon is not guaranteed to be epsilon-least-favorable. The paper reports a 'certified upper bound' on the worst-case risk in Figure 3 (using the Lipschitz property from Lemma 1), but it is unclear whether this certification strategy is still valid when the algorithm has been stopped early. The Lipschitz-based certification in Appendix D.3 bounds the gap between grid-evaluated risk and the true supremum, but it does not address the optimization error from incomplete iterations. The authors should clarify that the certified upper bound in Figure 3 applies only to the discretization/Monte Carlo gap, not to the optimization error, and that for time-limited runs the optimization error is uncontrolled."}],"tokens_in":33244,"tokens_out":846,"duration_ms":57251,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"This paper applies stochastic mirror ascent to the Bounded Normal Mean (BNM) problem and provides formal convergence guarantees — something prior numerical approaches (Feldman and Brown, Gourdin et al.) lacked. The core theoretical contribution is real and well-executed.","headline":"Solid theory, honest about its limitations, deserves a serious referee","tokens_in":34573,"tokens_out":702,"would_cite":true,"duration_ms":29819,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Algorithm computes near-minimax estimator for bounded normal mean","keywords":["Bounded Normal Mean","minimax estimation","stochastic mirror ascent","least-favorable distribution","statistical decision theory","impulse response estimation","Lipschitz risk function"],"falsifier":"If the risk function failed to be Lipschitz continuous in θ uniformly over decision rules, the discretization error would not be controllable by simply refining the grid, and the entire approach of reducing the infinite-dimensional maximin problem to a finite-dimensional one would break down. Additionally, if the empirical supergradient magnitudes were systematically larger than the worst-case bound suggests (rather than smaller, as observed), the practical performance would match the pessimistic theoretical iteration counts and the method would be computationally infeasible for moderate m.","tokens_in":33736,"feed_emoji":"📊","tokens_out":1235,"duration_ms":78295,"temperature":0.7,"pith_summary":"This paper presents a computational method for finding an approximately minimax estimator of a bounded normal mean — the classical problem of estimating the mean of a normal distribution when that mean is known to lie in a bounded interval. The authors use stochastic mirror ascent, an iterative optimization algorithm that exploits the geometry of the probability simplex, to find an approximately least-favorable distribution: the prior distribution that an adversary would choose to make the statistician's job hardest. The Bayes estimator under this approximate prior then serves as the approximately minimax estimator. The paper provides four main theorems. Theorem 1 shows that the algorithm, with explicitly specified step size, grid size, and iteration count, outputs a distribution whose Bayes risk is within a user-controlled tolerance of the maximin value, with high probability. Theorems 2 through 4 establish that as the tolerance shrinks, the approximate least-favorable distribution converges in Wasserstein distance to the true one, and the corresponding Bayes estimator converges uniformly over compact sets to the exact minimax estimator. A key technical ingredient is a new Lipschitz continuity result for the risk function (Lemma 1), which controls the discretization error introduced by restricting optimization to a finite grid. Simulations show risk improvements of 6–18% over the best linear minimax estimator, which is known to be at most 20% suboptimal. The method is applied to the practical problem of combining local projection and vector autoregression estimates of macroeconomic impulse responses, where the bounded normal mean problem arises naturally from bounding the bias of the VAR estimator.","feed_headline":"Algorithm computes near-minimax estimator for bounded normal mean","feed_subtitle":"Stochastic mirror ascent provably approximates the least-favorable distribution, yielding 6–18% risk gains over linear estimators and a new ","key_machinery":"Stochastic mirror ascent with negative-entropy mirror map on the probability simplex; Lipschitz continuity of the risk function R(d,θ) in θ with constant 3m²+4m; discretization of the parameter space to a finite grid; multiplicative weights update for prior distribution; Monte Carlo estimation of supergradients using a single draw per grid point per iteration.","core_discovery":"The central discovery is that the least-favorable distribution of the bounded normal mean problem — which has no closed-form solution for moderate bounds — can be approximated to arbitrary precision by running