acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Selective Distribution Comparison via Normal Log-Density Taylor Signatures

Abstract

MMD and KSD quantify distributional discrepancy but do not explicitly distinguish changes along local tangent and normal directions. We propose the Normal Log-Density Taylor Signature (NLTS), which uses a reference distribution to define local normal directions and compares reference log-density responses across samples from and . Under a local manifold model with small normal noise, we show that the negative-log-density Hessian identifies the local normal subspace, while, for spectrally isolated normal directions, projected scores provide the corresponding first-order Taylor coefficients. Synthetic experiments verify the intended normal–tangent selectivity. On CIFAR-10-C, NLTS remains effective with learned scores, while in single-cell perturbations its power is associated with geometric changes characterized separately by local PCA. Together, these results support the use of reference-defined local density geometry for selective high-dimensional distribution comparison. Code will be made publicly available.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
two-sample testing, nonparametric hypothesis testing, high-dimensional statistics, manifold geometry, distribution comparison

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AsQ1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper propose Normal Log-Density Taylor Signatures (NLTS) for comparing a reference distribution PP with another distribution QQ. Each observation is represented through the reference score by selected directions of the negative-log-density Hessian and their associated curvature values and finally the remaining score energy. The resulting are compared using MMD. The advantage is sensitivity to changes in normal geometry, together with a weaker response to redistribution along the data manifold. The paper provides a local theoretical interpretation and experiments on synthetic distributions, CIFAR-10-C.

Critical strengths

The distinction between different types of distributional change is very intesting, intuitive then well motivated! Using geometry to determine which changes should matter is a useful direction for distribution comparison particularly in high dimensional : measure concentration is at my knowledge not enough considered in litterature and I appreciate this concern. The construction is clear, understandable, and the Taylor expansion provides a good interpretation when the reference score and Hessian are exact. The synthetic evaluation includes both target changes and tangential controls, which is natural for the stated objective. The author are also honest about several important limitations. It distinguishes pointwise conditional invariance from invariance of the overall signature distribution, discusses repeated eigenvalues, and presents the real-data experiments conclusions as indirect. The CIFAR-10-C results show very promising sensitivity to corruptions.

Critical weaknesses
  1. It is unclear why two (or any arbitrary) directions are needed, or what they add compared with simpler calculations Indeed, if the data spread equally in every normal direction, can looking at one normal direction already detect this change ? The paper compares two directions with four, but may report the results with one direction. More importantly, the method could be compared with a test using only the squared length of the score, ∥sP(x)∥2\|s_P(x)\|^2. This quantity already measures how strongly the reference log density changes near an observation. The full method contains this information because
∥sP(x)∥2=a1(x)2+a2(x)2+r2(x).\|s_P(x)\|^2=a_1(x)^2+a_2(x)^2+r_2(x).

Comparing the full method with this simple quantity, and with r2r_2 alone, would show whether calculating the two directions provides a improvement and strengthen the contribution.

  1. The two selected directions may change substantially without a meaningful change in the data. In the affine Gaussian example, all the normal directions have the same eigenvalue. There is therefore no uniquely pair of directions which can be chosen. Moreover, different initializations or small numerical errors can produce different choices. Squaring the projections removes dependence on the sign of a direction is good but it does not remove dependence on which directions are selected. But I believe that this does not automatically mean that the final test is unstable. but the authours should explicit that : whether changing the initialization or the number of iterations changes the calculated numbers and the final decisions. I also recognize that the authors discuss the ambiguity in Appendix A.4.3, but its practical effect remains unclear to me.

  2. The calculation used on CIFAR differs from the calculation studied in the theory. The theory uses a Hessian, which is a symmetric matrix of second derivatives. On CIFAR, the authors instead approximate derivatives so the resulting operator is not necessarily symmetric. Appendix F.2 explains that the algorithm starts with random directions and updates them through eight iterations. The final directions are therefore not simply random but author should show that these eight iterations are sufficient or that the results are stable across initializations.

  3. The results using only eigenvalues in Table 10 need more clarification : for the affine Gaussian example described in Appendix D.1, the Hessian is the same at every observation so its eigenvalues are therefore also the same. A method using only these eigenvalues should have no information for distinguishing the two distributions. But Table 10 reports detection power of 0.98 for Local artifact and 0.78 for Single-dimension expansion using only eigenvalues. If these experiments use the stated affine Gaussian, the results are difficult to understand. Could the authors specify the distributions and numerical calculations used for these rows ?

