Pith. sign in

REVIEW 2 major objections 5 minor 12 references

By mapping the probability simplex to the positive orthant of a sphere, α-GaBO builds geometry-aware Gaussian-process kernels and information-geometric optimizers for Bayesian optimization, matching or beating constrained Euclidean BO on mi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

α-GaBO performs Bayesian optimization on the probability simplex by pulling spherical Matérn kernels through the Fisher-Rao sphere isometry and using α-connection optimizers, beating Euclidean baselines in several benchmarks and applications.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A useful framework for BO on the simplex, but the geometry write-up has some real errors and the empirical 'increased performance' claim is stronger than the results support. the 2 major comments →

arxiv 2603.09793 v2 pith:PDRYEL2L submitted 2026-03-10 cs.LG

Information Theoretic Bayesian Optimization over the Probability Simplex

Federico Pavesi , Antonio Candelieri , No\'emie Jaquier This is my paper
classification cs.LG
keywords Bayesian optimizationprobability simplexinformation geometryGaussian processesRiemannian kernelsacquisition function optimizationmixture optimizationsphere map
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Many expensive black-box objectives—mixture designs, classifier ensembles, task priorities in robot control—live on the probability simplex, the set of non-negative vectors summing to one. The paper proposes α-GaBO, a family of Bayesian optimization algorithms that treats this simplex as a curved information-geometric space instead of a flat constrained region. The construction has two parts: pull back valid spherical kernels through the sphere map x↦2√x (up to rescaling) onto the simplex, and maximize the acquisition function using exponential maps derived from a one-parameter family of α-connections. Empirically, the two instantiations (α=-1, exponential connection; α=0, Levi-Civita connection) match or outperform constrained Euclidean BO on benchmarks, mixture optimization, classifier mixtures, and a simulated robotic control task.

Core claim

The central claim is that the probability simplex can be treated as a Riemannian manifold with boundary by exploiting its isometry to the positive orthant of the sphere. Pulling spherical Matérn kernels back through the sphere map yields valid Gaussian-process covariance functions on the simplex, and optimizing the acquisition function via α-connection exponential maps respects the simplex geometry. The α parameter interpolates between the exponential connection (α=-1), whose exponential map is defined on the whole tangent space but cannot reach the simplex boundary, and the Levi-Civita connection (α=0), which can map iterates onto vertices and faces where mixture optima often lie. The paper

What carries the argument

The carrying mechanism is the sphere map φ(x)=2√x (element-wise), an isometry between the simplex equipped with the information-geometric metric and the positive orthant of the unit sphere. Because the simplex has a boundary, standard constructions of Riemannian kernels do not apply directly; the map sidesteps this by pulling back well-understood spherical Matérn kernels. The second ingredient is the α-connection family of exponential maps: α=-1 recovers the exponential connection with exponential map Exp_x(η)=exp(η)/(xᵀexp(η))⊙x, defined on all of the tangent space but unable to reach the boundary, while α=0 recovers the Levi-Civita connection with a sphere-like exponential map that can rea

Load-bearing premise

The construction assumes that the sphere-map isometry extends smoothly to the closed simplex, making the pulled-back kernels valid on boundary points, even though the square-root map is not differentiable at the boundary; if this extension fails, the algorithm's ability to find optima on vertices and faces is unsupported.

