acceptodds
Under review as a conference paper at ICLR 2027

Bayesian Optimization with Structured Measurements: A Vector-Valued RKHS Framework

Abstract

Bayesian optimization (BO) is an efficient framework for optimizing expensive black-box functions. However, it is typically formulated as learning an end-to-end mapping from inputs to scalar objectives, thereby discarding the potentially rich information whenever a structured system output is available. In this work, we study Bayesian optimization over a vector-valued operator with structured measurements, where each measurement observes multidimensional or functional outputs, e.g., trajectories or spatial fields, rather than a single scalar value. The objective is then defined as a linear functional of these measurements. This allows each observation to reveal substantially richer information about the underlying system compared to scalar observations. Assuming the unknown operator lies in a vector-valued reproducing kernel Hilbert space (RKHS), we derive high-probability concentration bounds for the kernel ridge regression (KRR) estimator directly in the measurement space, characterizing uncertainty in a general Hilbert space. Building on these results, we propose an algorithm based on the upper confidence bound (UCB) acquisition function with regret guarantees under mild assumptions, recovering sublinear rates for common kernels. Empirically, we demonstrate that leveraging structured measurements leads to improved sample efficiency by enabling efficient transfer of information across objectives and adaptation to time-varying settings.

Primary area:
optimization
Keywords:
Bayesian optimization, optimization, vector-valued RKHS

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qSuFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a framework for Bayesian optimization with vector-valued, and potentially infinite-dimensional, measurements, allowing one to exploit structure in complex measurements and efficiently adapt to changing objectives. Frequentist confidence bounds are derived under structural assumptions on the kernel and noise, yielding sublinear regret rates resembling those of scalar Bayesian optimization for popular kernels. The benefits of vector-valued measurements are illustrated in experiments on synthetic and simulation data.

Critical strengths
  • Relevant problem and framework. Exploiting structure in complex, possibly infinite-dimensional, observations is a practically relevant problem that the paper motivates well. The proposed framework is sound and models the problem well.
  • Sound contribution. The joint contribution of Theorems 1-3 showing that standard Bayesian optimization methods extend to infinite-dimensional measurements in a natural way is interesting and appears novel. I consider this conceptual contribution the strongest argument for acceptance. The derivations appear sound.
Critical weaknesses
  • Technical novelty and signposting. The proofs of Theorems 1-3 rely mostly on standard techniques and prior work, and the signposting and framing of the contribution overstate the technical contribution. In particular, Theorem 1 appears to be a direct specialization of Theorem E.23 in Massiani et al. (2025). I therefore find that the framing of the KRR uncertainty bound as a central contribution oversells the actual contribution. The prose in Appendix D stating that the proof merely “follows the same structure as Chowdhury & Gopalan (2021); Massiani et al. (2025), adapted to a general vector-valued setting” is misleading because Massiani et al. (2025) already operate in a general vector-valued setting and provide the main nontrivial step for the proof essentially verbatim. Once the uncertainty bound of Theorem 1 is established, the fact that the measurement space is infinite-dimensional is no substantial obstruction over the finite-dimensional case anymore, so concluding Theorem 2 appears straightforward with the established technique of Chowdhury & Gopalan (2021). Similarly, the UBD assumption makes lifting Lemma 1 of Chowdhury & Gopalan (2021) to the setting of Theorem 3 straightforward. Overall, I think the paper would benefit from explaining more clearly its technical contribution, or framing the contribution more explicitly around combining existing techniques resulting in a new framework. Combining known results/techniques in a new way can still be a substantial contribution but the signposting should reflect this accurately.
  • Comparison with prior work. While the significance of high-dimensional measurements is well-motivated, the discussion of related work is currently insufficient to convincingly support the novelty claims for the proposed framework. Most notably, the frameworks of [1] and [2] already target Bayesian optimization with high-dimensional measurements but are missing in the related work. An explicit discussion of these methods is warranted to judge whether the novelty is purely in the extension to the infinite-dimensional case, or for exploiting structure across high-dimensional measurements more broadly. Since the experiments appear to use up to 50-dimensional samplings of infinite-dimensional measurements, it seems that the finite-dimensional methods of [1] and [2] are directly applicable here. Seeing how these methods compare to the proposed algorithm would help to better illustrate the significance of the contribution.