  4. The experiments do not establish how reliably the method distinguishes the intended types of change. If we suppose that the reference data are concentrated near a curve due to measure concentration with a thin spread around one part and a thicker spread around another then increasing the proportion of observations in the thicker region changes the distribution of the calculated numbers, even if the spread at every location remains unchanged. Any experiment of this kind would help determine when the method responds to redistribution along the curve. The image experiments demonstrate detection of corruptions, but provide poor evidence about changes the method should ignore. and the paper should explain more precisely why this redistribution is considered a change to detect.

If the author carefully clarify all my weaknesses, I'm willing to raise my score as it is an very interesting idea.

Minor : the font of the title/sections seems different. I'm unsure about its importance but anyway, could the authors correct that in order to avoid any desk reject because of the ICLR guideline?

Questions for authors

See weaknesses.

7 comments

Q1. Why two directions, and what do they add beyond simpler score-based statistics?

We agree that k=1k=1 may suffice for isotropic changes. However, real-world shifts can be anisotropic, motivating k=2k=2 as the smallest multi-directional signature. To test this choice, we conducted two additional ablations.

(1) Directional ablation (controlled synthetic). We used Gaussian distributions with 8 tangent dimensions and 10 or 66 normal dimensions (D=18D=18 or 7474), with the first two normal directions given by the exact reference Hessian. We applied controlled variance-scaling shifts (not dimension expansion) and compared k=1k=1 and k=2k=2. The table reports empirical power averaged over n∈{16,32,64,128}n\in\{16,32,64,128\} and both dimensional settings (200 trials per setting, α=0.05\alpha=0.05).

Distribution shiftk=1k=1k=2k=2
Second normal direction only (SD ×1.55\times 1.55)22.44%58.88%
Uniform normal scaling (all normal SDs ×1.06\times 1.06)47.50%45.00%
First normal direction only (SD ×1.55\times 1.55)59.81%57.94%
Tangent-only scaling (SD ×1.35\times 1.35)4.75%4.69%

While both perform comparably on uniform normal and first-direction shifts, k=2k=2 improves second-direction detection by 36.44 percentage points, demonstrating the benefit of isolating directional information otherwise pooled in the residual.

(2) Score norm and residual-only comparison (CIFAR-10-C). As the reviewer notes, $ |s_P(x)|^2=a_1(x)^2+a_2(x)^2+r_2(x). $ NLTS retains this score energy while additionally representing its Hessian-aligned directional decomposition and curvature. We compared these representations across 342 matched settings (19 corruptions ×\times 3 severities ×\times 6 sample sizes):

MethodDetection power
NLTS (k=1k=1)94.14%
NLTS (k=2k=2)93.59%
Score norm squared (∥sP(x)∥2\lVert s_P(x)\rVert^2)82.60%
Residual-only (r2(x)r_2(x))82.61%

Both NLTS variants outperform the scalar baselines by approximately 11 percentage points, demonstrating the benefit of the structured score-curvature signature.

Together, these results support the full signature over simpler score-based statistics and k=2k=2 for anisotropic sensitivity, without claiming that two directions are universally necessary. We will incorporate both ablations and clarify the design choice.


Q2. Does eigenvector ambiguity compromise test stability?

We agree that squared projections do not resolve basis ambiguity, but as the reviewer suggests, this need not imply test instability.

To assess practical stability, we used the affine Gaussian synthetic model (D=256D=256, 8 tangent and 248 isotropic normal dimensions; normal standard deviation σn=0.05\sigma_n=0.05), with analytical scores and Hessians. We varied eigensolver initialization (6 random starts) and iteration count (T∈4,8,16T\in{4,8,16}) across 30 independent datasets (k=2k=2).

Distribution shiftDecision agreement across random startsDecision agreement across iteration countsMean ∣ΔMMD2∣\lvert\Delta\mathrm{MMD}^2\rvert across random starts
Null (P=QP=Q)90.4%100%0.0026
Isotropic scale (1.02$\times$)86.7%100%0.0039
Isotropic scale (1.50$\times$)100%100%0.0524
Anisotropic scale (4$\times$, one axis)92.0%100%0.0049

Test decisions were identical across all iteration counts, while initialization agreement reached 100% for the clear isotropic shift. Although weaker shifts showed some initialization sensitivity, the results confirm that eigenvector ambiguity does not necessarily compromise test-level stability. We will clarify this distinction and report the new results.

