acceptodds
Under review as a conference paper at ICLR 2027

Riemannian Isometric Kernels: Preserving Intrinsic Geometry of Hyperbolic Space

Abstract

Hyperbolic spaces, characterized by constant negative curvature, provide a natural geometric alternative for representing hierarchical data, while kernel methods further enhance the representation power by mapping hyperbolic embeddings into reproducing kernel Hilbert spaces (RKHSs). However, existing hyperbolic kernels enforce the isometry constraint only at the metric-space level, creating a substantial mismatch between the kernel-induced similarity used in practice and the manifold's theoretical geodesic distance. To address this limitation, we introduce Riemannian isometric kernels, whose canonical feature maps induce a metric that coincides exactly with the Riemannian metric of the input manifold, thereby intrinsically preserving its geometric structure. Starting from the definition of Riemannian isometry, we construct a family of Riemannian isometric kernels, including logarithmic kernel (RILog), binomial kernel (RIBin), exponential kernel (RIExp), and integral kernel (RIInt) variants, whose kernel-induced distances used in practice are monotonically increasing functions of the geodesic distance, ensuring consistency between the theoretical formulation and empirical evaluation. We evaluate the proposed kernels through geometric stress analysis and four diverse downstream tasks: zero-shot learning, few-shot learning, graph node classification, and semantic textual similarity. Across all benchmarks, the proposed kernels match or outperform state-of-the-art adaptive hyperbolic kernels, demonstrating their effectiveness and broad applicability.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hyperbolic kernel functions, Hyperbolic space, Riemannian isometry

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_ehNVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses geometric distortion in existing hyperbolic kernel methods, where mapping hyperbolic representations into an RKHS does not necessarily preserve the original Riemannian geometry.

To address this issue, the authors introduce Riemannian isometric kernels, whose canonical feature maps preserve the Riemannian metric of the input manifold. They establish theoretical conditions for Riemannian isometry and construct four positive-definite kernels: RILog, RIBin, RIExp, and RIInt. They also prove that the kernel-induced distances are monotonically increasing functions of the original geodesic distance.

The proposed methods are evaluated through geometric stress analysis and four downstream tasks: zero-shot learning, few-shot learning, graph node classification, and semantic textual similarity. The experimental results show that the proposed kernels generally outperform existing hyperbolic kernel methods.

Critical strengths

The paper provides a mathematically well-motivated approach to constructing hyperbolic kernels that preserve intrinsic Riemannian geometry. The authors establish sufficient conditions for Riemannian isometry and construct four positive-definite kernels satisfying these conditions. The theoretical development is a meaningful contribution.

The authors evaluate their methods across multiple tasks, including zero-shot learning, few-shot learning, graph node classification, and semantic textual similarity. The proposed kernels achieve competitive or superior performance across a broad range of benchmarks, suggesting their practical usefulness.

Critical weaknesses

1. About style

First, the font sizes in Tables 1–5 appear to have been noticeably reduced compared to the main text. This seems to violate the official ICLR 2027 formatting guidelines, which explicitly state: “Do not change any aspects of the formatting parameters in the style files” and “do not change font sizes (except perhaps in the References section).”

I strongly encourage the authors to revise these tables to comply with the official formatting requirements. If the tables do not fit within the available space, the authors should consider wrapping text onto multiple lines, restructuring the tables, or splitting them into smaller tables rather than reducing the font size.

2. The geometric preservation claim is not directly validated by the experiments.

The main motivation of this paper is to eliminate geometric distortion introduced by kernel mappings. However, the geometric stress experiment in Section 5.1 compares embeddings obtained from separately trained models using different similarity measures. This evaluates how similar the learned distance structures are, rather than directly measuring the geometric distortion introduced by the kernel mapping itself.

Consequently, although the proposed methods achieve lower stress values, these results do not directly demonstrate that the proposed kernels preserve the original Riemannian geometry better than existing kernels.

3. The relationship between Riemannian isometry and downstream performance improvements remains unclear.

Although the proposed kernels achieve promising results on several downstream tasks, the experiments do not establish whether these improvements are attributable to Riemannian isometry or simply to the different distance functions used during training.

In particular, the kernel-induced distances are monotonic transformations of the geodesic distance, and different transformations can influence optimization and predictive performance independently of exact Riemannian isometry.

Therefore, additional controlled comparisons would be necessary to clarify the specific practical advantage of the proposed isometric property.

4. Please check the mathematical notation

There appear to be several notation issues in the main text. Could the authors please carefully check and correct the following?

  1. Equations (2)–(3) (pp. 4–5): The local coordinates x~,y~\tilde{x},\tilde{y} are introduced, but the partial derivatives are written with respect to xi,yjx^i,y^j. The notation should be clarified and made consistent.

  2. Theorem 2 (p. 5, lines 213–220): The definition of F2F_2 includes the term k0k_0. Although k0=1k_0=1 follows from the definition of ktk_t, explicitly stating this convention would improve clarity.

  3. RIExp (p. 6, lines 235–238): The valid range of the parameter λ\lambda is not explicitly specified. The authors should clarify that λ>0\lambda>0, as required by the construction in Theorem 2.

  4. Theorem 3 (p. 6, lines 255–258): The notation dSi(x,y)d_{S_i}(x,y) is used, although Si=φi(M)S_i=\varphi_i(M) and x,y∈Mx,y\in M. Should this instead be dSi(φi(x),φi(y))d_{S_i}(\varphi_i(x),\varphi_i(y))?

Although the main ideas of the paper are understandable, the inconsistent notation causes considerable confusion and makes the paper unnecessarily difficult to follow. The authors are therefore encouraged to carefully review and correct the notation issues mentioned above.