[1] Maddox et al. “Bayesian optimization with high-dimensional outputs” (2021)

[2] Huang et al. “Combinatorial Bandit Bayesian Optimization for Tensor Outputs” (2026)

Questions for authors
  • Can you clarify whether your Theorem 1 is a direct consequence of Theorem E.23 in Massiani et al. (2025) once you identify H=HKM\mathcal{H}=\mathcal{H}_{K_{M}}, G=M\mathcal{G}=\mathcal{M}, h⋆=Mfh^{\star}=Mf, and Ln=KM(⋅,xn)∗L_{n}=K_{M}(\cdot,x_{n})^*, or whether your result requires a genuinely new argument to accommodate the measurement operator under your assumptions?
  • Can you explain how your results compare to those of [1,2] and what applications your framework supports that theirs does not?
  • You repeatedly call the assumptions “mild”, but argue mildness only for each assumption individually. Can you explain why the combination of assumptions, particularly UBD and the specific separability assumption around Theorem 3 together with realizability, is still mild?
Reviewer_DxH7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies BO when the objective is a linear functional of a structured measurement of an unknown vector-valued operator f (e.g., f(x) is a whole trajectory, and the cost is an integral over it). Instead of learning a direct map from inputs to a scalar, the paper learns f in a vector-valued RKHS and derives confidence bounds for the measured output Mf(x) directly in measurement space (Thm. 1). A UCB algorithm (vvBO) built on these bounds achieves sublinear regret for common kernels (Thms. 2–3), with extensions to time-varying objectives (Alg. 2) and nonlinear Lipschitz objectives (App. H.2). Experiments on synthetic benchmarks and a building-control case show gains from reusing structured observations across objectives.

Critical strengths
  • Useful unification. The measurement operator M covers all observation regimes in one framework: full observation (M=I), partial observations (projections/integrals), and classical scalar BO (M a functional). This cleanly separates what is measured from what is optimized — one learned model serves many objectives, which is the key mechanism behind the cross-objective transfer results.
  • Coherent theory. Confidence bounds live in measurement space (not in the full output space), avoiding conservative error propagation. Regret rates match classical scalar BO, showing vector-valued outputs do not degrade the rates.
  • Compelling experiments. Multi-phase benchmarks and the BOPTEST building study convincingly show reuse across objectives and tracking of time-varying prices; validation on an unseen objective (peak power) is a strong result.
  • Good reproducibility. Full proofs, implementation details, hyperparameters, and an AI-use statement are provided.
Critical weaknesses
  • Degree of theoretical novelty. Much of the technical machinery builds on established results: Lemma 1 corresponds to Proposition 6 of Carmeli et al. (2010), the proof of Theorem 1 draws on Corollary E.7 of Massiani et al. (2025) (where UBD kernels were introduced), and the regret argument follows the standard optimism decomposition. The distinct step of composing K with M and working in the induced RKHS is elegant; it would be helpful to see a clearer delineation of what is new in this work versus what is adapted from prior results.
  • Theory–practice gap on β. Regret guarantees require the theoretical β_t, but all experiments use a hand-tuned constant β (Table 3: values from 5 to 400 across operators/phases); λ and length scales are also manually tuned. Clarifying how the theoretical parameter relates to the one used in practice — or explicitly discussing this gap — would strengthen the connection between the guarantees and the algorithm that is actually evaluated.
  • Scope of the assumptions. Some structure is technically necessary (without the UBD kernel class, information gain can be infinite), but two conditions are strong: (i) noise must be sub-Gaussian in an infinite-dimensional space with covariance aligned to the kernel's isometry — practically unverifiable; (ii) explicit rates need separable kernels with finite-rank blocks and, for Matérn, ν > d(d+1)/2, excluding common smoothness choices in d ≥ 2. These limitations should be discussed in the main text.
  • Scalability. All experiments use at most 3-dimensional inputs and at most ~300 iterations; Table 2 reports ~10 ms/iteration on tiny problems. No evidence is given for higher-dimensional inputs, larger budgets, or many-basis (large n) measurement representations. The truncation scheme in Appendix I is heuristic, and there is no analysis of how truncation error propagates into the confidence bounds or regret.
  • Evaluation details. (i) no sensitivity analysis of λ or of the truncation level; (ii) Table 3's β column does not state which methods the values apply to, making the tuning protocol ambiguous; (iii) the hidden dependence of β_T on n = rank(B_M^0) (via ∑_i γ_T) is not discussed.
Questions for authors
  1. Do the advantages of vvBO survive a measurement-cost model, e.g., comparing methods under a matched budget of scalar observations?
  2. How can a practitioner verify the UBD kernel and the aligned sub-Gaussian noise assumptions? Are non-separable kernels (input-dependent output correlations) possible?
  3. Were all baselines given equal tuning effort, and are the β values in Table 3 method-specific or shared? Is there a sensitivity analysis for β and λ?
  4. How does the number of linear measurements affect the theory, and can its scalability be verified in practice?
Reviewer_RFs9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies Bayesian optimization when the black-box system produces structured outputs, such as trajectories or spatial fields, rather than a single scalar. The authors model the system as an unknown operator f:X→Yf:\mathcal{X}\to\mathcal{Y} in a vector-valued RKHS, observed through a known bounded linear measurement operator M:Y→MM:\mathcal{Y}\to\mathcal{M}, with the objective F(x)=⟨m,Mf(x)⟩MF(x)=\langle m, Mf(x)\rangle_{\mathcal{M}}. Working with the induced kernel KM=MKM∗K^M=MKM^*, they perform kernel ridge regression directly in the measurement space and derive a high-probability confidence bound for ∥Mf(x)−Mμt(x)∥M\|Mf(x)-M\mu_t(x)\|_{\mathcal{M}} under a uniform-block-diagonal (UBD) kernel assumption and trace-class sub-Gaussian noise. Based on this bound, they propose a UCB algorithm (vvBO) and prove a general regret bound. For separable kernels K=G⊗BK=G\otimes B, where BMB^M has a finite-rank block, they obtain explicit sublinear rates for linear, Gaussian, and Mat'ern kernels.

