acceptodds
Under review as a conference paper at ICLR 2027

Matched Score Queries for Curvature and Density at Branching Junctions

Abstract

At a junction a score field reveals weighted tangent rays, but these fix only first-order geometry. They say nothing about how a branch bends or how the density changes away from the center: the second-order structure needed to continue a branch past the junction. Small-noise expansions place this curvature and density in the first correction term, yet do not establish whether finite score observations can isolate the per-branch contributions when the center is also estimated. We answer this as an inverse problem, by matched queries at scales and . For a finite union of half-branches in the normalized score obeys ; matched subtraction cancels the leading field rather than estimating it, exposing , which is linear in the branchwise curvature and log-density slope. Given the tangent directions and weights on distinct rays, identifies all branch parameters. Point-coordinate queries attain this sharp scalar-information count, and any fixed continuous scheme that identifies every jet requires at least scalar outputs. This counts information rather than score-network evaluations. A geometry-dependent condition number sets stability as rays approach one another. Under coarse localization and full-rank calibration an center error contributes translation modes, so the augmented problem needs scalar observations except along a translation-invariant full line; we derive a perturbation bound and a conditional kernel-density rate. Across 180 population fits the fitted median exponent 0.998 matches the predicted order; matched-query systems remain full rank to with 16 branches, and a known-count frontend fitted to finite-noise scores composes with the inverse in – (135 population systems, median jet error 0.132); under strong population first-order error, matched responses reduce the median parameter error by a factor of 49.4 relative to naive tangent subtraction. The contribution is an information count: matched subtraction turns a weak cross-scale signal into a well-posed inverse whose second-order branch geometry is identifiable at an (resp. under center error) scalar-information cost.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
score-based models, geometric inference, inverse problems, identifiability, curvature estimation, manifold learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_961xOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses how to recover branch curvature at junctions from score fields beyond the tangent directions. It proposes matched queries at two noise scales to cancel the leading tangent field rather than estimate and provides theoretical results on identifying branch jet, scalar-information complexity and center error. The synthetic experiments evaluate on query conditioning, finite-sample KDE estimates and learned scores.

Critical strengths

This paper focus on providing a comprehensive theoretical analysis of recovering second-order geometry at branching functions from score fields.

Critical weaknesses
  1. The abstract is too long and introduces too many notations and theorem claims without first stating the main idea in the language that is easy enough to understand. Some concepts are not defined either in section 1, including "junction", "branch", "weighted tangent ray", "normalized score", "matched query" and "scalar component evaluation". It is difficult to understand the problem and main contributions without explicitly introducing notation and claims. It is not clear what message the Figure 1 intended to deliver, and "tangent blow-up" is not illustrated either. The "Evidence" paragraph at the end of the introduction is hard to understand because no experimental setting has been introduced.
  2. A major issue is the clarity: many notations are not well-defined in theorems, corollaries, including ρ\rho in equation 4, Q,F0,GQ, F_0, G in equation 15, etc.. Additionally, theorems are not clearly connected to the claims. Relationship between Theorem 3.1 and corollary 3.2 is explicitly explained, and the transition to section 4 is not very clear.
  3. The experimental section lacks experimental settings, evaluation metrics.
Questions for authors
  1. In equation 4, what is ρ\rho? How to understand this equation? What is wjw_j in equation 6?
  2. Notation Q,F0,GQ, F_0, G in theorem 3.1 are unclear in theorem 3.1. QQ is simply called as an compact query set, ad Cm,qC_{m, q} is not clearly defined, including mm.
  3. In theorem 4.2, does such sD always exist?
  4. More terminologies are not defined in section 5: "local calibration", "smooth localized score tests", "bounded reference offset". What's b in equation 24 exactly?
Reviewer_3TYCOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a "matched score query" procedure for recovering second-order geometric information—branch curvature and outward density change—at a branching junction from a score field. The core idea is a two-scale subtraction: querying the score at two bandwidths (σ and λσ) around a common provisional center so that the shared first-order tangent field cancels, leaving a residual that encodes the branch jets. The authors supply a two-term asymptotic expansion, identifiability and perturbation results, and a set of synthetic experiments comparing a population score against a KDE-based score.

