acceptodds
Under review as a conference paper at ICLR 2027

Beyond Unimodal Bases: Pullback Geometry for Multimodal Data

Abstract

Data-driven Riemannian geometry provides nonlinear interpolation and geometric representations of high-dimensional data. For these operations to be statistically meaningful, paths between observations should preferentially traverse high-likelihood regions. Existing scalable pullback constructions typically use a unimodal Gaussian latent distribution, assuming that the data reside close to a single manifold. For multimodal data, mapping separated modes or local structures into one Gaussian region can require substantial transport deformation and compromise the resulting geometry. We introduce a pullback geometry for data supported on mixtures of manifolds. Using a latent Gaussian mixture, we define its Riemannian metric as the matrix square of the responsibility-weighted expected component precision. The metric is smooth and positive definite and recovers the existing Gaussian construction in the single-component limit. For structured overlapping mixtures, we establish conditions under which the log-density is concave along geodesics, providing a formal connection between the proposed geometry and paths through high-likelihood regions, and derive the corresponding local curvature relations. We instantiate this geometry in a normalizing flow with adaptive mixture learning, allowing the number of active components to emerge from the data and supporting component-wise reconstruction and local effective-dimension estimation. Experiments on synthetic geometric data, a controlled multi-view image setting with a known reference trajectory, and MNIST show reduced transport distortion, competitive path support, close reference-trajectory recovery, and improved interpolation realism. These results extend scalable pullback geometry beyond datasets that reside close to a single manifold while retaining tractable and interpretable local structure.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Pullback Riemannian geometry, normalizing flows, Gaussian mixture models, density-aware geodesics, multimodal data

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_Q5SYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

A multimodal extension of pullback geometry is introduced. It replaces the unimodal Gaussian latent distribution, which was used in prior work, by a Gaussian mixture model and uses the a responsibility-weighted expected component precision to define a Riemannian metric. The resulting geometry is positive definite, and experiments indicate that is can better represent multimodal datasets than unimodal constructions applied in prior work. The construction is supported by theoretical analysis.

Critical strengths
  1. A important limitation of pullback geometries is addressed. A single Gaussian latent space is not enough for data that consists of multiple manifold-like structures. The proposed step from unimodal Gaussians to mixtures is natural and allow for an interpretation as going from a single chart to multiple local charts.

  2. The proposed Riemannian metric using P^2 is elegant, has desired properties (smoothness, positive definiteness) and agrees with the existing Gaussian metric from Diepeveen et al. (2025) for K=1. Experiments demonstrate befits from the proposed construction compared to prior work.

Critical weaknesses
  1. The exact score-based metric is derived but the introduced metric discardes one of the terms, the covariant term. It is a reasonable construction and produces a valid metric. However, this is a design choice and the metric is not uniquely derived from first prinicples. This means that alternative choices of a metric are possible but an exploration of alternative design choices is missing. While the paper demonstrates that the proposed metric works, it is not clear how much of the observed improvement comes from learning a Gaussian mixture, and how much comes specifically from choosing the metric to be P^2. For example, would using P^a or (P-Cov)^2 yield similar (or worse or better) results?

  2. While the Cov term cannot be guaranteed to be positive definite, it is not clear to me whether or not the Cov term in applications is actually harmful. Even though positive definteness cannot be established it still may be of benefit in applications. I am missing some experiments including Cov term and shed light on these questions.

  3. The theoretical results are established under strong assumptions that I think will not hold in applications. Therefore, it merely serves as a motivation for the approach than establishing guarantees.

Questions for authors

The concavity results are proved for certain star-shaped mixtures. To what extent do you expect the guarantees provided by the theory hold for the mixtures learned in the experiments?

2 comments
AuthorsOpenReview ↗

We clarify the motivation for , the effect of component-score covariance, and the scope of the theoretical guarantees for learned mixtures.

Thank you for the constructive review.

Response to Weakness 1: Metric choice and alternative constructions

We agree that P2P^2 is not uniquely forced by the exact score-Hessian construction. It is a deliberate design choice motivated by several properties: it is smooth and strictly positive definite for every nondegenerate GMM, explicit and tractable to evaluate, and exactly recovers the original SBPRG metric when K=1K=1. Within the family PaP^a, exact recovery of the single-Gaussian metric requires a=2a=2. This exponent also yields the algebraic cancellation used to derive the explicit GLC condition in our structured theoretical analysis; we will add the detailed derivation and a comparison with general aa to the appendix.