Critical strengths

I think the main strength of the paper lies in its formulation. Introducing an explicit measurement operator MM provides a framework that covers both scalar BO (when MM is a linear functional) and full functional observation (when M=IM=I). Moving to the measurement space through the induced kernel is a sensible design choice, because it avoids paying for uncertainty in directions of Y\mathcal{Y} that the learner can never observe. Separating the model of ff from the objective functional mm is what allows the method to reuse data when the objective changes, and this is well motivated by the building-control application.

Critical weaknesses

My main concern is that the theoretical contribution is limited compared with existing work and does not address the paper's central motivation. Theorem 1 is essentially the confidence bound of Chowdhury and Gopalan (2021) applied to the induced kernel given by Proposition 6 of Carmeli et al. (2010). The regret analysis in Theorems 2 and 3 also follows the standard argument of Chowdhury & Gopalan (2021). Under the finite-rank assumption on B0MB_0^M, the information gain reduces to a sum of nn scalar information gains, so the infinite-dimensional part of the problem enters only through Tr(RV)\mathrm{Tr}(R_V). The constants nn and Tr(RV)\mathrm{Tr}(R_V) are hidden in the O(⋅)O(\cdot) notation of Theorem 3, although they are exactly what should change with the choice of MM. More importantly, the paper repeatedly claims that richer measurements reveal more information, but no result compares the bounds for different measurement operators, so the theory does not show when or by how much a richer MM helps. A richer MM could even increase Tr(RV)\mathrm{Tr}(R_V). Finally, the Mat'ern rate uses the old information-gain bound that requires ν>d(d+1)/2\nu>d(d+1)/2, whereas Vakili et al. (2021) give γT=O~(Td/(2ν+d))\gamma_T=\tilde{O}(T^{d/(2\nu+d)}) for all ν>1/2\nu>1/2. Similarly, recent improvements in the information gain reported by Iwazaki (2025, 2026) have not been accounted for.

Questions for authors

My main question is whether the authors can formally characterize the benefit of richer measurements. For example, for two measurement operators M1M_1 and M2M_2 where M2M_2 refines M1M_1, can they compare the posterior variance of F(x)F(x) or the resulting regret bounds? How does Tr(RV)\mathrm{Tr}(R_V) behave as the measurement becomes richer, for instance, under a finer discretization of the trajectory?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.