Pith. sign in

REVIEW 4 major objections 5 minor 17 references

Principal Curvatures Estimation with Applications to Single Cell Data

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read AdaL-PCA estimates principal curvatures of point-cloud surfaces by choosing a per-point scale automatically, then uses those curvatures to identify branch points and differentiation directions in single-cell data.

desk verdict Adaptive scale selection is a real idea, but the paper doesn't prove it isn't just picking the search boundary; the torus numbers are strong, the ellipsoid numbers are not, and the absence of any sensitivity analysis makes the central claim unsubstantiated. read the letter →

arxiv 2502.03750 v1 pith:2J3IYCOL submitted 2025-02-06 cs.LG cs.AI

Yanlei Zhang , Lydia Mezrag , Xingzhi Sun , Charles Xu , Kincaid Macdonald , Dhananjay Bhaskar , Smita Krishnaswamy , Guy Wolf
show 1 more author
Bastian Rieck
This is my paper · ORCID
classification cs.LGcs.AI
keywords principalcurvaturesGaussiancurvaturelocalPCAexplainedvarianceratiomanifoldlearningsingle-cellRNAsequencingPHATEpointcloudgeometry
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that principal curvatures of a point-cloud surface can be recovered by adapting the neighborhood scale locally through the explained-variance ratio of a local PCA: the curvature scale is read off from the radius at which two principal components fail to explain the local variance. On sampled torus, ellipsoid, and hyperbolic paraboloid surfaces, the resulting estimator tracks Gaussian and mean curvature and keeps its accuracy as noise is added. Because the scale is chosen per point, the method removes the manual tuning that fixed-scale curvature estimators require. The paper applies the same machinery to single-cell RNA-sequencing data, where curvature values and principal directions mark the branching of cell lineages and point toward the states into which cells differentiate.

What carries the argument

The machinery is the explained-variance ratio of a local PCA: $\rho(r) = (\sigma_1(r)^2 + \sigma_2(r)^2) / (\sigma_1(r)^2 + \sigma_2(r)^2 + \sigma_3(r)^2)$. For a surface in $\mathbb{R}^3$, the first two singular vectors span the tangent plane and the third gives the normal, so $\rho(r)$ measures how much of the local variance the tangent plane explains. The algorithm reads two scales from this one curve: the largest radius with $\rho(r)$ above a threshold $\gamma$ sets the tangent-plane neighborhood $\epsilon_{PCA}$, while $\tau = \arg\min_r \rho(r)$ is declared the curvature scale at which the surface bends away from the tangent plane most strongly. Directional curvatures computed in that $\tau$-neighborhood are sorted and extremal averages become the principal curvatures $\kappa_1, \kappa_2$, with Gaussian curvature $K = \kappa_1 \kappa_2$ and mean curvature $H = \kappa_1 + \kappa_2$.

What would settle it

On a known surface with fixed curvature, generate point clouds at two very different sampling densities and compare the $\tau$ values chosen by $\arg\min_r \rho(r)$; if $\tau$ changes with density rather than staying tied to the bending scale, the scale-selection mechanism is not tracking curvature.

Watch

Extended reading notes

Core claim

AdaL-PCA treats curvature estimation as a two-scale local PCA problem. For each point, a first radius $\epsilon_{PCA}$ is chosen as the largest neighborhood whose first two singular vectors explain a fraction $\gamma$ of the variance, fixing the tangent plane; a second radius $\tau$ is the minimizer of the same explained-variance ratio, marking where the surface bends away from that plane. In the $\tau$-neighborhood, directional curvatures are approximated by $\kappa_q = 2 (O_2 \cdot v_q) / \|v_q\|^2$, where $v_q$ is a neighbor vector and $O_2$ is the normal from the local PCA frame; the largest and smallest directional curvatures are kernel-averaged to give $\kappa_1$ and $\kappa_2$. The paper claims this adaptive procedure recovers Gaussian and mean curvature more faithfully than fixed-parameter baselines and that the resulting signed curvature and principal directions expose branching structure in single-cell differentiation.

Load-bearing premise

The paper assumes the radius at which the explained-variance ratio bottoms out is the scale where the surface actually bends; if that minimum is caused by noise or sampling density instead, the curvature estimates are biased.