What would settle it

Evaluate the pulled-back kernel on a set of simplex points that includes vertices and face points, form the Gram matrix, and check its smallest eigenvalue; if it is negative or if Gaussian-process posterior predictions at boundary points contradict the corresponding sphere-kernel predictions under the isometry, the boundary claim collapses. A more targeted check is to test whether the covariance between a vertex and a nearby interior point matches the spherical covariance through the map.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A geometry-aware BO over the simplex can be assembled from off-the-shelf spherical kernels through the sphere map, without inventing a new theory of kernels on manifolds with boundary.
  • The α=0 variant can recommend vertices and faces of the simplex, which is essential when the optimal mixture concentrates on a subset of components; α=-1 is suited to interior optima.
  • The method extends naturally to product spaces, such as a simplex of mixture fractions combined with a Euclidean time parameter, as used in the concrete-strength experiment.
  • On the tested mixture-design and robot-control tasks, α-GaBO reaches lower function values with lower variance than constrained Euclidean BO.
  • The paper's reparametrization of soft task priorities as elements of the simplex guarantees that at least one robot task is active at every time, embedding that constraint directly into the search space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the boundary extension of the pulled-back kernel is well behaved, the same isometry-plus-pullback recipe could supply kernels for other information manifolds with boundary, and for categorical or discrete settings via simplex relaxations; the paper only sketches these directions.
  • The α parameter could serve as a tunable prior over where the optimum is expected: α=-1 encodes a search interior to the simplex, α=0 allows boundary optima; no experiment in the paper varies α as a learned or adaptive hyperparameter.
  • A direct numerical check of positive definiteness on boundary points (vertices and faces) would isolate whether the apparent gains on boundary optima come from the geometry or from an implicit regularization in the kernel.
  • For any new application, the choice between α=-1 and α=0 could be guided by a cheap diagnostic: if the known optimum of a proxy task lies on a vertex, α=0 should dominate; if interior, α=-1 may converge more smoothly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. This paper proposes α-GaBO, a family of Bayesian optimization algorithms on the probability simplex Δ^d. The authors use the Fisher–Rao metric and the sphere map φ(x)=2√x to pull back Matérn kernels defined on the hypersphere (Borovitskiy et al.) to the simplex. For acquisition-function maximization, they use exponential maps associated with the α-connection from information geometry, instantiating two members: α=-1 (exponential connection) and α=0 (Levi-Civita connection). Experiments on three synthetic benchmarks, three optimal-mixture tasks, a mixture-of-classifiers task, and a simulated robotic multi-task control problem compare α-GaBO with constrained Euclidean BO, including BORIS. The paper reports that α-GaBO generally matches or outperforms baselines, while acknowledging some cases where Euclidean BO is competitive or better.

Significance. If the construction is made rigorous at the boundary, the paper provides a useful and well-motivated approach to geometry-aware BO on the simplex, a domain that arises in numerous applications. The use of information-geometric α-connections to parametrize the acquisition optimizer is a novel contribution, and the closed-form exponential maps for α=-1,0 are practically valuable. The paper builds on solid prior work (spherical Matérn kernels, GaBO) and includes a diverse evaluation suite. However, the empirical claims are weakened by post-hoc model selection, lack of statistical testing, and absence of released code; the theoretical treatment of the simplex boundary is also incomplete.

major comments (2)
  1. [Sec. 4.1, Eq. (11)] The pullback kernel is defined via the sphere map φ(x)=2√x. The text states that 'the smoothness of the isometry guarantees the differentiability of k', but φ is not differentiable at the boundary of the simplex. Since the paper explicitly motivates optima lying on vertices/faces, the boundary behavior of k is load-bearing. Injectivity of φ ensures the pullback is positive definite and continuous on the closed simplex, but the claimed differentiability and the validity of the spectral expansion at boundary points are not established. Please replace the assertion with a precise statement, e.g., prove continuity/positive-definiteness on the closure via the sphere representation and state the actual inherited regularity, or restrict the claims to the interior and describe boundary handling separately.
  2. [Sec. 5, first paragraph and Figs. 3–4] The experimental section reports that the kernel was changed to ν=5/2 and the acquisition to LCB for Olympus 'where we found it beneficial' and 'where we observe better performances'. This post-hoc selection based on observed results undermines the claim of a generally superior method. Moreover, the paper's own results show constrained Euclidean BO slightly outperforming α0-GaBO on Concrete (Fig. 3a) and Euclidean BO on S7 leading to slightly lower function values in the mixture-of-classifiers task (Fig. 4). No statistical significance tests are provided, and no code is released. The abstract's 'increased performance compared to constrained Euclidean approaches' is therefore not supported as a general claim. Please: (a) report a pre-specified protocol or an ablation demonstrating that the kernel/acquisition choices are not dataset-specific; (b) provide paired statistical tests (e.g., Wil
minor comments (5)
  1. [Sec. 4.1 and App. A.2] The statement that φ is a diffeomorphism between the probability simplex and the positive orthant is only correct for the interiors; on the closed sets it is a homeomorphism. Please clarify.
  2. [Eq. (11)] The notation '1/2 φ*' is nonstandard and ambiguous. Clarify the pullback operation and the factor 1/2.
  3. [Fig. 2] The y-axis is labeled 'log10(F)' while the caption and text refer to 'logarithm of the regret'. Please make the quantity plotted explicit and consistent.
  4. [Sec. 3.3] Minor wording issue: 'surrogate a random field f' should likely be 'surrogate: a random field f' or 'surrogate random field f'.
  5. [Sec. 5.2] The observation that the Concrete optimum lies on the border, which α-1-GaBO cannot reach, is an expected limitation. Consider reflecting this explicitly in the conclusions.

