acceptodds
Under review as a conference paper at ICLR 2027

Parametric APUB for Decision Making under Parameter Uncertainty

Abstract

Parametric optimization can exploit a known distribution family but usually treats its fitted parameters as certain. We develop parametric Average Percentile Upper Bound (APUB) optimization to adjust a decision and certify its expected cost against this epistemic uncertainty. APUB refits the parametric model on bootstrap datasets and minimizes an average of upper percentiles of the resulting expected-cost functions. We study generative parametric bootstrap and weighted-likelihood bootstrap as two prior-free constructions in one framework. The theory establishes decision consistency under construction-specific conditions, regular fixed-decision coverage, and finite-ensemble error control; an adaptive algorithm grows and reuses the fitted ensemble. In paired, fixed-truth mixed-normal and mixed-log-normal newsvendor experiments, positive-level PB- and WLB-APUB have lower mean regret than SAA and fitted-model optimization at small samples, while the adjustment contracts as the sample size increases. Only APUB displays approximately level-tracking coverage curves, giving its nominal level an interpretable one-sided confidence meaning. APUB avoids this prior-specification risk through a data-driven resampling adjustment.

Primary area:
optimization
Keywords:
Average Percentile Upper Bound, Epistemic Uncertainty, Parameter Uncertainty, Optimization under Uncertainty

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_L79uOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the parametric Average Percentile Upper Bound (APUB) optimization, which uses the conditional value at risk (CVaR) as the objective function. The paper considers two bootstrap methods, generative parametric bootstrap and weighted likelihood bootstrap, for constructing the objective function. Under correct model specification, the paper establishes the convergence of loss functions in probability, the asymptotic coverage at the ideal APUB optimizer, and a uniform error bound for the loss function conditional on the data. The paper demonstrates the excellent performance of the proposed two bootstrap APUB in comparison with the nonparametric bootstrap APUB, Bayesian DRO and moment DRO methods.

Critical strengths

The paper is generally well written. The theory developed in this paper studies the APUB optimization previously proposed in Liu et al. (2026), for two extended versions of parametric bootstrap. The theory is generally solid and technically substantial, covering the convergence of loss function under parametric bootstrap estimators, the coverage of selected decisions, and uniform error bound, though they depend on strong high-level assumptions. The proofs seem mostly correct. The empirical results demonstrate that the APUB with parametric bootstrap methods perform well and better than the Bayesian DRO under misspecified priors and the moment DRO methods.

Reference

Shixin Liu, Ming Gao, and Jian Hu. (2026). Minimizing upper confidence bounds: A data-driven framework for stochastic programming. In Proceedings of the 43rd International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research.

Critical weaknesses
  1. The current APUB with parametric bootstrap is a natural extension of the nonparametric bootstrap version of APUB recently proposed in Liu et al. (2026). The objective functions using CVaR are the same in the two papers. The main difference is that Liu et al. (2026) uses a nonparametric bootstrap that resamples the original data directly, while the current paper uses a parametric bootstrap that samples from the parametric model and a weighted likelihood bootstrap.

Regarding the theory, I feel that the theory established in the paper is standard and not surprising, as the results on convergence of loss in probability, asymptotic coverage of the ideal optimizer, and uniform error bound are generally expected based on the theoretical properties of bootstrap estimation.

More importantly, some theorems rely on strong high-level assumptions, which probably should be verified using low-level assumptions on the model and the data generation process. Such assumptions include

  • Assumption (A4) in Proposition 1, regarding the uniform convergence of MNW(θ~NW)M_N^{\text{W}}(\tilde \theta_N^{\text{W}}). This generally consists of two convergence results: the uniform convergence of θ~NW\tilde \theta_N^{\text{W}} to θ\theta, and the convergence of the MNWM_N^{\text{W}} function. Both convergence results should be proved under low-level assumptions.

  • Assumption (P3) in Proposition 2, for a similar reason to Assumption (A4) above.

  • Assumptions (9), (10), and (11) in Theorem 2: The three convergence results should be proved, rather than being assumed. The form of the function s(x)s(x) is also not given.

In addition, the propositions and theorems rely on some strong and unconventional assumption, such as Assumption (P2) in Proposition 2. I hope the authors can clarify when such assumptions can hold, and it will be best if all the assumptions can be directly verified for the models in Section 5.