Editorial extensions

If this is right

  • On the torus, ellipsoid, and hyperbolic paraboloid, AdaL-PCA's Gaussian-curvature estimates have lower RMSE than the fixed-parameter baseline [10] at every listed noise level, and higher Pearson correlation than the diffusion-curvature baseline [9] on the torus and hyperbolic paraboloid.
  • Because both neighborhood radii are selected per point from the explained-variance curve, the method carries no hand-set scale; the paper demonstrates this on surfaces with different curvature profiles and under additive Gaussian noise.
  • In the embryoid-body single-cell data, the method assigns near-zero curvature to the stem-cell stage and pronounced curvature changes at the stage where endoderm and ectoderm lineages separate, so curvature values can serve as a differentiation marker.
  • The principal directions recovered alongside the curvatures point along the lineage divergence, giving not just the location but the orientation of cell-fate decisions; the paper shows the same branching structure in an iPSC reprogramming dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same explained-variance-minimizing scale could be used to choose neighborhoods for other geometric quantities, such as normal variation or scalar curvature, but the paper does not test these extensions.
  • Editorial inference: since the method is applied after a PHATE embedding, the numerical curvatures depend on whether that embedding preserves distances; the paper does not quantify how metric distortion in the embedding changes the curvature magnitudes.
  • Editorial inference: the averaging of the top and bottom 20% of directional curvatures is a heuristic; using different percentiles or robust averaging would change the estimates, which is a testable modification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes AdaL-PCA, a method for estimating principal, Gaussian, and mean curvature of a point cloud sampled from a two-dimensional surface. The method first estimates a tangent plane by weighted local PCA, with the neighborhood radius ϵ_PCA chosen as the largest radius for which the explained variance ratio ρ(r) exceeds a threshold ρ0. It then selects a second, larger radius τ as the minimizer of ρ(r), and computes directional curvatures of the surface via the normal component of neighbor displacement vectors, averaging the extreme directional curvatures to approximate principal curvatures. The method is tested on torus, ellipsoid, and hyperbolic-paraboloid point clouds with additive Gaussian noise and compared against two existing curvature-estimation methods. It is also applied to single-cell RNA-seq and mass cytometry data embedded with PHATE, where the estimated Gaussian curvature and principal directions are interpreted in terms of cellular differentiation.

Significance. The paper provides a concrete algorithm with an open-source implementation and validates it against ground truth on synthetic surfaces, which is a strength. At zero noise, the torus and ellipsoid results (Pearson 0.996 and 0.988) are impressive. However, the adaptive scale-selection heuristic at the core of the method is not theoretically justified, and the method's performance degrades sharply on the ellipsoid with modest noise (Pearson 0.325 at σ=0.1). The significance of the contribution is therefore conditional on resolving these issues; if the heuristic is replaced or validated, the method could be a useful parameter-light tool for curvature estimation in low-dimensional embeddings.