These properties justify our choice but do not show that P2P^2 is empirically optimal. Other constructions, including regularized versions of the score-Hessian metric, are possible. The current submission does not systematically compare all such choices and therefore does not completely separate the benefit of learning a GMM from the benefit of the specific P2P^2 metric. We will state this limitation explicitly.

Response to Weakness 2: Effect of the component-score covariance

Let sk(z)=∇log⁡N(z;μk,Ak)s_k(z)=\nabla\log\mathcal N(z;\mu_k,A_k) denote the score of component kk, and define its responsibility-weighted covariance as Cs(z)=Cov⁡γ(z)[sk(z)].C_s(z)=\operatorname{Cov}_{\gamma(z)}[s_k(z)].

This is the covariance of the component scores, not the covariance AkA_k of an individual Gaussian. With the expected precision P(z)=∑kγk(z)Ak−1,P(z)=\sum_k\gamma_k(z)A_k^{-1}, the exact mixture log-density Hessian is H(z)=−P(z)+Cs(z).H(z)=-P(z)+C_s(z).

Although Cs(z)C_s(z) is positive semidefinite, it is added to the negative-definite term −P(z)-P(z). It can therefore partially or completely cancel expected precision in some directions. If a nonzero vector vv satisfies H(z)v=(−P(z)+Cs(z))v=0,H(z)v=\bigl(-P(z)+C_s(z)\bigr)v=0, then v⊤H(z)⊤H(z)v=0.v^\top H(z)^\top H(z)v=0.

Thus, the direct score-Hessian tensor H(z)⊤H(z)H(z)^\top H(z) is only positive semidefinite at that point and does not define a Riemannian metric. Near cancellation can likewise cause poor conditioning. This is why the exact score-Hessian construction is theoretically unsafe for the role required in this work. We will add a detailed derivation of this cancellation and an explicit two-component example to the appendix.

We agree, however, that this argument does not establish that the covariance term is empirically harmful whenever H(z)H(z) remains nonsingular. It may contain useful information in such regions. The submitted experiments do not include an ablation retaining this term, so we do not claim to have isolated its empirical effect. A systematic comparison of P(z)2P(z)^2 and H(z)⊤H(z)H(z)^\top H(z), including conditioning and path quality, remains important future work.

Response to Weakness 3 and Question 1: Scope of the theoretical guarantees

We agree that the GLC and curvature results require strong structure. The shared-mean and common-eigenbasis assumptions are used to derive explicit conditions. The general guarantee is instead that P2P^2 defines a smooth, strictly positive-definite metric for every nondegenerate GMM.

The appendix partially extends GLC to mixtures with arbitrary separated means and shared covariance, but we do not have corresponding guarantees for mixtures with both arbitrary means and arbitrary covariances. Moreover, global GLC cannot generally hold for sufficiently separated modes because every continuous path connecting separated high-density regions must pass through lower density.

The exact star guarantees do not apply directly to all mixtures learned in the experiments. Wavy-X and BlueWhale use a common component mean. For MNIST, we relax the shared-mean condition by allowing component means to vary within a dataset-specific radius. The star theory therefore applies most directly to the common-mean settings and motivates structured post-training diagnostics, while path behaviour for the relaxed MNIST mixture is assessed empirically.

The Gaussian-star class was chosen because it provides greater flexibility than a single Gaussian while remaining sufficiently structured for explicit analysis. This balance is relevant because our flow is constrained by isometry regularization and cannot be assumed to transform arbitrary data into any desired latent configuration.

For learned mixtures with relaxed assumptions, the theory provides a structured reference for post-training diagnostics, but not a theorem-level guarantee. For mixtures with relaxed means, the reported star–Ricci quantity should therefore be interpreted as a structured post-training diagnostic rather than the exact Ricci curvature of the general learned metric.

We will distinguish clearly between general metric validity, exact star guarantees, and theory-informed diagnostics and empirical validation under relaxed assumptions.

Thank you for the detailed responses and the willingness to communicate the identified weaknesses more clearly in the paper. Overall, my “weak accept” rating remains unchanged.

Reviewer_3kEhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper extends score-based pullback geometry to settings where the data-distribution is multimodal. That is, they learn a Riemannian metric on the data-space which can be used for Riemannian machine learning. The metric is used and evaluated by recovering interpolating paths between data points that traverse high-likelihood regions.