  1. I feel that the paper has not adequately demonstrated the advantage of APUB over the competitors. The paper starts with a discussion on the distributionally robust optimization literature. However, the main text has not investigated the robustness of APUB optimization to model misspecification and/or data contamination. For all the theory and numerical experiments in the paper, the model is always correctly specified, with a true parameter. What if the model specification is incorrect, or the data contains some outliers? I think one potential advantage of using bootstrap methods is that they can possess some robustness in such scenarios, but this robustness issue has been completely neglected in the current paper.
Questions for authors
  1. The theory only concerns the convergence of the loss function HH, but it does not seem to establish the convergence for the estimators of θ\theta and the APUB optimal solution of xx. It seems some of these convergence results have been proved en route. Perhaps the authors can state the results more clearly. For example, in Theorem 2, the convergence of xN,γx_{N,\gamma} to x⋆x^{\star} should be a proved result, but it reads like an assumption.

  2. The definition of KhK_h in Theorem 3 needs to be clearer.

  3. The numerical experiments do not take into account the computation time. The bootstrap-based APUB is already computationally intensive. I wonder if the nonparametric bootstrap APUB may be computationally slightly more efficient than the parametric version. The authors should report the computation time for all the competing methods under consideration.

  4. In practice, how do one choose the value of γ\gamma? From the numerical results, it seems that the regret does not change much, but only the observed coverage gets better when γ\gamma increases. However, due to the limited sample size, we cannot increase γ\gamma arbitrarily close to 1, as fewer data points in the tail may increase random error. What is a suggested method to choose the best value for γ\gamma in practice?

Reviewer_21q9OpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors consider a setting where one probability distribution Pθ0P_{\theta_0} is taken from a family of distributions indexed by θ\theta. The setting is to predict, i.e., get an estimate θ^N\hat \theta_N, and then optimize, i.e., minimize H(x,θ)=Eξ∼Pθ(ℓ(x,ξ))H(x, \theta) = E_{\xi \sim P_{\theta}}(\ell (x,\xi)). They consider two bootstrap methods; PB simulates samples from Pθ^NP_{\hat \theta_N} as a dataset, and WLB, which re-weights the data proportional to an exponential(λ=1\lambda = 1) draw. Upper confidence bounds are given for both methods. Theorem 1 shows convergence in probability of a uniform bound on the expected change in loss. Theorem 2 gives the asymptotic limit of the confidence of their upper bound. Theorem 3 gives a bound on the computable bootstrap error against the ideal percentile. The numerical results show the regret of their proposed method and baselines, as well as well as the coverage provided by the bounds.

Critical strengths
  • This paper has an interesting problem formulation, and presents a way to combine bootstrap methods with confidence bounds. The idea is interesting and has a lot of potential.
  • The presentation of the bootstrap methods is very clear and I very easily learned about bootstrap methods that I was not previously familiar with.
  • The paper is well-written and generally has a very clean problem statement.
Critical weaknesses
  • Although I am not deeply familiar with all the methods in upper and lower confidence bounds in bootstrap methods, the literature review in this paper seems very limited. There are 3 landmark papers (pre-1997) cited, and 1 recent paper, regarding APUB. The background section contrasts with distributionally-robust optimization and Bayesian optimization, but it needs to situate itself more firmly into methods that use confidence bounds in optimization, e.g., recently popular conformal prediction methods. Additionally, contextualizing this paper in the previous works on using upper confidence bounds would greatly improve readability.
  • The above point also leads to a stronger concern: it is difficult for me to assess the novelty of the paper as it is not situated within the literature at all. What are the existing works most similar to Theorem 1, and how does Theorem 1 differ? The authors present their APUB method as though there is no significant work on confidence bounds in optimization. The propositions and theorems have little to no discussion of proof sketches as to what methods are used, which makes it hard to assess the theoretical contribution.
  • While I praised the presentation of the bootstrap methods, the upper bounds in Equation (5) have no interpretation, discussion on novelty/related work.
  • The theorems only provide asymptotic limits for the methods presented here. Beyond consistency, is there a theoretical reason for using APUB?
  • The numerical methods were difficult to interpret for me, and it was not obvious that the proposed methods provide benefits over the alternatives.
Questions for authors
  • Can you contextualize the literature within the upper confidence bounds in optimization literature more explicitly? Additionally, for any definitions, can you provide more details on the originality or source (e.g., the bootstrapping methods, the upper bounds used). Can you state known theorems regarding the bootstrap methods as propositions, so the logic of the paper is more clear?
  • How do the theorems compare to the existing literature, and which methods are combined in order to prove the theorems?
  • Can you more explicitly state the goals these methods in practice, and why the figures demonstrate that this method is useful in these toy examples? Additionally, can you clarify and interpret the figures in the numerical results section, as to what the relative performance of the stated methods and the baselines are meant to convey?
2 comments
AuthorsOpenReview ↗

Authors’ Second-Round Response

We thank the reviewer for the constructive comments. We have submitted the revised manuscript, with major changes marked in blue. The changes addressing your concerns are summarized below.