Circularity Check

0 steps flagged

No significant circularity: the kernel is a genuine pullback construction and the empirical comparisons are external; minor self-citations are foundational, not self-referential.

full rationale

The paper's central derivation is the construction of kernels on the probability simplex by pulling back spherical Matérn kernels through the sphere map φ(x)=2√x (Eqs. 9–11). This is a constructive definition: positive definiteness of the pullback follows from injectivity of φ and positive definiteness of the sphere kernel, not from the benchmark results or from the paper's own conclusions. The acquisition-function optimizers use exponential maps of the α-connection (Eqs. 30–32), which are standard information-geometry results cited to Amari, Ay et al., and Pistone; they are not fitted to the data. Self-citations to Jaquier et al. (GaBO) and Candelieri et al. (BORIS) occur, but they serve as foundations and baselines rather than as evidence that α-GaBO's predictions are forced. The statement in Sec. 4.1 that smoothness of the isometry guarantees differentiability is questionable at the boundary, but that is a correctness/regularity concern, not a circularity. Likewise, the per-dataset choices of Matérn ν=5/2 and LCB for Olympus (Sec. 5) and the absence of released code or significance tests weaken the empirical superiority claim, but they do not make the derivation equivalent to its inputs. Overall, no load-bearing prediction reduces by construction to a fitted parameter or to a self-citation chain.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The method builds on standard information geometry and spherical kernel results; it introduces no new entities. The main non-standard assumption is the boundary extension of the isometry and kernel regularity, plus standard BO convergence assumptions.

free parameters (6)
  • α-connection parameter α = -1 and 0 tested
    Chosen by the user to define the optimizer family; not fitted, but it is the parameter of the proposed method.
  • GP kernel lengthscale κ = per-experiment, not reported
    Standard GP hyperparameter tuned via marginal likelihood; central method requires it but no specific value is load-bearing.
  • GP output variance σ² = per-experiment, not reported
    Standard GP hyperparameter tuned via marginal likelihood.
  • GP observation noise variance = per-experiment, not reported
    Standard GP hyperparameter.
  • Spherical kernel series truncation order = not reported
    Chosen for numerical approximation of Eq. (10); affects accuracy but not method logic.
  • Robotics objective weights = a1=a2=10, a3=1, λ=0.01, w1=w3=w4=1, w2=1e-3, b1=b2=1, b3=10
    Hand-chosen constants in Eq. (39) defining the benchmark loss; not part of the algorithm.
axioms (6)
  • standard math The sphere map φ is an isometry between the probability simplex with Fisher-Rao metric and the positive orthant of the sphere.
    Sec. 4.1 and App. A.2, citing Amari 2016 and Ay et al. 2017.
  • domain assumption The metric and isometry extend to the closed simplex including the boundary.
    App. A.2 uses Ay et al. Proposition 2.1 to extend the metric to the boundary; this is needed for kernels on vertices/faces, where φ is not differentiable.
  • standard math Matérn kernels on the sphere are positive definite and the restriction to the positive orthant pulls back to a valid kernel on the simplex.
    Uses Borovitskiy et al. 2020, Eq. (7)/(10); restriction and pullback preserve positive definiteness.
  • standard math The α-connection exponential maps (31)-(32) are correct and can be used inside Riemannian trust-region acquisition optimization.
    App. A.1, from Amari 2016 and Ay et al. 2017; trust-region machinery from Absil et al.
  • domain assumption Acquisition function optimization reaches a near-global maximizer of the acquisition function.
    Standard BO assumption; not proven for nonconvex acquisition on simplex with α-connections.
  • domain assumption The ANN trained on concrete data is a faithful oracle for the true compressive strength function.
    Sec. 5.2; if the ANN is a poor approximation, reported performance comparisons may not reflect the real objective.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Information Theoretic Bayesian Optimization over the Probability Simplex." pith.science (2026). https://pith.science/paper/PDRYEL2L