major comments (4)
  1. [II-B, Eq. (3), Algorithm 1] The selection τ = argmin_r ρ(r) is the load-bearing step for the claimed adaptivity, but the manuscript provides no argument that the minimum of ρ corresponds to the curvature scale, and the standard small-scale behavior of ρ contradicts it. For a C^2 surface written as a height function h(u) = ½ u^T II u + O(‖u‖^3) over the tangent plane, the tangent-plane variance grows as c r^2 while the normal variance grows as C(κ1²+κ2²) r^4, so ρ(r) = 1 − C'(κ1²+κ2²) r² + O(r³) is strictly decreasing for small r. Hence the argmin in the search range (0, 0.2δ] is generically the boundary r = 0.2δ, not an interior curvature scale. Moreover, since ϵ_PCA is defined as the largest r with ρ(r) > ρ0, a decreasing ρ implies τ > ϵ_PCA, and the normal vector obtained from the small-scale PCA is then used at points q with ‖q−p‖ < τ, outside the region where the O(t) error in Eq. (4) is controlled. The paper needs either a theoretical justification that the argmin identifies the bending scale or an empirical check that the selected τ is an interior point for the reported datasets.
  2. [III-A, Table II] The results on the ellipsoid undermine the claim of robust performance. The Pearson correlation between estimated and ground-truth Gaussian curvature drops from 0.988 at zero noise to 0.325 at σ=0.1, and becomes negative (−0.153, −0.131, −0.018) for σ=0.3–0.5. This behavior is not discussed anywhere in the text, yet the conclusion states the method 'demonstrates its ability to recover Gaussian and mean curvatures even in noisy settings.' At minimum, the authors should report and explain this failure or restrict the robustness claim to the surfaces where the method works.
  3. [Algorithm 2, local PCA step] The pseudocode sets O ← U[:3, :] (or U[:, :3]) after SVD(B) = UΣV^T. Since B is an N×3 matrix of centered neighborhood points, U is N×N (left singular vectors) and does not live in R^3; the tangent-plane basis and normal should be taken from the right singular vectors V (rows of V^T). As written, the algorithm is not a valid local PCA and cannot reproduce the reported results. Please correct the pseudocode and verify that the implementation uses the right singular vectors.
  4. [III-A, Tables I and II] The comparison against Hickok & Blumberg and Diffusion Curvature reports single-run RMSE, energy distance, and Pearson correlations with no error bars or repeated trials. Because the data are generated with random noise, the differences between methods (e.g., Torus σ=0.1: RMSE 1.391 vs 7.489) may be statistically meaningful, but the reader cannot assess variability. Additionally, the only baselines are the two cited works; the introduction's discussion of second fundamental form-based methods (Sritharan et al., [11]) suggests a natural comparison that is missing. The 'state-of-the-art' claim requires either error bars or a broader baseline set.
minor comments (5)
  1. [III-A] The text refers to 'Fig. I and Fig. II' but the corresponding items are Tables I and II; please correct the cross-reference.
  2. [Algorithm 2] The meaning of O[2] and the indexing of κ1/κ2 should be clarified; as written, C is sorted in ascending order, so C[:k] contains the most negative (smallest) directional curvatures and C[-k:] the most positive, but the text calls these 'highest' and 'lowest' without specifying sign convention.
  3. [II-C, Eq. (4)] The O(t) error term should be stated more precisely; the approximation is valid for small t and depends on the normal orientation, which should be fixed consistently for signed curvatures.
  4. [III-B] The application to single-cell data assumes that the PHATE embedding preserves the curvature structure of the original high-dimensional data, but this assumption is not stated or tested; a brief discussion would help.
  5. [II, Algorithm 1] The kernel function K, the threshold ρ0, the percentage p, and the search bound 0.2δ are not specified in the text; these are user-chosen parameters that qualify the claim of no hand-tuning.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: AdaL-PCA's curvature estimates are validated against independent ground truth, and its adaptive scale heuristic is an assumption rather than a fitted parameter.

full rationale

The paper's derivation chain is self-contained. The curvature estimator in Eq. (4) is an established external approximation (Taubin), and the principal-curvature extraction in Algorithm 2 is a standard averaging procedure. The adaptive radius τ in Eq. (3) is chosen from the explained-variance ratio ρ(r), but it is not defined in terms of the ground-truth curvature; it is a heuristic bandwidth-selection rule. Even though the same local PCA is used both for scale selection and for tangent/normal estimation, this does not make the curvature estimate equivalent to its inputs by construction. The validation on tori, ellipsoids, and saddles compares against known ground-truth Gaussian and mean curvatures, so there is no fitted-input-called-prediction circularity. The only self-citations are [9] (Diffusion Curvature, used as a baseline competitor) and [16] (a scalability pointer); neither is load-bearing for the paper's central claim. Potential weaknesses of the scale-selection heuristic are correctness concerns, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. Its free parameters are the kernel, the variance threshold, the extreme-percentage p, and search bounds. The main ad hoc axiom is the argmin-ρ scale-selection heuristic, which is not derived.

free parameters (4)
  • Kernel function K and its bandwidth scaling
    The choice of kernel is not specified in the paper; the method depends on it for both PCA weighting and curvature averaging.
  • Ratio bound ρ0 (γ) for PCA neighborhood selection
    The threshold for explained variance ratio in Eq. 2 is not given; it is a hand-chosen hyperparameter.
  • Percentage p for averaging extreme directional curvatures = 0.2
    The paper states a default of 20% of largest and smallest directional curvatures are averaged; this is an arbitrary choice.
  • Search radius upper bound for neighborhood selection = 0.2δ (δ = max pairwise distance)
    Algorithm 1 explores r in (0, 0.2δ]; this bound is chosen without justification.
assumptions (4)
  • domain assumption Point cloud data lie on a smooth 2-dimensional manifold embedded in R^3
    Section II-C and Algorithm 1 operate on points in R^3 and use a 2D tangent plane; this is assumed throughout.
  • standard math The directional curvature approximation κ ≈ 2 N·T / ||T||^2 (Eq. 4) holds for the sampled neighborhoods
    This is a standard second-order approximation from differential geometry, cited to Taubin [14].
  • ad hoc to paper The radius minimizing the explained variance ratio ρ(r) is the appropriate scale for estimating curvature
    Eq. 3 defines τ as argmin ρ(r); Figure 1 provides an example but no proof. This is the central heuristic of the method.
  • domain assumption The PHATE embedding preserves the curvature structure necessary for biological interpretation
    Section III-B applies AdaL-PCA to single-cell data via PHATE without stating or testing this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Principal Curvatures Estimation with Applications to Single Cell Data." pith.science (2026). https://pith.science/paper/2J3IYCOL