Response to Weakness 1 and Question 1: Literature and attribution — Section 2, p. 2; Section 3.1, p. 3; Appendix F.1, pp. 38–39

We expanded the review to include 14 complementary studies covering empirical confidence penalties, statistically calibrated DRO, bootstrap optimization, parametric confidence regions, conformal optimization, and sequential UCB methods, including direct UCB minimization.

The revision distinguishes their statistical targets and decision settings. The cited conformal methods calibrate predictive outcome regions, whereas APUB addresses estimation uncertainty in expected cost. We acknowledge that confidence-based optimization, parametric uncertainty protection, and bootstrap fitting predate our work.

Section 3.1 attributes PB and WLB to their sources. Section 3.2 attributes upper-percentile aggregation to nonparametric APUB and its positive-part representation to established CVaR theory. Propositions 1–2 state sufficient conditions for the PB/WLB fitted-cost laws used in our framework.

Response to Weakness 2 and Question 2: Contribution and proof methods — Section 4, pp. 4–6; Appendix F.2 and Table 7, p. 39

We added a proof roadmap and a table distinguishing established techniques from what each result verifies here.

Theorem 1 uses upper-average stability and approximate-minimizer arguments, related to existing nonparametric APUB and Bayesian risk analyses. The construction-specific task is establishing uniform fitted-cost contraction after nonlinear parameter refitting. Proposition 1 handles weighted likelihood fitting; Proposition 2 additionally controls PB’s random fitted generating distribution.

The revised Theorem 2 derives the statistical limits needed for selected-decision coverage under explicit regularity conditions. Its proof combines likelihood expansions, moment control, a uniform delta method, and local equicontinuity. Theorem 3 combines DKW concentration, a finite-net union bound, and Lipschitz continuity.

This clarifies the contribution without claiming that bootstrap fitting, upper-tail aggregation, or the general proof techniques are new.

Response to Weakness 3: Equation (5) — Section 3.2, pp. 3–4

The paragraph following Equation (5) explains its purpose and origin. APUB protects against decisions that appear attractive under one fitted model but become costly under estimation perturbations. The upper tail is over fitted expected costs, not individual realized market losses. Zero-level aggregation gives bootstrap averaging; positive levels emphasize unfavorable fits.

The extension concerns the PB/WLB fitted-cost laws and their statistical justification, rather than a new risk functional.

Response to Weakness 4: Beyond consistency — p. 6; Appendix F.3, pp. 39–40; Appendix G, pp. 41–44

We explain three additional properties: preservation of convexity when fitted costs are convex; an estimation-standard-error adjustment in the regular normal limit; and coverage at the ideal data-selected optimizer under the stated assumptions.

Theorem 3 is a finite-ensemble probability bound, conditional on the observed data, rather than an asymptotic result. It controls bootstrap simulation error, not population coverage.

The new Appendix G compares APUB against two matched normal-approximation penalties. Decision quality is similar, while APUB is approximately three times faster under the reported eight-start penalty protocol and provides numerical global optimality-gap certificates. We do not claim universal regret superiority.

Response to Weakness 5 and Question 3: Practical goals and results — Section 5, pp. 8–9; Appendix G.2–G.3, pp. 41–42

We clarified the practical question: how costly is it to act on an unlucky training sample? The text distinguishes mean regret, variability across training samples, and coverage of the reported cost at the recommended decision.

We explain each comparison: SAA versus FMO assesses family information; BAG versus APUB assesses upper-tail adjustment; NP-APUB assesses aggregation without parametric fitting; and the two Bayesian priors assess prior sensitivity.

At N=100 and level 0.5, mixed-normal mean regrets are 7.14 for WLB-APUB and 7.28 for PB-APUB, versus 7.51 for FMO and 8.81 for SAA. The aligned Bayesian prior achieves 6.30, while the misspecified prior gives 15.50. These results support a useful prior-free alternative, not dominance over reliable prior knowledge.

