acceptodds
Under review as a conference paper at ICLR 2027

Recovering Weighted Tangent Geometry from a Single-Scale Score Field

Abstract

Score fields learned by diffusion models offer a direct probe of local data geometry, but the usual tangent-space description is insufficient at branch points, where the first-order object is a weighted measure over branch directions. Recovering this geometry matters because junctions arise when trajectories merge, road segments intersect, or strata meet, and their directions and relative masses determine what a learned field represents. We ask whether a score field at one known noise level can recover the branch center, local homogeneity degree, directions, and relative masses. This inverse problem is difficult because the observed score is an uncentered vector field, a one-point Hessian cannot distinguish even regular junctions with different branch counts, and ordinary score accuracy need not preserve the angular information required for geometry recovery. We address these challenges by using homogeneity under Gaussian smoothing to obtain a linear, score-only calibration of the unknown center and degree. Once centered, the tangential score on one shell integrates to a scalar transform with strictly positive spherical-harmonic multipliers. This yields exact identification of the normalized angular measure in any ambient dimension and constructive recovery of finite positive rays. In the plane, it further yields sharp moment requirements, finite-query certificates, and an tangent-score approximation for finite branches. Experiments apply the full chain to population, kernel-density, and learned score fields, showing across four geometries that lower validation normalized-score error can coincide with higher angular-moment error, thereby distinguishing ordinary score fit from recoverable junction geometry.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
score-based generative models, tangent measures, singular geometry, inverse problems, spherical harmonic analysis

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_9KByOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies whether a score field at one known Gaussian noise level determines the local weighted geometry of a junction namely its center, homogeneity degree, branch directions, and relative masses. For an exact homogeneous tangent measure, the authors derive a derivative-free linear calibration of center and degree. Once centered, integrating the tangential score over a sphere recovers a scalar angular transform. Strictly positive harmonic multipliers establish injectivity, and classical moment methods then recover finite collections of rays. The planar analysis adds sharp consecutive-moment requirements, finite-query error bounds, conditional certification, and an $O(\sigma^\beta)$ tangent-score approximation for specified curved branches. Synthetic population, KDE, and neural-score experiments examine these stages and report that improved score RMS can coexist with worse angular-moment recovery. Exact identification concerns the homogeneous model; recovery from finite-noise curved data and learned fields remains conditional on approximation, conditioning, and error bounds.

Critical strengths
  1. I liked that the paper shows how to recover both branch directions and their relative weights from scores at one noise level. Theorem 1 addresses the unknown center and homogeneity degree, and Theorem 2 shows that one centered shell identifies the angular measure. I think the main contribution is connecting the score field to the moments needed by existing reconstruction methods.

  2. I also liked the explanation of why a good score estimate does not necessarily give accurate junction geometry. The example of different junctions sharing the same vertex Hessian makes this limitation concrete. The stability analysis and Table 7 further show why recovering the correct branch count can be easier than recovering accurate directions and weights.

Critical weaknesses
  1. I had a question about whether the finite-sample guarantee covers estimating the center and constructing the KDE from the same data. In Appendix I, the concentration argument leading to Equation (98) appears to treat the query locations as fixed. However, estimating the center from the data also makes the shell locations depend on those data. I could not find an argument accounting for this dependence. This matters for the probability guarantee in Equation (17), although it does not directly affect the exact identifiability results. An independence assumption or a bound that holds uniformly over the possible query locations could resolve the concern.

  2. I had trouble determining whether the experiments evaluate the complete procedure starting from an unknown center and degree. Table 1 reports calibration and reconstruction separately, and the protocols in Appendices K.3–K.6 do not clearly explain whether the estimated center and degree are passed into reconstruction. I think this needs clarification because the reported recovery rates may describe a simpler setting than the full problem introduced in the paper.

Questions for authors
  1. Could you clarify whether the center estimate in the finite-data analysis uses the same samples as the KDE? If it does, how does the concentration argument account for the resulting data-dependent query locations?

  2. Could you identify which reported experiments pass both the estimated center and estimated degree into branch reconstruction? I would find the joint recovery results helpful for understanding how errors from the different stages interact. If this has not been tested, a small synthetic evaluation would help support the claim that the full procedure has been validated.

Reviewer_WX9mOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether the local geometry of a junction can be recovered from a score field sσ(x)=∇xlog⁡(μ∗ϕσ)(x)s_\sigma(x)=\nabla_x \log (\mu * \phi_\sigma)(x) at a single known noise scale. While previous score-geometry analyses typically characterize smooth supports through a single tangent space, a branch point generally has no unique tangent space. The paper instead represents its first-order geometry by a normalized angular tangent measure, whose atoms encode branch directions and whose weights describe their relative local mass. The goal is to recover the junction center x0x_0, the local homogeneity degree dd, and this weighted tangent geometry using only local score queries.