Q3. Nonsymmetric CIFAR operator and numerical stability

CIFAR's finite-difference (FD) operator is not guaranteed to be symmetric. We evaluated test stability across power iterations (T∈{4,8,16,32}T\in\{4,8,16,32\}) and three initialization seeds on CIFAR-10-C.

IterationsGaussian noiseJPEG compressionBrightness
4100% (100%)97.2% (94.4%)27.8% (77.8%)
8 (default)100% (100%)100% (100%)36.1% (77.8%)
16100% (100%)100% (100%)30.6% (77.8%)
32100% (100%)100% (100%)27.8% (77.8%)

Detection rate (decision agreement across initializations); severity 1, n=32n=32, 12 tests/seed, α=0.05\alpha=0.05.

At the default T=8T=8, Gaussian noise and JPEG achieved 100% detection and initialization agreement, maintained at T=16,32T=16,32. Brightness remained challenging (36.1% detection; 19.4% null rejection).

Importantly, in practice, FD achieved substantially faster NLTS feature extraction (approximately 4.5$\times$ under our benchmark settings) without loss in detection performance relative to the symmetrized operator for Gaussian noise and JPEG compression (100% for both). See our response to Reviewer kbGz, Q1 for the detailed comparison.

Together, these findings demonstrate stable test decisions for the strongly detected shifts without requiring a symmetric operator, while not implying exact eigendirection recovery or universal stability. We will clarify these empirical findings and the scope of the theoretical guarantees in the revision.


Q4. Inconsistent eigenvalue-only results in Table 10

The reviewer is correct: the affine Gaussian Hessian is constant, so its eigenvalues cannot distinguish distributions.

Our investigation identified an assembly error: results from a different method explored in our subsequent research were inadvertently included in Table 10. We found four affected entries, including two beyond those identified by the reviewer. We sincerely apologize for this error, which overstated eigenvalue-only performance.

The corrected results (Cohen's dd / empirical power) are:

ScenarioReportedCorrected
Dimension-mixture shift1.02 / 0.480.00 / 0.000
Local artifact2.36 / 0.980.02 / 0.06
Curved local artifact3.29 / 0.950.48 / 0.10
Single-dimension expansion1.28 / 0.780.01 / 0.03

We recomputed the eigenvalue-only test using MMD on the top two absolute eigenvalues of the fixed-reference Hessian and verified their constancy in the affine Gaussian setting. An audit of the remaining reported numerical results found no further discrepancies.

These corrections do not affect the NLTS results or the paper's main conclusions; rather, they clarify the limitations of eigenvalue magnitudes alone. We will correct Table 10 accordingly.

Q5. What changes should NLTS detect, and when should it ignore tangential redistribution?

We thank the reviewer for the concrete example, which helps clarify the intended scope of NLTS.

Motivation and detection target. Our goal is to detect changes in off-manifold variation across the population. Such changes can arise in two ways:

  • Case 1: The conditional distribution of normal displacements changes at a manifold location, ρQ(ξ∣m)≠ρP(ξ∣m)\rho_Q(\xi\mid m)\neq\rho_P(\xi\mid m).
  • Case 2: The conditional distributions remain unchanged, but the sampling frequencies of manifold locations change, qQ(m)≠qP(m)q_Q(m)\neq q_P(m).

The reviewer's example is Case 2. When more samples come from regions with greater normal thickness, the population contains more samples with large off-manifold deviations, even though local normal variation remains unchanged. For applications monitoring departures from the reference manifold, this change in the population is important to detect.

Conversely, NLTS aims to suppress responses to tangential redistribution that leaves population-level normal variation unchanged, as examined by the six tangential controls in Fig. 3(b) and Appendix D.4.

Already established in the paper. Proposition 3 shows that tangential reweighting preserves the conditional normal distribution at each location. Importantly, we already state immediately after Proposition 3 that this does not imply invariance of the population-level signature distribution (Case 2). Eq. (18) gives the precise condition: NLTS has a positive population discrepancy when the reference-defined signature distributions differ.