Appendix G additionally reports average regret among the worst 50% and 20% of training replications. We distinguish these tail metrics from central percentile bands and avoid interpreting improved coverage as proof of improved tail regret.

Thank you for the detailed response to my comments. I have raised my score to 3 accordingly.

Reviewer_zgPHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper extends Average Percentile Upper Bound (APUB) optimization to parametric expected-cost models. It generates fitted models using either generative parametric bootstrap (PB) or weighted-likelihood bootstrap (WLB), then minimizes the upper-tail average of their expected-cost functions. The outer functional is CVaR over estimation uncertainty. The paper establishes construction-specific fitted-cost contraction, decision consistency, asymptotic coverage at an ideal data-selected optimizer under local regularity conditions, and a conditional finite-ensemble approximation bound. An adaptive implementation reuses bootstrap fits across stages and risk levels. Experiments use two fixed 20-product, three-component mixture populations, Gaussian and log-normal, with 30 paired training replications and sample sizes 100, 200, and 1000. Parametric APUB generally improves mean regret over SAA and fitted-model optimization, while increasing the coverage of the reported cost as the aggregation level increases.

Critical strengths
  1. The paper clearly distinguishes uncertainty about the parameter from randomness of realized outcomes. Also, incorporating a known distribution family without specifying a parameter prior is a reasonable practical objective. The proposed objective is somewhat natural.

  2. Sound theoretical and empirical results.

Critical weaknesses
  1. Coverage result needs further analysis, for example, sharpness or calibration. The coverage guarantee is somewhat conservative: Theorem 2's resulting level is strictly larger than γ\gamma. Also, direct evaluation gives approximately 57.7%, 78.8%, and 91.9% coverage at γ=.1,.5,.8\gamma=.1,.5,.8, respectively. This implies that the coverage guarantee is over-conservative and probably not sharp. Some discussions or some extensions on how to calibrate the coverage would be useful.

  2. The paper appropriately cites its predecessors. Nevertheless, upper-percentile averaging and bootstrap UCB optimization are already the core of Liu et al., Minimizing Upper Confidence Bounds. The Bayesian-risk formulation in Wu, Zhu, and Zhou already applies CVaR to parameter-dependent expected costs and studies consistency and normal asymptotics; its CVaR limit contains the same normal-tail coefficient. The distinctive contribution here is therefore the PB/WLB fitting construction, its conditions and transfer results, and its practical optimization workflow. That can be useful, but the paper should explain more precisely which technical obstacle is resolved beyond composing established bootstrap fitting and CVaR stability.

Questions for authors
  1. Is γ\gamma intended as a conservative coverage lower bound, a target actual coverage, or simply a tail-aggregation parameter? How should the level-tracking claims be reconciled with Φ(κγ)\Phi(\kappa_\gamma)?

  2. Which additional result or practical advantage is unavailable from existing APUB and posterior-CVaR analyses? Can the gain over a matched normal-approximation uncertainty penalty be demonstrated?

Comments on the paper

10 comments
AuthorsOpenReview ↗

Authors’ Second-Round Response

We thank the reviewer for the constructive follow-up. We have submitted the revised manuscript, with major changes marked in blue. The revision addresses the remaining questions as follows.

1. New experiment addressing Question 2 and your follow-up — Appendix G, pp. 41–44

We completed the promised comparison against both original-fit-centered and bootstrap-mean-centered normal penalties, using identical bootstrap ensembles. The experiment includes 30 paired replications, three sample sizes, three aggregation levels, and both PB and WLB. Both penalty methods use eight-start optimization. Page 41 defines the objectives and protocol; page 42 summarizes the findings; pp. 43–44 provide the complete results.

The findings are:

  • Decision performance: APUB and the bootstrap-mean-centered normal penalty have similar regret and coverage. This experiment does not establish a substantial advantage from capturing asymmetry or upper-tail shape. We state this explicitly on p. 42.
  • Optimization time: Across 540 problems per method, average optimization times are 1.54 seconds for APUB, 4.64 seconds for the original-fit-centered penalty, and 4.61 seconds for the bootstrap-mean-centered penalty. These times exclude shared bootstrap fitting.
  • Solution quality: All 540 APUB solves meet the prescribed numerical global optimality-gap tolerance. The eight-start penalty optimizations do not provide corresponding global optimality certificates.