Critical strengths

The mathematical development is careful, and the reproducibility statement and experiment-to-claim map are commendable. However, in its current form the paper does not give a reader enough grounding to understand what problem is being solved, why it matters, or whether the proposed method actually works better than alternatives.

Critical weaknesses
  1. Insufficient background and motivation. The paper opens directly inside a fairly narrow technical niche (weighted tangent rays, tangent blow-ups, boundary-layer expansions) without establishing the broader context that would make the problem legible. It is never clearly explained what a "branching junction" corresponds to in practice, what application would consume curvature and density estimates at such a junction, or how this fits within the existing literature on score matching, geometric measure recovery, or rectifiability.

  2. The purpose and use-case of the algorithm are underspecified. Algorithm 1 is stated, but the paper does not describe a realistic end-to-end setting in which it is deployed. It is unclear what the "first-order estimate" it requires is assumed to come from, how the method is meant to be embedded in a learned-score or generative-model workflow, or what decisions the recovered jets would drive.

  3. Experimental design is too limited. The evaluation is almost entirely synthetic: population scores versus KDE scores on hand-designed ray configurations in low dimensions (D = 3–5), reported through medians and proxy errors. There is no comparison against any competing baseline, no real-data experiment, and no learned-score model in the loop despite the theoretical treatment of learned-score error in Section E.3.

Questions for authors

In its present state the contribution is not yet evaluable on the two dimensions that matter—novelty and effectiveness. I would encourage the authors to add a motivating application and literature positioning, an explicit use-case, a baseline comparison, and at least one realistic or learned-score experiment before the significance of the work can be fairly judged.

  1. Could the authors clarify the concrete application context that motivates recovering second-order information (branch curvature and outward density change) at a branching junction?
  2. Could the authors explain where these are assumed to come from in a realistic end-to-end workflow, and how the method would be embedded in a learned-score or generative-model setting?
  3. What is the justification for omitting a comparison against simpler two-scale or finite-difference alternatives, and can the authors supply at least one baseline comparison plus a realistic or learned-score experiment?
Reviewer_Ugv5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies a junction of ss one-dimensional branches in RD\mathbb{R}^D, each with a tangent direction, weight, curvature vector, and log-density slope, smoothed by a Gaussian of scale σ\sigma. It shows that the normalized score Fσ(z)=σsσ(x0+σz)F_\sigma(z) = \sigma s_\sigma(x_0 + \sigma z) expands as F0+σG+O(σ1+α)F_0 + \sigma G + O(\sigma^{1+\alpha}), where F0F_0 depends only on the tangents and weights and GG is linear in the per-branch curvature and density slope ("jets"). Querying at two scales σ\sigma and λσ\lambda\sigma and subtracting cancels F0F_0 and exposes GG. The main results are that GG identifies all jets given the first-order geometry (Thm 4.1), that sDsD scalar evaluations are necessary and sufficient (Thm 4.2), that an O(σ2)O(\sigma^2) center error adds DD translation modes and (s+1)D(s+1)D evaluations suffice except for a translation-invariant line (Thm 5.3), and a least-squares perturbation bound with a KDE rate of N−α/(2α+3)N^{-\alpha/(2\alpha+3)} (Thm 6.1, Cor 6.2). Experiments are synthetic: population scores by quadrature, Gaussian KDE scores, and a small learned score on planar junctions.

Critical strengths
  • The theorems are stated carefully and the proofs, which I checked for Thms 3.1, 4.1, 5.2, 5.3, 6.1 and Cor 6.2, are correct.
  • The experiments are organized as one test per claim (App. F.1), with configs and seeds in the supplement; the appendix tables report the unfavorable cases rather than hiding them.
  • The two-scale cancellation is a clean idea, and the translation-mode analysis in §5 is a nice touch.
Critical weaknesses
  1. The theory is elementary relative to its framing. The forward expansion is already available (the paper credits Brosse & Dalalyan 2026 with score, log-Hessian, and scale-derivative expansions, L069–L071, L454–L458). The matched difference is a one-sided finite difference in σ\sigma, as the paper notes at L210 (G=∂σ[σsσ]G = \partial_\sigma[\sigma s_\sigma] at 0+0^+). Thm 4.1 is Gaussian-convolution injectivity plus localized test functions; Thm 4.2 is a parameter count by invariance of domain. Calling the latter "sharp scalar-information complexity" oversells a dimension count. Prop 5.1, on which the (s+1)D(s+1)D claim rests, is only sketched (App. D.1): the "localized score tests" and the "exact tangent identity" are never stated.