@misc{pith2026250203750,
  author       = {Pith},
  title        = {Pith review of: Principal Curvatures Estimation with Applications to Single Cell Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2J3IYCOL}},
  note         = {Machine review of arXiv:2502.03750}
}
read the original abstract

The rapidly growing field of single-cell transcriptomic sequencing (scRNAseq) presents challenges for data analysis due to its massive datasets. A common method in manifold learning consists in hypothesizing that datasets lie on a lower dimensional manifold. This allows to study the geometry of point clouds by extracting meaningful descriptors like curvature. In this work, we will present Adaptive Local PCA (AdaL-PCA), a data-driven method for accurately estimating various notions of intrinsic curvature on data manifolds, in particular principal curvatures for surfaces. The model relies on local PCA to estimate the tangent spaces. The evaluation of AdaL-PCA on sampled surfaces shows state-of-the-art results. Combined with a PHATE embedding, the model applied to single-cell RNA sequencing data allows us to identify key variations in the cellular differentiation.

Figures

Figures reproduced from arXiv: 2502.03750 by the authors.

Figure 1
Figure 1. Comparison of the explained variance ratio of the top two singular values and accuracy (RMSE) of Gaussian curvature estimation w.r.t. increas￾ing radii of ϵ-neighborhood and τ-neighborhood around p on torus [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Directional curvatures in an ϵ-PCA neighborhood of p. Algorithm 2 Estimation for Principal Curvature, Gaussian Curvature, and Mean Curvature Input: Point cloud data x1, . . . , xm ∈ R 3 , query point p, kernel function K with supports in [0,1], the pair (ϵPCA, τ ), the percentage p ∈ (0, 1) of total number of points for which the largest (smallest) directional curvature κ1 (κ2) is computed. (Np,ϵPCA , DϵPCA ) ← {(q,… view at source ↗
Figure 3
Figure 3. Comparison of AdaL-PCA against ground truth for mean curvature on three toy datasets. Corr stands for Pearson correlation and RMSE stands for the root means squared error. TABLE I Root Mean Square Error (RMSE) and Energy Distance (Eng. Dist) of Gaussian curvature estimation for different noise levels. For RMSE and Eng. Dist, smaller is better. data noise Ours H. & B. RMSE Eng.Dist RMSE Eng.Dist 0.0 0.462 0.462 1.302… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Gaussian curvature and principal directions on IPSC dataset using AdaL-PCA. into several lineages (cell types), the principal directions in Fig. 4C, D obtained from projecting the three-dimensional principal directions using the PHATE embedding allow us to track the st…
Figure 4
Figure 4. Figure 4: Gaussian curvature and principal directions of embryonic stem cell differentiation. (A) PHATE visualization of scRNA-seq data color-coded by time intervals. (B) PHATE plot colored by Gaussian curvature values. (C, D) Principal directions at different stages of developm…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages

  1. [11]

    Computing the riemannian curvature of image patch and single-cell rna sequencing data manifolds using extrinsic differential geometry,

    D. Sritharan, S. Wang, and S. Hormoz, “Computing the riemannian curvature of image patch and single-cell rna sequencing data manifolds using extrinsic differential geometry,” Proceedings of the National Academy of Sciences , vol. 118, no. 29, p. e2100473118, 2021

    work page 2021

  2. [1]

    Curvature estimates for submanifolds with prescribed gauss image and mean curvature,

    Y . L. Xin, “Curvature estimates for submanifolds with prescribed gauss image and mean curvature,” Calculus of Variations and Partial Differential Equations, vol. 37, no. 3, pp. 385–405, Mar 2010. [Online]. Available: https://doi.org/10.1007/s00526-009-0268-8

    work page doi:10.1007/s00526-009-0268-8 2010

  3. [2]

    A comparison of gaussian and mean curvatures estimation methods on triangular meshes,

    T. Surazhsky, E. Magid, O. Soldea, G. Elber, and E. Rivlin, “A comparison of gaussian and mean curvatures estimation methods on triangular meshes,” in 2003 IEEE International Conference on Robotics and Automation (Cat. No.03CH37422) , vol. 1, 2003, pp. 1021–1026 vol.1

    work page 2003

  4. [3]