Thus, the demonstrated advantage is comparable decision quality with approximately threefold lower optimization time under the reported protocol, together with numerical global optimality certificates. APUB preserves convexity when fitted costs are convex, whereas neither normal-penalty formulation generally preserves it. We distinguish this established structural advantage from potential statistical gains that this experiment does not demonstrate.

2. Clearer technical contribution addressing Weakness 2 — pp. 2, 4–6, and 39

We revised the contribution statement and theoretical discussion to identify precisely what is new. We do not claim novelty for upper-percentile averaging or the normal-tail coefficient.

The construction-specific contribution is establishing uniform fitted-cost contraction after nonlinear parameter refitting: Proposition 1 treats WLB, while Proposition 2 additionally handles PB’s random fitted generating distribution. Theorems 1–2 connect these properties to decision consistency and coverage at the ideal data-selected optimizer. The revised Theorem 2 states regularity conditions from which the required statistical limits are derived.

Page 39 explicitly compares these results with nonparametric APUB and Bayesian risk optimization and summarizes the proof ingredients. The resulting framework exploits known distribution-family information without requiring a parameter prior; existing empirical-cost or posterior-law analyses do not directly establish the properties of these PB/WLB fitted-cost laws.

3. Calibration clarification addressing Weakness 1 and Question 1 — p. 6 and Appendix B.6, p. 24

Although you indicated that these concerns were resolved, we have incorporated the clarification into the manuscript. The revision distinguishes the aggregation level from actual limiting coverage and gives the calibration relationship, its uniqueness argument, and numerical examples. It also states that this calibration is asymptotic rather than a finite-sample guarantee.

These revisions clarify both the scope of the theoretical contribution and the practical advantage supported by the new evidence.

AuthorsOpenReview ↗

General Response and Revision Summary

We thank all three reviewers for their careful evaluation and constructive suggestions. We have submitted a revised manuscript, with major changes marked in blue. The revision strengthens the theory, clarifies the contribution, and adds the requested computational comparison.

1. Stronger theoretical justification

Responding to L79u (Weakness 1, Question 1, and follow-up), 21q9 (Weakness 2 and Question 2), and zgPH (Weakness 2).

The revised Theorem 2 derives the sampling limit, local equicontinuity, and conditional bootstrap approximation from explicit regularity and fitting conditions. The proof establishes these limits and transfers them to coverage at the ideal data-selected optimizer.

We give explicit formulas for Fisher information, the estimator’s asymptotic covariance, and fitted-cost variance. We distinguish statistical regularity from numerical fitting accuracy: exact global likelihood maximizers satisfy the fitting-gap conditions automatically, whereas local numerical convergence does not certify them. Appendix B.5.4 explains applicability to the mixture models and the remaining requirements.

2. Clearer contribution and broader literature review

Responding to 21q9 (Weaknesses 1–3 and Questions 1–2), zgPH (Weakness 2 and Question 2), and L79u (the novelty concern in Weakness 1).

The expanded review discusses 14 complementary studies on confidence-based optimization, bootstrap methods, parametric confidence regions, conformal optimization, and sequential UCB methods, including direct UCB minimization.

We attribute bootstrap fitting, upper-percentile aggregation, and the positive-part representation to existing work. Our contribution is the prior-free parametric APUB framework, its construction-specific PB/WLB fitted-cost analysis, and the transfer to decision consistency, selected-decision coverage, and finite-ensemble accuracy. Table 7 on p. 39 summarizes the proof ingredients.

3. Coverage interpretation and calibration

Responding to zgPH (Weakness 1 and Question 1), L79u (Question 4), and 21q9 (Weaknesses 3–4).

The revision distinguishes the aggregation level from actual coverage. Theorem 2 gives an exact regular asymptotic coverage limit, exceeding the aggregation level because APUB averages upper percentiles. Appendix B.6 derives calibration and provides numerical examples.

Calibration is asymptotic, not a finite-sample guarantee. Improved coverage does not itself establish lower decision regret.

4. New matched normal-penalty experiment

Responding to zgPH (Question 2 and follow-up), L79u (Question 3), and 21q9 (Weaknesses 4–5 and Question 3).

We compare APUB with original-fit-centered and bootstrap-mean-centered normal penalties using identical bootstrap ensembles. The experiment includes 30 paired replications, three sample sizes, three aggregation levels, and both PB and WLB. Both penalty methods use eight-start optimization.

