acceptodds
Under review as a conference paper at ICLR 2027

Optimal Bayesian Optimization with Stereographic Functions

Abstract

In Bayesian optimization, the reward-maximizing policy solves a Bellman equation whose state is an entire posterior distribution, and that recursion is intractable; practice substitutes acquisition functions that look a single query ahead. We give a setting in which the optimal policy can instead be written down. The unknown function is linear in the stereographic projection of its input into , which makes it a smooth function of the input coordinates with one peak and one valley; evaluations are priced, and probing one known direction costs more than probing any other; and the learner is scored on a single final query, against a point drawn at random from those it could afford. The terminal reward turns out to be a quadratic form in the posterior mean whose weight matrix annihilates the expensive direction, so information about that direction has no value. The Bellman equation then admits a closed-form solution, and the optimal query schedule is non-adaptive: it can be computed in full before any data is seen. Choosing that schedule reduces to water-filling over the eigenvalues of the prior, which we solve in time, and the resulting design puts no weight on the expensive direction, so every query it issues costs only the base price. Experiments check the closed-form value function against Monte Carlo estimates and measure how much exact optimality is worth relative to structured baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian optimization, experimental design, optimality, stereographic, dynamic programming

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Q1ZVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper seeks a Bayesian optimization setting in which the reward-maximizing policy can be computed exactly rather than approximated by myopic acquisition functions. The unknown function is linear in the stereographic projection of the input, f(θ)=ϕ(θ)⊤zf(\theta) = \phi(\theta)^\top z, so each function has one peak and one valley and inference is Bayesian linear regression; queries are priced as κ+(m⊤c)2\kappa + (m^\top c)^2 with a known expensive direction c; and the learner is scored on one final query by the squared posterior-mean gap between its query and a random point drawn on the boundary of the affordable region. Under an assumption on the posterior mean (A3), the terminal value is μ⊤Pμ\mu^\top P\mu with P=p(I−c^c^⊤)P = p(I - \hat c\hat c^\top), which annihilates c. The authors then show that the Bellman equation has the closed-form solution V=μ⊤Pμ+tr⁡(PΛ−1)−J(Λ,B)V = \mu^\top P\mu + \operatorname{tr}(P\Lambda^{-1}) - J(\Lambda, B), so the optimal schedule is non-adaptive; that it reduces to an exact design problem over information matrices characterized via Schur–Horn; and, when c is an eigenvector of the prior precision, that the optimal design is water-filling on the prior spectrum with no weight on c, computable in O(n3)O(n^3). Synthetic experiments verify the value function against Monte Carlo and compare with cost-blind A-optimal design, uncertainty sampling and random queries for n = 3 to 128.

Critical strengths

(1) The goal of identifying a nontrivial sequential design problem where the Bayes-optimal policy can be written down, and using it to explain when adaptivity has value, is a worthwhile and clearly argued research direction. (2) The derivations are clean and, under the stated assumptions, correct. (3) The paper is reasonably well written, with good intuition for each step.

Critical weaknesses

My main concern is the practicality of the problem. It seems the problem setting is engineered so that the main conclusions hold, and the link to Bayesian optimization is weak. The random baseline on the boundary slice is chosen so that its mean exactly cancels the expensive-direction component of the final query, which is what makes information along c worthless. Also, the squared reward with adversarial sign rewards peaks and valleys alike and even rewards a query orthogonal to μ\mu on the degenerate branch. None of this corresponds to a standard BO objective such as the expected value of f at the recommended point or simple regret. Once the stereographic projection drops out (Section 6), it seems that the problem is Bayesian L-optimal linear design with linear costs, for which non-adaptivity under a quadratic terminal reward is well known (posterior covariance independent of responses, law of total variance), so the conceptual novelty is limited.

Questions for authors

So my questions are: Can the authors motivate the scoring rule (squared posterior-mean gap against a random point on the affordable boundary, with adversarial sign) from a BO perspective, and say which conclusions survive under a standard objective such as the expected value of f at the final query or simple regret? How do your results differ from classical Bayesian L-optimal design for linear-Gaussian models, where open-loop optimality under quadratic terminal reward is known?

1 comment

Thanks for the detailed response, and I do not have further questions. Part of my concerns are solved after reading the "What is classical and what is new" section, but I still feel that the square score is considered only because it makes the Bellman equation solvable. So the practicality of the work is a little weak, and I will keep my score unchanged.

Reviewer_UAi9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies a structured Bayesian optimization problem with a function linear in stereographic features, a known direction-dependent query cost, and a terminal max-min criterion based on the squared expected gap to a random budget-boundary baseline. The authors derive a reduction to static experimental design for a quadratic terminal reward and use Assumption A3 to connect that reward to the original score. When the prior precision commutes with the reward-weight matrix P, the design is obtained by water-filling over the prior spectrum and reconstructed as an integer number of unit-norm queries. Synthetic experiments compare the design with three baselines and separately check the quadratic value-function identity by Monte Carlo simulation. The stated running time is cubic in the feature dimension plus a query-reconstruction term.

Critical strengths

1.The experimental-design derivation for the quadratic objective is well organized. The connection between deterministic precision updates, information matrices, water-filling, and constructive recovery of unit-norm queries makes the solution interpretable. Treating an integer number of queries rather than only a fractional design is a useful part of this construction. 2.The paper is explicitly positioned as an analytically tractable diagnostic example, not a competitive general-purpose Bayesian optimization method. The 28,000 paired synthetic instances and the Monte Carlo check provide useful numerical evidence within the tested setting. Section 6 also explicitly discusses commutation, known costs, and the single-terminal-query restriction.