Critical strengths
  1. The paper is a natural extension of score based pullback geometry, and fits nicely in that literature.
  2. The problem setting of capturing mixtures of data-manifold is interesting and challenges the common manifold hypothesis, which typically assume that the data comes from a single manifold which is typically assumed to be connected, with no intersection points, and a fixed global dimension.
Critical weaknesses
  1. It is unclear what the authors mean by "manifold". The manifold hypothesis often referrs to when the data distribution is assumed to be strictly lower dimensional and therefore singular (e.g. not absolutely continuous w.r.t Lebesgue measure). However, the setting of pullback geometry using normalizing flows does not cover such case. Indeed, by definition a normalizing flow is a diffeomorphism from R^d to R^d, it is therefore invertible and thus the resulting pushforward density is not singular and its support include all of the space. The geometry therefore only comes from the resulting non-singular density, and does not fit the standard manifold-hypothesis picture in which the geometry (the manifold) is separate from the density. Clarifications of what is meant by manifold here is necessary, as well as what is expected from a geometry coming from such full-dimensional geometry. This is particularly relevant given the problem of "manifold overfitting" that plagues normalizing flows [1].
  2. The paper is very hard to read. The technical terminology is often not appropriately introduced. The paper also assumes familiarity with unimodal pullback geometry.

[1] Deep Generative Models through the Lens of the Manifold Hypothesis: A Survey and New Connections by Loaiza-Ganem et al

Questions for authors
  1. What is meant by "tractable latent space" in the context of "scalable pullback geometry"? Particularly given that for normalizing flows, the latent space is the same dimension as the original data space, so I struggle to see what computations can be made faster in that space. Additionally, has pullback geometry been scaled beyond MNIST?
  2. Could the authors clarify when the covariance term is dropped from equations 4 to 5?
  3. As a simple geodesic baseline, could the authors consider an energy defined simply as the likelihoods of the midpoints under the learned density? I would like to see how this differs from the proposed method.
  4. How is the number of components of the GMM model set and tuned in practice? If one such component is required for every mode, this seems to be a limitation of the method's scalability.
3 comments
AuthorsOpenReview ↗

We clarify our use of “manifold” and outline concrete revisions to make the paper more self-contained and easier to follow.

Thank you for the detailed review. We agree that several central terms and design choices require clearer explanations.

Response to Weakness 1: Meaning of “manifold”

Here, a manifold refers to a lower-dimensional smooth set in the differential-geometric sense, and a mixture of manifolds refers to a union of such sets. In our framework, we relax exact support on these manifolds by assuming that observations are concentrated near them, resulting in a full-dimensional statistical model. We do not aim to extract the underlying lower-dimensional sets explicitly.

Instead, the latent GMM encodes the geometry around these structures through its component covariances and responsibilities. In the original score-based Gaussian metric, high-variance directions have lower geometric cost, while low-variance directions have greater cost. In our mixture model, component covariances encode different local directional structures, and responsibilities determine which structures are locally relevant. The pullback transfers this mixture geometry to data space.

Unlike the earlier Riemannian autoencoder construction, our flow remains full-dimensional and does not explicitly extract the underlying manifolds. The component-wise reconstruction and effective-dimension estimate in Equation (16) are post-training local analyses. We will clarify this interpretation in the abstract and background and state that exact manifold extraction and manifold overfitting are outside the scope of this work.

Response to Weakness 2: Presentation and readability

We agree that the paper assumes too much familiarity with unimodal score-based pullback geometry. The reviewer’s comments show that the original SBPRG construction, the role of the latent distribution, and the distinction between Equations (4) and (5) were not sufficiently clear; our own rereading confirms this.

To streamline the paper around its intended contribution, we will first introduce the single-Gaussian SBPRG metric and explain how isometry regularization makes the choice of latent distribution important. We will then explain why a single Gaussian can be restrictive, define GMM responsibilities and expected precision, and introduce the proposed metric. After establishing the geometry, we will present the adaptive mixture-learning procedure. Detailed derivations and secondary implementation details will be moved to the appendix. We will also define all technical terms before use and include a concise overview of the most relevant prior work in the main text.

AuthorsOpenReview ↗

We clarify tractability, the transition to the stabilized metric, the proposed density-only baseline, and adaptive component selection and scalability.

Response to Question 1: What is meant by “tractable,” and has pullback geometry scaled beyond MNIST?

By “tractable geometry,” we mean that geodesics can be computed stably and efficiently, while the geometry remains scalable to learn and evaluate. We will replace the ambiguous phrase “tractable latent space” accordingly.