Additional experiment setting. Following the reviewer's suggestion, we constructed a curved manifold with location-dependent normal thickness. We varied the sampling distribution along the manifold while keeping the conditional normal distributions unchanged. For manifold location θ∈[−π,π)\theta\in[-\pi,\pi), we used qP(θ)=12πq_P(\theta)=\frac{1}{2\pi} and qQ(θ)=1+βcos⁡θ2π.q_Q(\theta)=\frac{1+\beta\cos\theta}{2\pi}. Normal thickness was greatest near θ=0\theta=0, so increasing β\beta increased the proportion of samples from thicker regions. We also tested two tangential controls: uniform normal thickness, and heterogeneous thickness with reweighting proportional to 1+βsin⁡θ1+\beta\sin\theta, which preserves the population distribution of normal thickness by symmetry.

Two diagnostic checks were included: a conditional-normal oracle using known manifold coordinates and normalized normal displacements, and a tangent-location control using (cos⁡θ,sin⁡θ)(\cos\theta,\sin\theta). The former checks responses to local normal changes; the latter verifies redistribution along the manifold. Neither is a competing baseline. We used analytical scores and Hessians (D=256D=256, k=2k=2, nP=nQ=128n_P=n_Q=128; 300 trials; permutation tests at α=0.05\alpha=0.05). The main reweighting experiments used β=0.85\beta=0.85.

Distribution changeNLTSConditional-normal oracleTangent-location control
No change (PP vs. PP)4.7%4.0%5.7%
Reweighting, uniform thickness6.0%5.7%99.3%
Reweighting, heterogeneous but symmetry-matched thickness5.3%5.0%99.3%
Reweighting toward thicker regions99.0%6.0%99.3%
Increased normal thickness (1.8$\times$)99.3%100%5.3%

Empirical rejection rates; significance level 0.05.

Results: NLTS detects the intended population-level change. NLTS remains near the nominal rejection level for uniform-thickness and symmetry-matched reweighting (6.0% and 5.3%), but detects the reviewer's proposed redistribution toward thicker regions (99.0%). The conditional-normal oracle remains near nominal (6.0%), consistent with the unchanged local conditional distributions. NLTS also detects actual normal-thickness changes (99.3%).

We further varied the redistribution strength β∈{0,0.2,0.4,0.6,0.85}\beta\in\{0,0.2,0.4,0.6,0.85\}. NLTS rejection increased from 5.5% to 14%, 50%, 91%, and 100%, respectively (200 trials per setting), while the conditional-normal oracle remained near nominal (3.5%–8.5%).

Revision to the Introduction. We will make the detection target explicit by adding the following explanation:

Our goal is to detect changes in off-manifold variation across the population. Such changes can arise either from changes in local conditional normal distributions or from changes in the frequency of sampling regions with different normal thickness. In the latter case, the population-level prevalence of off-manifold deviations can change even when local conditional distributions remain unchanged. NLTS aims to capture both types of change while limiting sensitivity to tangential redistribution that leaves population-level off-manifold variation unchanged.

Reviewer_kbGzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the Normal Log-Density Taylor Signature (NLTS) for comparing high-dimensional distributions. The authors aim to distinguish distributional changes along tangent and normal directions. The method uses the reference score and negative-log-density Hessian to construct local geometric features. Propositions 1 and 2 establish the relationship between the dominant Hessian eigenspace and local normal directions under a small-noise manifold model. The authors then use maximum mean discrepancy (MMD) to compare the resulting signatures. Numerical results show that NLTS performs well for several distributional changes compared with existing methods.

Critical strengths

The distinction between tangent and normal changes is interesting. The authors use local density geometry to study this problem. Propositions 1 and 2 in Subsection 3.1 (p. 4) provide some theoretical support for the proposed method. I also find the combination of first- and second-order information useful.

The numerical results are good. In Subsection 4.1 (pp. 6-7), NLTS achieves high detection power for the selected normal changes while maintaining low rejection rates for tangential changes. The CIFAR-10-C results in Figures 4 and 5 (p. 8) are also promising. The method performs well across different corruption types. The authors also include several additional experiments in the appendix.

Critical weaknesses

I have two main concerns.