  2. With estimated scores the method does not work at the tested sizes, and the abstract reports the population numbers. Table 1: population jet error 0.124–0.142 in D=3D = 3–55, KDE jet error 0.927–0.975 at N=131,072N = 131{,}072; Table 6: KDE 1.77–1.80, i.e., worse than predicting zero; Table 7: learned scores 0.34–0.97. The abstract states "median jet error 0.132" and "finite-noise scores," where finite noise means finite σ\sigma with the exact population score, not finite samples.

  3. Matching is not better than naive subtraction where it matters. Table 9: the factor 49.4 is population + strong first-order perturbation. For KDE scores at exact or mild perturbation the naive/matched ratio is 0.863/0.893, so naive subtraction is better; matching wins only at strong perturbation (1.96).

  4. Scale-up numbers are not usable recovery. Table 10: at D=10D = 10 and 2020 under "strong" perturbation (tangent RMS ≈3.4∘\approx 3.4^\circ) the relative error is 79–86 with condition numbers ∼104\sim 10^4. "Remain full rank to D=20D = 20 with 16 branches" (abstract) is true and uninformative. The N−1/5N^{-1/5} rate is verified only on planar junctions (Fig. 5), while Eq. 34 has no DD-dependence and the KDE error jumps from ∼0.25\sim 0.25 (D=2D = 2) to ∼1.8\sim 1.8 (D=3D = 3–55) at the same NN, which is not explained.

  5. Assumptions versus motivation. The paper is motivated by score-based generative models on data, but requires zero-thickness one-dimensional branches, exact Gaussian smoothing, known tangents and weights, a center known within O(σ)O(\sigma), and branch correspondence across scales. There is no real data, no pretrained diffusion model, and no downstream use of a recovered jet. The limitations paragraph (L465–L471) acknowledges this; the question is whether an identifiability note on this model is an ICLR contribution.

  6. Related work. The filament and metric-graph literature, where curvature and density of one-dimensional structures are recovered from smoothed densities, is not discussed: ridge estimation (Genovese, Perone-Pacifico, Verdinelli, Wasserman 2014; Chen, Genovese, Wasserman 2015), metric graph reconstruction (Aanjaneya et al. 2012; Lecci, Rinaldo, Wasserman 2014; Chazal and Sun 2015). The mean-shift identity σ2sσ\sigma^2 s_\sigma would also place GG as a scale-space derivative of the mean-shift vector in one sentence.

  7. Presentation. The abstract is about 330 words and lists every number in the paper. Code names leak into the text ("E4 diagnostic," "paper configuration"). "Remainder exponent 0.998" (L375) refers to the slope of the parameter error in σ\sigma, not to the O(σ1+α)O(\sigma^{1+\alpha}) remainder.

Questions for authors
  1. State Prop 5.1 precisely: what are the localized score tests, what is the exact tangent identity, and how is the O(σ)O(\sigma) coarse localization obtained in practice?
  2. Put the KDE and learned-score jet errors in the abstract. At what NN does the KDE pipeline reach, say, 20% relative error in D=3D = 3–55?
  3. Why does the KDE error jump from ∼0.25\sim 0.25 in D=2D = 2 to ∼1.8\sim 1.8 in D=3D = 3–55 at the same NN when Eq. 34 has no DD-dependence?
  4. Table 9: when should a practitioner match scales rather than subtract a fitted tangent field, given that naive subtraction is better for KDE scores at mild perturbation?
  5. What happens for branches of positive thickness ε\varepsilon, or for two-dimensional strata meeting along a curve, when ε≈σ\varepsilon \approx \sigma?
  6. Is there any evidence on a pretrained score model, even a toy one with a known branching structure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.