The proposed recovery proceeds in two main stages. First, the authors exploit homogeneity under Gaussian smoothing to derive a score-only linear calibration procedure for jointly estimating the unknown center x0x_0 and homogeneity degree dd, without differentiating the score. After recentering, the tangential component of the score on a single shell is integrated to recover a scalar angular density. Its spherical-harmonic coefficients provide angular moments of the tangent measure, from which finite collections of branches can be reconstructed using classical finite-rank moment methods. The paper proves exact identifiability of the normalized angular measure for arbitrary ambient dimension D≥2D\geq 2. For at most KK finite positive rays, moments through degree 2K−12K-1 give constructive recovery in general dimension, while in the planar case moments through degree KK are sufficient and degree K−1K-1 is insufficient.

The paper further provides finite-query and perturbation guarantees and validates the full recovery pipeline on population, KDE, perturbed, and learned score fields. The synthetic experiments show accurate recovery under increasing local coverage, while the learned-score experiments reveal an important mismatch: lower validation score error can coincide with higher angular-moment error and incomplete geometric recovery.

Critical strengths
  1. A novel and well-motivated extension of score-based geometric analysis beyond smooth manifolds.
    The paper addresses an important limitation of the usual tangent-space view of score geometry: at junctions or stratified singularities, a single tangent space is not sufficient to describe the local structure. Replacing it with a weighted angular tangent measure is a natural and conceptually appealing formulation, as it captures branch count, directions, and relative local mass within a unified first-order object.

  2. Strong and technically nontrivial identifiability results from a single score scale.
    A central contribution is the result that the junction center, homogeneity degree, and normalized angular tangent measure can be recovered from local score queries at a single known noise scale. The score-only calibration of x0x_0 and dd avoids differentiating the score, and the one-shell construction connects tangential score information to angular moments through spherical harmonics. The arbitrary-dimensional identifiability result and the sharp planar moment requirement provide a particularly clean theoretical characterization of what information is recoverable.

  3. The paper carefully distinguishes identifiability from stability and finite-sample recovery.
    Rather than stopping at an exact inverse result, the authors analyze how errors propagate through center calibration, shell integration, harmonic inversion, branch-count estimation, direction recovery, and weight estimation. The conditional certification framework is also explicit about when the available information is insufficient and allows the procedure to abstain instead of making an uncertified geometric claim. This makes the theoretical treatment substantially more informative than an idealized noiseless identifiability result alone.

  4. The learned-score experiments reveal an interesting distinction between ordinary score accuracy and geometric fidelity.
    The observation that lower validation score error can coincide with larger angular-moment error and incomplete junction recovery is particularly interesting. It suggests that conventional score-matching metrics may not fully capture whether a learned score field preserves fine local geometric structure. This finding provides a useful motivation for geometry-aware diagnostics of learned score models and opens several directions for further investigation.

Critical weaknesses

Weaknesses

  1. Limited validation on high-dimensional and pretrained score models.
    While the theoretical results apply to arbitrary ambient dimension D≥2D \geq 2, the learned-score experiments are restricted to low-dimensional synthetic settings, primarily D=2D=2 and D=3D=3. The D=5D=5 experiments use exact or artificially perturbed score fields rather than learned scores. It therefore remains unclear how well the proposed recovery procedure scales statistically and computationally to realistic high-dimensional score models. This is particularly relevant because the required number of scalar tangential observations scales with both the ambient dimension and the branch bound (KD−1KD-1 for the fixed-query setting), while the higher-dimensional reconstruction relies on spherical-harmonic moment recovery. An evaluation on a higher-dimensional learned score field, or ideally on a pretrained model, would substantially strengthen the empirical support for the general-dimensional theory.

  2. The local geometric model is relatively idealized.
    The framework assumes that the local geometry admits an approximately homogeneous tangent description, and the constructive finite-ray setting represents the junction as a finite collection of tangent rays with associated weights. This is natural as a local blow-up model for sufficiently regular branches, but the strongest finite-noise guarantees are established for zero-thickness planar C1,βC^{1,\beta} branches with positive C0,βC^{0,\beta} densities. It is less clear how the method behaves for thicker supports, heterogeneous-dimensional strata, or junctions that cannot be well represented by a small number of discrete tangent directions. Clarifying the extent to which the proposed recovery remains meaningful beyond the finite-ray setting would improve the connection to more realistic stratified data.