Critical weaknesses

1.The connection between the original terminal score and the Bellman problem needs a substantive correction. Equation (10) and Appendix D.2 use a quadratic stopping reward, whereas Equation (28), derived under the first inequality of A3, also includes a degenerate branch. Appendix E reports the original max-min score. Under the experimental initialization, where the prior mean is the unit vector in the expensive direction and p = 0.91, Equation (28) gives an initial score of 0.09 rather than the zero stated in Section 5. The second condition of A3 therefore already fails at the initial state. Either condition of A3 can fail, so a complete treatment of Equation (4) must also cover a slack final-query budget constraint. The quadratic Monte Carlo check does not establish optimality for the full objective. This concern does not by itself invalidate Proposition 2 for a globally quadratic reward, but it affects its application to Equation (4) and is the main reason for my recommendation. 2.The numerical study is confined to priors for which the expensive direction is an eigenvector, matching the commuting setting of the theorem. Section 6 already explains that prior correlations can make measurements along the expensive direction informative about rewarded directions. A small-dimensional sensitivity analysis could illustrate this boundary quantitatively; it would be an optional extension, not a requirement for the stated commuting-case result.

Questions for authors

1.Please reconcile the nonzero initial score implied by Equation (28) with Section 5, and specify which objective Proposition 2 and Algorithm 1 optimize. Can the stopping value and continuation expectation be written for all relevant posterior states, including those where the final-query budget constraint is slack, or can conditions be given that preserve A3 under the Gaussian updates? Please separate verification of the quadratic identity from evidence about the full terminal objective, and clarify whether the reported A3 frequencies concern final states or entire trajectories. 2.As an optional boundary analysis, could a low-dimensional study vary the angle between the expensive direction and a chosen prior eigenvector while fixing the spectrum, budget, and noise? If such a study is included, please compare budget-feasible designs under the same objective and state the numerical solver's search scope and optimality limitations. This extension is not necessary to validate the explicitly commuting design theorem.

Reviewer_WE2ROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses a Bayesian optimization problem in which certain restrictions are applied: 1) the unknown function is linear in the stereographic projection, 2) the evaluation cost depends on the query direction, and 3) the reward is scored on a final query against a random baseline. Through a quadratic terminal reward and the properties of Gaussian linear models, the authors provide a theoretical analysis showing that the exploration schedule can be determined in advance, without adapting to intermediate observations. The authors assume a commuting assumption and have the problem reduced to an experimental design problem with a water-filling solution. The experiments compare the proposed design against several baselines and use Monte Carlo estimates to validate the analytical value function.

Critical strengths
  1. The theoretical contribution is elegant and thoroughly conducted. The authors give a good motivation for the reduction from sequential-decision making to optimal experimental design, and the mathematics is appealing. The insight that the contribution of the posterior mean can be separated from the uncertainty, which can be resolved by exploration, is particularly interesting.

  2. The benchmark is useful for the questions arised. A computable optimal policy is available for the problem used in the benchmark, and this is a useful testbed for studying non-myopic Bayesian optimization methods, as well as for understanding when adaptivity is unnecessary.

  3. Overall, the paper has a clear interpretation and implementation. The water-filling solution gives an intuitive explanation of how the query budget should be allocated. In addition, the paper also addresses how to reconstruct an actual query sequence from the optimal information matrix.

Critical weaknesses
  1. The assumed function class and other structural conditions limite the applicability of the results. Although the authors explain the geometry of stereographic functions, no concrete examples of real-world problems have been provided to demonstrate these properties. It would be extremely helpful to dicuss potential applications of the results.

  2. The empirical validation seems limited in its scope. All experiments use synthetic instances constructed to satisfy the assumptions. Obviously, this is reasonable for the purpose of validating the theory, but there are no real-world examples demonstrating the relevance of the proposed formulation. Additional experiments in this regard would help establish the significance of this work.

  3. The practical relevance of the commuting assumption is unclear. Although its mathematical role is explained and the authors acknowledge its limitations, they do not discuss concrete settings in which the expensive direction would naturally align with a principal direction of the prior precision. The epxeriments then enforece this alignment by construction. Providing motivating examples or discussing how often this assumption is reasonable in practice would help clarify the applicability of the proposed framework.

Questions for authors
  1. Do the theoretical results cover the full stochastic terminal objective, including trajectories where A3 fails, or only the regime in which the quadratic reward applies?

  2. Can the authors provide examples of real-world optimization problems that might be reasonably approximated by the proposed function class, or explain how they envision this theoretical result being used in practice?

Given sufficient real-world experiments and meaningful analysis provided to demonstrate the applicability of the results, I am willing to raise my rating appropriately.

1 comment

Thank you for the thorough clarifications. After examining the responses as well as the modifications in the manuscript, I will keep my rating as is. While the authors have explained the assumed function class and the commuting assumption, it is hard to determine the contribution scope of the manuscript, as a testbed or a structural statement. For example, if it is a testbed for non-myopic acquisition rules for cases when the Bayes-optimal policy is known, a straightforward step, in addition to synthetic benchmarks, is to perform the tests on at least 2-3 real-world applications. The applications should be thoroughly defined and explained how they align with the function class and assumptions, not just through a few brief sentences. As such, I find the manuscript not ready for publication in its current form.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.