The latent metric is explicit:

\qquad g_Z(z)=P(z)^2.$$ Its discrete path energy can be evaluated directly from the fitted mixture parameters; for diagonal covariances, a metric-vector product costs $\mathcal O(Kd)$. The optimized latent geodesic is then mapped through the inverse flow to obtain the corresponding data-space geodesic under the pullback metric. Thus, the computational advantage comes from the explicit metric and efficient latent geodesic computation, although the data and latent spaces have the same dimension. The evaluations discussed in our paper remain limited to synthetic or controlled image data and MNIST. These settings allow transport distortion and path recovery to be assessed against known geometric structure, making them a necessary intermediate step. Scalability to high-resolution natural images has not yet been established, and we will state this limitation explicitly. **Response to Question 2: When is the covariance term dropped between Equations (4) and (5)?** Equation (5) defines a new stabilized metric; it is not Equation (4) with the covariance term assumed to vanish. The omitted term is the component-score covariance $C_s(z)=\operatorname{Cov}_{\gamma(z)}[s_k(z)]$ in Equation (4). This term can cancel the expected precision in some directions, making the mixture Hessian singular and its Gram matrix only positive semidefinite. The resulting construction may therefore fail to define a Riemannian metric. We instead define $g_Z(z)=P(z)^2$, which is smooth, strictly positive definite for every nondegenerate GMM, and exactly recovers SBPRG when $K=1$. The component-score covariance is excluded globally from the definition and evaluation of the proposed metric; it is not removed only during training or in selected regions. We will clarify this design step in the main text and add the full derivation and an explicit example to the appendix. **Response to Question 3: Density-only midpoint baseline** We agree that midpoint likelihood provides a useful density-only baseline because it separates scalar likelihood from the displacement and directional covariance information contained in the expected-precision metric. It evaluates whether sampled segment midpoints lie in high-density regions, whereas the proposed Riemannian energy additionally accounts for segment displacement and local directional geometry. We will compare the two using the same learned density, endpoints, discretization, initialization, and optimization budget. We will report the midpoint objective as a density-only path score rather than as a Riemannian geodesic energy. **Response to Question 4: How is the number of components set and tuned, and is one component required for every mode?** We initialize an overcomplete mixture with $m_0$ components, chosen as a computationally feasible upper bound rather than as an estimate of the number of modes or classes. The final active count $K_{\mathrm{eff}}$ is not tuned separately by hand; it emerges through responsibility-based updates, sparse Dirichlet regularization, merging, and pruning. Gaussian components are not assigned one-to-one to statistical modes, i.e., local maxima of the complete mixture density. Although an individual Gaussian is unimodal, one component may support several nearby empirical modes or structures with similar covariance patterns, while multiple overlapping components may contribute to the same density region. Gaussian components therefore need not correspond directly to statistical modes, semantic classes, or geometric branches. Empirically, the adaptive procedure retains substantially fewer components than the overcomplete initialization: Wavy-X reduces from 20 candidates to two active components, while MNIST reduces from 30 to 13, without using arm or digit labels. The scalability concern remains relevant when the data contain many well-separated modes or local structures, since these may require many active components. Computational cost therefore grows with $K_{\mathrm{eff}}$. Merging and pruning remove unnecessary components but cannot eliminate this dependence when a large mixture is genuinely required. We will clarify this limitation and, where possible, report sensitivity to the initial count $m_0$.

I thank the authors for their rebuttal and clarifications. However, my concerns regarding the exposition and experimental validation remain insufficiently addressed by the material currently provided. The proposed revisions and additional experiments could help resolve these concerns, but their impact cannot yet be assessed. I therefore maintain my score.

Reviewer_4ovnOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work extends score-based pullback Riemannian geometry to multimodal data distributed across multiple manifolds, addressing the severe transport distortion induced by the standard unimodal Gaussian latent base.

Key contributions:

  1. A smooth, positive‑definite Riemannian metric defined as the squared responsibility‑weighted expected precision matrix, which recovers the existing single‑Gaussian SBPRG construction in the one‑component limit.
  2. Theoretical analysis of geodesic log‑concavity conditions and local curvature relations for structured Gaussian star distributions, formally connecting the induced geometry to high‑likelihood regions.
  3. A stabilized adaptive learning scheme that jointly optimizes a normalizing flow and an over‑complete Gaussian mixture, allowing the number of active components to emerge automatically from data, with component‑wise local reconstruction support.