  3. Single-scale identifiability does not eliminate the practical scale-selection problem.
    A key strength of the paper is that one known noise scale is sufficient for identifiability. However, practical recovery can still depend strongly on the choice of σ\sigma. Smaller σ\sigma improves the tangent approximation but reduces effective local sample coverage, whereas larger σ\sigma improves coverage at the cost of increased geometric bias. The shell radius RR introduces an additional conditioning choice. Thus, “a single scale is sufficient” is different from “a suitable single scale can be selected reliably in practice.” A data-driven criterion for choosing σ\sigma (and possibly RR) would make the proposed framework substantially more deployable.

  4. The cause of the gap between score accuracy and geometric accuracy remains unclear.
    One of the most interesting empirical findings is that lower validation normalized-score error can coincide with larger angular-moment error and poor full-geometry recovery. However, the mechanism behind this divergence is not yet well characterized. For example, it is unclear whether training itself fails to preserve geometry-sensitive angular harmonics, whether the score-to-moment inverse amplifies otherwise small score errors, or whether the effect is primarily due to finite sampling, architecture, or optimization. Since this observation is central to the paper's motivation for auditing learned score fields, a more detailed analysis of where the geometric information is lost would considerably strengthen the empirical story.

Questions for authors
  1. How do the authors expect the proposed recovery procedure to scale to high-dimensional learned score fields?
    The theoretical results apply to arbitrary ambient dimension, but the learned-score experiments are limited to low-dimensional synthetic settings. In higher dimensions, both the number of required tangential observations and the complexity of spherical-harmonic reconstruction may increase substantially. Do the authors expect any fundamental statistical or computational bottleneck as DD grows, and is there a practical strategy for applying the method to high-dimensional pretrained score models?

  2. How should the noise scale σ\sigma be selected in practice?
    The paper shows that a single known scale is sufficient for identifiability, but the finite-data analysis suggests a trade-off between tangent-approximation bias at larger σ\sigma and reduced local sample coverage at smaller σ\sigma. Is there a data-driven criterion for selecting σ\sigma (and possibly the shell radius RR) when the local geometry and intrinsic dimension are unknown? In particular, could the proposed calibration or certification diagnostics themselves be used for scale selection?

  3. What causes the observed divergence between score accuracy and geometric accuracy?
    The finding that validation normalized-score error can improve while angular-moment error worsens is particularly interesting. Do the authors have evidence indicating whether this degradation is already present in specific angular harmonics of the learned score field, or whether it mainly arises from error amplification during the score-to-moment inversion? Understanding this distinction would help clarify whether the limitation lies primarily in score learning or in the inverse reconstruction procedure.

  4. How far can the framework be extended beyond finite-ray junctions?
    The weighted tangent-measure formulation is more general than the finite-ray case, but the constructive recovery and strongest finite-noise guarantees focus on discrete branches with relatively regular local geometry. What aspects of the method would remain applicable when the tangent measure has a continuous angular component, when branches have non-negligible thickness, or when strata of different intrinsic dimensions meet?

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

This paper studies whether the score function of a distribution supported on a low-dimensional manifold can recover the local data geometry at non-smooth junctions of the manifold. Specifically, the authors show that the branch center, local homogeneity degree, directions and relative masses can be recovered from the score field. Theory is developed to understand the error and stability of the estimations, and experiments are performed to test the proposed procedures in practice.

Critical strengths
  • The problem of recovering data geometry from learned score functions is important and interesting.
Critical weaknesses

I'm very familiar with score functions and diffusion models and know the fundamentals of measure theory. However, this paper, to me, is very hard to follow. The language is quite abstruse. Many terms are too abstract and vague to understand what they refer to. Quite a few mathematical symbols are not defined the first time they are introduced. Theorems are not presented in a self-contained manner. Below are a few examples:

  • Line 43: "We call the local scaling exponent the homogeneity degree, the local measure dimension in our tangent model." What does the "local scaling exponent" refer to?
  • Line 52: "The score is an uncentered vector field with an unknown reference point..." What does "uncentered" mean here?
  • Line 89: "Homogeneity makes the smoothed conic density an Ornstein–Uhlenbeck eigenfunction;" What do "homogeneity" and the "smoothed conic density" refer to?
  • Line 104: "The error chain exposes tangent bias, local coverage, calibration error, harmonic amplification, network error, and query aliasing." What does the "error chain" refer to?
  • Line 114: "Our scope is local identification from one score slice;" What do "local identification" and "one score slice" mean here?
  • Line 36: σ\sigma, μ\mu and ϕ\phi are undefined.
  • Line 26: σ\sigma is undefined.
  • Lines 86 and 94: DD is undefined.
  • Line 173: What do the "strong and weak identities" refer to respectively?
Questions for authors

Please see weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.