    On reliable curvature estimation,

    P. Flynn and A. Jain, “On reliable curvature estimation,” in Proceedings CVPR ’89: IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 1989, pp. 110–116

    work page 1989

  5. [4]

    Simplified gaussian and mean curva- tures to range image segmentation,

    C. Zhao, D. Zhao, and Y . Chen, “Simplified gaussian and mean curva- tures to range image segmentation,” in Proceedings of 13th International Conference on Pattern Recognition , vol. 2, 1996, pp. 427–431 vol.2

    work page 1996

  6. [5]

    Estimating curvatures and their derivatives on triangle meshes,

    S. Rusinkiewicz, “Estimating curvatures and their derivatives on triangle meshes,” in Proceedings. 2nd International Symposium on 3D Data Processing, Visualization and Transmission, 2004. 3DPVT 2004. IEEE, 2004, pp. 486–493

    work page 2004

  7. [6]

    Ball pivoting algorithm and discrete gaussian curvature: A direct way to curved nanographene circularly polarized luminescence spectral simulation,

    G. Bella, G. Bruno, and A. Santoro, “Ball pivoting algorithm and discrete gaussian curvature: A direct way to curved nanographene circularly polarized luminescence spectral simulation,” FlatChem, vol. 42, p. 100567, 2023. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S2452262723000995

    work page 2023

  8. [7]

    A cooperative regulation method for greenhouse soil moisture and light using gaussian curvature and machine learning algorithms,

    J. Hou, Y . Li, Z. Sun, H. Wang, M. Lu, J. Hu, and H. Wu, “A cooperative regulation method for greenhouse soil moisture and light using gaussian curvature and machine learning algorithms,” Computers and Electronics in Agriculture , vol. 215, p. 108452, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0168169923008402

    work page 2023

Show all 17 references
  1. [8]

    Curvanet: Geometric deep learning based on directional curvature for 3d shape analysis,

    W. He, Z. Jiang, C. Zhang, and A. M. Sainju, “Curvanet: Geometric deep learning based on directional curvature for 3d shape analysis,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , ser. KDD ’20. New York, NY , USA: Associ...

    2020

  2. [9]

    Diffusion curvature for estimating local curvature in high dimensional data,

    D. Bhaskar, K. MacDonald, O. Fasina, D. S. Thomas, B. Rieck, I. Adelstein, and S. Krishnaswamy, “Diffusion curvature for estimating local curvature in high dimensional data,” in Advances in Neural Information Processing Systems , A. H. Oh, A. Agarwal, D. Belgrave, and K. Cho, ...

    2022

  3. [10]

    An intrinsic approach to scalar-curvature estimation for point clouds,

    A. Hickok and A. J. Blumberg, “An intrinsic approach to scalar-curvature estimation for point clouds,” arXiv preprint arXiv:2308.02615 , 2023

    2023 arXiv

  4. [12]

    M. P. do Carmo, Differential geometry of curves & surfaces , 2nd ed. Dover Publications, Inc., Mineola, NY , 2016

    2016

  5. [13]

    Vector diffusion maps and the connection laplacian,

    A. Singer and H.-T. Wu, “Vector diffusion maps and the connection laplacian,” Communications on pure and applied mathematics , vol. 65, no. 8, pp. 1067–1144, 2012

    2012

  6. [14]

    Estimating the tensor of curvature of a surface from a polyhedral approximation,

    G. Taubin, “Estimating the tensor of curvature of a surface from a polyhedral approximation,” in Proceedings of IEEE International Conference on Computer Vision . IEEE, 1995, pp. 902–907

    1995

  7. [15]

    Fast approximate truncated svd,

    S. L. Shishkin, A. Shalaginov, and S. D. Bopardikar, “Fast approximate truncated svd,” Numerical Linear Algebra with Applications , vol. 26, no. 4, p. e2246, 2019

    2019

  8. [16]

    Fast algorithms for singular value decomposition and the inverse of nearly low-rank matrices,

    C. Xu, W. Xu, and K. Jing, “Fast algorithms for singular value decomposition and the inverse of nearly low-rank matrices,” National Science Review, vol. 10, no. 6, p. nwad083, 2023

    2023

  9. [17]

    Embryoid body data for phate,

    K. Moon, “Embryoid body data for phate,” Mendeley Data, vol. 1, 2018

    2018

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.