Validation:

The method is validated both theoretically and empirically. Experiments on synthetic Wavy‑X data, the BlueWhale multi‑view dataset with metadata‑defined reference trajectories, and MNIST demonstrate reduced transport distortion, improved path support, close reference‑trajectory recovery, and more realistic interpolation compared to single‑Gaussian SBPRG and fixed‑mixture baselines.

Critical strengths
  1. Rigorous theoretical foundation with guaranteed compatibility and formal geometric properties

The paper proves the proposed expected-precision metric is smooth, positive definite, and exactly recovers the established SBPRG construction in the single-component limit, ensuring backward compatibility. It further derives formal geodesic log-concavity conditions and local curvature relations for structured Gaussian star distributions, establishing a principled connection between the induced Riemannian geometry and high-likelihood regions of the data distribution. This theoretical rigor is critical for a geometry-learning paper: it grounds the method in formal Riemannian geometry rather than presenting an ad-hoc empirical construction, which substantially strengthens confidence in the generality of the results.

  1. Practical adaptive learning framework with comprehensive, complementary empirical validation

The work delivers a complete stabilized training algorithm that jointly learns a normalizing flow and an over-complete Gaussian mixture, allowing the number of active modes to emerge automatically from data without pre-specification — a key practical advantage for real multimodal data where mode count is unknown. Experiments across three complementary settings (synthetic Wavy-X, BlueWhale with ground-truth reference trajectories, and MNIST) consistently demonstrate reduced transport distortion, better path support, closer trajectory recovery, and more realistic interpolation compared to unimodal and fixed-mixture baselines. The breadth of validation across geometric, trajectory, and perceptual criteria provides strong empirical support for the claimed benefits.

Critical weaknesses
  1. Theoretical guarantees are restricted to highly structured synthetic mixtures and do not generalize to real multimodal data

The paper’s core theoretical results — geodesic log-concavity conditions, local curvature relations, and the formal link between the induced metric and high-likelihood regions — are derived exclusively for Gaussian star distributions with shared means and common eigenbases. These highly symmetric, structured configurations are not representative of real-world datasets such as MNIST or even the Blue Whale data used in the experiments. This means the key geometric motivation — that geodesics will preferentially traverse high-density regions — remains empirically motivated rather than theoretically justified for general multimodal data. The gap between the restrictive theoretical assumptions and the more realistic experimental settings weakens the conceptual rigor of the contribution.

  1. Empirical evaluation is limited in scope and lacks comparison to competing manifold/geometry learning methods

The experiments primarily compare against the unimodal SBPRG baseline and fixed-mixture controls, but do not benchmark against other established approaches for multi-manifold geometry learning, such as multi-chart flows, diffusion-based geometry, or density-aware distance metrics. It is therefore unclear whether the proposed method offers advantages relative to alternative techniques for multimodal data. Additionally, the MNIST evaluation relies on an indirect classifier-feature Fréchet distance measure of perceptual realism, without direct validation of geometric accuracy or downstream task utility (e.g., retrieval, classification, optimization over the learned manifold). This limits the evidence for practical impact.

Questions for authors

The core theoretical results — geodesic log-concavity conditions and local curvature relations — are derived only for Gaussian star distributions with shared means and common eigenbases. Could the authors clarify whether corresponding theoretical guarantees exist for general Gaussian mixtures with separated means and arbitrary covariances, to ensure that geodesics preferentially traverse high-likelihood regions beyond this highly structured class? Alternatively, is this property currently only theoretically justified for star configurations and empirically observed for more general multimodal data?

The empirical evaluation primarily benchmarks against unimodal SBPRG and fixed-mixture controls, but does not include comparisons to other established multi-manifold geometry learning approaches such as multi-chart flows, diffusion-based geometry, or density-aware distance metrics. Could the authors explain the choice of baselines and whether supplementary results exist demonstrating advantages relative to these alternative methods for multimodal data? Additionally, the MNIST evaluation relies on an indirect classifier-feature Fréchet measure of perceptual realism; is there validation of geometric accuracy or downstream task utility (e.g., classification, optimization over the learned manifold) to support practical impact?

2 comments
AuthorsOpenReview ↗

We clarify that the metric-validity result applies to general nondegenerate GMMs, while the explicit GLC and curvature guarantees require the structured star assumptions; learned mixtures with relaxed assumptions are assessed through theory-informed diagnostics and empirical path evaluation. We also explain our controlled baseline choices, acknowledge the absence of broader method and downstream comparisons, and will add a density-only baseline in the revised version.