@misc{pith2026260309793,
  author       = {Pith},
  title        = {Pith review of: Information Theoretic Bayesian Optimization over the Probability Simplex},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDRYEL2L}},
  note         = {Machine review of arXiv:2603.09793}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Bayesian optimization is a data-efficient technique that has been shown to be extremely powerful to optimize expensive, black-box, and possibly noisy objective functions. Many applications involve optimizing probabilities and mixtures which naturally belong to the probability simplex, a constrained non-Euclidean domain defined by non-negative entries summing to one. This paper introduces $\alpha$-GaBO, a novel family of Bayesian optimization algorithms over the probability simplex. Our approach is grounded in information geometry, a branch of Riemannian geometry which endows the simplex with a Riemannian metric and a class of connections. Based on information geometry theory, we construct Mat\'ern kernels that reflect the geometry of the probability simplex, as well as a one-parameter family of geometric optimizers for the acquisition function. We validate our method on benchmark functions and on a variety of real-world applications including mixtures of components, mixtures of classifiers, and a robotic control task, showing its increased performance compared to constrained Euclidean approaches.

Figures

Figures reproduced from arXiv: 2603.09793 by Antonio Candelieri, Federico Pavesi, No\'emie Jaquier.

Figure 1
Figure 1. Figure 1: α-GaBO leverages the sphere map φ, which estab￾lishes an isometry between the probability simplex ∆d and the positive orthant S d ≥0 of the sphere. BO on the simplex is performed via equivalent representations on the sphere. practical success across a wide range of applications, in￾cluding hyperparameter tuning in machine learning [Wu et al., 2019], robotics [Jaquier et al., 2020, 2022a], exper￾imental des… view at source ↗
Figure 2
Figure 2. Figure 2: Logarithm of the regret (median and quartiles) and distribution over the final recommendation for [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Regret (median and quartiles) and distribution over the final recommendation for [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Regret (median and quartiles) and distribution over [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Left: Regret (median and quartiles) and distri￾bution over the final recommendation for α0-GaBO ( ), α-1-GaBO ( ), S d -Eucl. BO ( ), and BORIS ( ) for robotic multi-task control. Right: Snapshots of the optimal trajectory with target left ( ) and right ( ) hand positions. avoiding collisions and sudden potentially-damaging move￾ments. We compose 4 elementary tasks: two reaching tasks for the left and righ… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 1 canonical work pages

  1. [6]

    Martin Jaggi

    doi: 10.1088/2632-2153/abedc8. Martin Jaggi. Revisiting Frank-Wolfe: Projection- free sparse convex optimization. InIntl. Conf. on Machine Learning (ICML), pages 427–435, 2013. URL https://proceedings.mlr.press/v28/ jaggi13.html. Noémie Jaquier and Leonel Rozo. High-dimensional Bayesian optimization via nested Riemannian man- ifolds. InNeural Information ...

    doi:10.1088/2632-2153/abedc8 2013

  2. [7]

    cc/paper_files/paper/2020/file/ f05da679342107f92111ad9d65959cd3-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ f05da679342107f92111ad9d65959cd3-Paper. pdf. Noémie Jaquier, Leonel Rozo, Sylvain Calinon, and Math- ias Bürger. Bayesian optimization meets Riemannian manifolds in robot learning. InProceedings of the Con- ference on Robot Learning, volume 100, pages 233–246,

    2020

  3. [8]

    Noémie Jaquier, Viacheslav Borovitskiy, Andrei Smolen- sky, Alexander Terenin, Tamim Asfour, and Leonel Rozo

    URL https://proceedings.mlr.press/ v100/jaquier20a.html. Noémie Jaquier, Viacheslav Borovitskiy, Andrei Smolen- sky, Alexander Terenin, Tamim Asfour, and Leonel Rozo. Geometry-aware Bayesian optimization in robotics using Riemannian matérn kernels. In Conference on Robot Learning (CoRL), pages 794– 805, 2022a. URL https://proceedings.mlr. press/v164/jaqui...