stochastic mirror ascent on a discretized version of the maximin problem, and that both sources of error (discretization and optimization) can be explicitly bounded and controlled. The risk function's Lipschitz continuity in the parameter, with a constant that is uniform across all decision rules, is the key property that makes the discretization error tractable. This, combined with the well-understood convergence theory of mirror ascent on the simplex, yields a procedure with end-to-端","pith_inferences":["The Lipschitz constant 3m²+4m used in the analysis is derived from a total variation bound on normal distributions that is known to be improvable by a factor of roughly √(2/π); adopting the sharper bound would reduce both the required grid size and iteration count.","The observation that empirical supergradient magnitudes are 2–5x smaller than the worst-case Lipschitz bound suggests that data-adaptive step sizes (as in AdaGrad-style methods) could close much of the gap between theoretical and practical iteration counts, though extending high-probability guarantees to data-dependent step sizes would require new concentration arguments.","The restriction to an ex ante fixed grid could be lifted by algorithms that incrementally add support points, potentially exploiting the known finite-support property of the true least-favorable distribution to reduce dimensionality."],"forward_implications":["The algorithm makes minimax estimation practically accessible for the bounded normal mean problem, which serves as a building block in nonparametric estimation and robust inference beyond econometrics.","The framework of controlling discretization error via Lipschitz continuity of the risk function, plus optimization error via mirror ascent theory, could be applied to other statistical decision problems where the least-favorable distribution lacks a closed form.","The application to combining LP and VAR estimators suggests that nonlinear bias correction of VAR estimates can systematically outperform linear shrinkage, with the degree of improvement tied to the tightness of the assumed bias bound.","The gap between the theoretically required iteration count (millions for moderate m) and the practically achieved accuracy suggests that adaptive step-size methods could substantially reduce computational cost while maintaining guarantees."],"fun_headline_variants":["Stochastic mirror ascent approximates least-favorable distribution for bounded normal mean","Estimating bounded normal means via stochastic mirror ascent","Bounded normal mean minimax estimation approximated via stochastic mirror ascent","Computing approximate minimax estimators for bounded normal means via mirror ascent","Stochastic mirror ascent yields 6-18% risk gains for bounded normal mean estimation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The convergence guarantee requires a number of iterations that scales as m⁴ ln(m), reaching roughly 2.3 million for m=3, driven by a worst-case Lipschitz constant M=4m² that the authors' own simulations suggest overstates the required computational effort by a factor of 2–5. The authors note this bound is sufficient but not necessary, and that far fewer iterations suffice in practice.","fun_headline_variants_meta":{"raw":{"variants":["Stochastic mirror ascent approximates least-favorable distribution for bounded normal mean","Estimating bounded normal means via stochastic mirror ascent","Bounded normal mean minimax estimation approximated via stochastic mirror ascent","Computing approximate minimax estimators for bounded normal means via mirror ascent","Stochastic mirror ascent yields 6-18% risk gains for bounded normal mean estimation","Approximating least-favorable distributions for bounded normal means via mirror ascent","Stochastic mirror ascent for approximate minimax estimation of bounded normal means"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1104,"prompt_tokens":471,"completion_tokens":633,"prompt_tokens_details":null},"tokens_in":471,"tokens_out":633,"duration_ms":7875,"temperature":1.0,"reasoning_tokens":593,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-07T15:27:43.219598+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the risk function failed to be Lipschitz continuous in θ uniformly over decision rules, the discretization error would not be controllable by simply refining the grid, and the entire approach of reducing the infinite-dimensional maximin problem to a finite-dimensional one would break down. Additionally, if the empirical supergradient magnitudes were systematically larger than the worst-case bound suggests (rather than smaller, as observed), the practical performance would match the pessimistic theoretical iteration counts and the method would be computationally infeasible for moderate m.","supporting_citations":[],"review_version":1}