APUB and the bootstrap-mean-centered penalty have similar regret and coverage. Across 540 problems per method, average optimization times are 1.54 seconds for APUB, 4.64 seconds for the original-fit-centered penalty, and 4.61 seconds for the bootstrap-mean-centered penalty, excluding shared fitting costs. All APUB solves meet the numerical global optimality-gap tolerance; the penalty optimizations provide no corresponding global certificate.

These results support comparable decision quality with lower optimization time under the reported protocol. The experiment does not demonstrate a substantial statistical advantage from retaining nonnormal bootstrap tails.

Appendix G.4 separately reports an NP/PB/WLB timing pilot. Nonlinear refitting can make parametric approaches substantially more expensive; we do not claim universal computational superiority.

5. Clearer experimental goals and limitations

Responding to 21q9 (Weakness 5 and Question 3), L79u (Weakness 2 and Question 4), and zgPH (Question 2 and follow-up).

The practical question is how costly it is to act on an unlucky training sample. We distinguish mean regret, variability across training datasets, upper-tail regret, and coverage at the recommended decision. The new comparison reports average regret among the worst 50% and 20% of training replications.

We explain which comparisons assess family information, upper-tail adjustment, and prior sensitivity. Prior independence does not imply robustness to an incorrect family or contaminated observations. APUB does not universally dominate reliable prior information or improve every regret summary.

We appreciate the reviewers’ suggestions. Reviewer-specific responses provide further details and page references.

AuthorsOpenReview ↗

1. Stronger theoretical justification — Section 4, pp. 4–6; Appendix B.5, pp. 20–24

The principal theoretical revision addresses the concern that the previous coverage theorem assumed the statistical convergence properties it needed. The revised Theorem 2 states explicit regularity and fitting conditions and derives those convergence properties.

The assumptions concern identifiability, likelihood smoothness, integrable derivative envelopes, nonsingular Fisher information, numerical fitting accuracy, and regularity of the expected-cost function. The proof proceeds in two steps.

Step 1: Derive estimator and bootstrap convergence.

The proof establishes parameter consistency and the sampling limit

N(θ^N−θ0)⇒N(0,V)\sqrt{N}(\widehat\theta_N-\theta_0)\Rightarrow\mathcal{N}(0,V).

It also establishes conditional first-moment Wasserstein convergence of the bootstrap estimation error to the same normal limit. Applying these results to the fitted expected-cost function yields the sampling approximation, local equicontinuity, and bootstrap approximation previously assumed in the coverage theorem.

First-moment control is necessary because weak convergence alone does not control an upper-tail average: rare large values can affect that average without changing the weak limit.

The theorem explicitly defines the score and Fisher information:

uθ(ξ)=∇θlog⁡pθ(ξ)u_\theta(\xi)=\nabla_\theta\log p_\theta(\xi),

I(θ0)=Eθ0[uθ0(ξ)uθ0(ξ)⊤]I(\theta_0)=\mathbb{E}_{\theta_0}[u_{\theta_0}(\xi)u_{\theta_0}(\xi)^\top].

For the correctly specified regular likelihood model,

V=I(θ0)−1V=I(\theta_0)^{-1}.

Writing g(x)=∇θH(x,θ0)g(x)=\nabla_\theta H(x,\theta_0), the fitted-cost asymptotic variance is

s(x)2=g(x)⊤Vg(x)s(x)^2=g(x)^\top Vg(x).

Step 2: Establish coverage at the selected decision.

Consistency and local equicontinuity transfer the fitted-cost approximation to the ideal APUB optimizer selected using the same data. Thus optimizer consistency and the coverage limit are conclusions of the theorem, rather than separately assumed properties.

Numerical fitting accuracy is distinguished from statistical regularity.

Conditions (A4) and (P3) concern global likelihood-maximization gaps. These gaps are zero for exact global maximizers. Approximate numerical fits require explicit accuracy: vanishing normalized gaps suffice for the consistency argument, while the coverage theorem imposes the stronger accuracy needed for first-order asymptotics.

Smoothness and identifiability cannot guarantee that an arbitrary numerical routine finds a global maximum. The revision therefore does not claim that local convergence diagnostics certify the theoretical fitting requirements.

Appendix B.5.4 explains how envelope and continuity conditions apply to Gaussian and lognormal mixtures. It also identifies additional requirements, including distinct components, an identifiable parameterization, interior truth, nonsingular information, and fitting accuracy.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.