The first concern is about the theoretical results. Propositions 1 and 2 (p. 4) require a small-noise manifold model and suitable eigenvalue separation. These conditions may not hold in real applications. In Appendix F.2 (pp. 32-34), the authors use an approximate operator that is not necessarily symmetric. However, the theoretical results are based on a symmetric Hessian. I am not sure whether the resulting directions still represent the normal space. This makes it difficult to understand how the theoretical results support the practical implementation.

The second concern is about the numerical experiments. The synthetic experiments are well designed, but the true tangent and normal directions are known in these settings. For CIFAR-10-C and the single-cell data, these directions are unknown. Thus, good detection performance does not necessarily mean that the method is detecting changes in normal directions. I am also not sure whether the Hessian-based construction is necessary for the reported improvements. In Eq. (16) (p. 5), the proposed signature contains projected-score energies, Hessian eigenvalues, and residual score energy. Appendix E.2 (pp. 30-31) compares the first two components, but does not separately study the residual score energy.

Questions for authors

My questions are related to the two concerns above.

(1) Propositions 1 and 2 (p. 4) assume an exact reference Hessian with suitable eigenvalue separation. However, Appendix F.2 uses an approximate operator that is not necessarily symmetric. How does this affect the estimated normal directions? Can the authors provide numerical results comparing the estimated and true normal directions when the reference score is learned?

(2) In Eq. (16) (p. 5), the NLTS signature includes residual score energy. What is the contribution of this component? Can the authors compare NLTS with a simpler method based on the score norm or score features alone? I would like to understand whether the Hessian-based decomposition is needed to obtain the reported improvements.

(3) The CIFAR-10-C results in Figures 4 and 5 (p. 8) are quite good. However, the true tangent and normal directions are unknown. How can we tell whether the improvement really comes from normal-direction sensitivity or other distributional differences? I also wonder about the computational cost of estimating the Hessian directions. Can the authors provide runtime comparisons with the baseline methods?

2 comments

(The experiment for Q1 is still ongoing, so we will address Q2 and Q3 first.)


Response to Q2: Contribution of the residual and Hessian components.

To isolate the contributions of NLTS components, we performed matched ablations on CIFAR-10-C across 342 settings (19 corruptions × 3 severities × 6 sample sizes), using multiscale RBF-MMD throughout.

FeatureMean rejection rate
Full NLTS0.936
Score norm squared0.826
Residual energy only0.826
Without Hessian eigenvalues0.793
Without residual energy0.874

Full NLTS improves upon the score-norm baseline by 11.0 percentage points. More importantly, removing residual energy reduces rejection by 6.2 points, while removing Hessian eigenvalues reduces it by 14.3 points. Although residual energy alone performs similarly to the score norm, it contributes when combined with the other NLTS components.

These results demonstrate complementary contributions from the Hessian eigenvalues and residual score energy, providing discriminative information beyond the scalar score norm.


Response to Q3: Geometric interpretation and computational cost.

As discussed in the paper, CIFAR-10-C corruptions introduce variations absent from the clean reference distribution, which we expect to displace samples away from the clean-data manifold, with substantial normal-direction components. Our synthetic experiments validate tangent–normal selectivity, while our CIFAR-10-C experiments demonstrate the practical utility of NLTS in real-world distribution shifts.

Regarding computational cost, our A100 benchmark shows 80.9 ms per image for NLTS feature extraction (k=2k=2, 8 power iterations, batch size 8), compared with 1.29 ms for score-only evaluation. While NLTS incurs additional computational overhead, our focus is on geometry-selective distribution comparison rather than computational efficiency.

Response to Q1: Learned-score recovery and approximate Hessian operators

(1) CIFAR-10-C

We directly evaluated the effect of operator approximation on CIFAR-10-C. We compared the original finite-difference (FD) operator, with and without clipping, against the symmetrized alternative,

Hsym=−12(Js^+Js^⊤).H_{\mathrm{sym}}=-\frac12(J_{\hat{s}}+J_{\hat{s}}^\top).

Using matched samples (severity 1, n=32n=32, k=2k=2, T=8T=8, 3 seeds, 12 tests/seed, α=0.05\alpha=0.05), we obtained:

CorruptionFDSymmetrizedMin. principal cosine
Gaussian noise100%100%0.519
JPEG100%100%0.740
Brightness36.1%33.3%0.695

Detection rates and median minimum principal cosines.