    2022

  4. [10]

    Saeed Masoudnia and Reza Ebrahimpour

    doi: 10.1007/s00245-019-09564-3. Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey.Artificial Intelligence Re- view, 42(2):275–293, 2014. doi: https://doi.org/10.1007/ s10462-012-9338-y. Jonas Mockus. Bayesian approach to global optimization. InMathematics and its Applications, 1989. doi: https: //doi.org/10.1007/978-94-009-090...

    arXiv 2014

  5. [1995]

    cc/paper_files/paper/1995/file/ 7cce53cf90577442771720a370c3c723-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/1995/file/ 7cce53cf90577442771720a370c3c723-Paper. pdf. Jia Wu, Xiu-Yun Chen, Hao Zhang, Li-Dong Xiong, Hang Lei, and Si-Hao Deng. Hyperparameter optimization for machine learning models based on Bayesian optimization. Journal of Electronic Science and Technology, 17(1):26– 40, 2019. doi: https://doi.o...

    doi:10.11989/jest.1674-862x 1995

  6. [1998]

    Adam Kalai and Santosh Vempala

    doi: https://doi.org/10.1023/A:1008306431147. Adam Kalai and Santosh Vempala. Efficient algorithms for universal portfolios.Journal of Machine Learning Research, 3:423–440, 2002. doi: 10.1109/SFCS.2000. 892136. Oussama Khatib. The potential field approach and oper- ational space formulation in robot control. InAdap- tive and Learning Systems: Theory and A...

    doi:10.1023/a:1008306431147 2002

  7. [2011]

    Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P Adams, and Nando De Freitas

    doi: 10.1109/icra.2011.5980202. Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P Adams, and Nando De Freitas. Taking the human out of the loop: A review of Bayesian optimization. Proceedings of the IEEE, 104(1):148–175, 2015. doi: 10.1109/JPROC.2015.2494218. Patricia Kaori Soares, Roy Edward Bruns, and Ieda Spacino Scarminio. Statistical mixture design–p...

    arXiv 2011

  8. [2013]

    Immanuel M Bomze

    URLhttp://archive.ics.uci.edu/ml. Immanuel M Bomze. Regularity versus degeneracy in dy- namics, games, and optimization: a unified approach to different aspects.SIAM review, 44(3):394–414, 2002. doi: https://doi.org/10.1137/S00361445003756. Viacheslav Borovitskiy, Alexander Terenin, Peter Mostowsky, and Marc Peter Deisentoth. Matérn Gaussian processes on ...

    work page doi:10.1137/s00361445003756 2002

  9. [2015]

    Riley Hickman, Priyansh Parakh, Austin Cheng, Qianx- iang Ai, Joshua Schrier, Matteo Aldeghi, and Al’an Aspuru-Guzik

    doi: 10.1098/rspa.2015.0142. Riley Hickman, Priyansh Parakh, Austin Cheng, Qianx- iang Ai, Joshua Schrier, Matteo Aldeghi, and Al’an Aspuru-Guzik. Olympus, enhanced: benchmarking mixed-parameter and multi-objective optimization in chemistry and materials science.ChemRxiv, 2023. doi: 10.26434/chemrxiv-2023-74w8d. Florian Häse, Matteo Aldeghi, Riley J. Hick...

    arXiv 2015

  10. [2020]

    cc/paper_files/paper/2020/file/ 92bf5e6240737e0326ea59846a83e076-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 92bf5e6240737e0326ea59846a83e076-Paper. pdf. Nicolas Boumal.An introduction to optimization on smooth manifolds. Cambridge University Press, 2023. URL http://www.nicolasboumal.net/book. Antonio Candelieri. Resource allocation via Bayesian op- timization: an efficient alternative to semi-band...

    2020

  11. [2021]

    cc/paper_files/paper/2021/file/ 44e76e99b5e194377e955b13fb12f630-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2021/file/ 44e76e99b5e194377e955b13fb12f630-Paper. pdf. Aasa Feragen, Francois Lauze, and Soren Hauberg. Geodesic exponential kernels: When curvature and lin- earity conflict. InIEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3032–3042, 2015. doi: 10.1109/CVPR.2015.7298922. Margueri...

    arXiv 2021

  12. [2023]

    James Chok and Geoffrey M

    doi: 10.1007/s10472-023-09883-w. James Chok and Geoffrey M. Vasil. Optimization over a probability simplex.Journal of Machine Learning Re- search, 26(73):1–35, 2025. URL http://jmlr.org/ papers/v26/23-1166.html. Oscar Davis, Samuel Kessler, Mircea Petrache, ˙Ismail ˙I Ceylan, Michael Bronstein, and Avishek J Bose. Fisher flow matching for generative model...

    arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.