Thank you for the positive and careful assessment. We agree that the scope of the theoretical guarantees and empirical comparisons should be stated more explicitly.

Response to Weakness 1 and Question 1: Scope of the theoretical guarantees

The expected-precision metric gZ=P2g_Z=P^2 is smooth and strictly positive definite for every nondegenerate GMM, including mixtures with arbitrary means and positive-definite covariances. This general metric-validity result applies to all learned experimental mixtures.

The explicit geodesic log-concavity (GLC) and curvature guarantees are established for shared-mean Gaussian stars whose component precisions share a common eigenbasis. The appendix partially extends GLC to mixtures with arbitrary component means and shared covariance AA. Writing

Σμ(z)=Cov⁡γ(z)[μk],\Sigma_\mu(z)=\operatorname{Cov}_{\gamma(z)}[\mu_k],

GLC at zz in every tangent direction holds if and only if

\left(A^{-1/2}\Sigma_\mu(z)A^{-1/2}\right) \leq 1.$$ This condition shows how increasing separation of the locally relevant means can cause GLC to fail. We do not currently have corresponding GLC or curvature guarantees for mixtures with both arbitrary means and arbitrary covariances. Moreover, global GLC cannot generally hold for strongly separated modes because any continuous connecting path must pass through lower-density regions. The synthetic point-cloud Wavy-X and BlueWhale experiments use a common component mean. For MNIST, we relax the shared-mean condition by allowing component means to vary within a dataset-specific radius. The star theory both informs the metric construction and motivates the post-training GLC and star–Ricci diagnostics. We do not interpret these diagnostics as a direct theorem-level guarantee for the relaxed MNIST mixture; path behaviour is evaluated empirically. We will revise the paper to distinguish clearly between the general metric-validity result, the exact guarantees under the star assumptions, and the theory-informed diagnostics and empirical validation used under relaxed assumptions. **Response to Weakness 2 and Question 2: Scope of the empirical evaluation** We agree that the submitted experiments do not establish superiority over all manifold- or geometry-learning methods. The baselines were selected as controlled comparisons under a common flow architecture. SBPRG is the existing single-Gaussian method extended by our approach. The fixed-$K$ models distinguish introducing a mixture from adaptively learning its component structure, while straight latent interpolation provides a non-geometric reference. All flow-based models use the same RealNVP architecture and comparable training settings. We do not currently have supplementary comparisons with multi-chart flows or diffusion-based geometry. These methods use different representations, training objectives, and geometric constructions, so they constitute broader method comparisons rather than controlled ablations. We agree that such comparisons would broaden the evaluation and will state their absence as a limitation. A density-only path objective provides the most direct additional comparison because it can use the same trained flow, GMM, endpoints, discretization, initialization, and optimization setting. It separates scalar likelihood from the directional information contained in the expected-precision metric. We will add the midpoint-likelihood baseline in the revised version. The three datasets evaluate complementary properties. Wavy-X evaluates transport distortion and adherence to the observed data structure. BlueWhale evaluates recovery of a metadata-defined reference trajectory. These provide more direct evaluations of geometry and reference-path recovery. MNIST has no prescribed ground-truth interpolation path, so classifier-feature Fréchet distance assesses only the realism of interior path samples in the selected feature space. We do not interpret it as a direct measure of geometric accuracy. The current submission does not evaluate downstream classification, retrieval, or optimization using the learned metric. We agree that these tasks would provide useful additional evidence of practical utility. We will clarify that the MNIST experiment supports interpolation realism, while broader method comparisons and downstream validation remain future work.

I have reviewed the authors’ response and planned revisions.

On theoretical scope: The authors clarify three tiers of results: general metric validity for all nondegenerate GMMs, formal GLC/curvature guarantees only under structured star-distribution assumptions, and theory-guided diagnostics plus empirical evaluation for relaxed realistic mixtures. This resolves my concern about overgeneralized claims, though full guarantees for arbitrary GMMs remain absent as a reasonable limitation.

On empirical scope: The rationale for controlled same-architecture baselines is methodologically justified. The planned density-only baseline and explicit statement of missing broader comparisons and downstream validation improve transparency, but the method’s relative standing against alternative geometry-learning paradigms remains unestablished.

I maintain my rating of 3 (Weak Acceptance).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.