Despite differences in estimated directions, both operators achieved 100% detection for Gaussian noise and JPEG compression. For brightness, FD slightly outperformed symmetrization (36.1% vs. 33.3%); however, given the low detection rates for both, we regard their performance as comparable rather than evidence of an FD advantage. See also our response to Reviewer AsQ1, Q3 for initialization and iteration stability.

Importantly, FD offers a substantial computational advantage by avoiding the reverse-mode differentiation required by symmetrization. On an NVIDIA A100 (k=2k=2, T=8T=8), we measured:

DatasetFDSymmetrizedSpeedup
CIFAR-10 clean1.99 s9.04 s4.53$\times$
CIFAR-10-C JPEG2.03 s9.21 s4.55$\times$

NLTS feature extraction for 8 images; median of three runs. FD batch size 4; symmetrized VJP microbatch size 1.

(2) Synthetic data

We verified recovery against ground-truth normal directions using a synthetic Gaussian tube (D=12D=12: 8 tangent and 4 normal dimensions). A linear score model trained via denoising score matching yielded:

Training samplesMean minimum principal cosine
5120.951
2,0480.981
8,1920.995

These results demonstrate accurate normal-space recovery from learned scores under controlled conditions. Together with the CIFAR experiments, they support the practical value of FD-based NLTS: stable detection for strongly detected shifts at substantially lower computational cost, while the theoretical normal-direction guarantee remains conditional on the exact-Hessian assumptions. We will clarify these distinctions in the revision.

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

This paper proposes the Normal Log-Density Taylor Signature (NLTS), a fixed-reference framework for comparing probability distributions PP and QQ on RD\mathbb{R}^D while explicitly emphasizing changes along local normal (off-manifold) directions. Assuming PP concentrates near a smooth dd-dimensional manifold, NLTS leverages the ambient reference score sP(x)s_P(x) and negative-log-density Hessian HP(x)H_P(x). Under a small-noise local model, the dominant spectral subspace of HP(x)H_P(x) recovers the local normal space, while projected scores along spectrally isolated directions encode leading first-order Taylor coefficients.For a sample xx, NLTS constructs a (2k+1)(2k+1)-dimensional signature consisting of top-kk Hessian eigenvalues λj(x)\lambda_j(x), squared projected scores aj(x)2a_j(x)^2, and the residual unselected score energy rk(x)r_k(x). The distributions of these signatures under PP and QQ (standardized using reference statistics) are compared via multiscale RBF-MMD. Across synthetic benchmarks, NLTS demonstrates strong sensitivity to normal-geometry changes while maintaining low response to tangential perturbations. Empirical evaluations on CIFAR-10-C corruptions and single-cell RNA-seq drug perturbations further show competitive shift detection and alignment with local-PCA geometric descriptors.

Critical strengths

Principled Geometry-Selective Theoretical Foundation: The paper offers a formal connection between ambient differential quantities (score and negative-log-density Hessian) and local normal/tangent coordinate manifolds. Propositions 1 and 2 establish that dominant Hessian eigenspaces track local normal spaces and that score projections yield first-order Taylor coefficients under small noise, providing a clear justification for directional selectivity.

Critical weaknesses

Computational Bottleneck Retaining top-kk spectral components at every evaluation point requires running iterative matrix-free eigensolvers on Hessian-vector or Jacobian-vector products. For large pretrained score models, performing multiple forward/backward passes per sample incurs substantial computational overhead. The paper lacks a concrete empirical breakdown of query-time latency, memory footprint, or wall-clock time comparison against baselines.

Questions for authors

None

2 comments

We thank the reviewer for the positive assessment of our theoretical foundation and the constructive feedback on computational efficiency.

Computational overhead. We acknowledge the additional cost of estimating local Hessian directions. In our A100 GPU benchmark with the CIFAR-10 DDPM, NLTS feature extraction takes approximately 80.9 ms per image on k=2, 8 power iterations, batch size 8).

We will include this measurement and explicitly discuss the computational trade-off in the revised paper. Our focus is on geometry-selective distribution comparison rather than computational efficiency.

Thank you for your response. I retain my positive assessment of this work.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

We have uploaded a revised manuscript incorporating the rebuttal feedback and additional experiments. The revision clarifies NLTS's geometric claims and strengthens the empirical evidence. All changes are highlighted in blue for ease of review.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.