In addition, the authors are encouraged to carefully check the appendix for similar notation issues and ensure consistency throughout the paper.

Questions for authors

1. Regarding the geometric stress experiment in Section 5.1

As I understand it, the authors compare the distances between embeddings obtained from separately trained models, rather than directly evaluating the distortion introduced by the kernel mapping. Is this interpretation correct? If so, could the authors explain how this experiment directly supports the geometric preservation claim?

I believe a more direct experiment would be useful. For example, the authors could compare the original Riemannian metric with the kernel-induced metric at sampled points, or compare curve lengths before and after the kernel mapping. These evaluations are theoretically feasible and would provide more direct evidence for the central claim of the paper.

2. Regarding the downstream experiments

Could the authors provide controlled experiments to clarify whether the performance improvements are specifically due to Riemannian isometry? For example, comparing the proposed kernels with non-isometric transformations of the same kernel distances under matched training conditions would help distinguish the effect of geometric preservation from that of changing the distance function.

3. Regarding the differences between RILog and RIBin

According to the kernel definitions, the squared kernel distances of RIBin approach those of RILog as the parameter t approaches zero. Appendix E states that t = 0.001 is used, and Appendix D.4 reports the same curvature c = 0.001 for both methods on AWA1 and AWA2. Under the clipping procedure described in Appendix D.1, their distances should therefore be extremely close. However, Table 2 reports a substantial difference between RILog and RIBin, particularly on AWA2 (76.3 versus 63.7 in harmonic mean). Could the authors clarify the reason for this discrepancy and provide results across multiple random seeds?

5 comments

We sincerely thank the reviewer for the thoughtful and detailed feedback. We particularly appreciate the reviewer’s recognition of the mathematical motivation and theoretical contributions of our work. The comments have helped us identify several aspects that require clearer presentation and stronger empirical validation.

We try our best to address each concern below.

W1 & W4. Formatting Issues and Mathematical Notation

Response: We thank the reviewer for pointing out the formatting and notation issues. We agree that both the presentation of the tables and the consistency of the mathematical notation should be improved.

Regarding the formatting, we will revise Tables 1–5 to comply with the official ICLR formatting requirements by restoring the default font settings of the official style file. For tables that do not fit within the available space, we will restructure the columns, wrap long entries, or split the tables into smaller ones rather than reducing the font size.

Regarding the mathematical notation, we will revise the relevant equations and definitions to improve clarity and consistency. Specifically, we will:

  1. Clarify the distinction between manifold points and their local coordinate representations in Equations (2)–(3). In particular, xix^i and yjy^j denote the ii-th and jj-th coordinate components of the local chart representations of xx and yy, respectively; the partial derivatives are indicated explicitly by ∂\partial.

  2. Explicitly define the zeroth-order kernel term in Theorem 2 as the constant kernel k0(x,y)≡1k_0(x,y)\equiv 1.

  3. State the valid parameter range for RIExp explicitly, including λ>0\lambda>0.

  4. Correct and standardize the distance notation in Theorem 3, clearly distinguishing distances on the original manifold, the embedded submanifold, and the RKHS.

We will also review the appendix to ensure consistent notation throughout the manuscript.

W2 & Q1. Direct Validation of Riemannian Geometry Preservation

Response: We thank the reviewer for this insightful comment. We agree that our original geometric stress experiment does not directly measure the geometric distortion introduced by the canonical feature map. We clarify its motivation below and introduce additional experiments to directly evaluate geometric preservation.

1. Clarification of the motivation and scope of the original stress experiment.

The original stress experiment was designed to investigate whether different similarity measures lead to learned representations whose pairwise distance structures remain consistent with the intrinsic hyperbolic geometry. The motivation is to assess the extent to which the geometry of the learned representations deviates from the reference hyperbolic representation after neural network training and optimization. In this sense, the experiment provides a representation-level evaluation of geometric consistency across different similarity measures.

However, we acknowledge that the reviewer is correct: since the embeddings are obtained from separately trained models, the stress values reflect both the effects of the similarity measure and the representation learning process. Therefore, this experiment does not isolate the distortion introduced by the canonical feature map itself, and should not be interpreted as a direct validation of the Riemannian isometry property. We will clarify this distinction in the revised manuscript.

2. Why we do not numerically compare the Riemannian metric tensors.

The local metric preservation property is already established analytically in our paper. Specifically, for a smooth kernel kk, the Riemannian metric induced by its canonical feature map is

\left. \frac{\partial^2 k(x,y)} {\partial x^i\partial y^j} \right|_{y=x}.$$ For the Poincaré ball, the original Riemannian metric is $$h_{ij}(x)= \frac{4\delta_{ij}} {(1-c\|x\|^2)^2}.$$ Our theoretical results establish that the proposed kernels satisfy $$g^k_{ij}(x)=h_{ij}(x)$$ at every point in the domain. Together with the required smoothness and embedding conditions, Theorem 1 establishes that the canonical feature map is a Riemannian isometry. Consequently, the equality of the two metric tensors is an exact theoretical result, rather than a numerical property that needs to be established through sampling. Numerically evaluating the metric tensors at sampled points would only provide a finite-precision verification of an equality already guaranteed by the theorem. Instead, we provide additional numerical experiments on curve lengths and angles, which offer complementary and more intuitive evaluations of the geometric consequences of this isometry.

3. Geodesic curve-length preservation.

A direct consequence of Riemannian isometry is that the length of every smooth curve is preserved under the canonical feature map. We numerically evaluate this property using the kernel-induced RKHS distance.

We consider a 5-dimensional Poincaré ball with curvature parameter c=1c=1 and generate 500 random geodesics. The hyperbolic radial distance of each starting point from the origin is sampled uniformly from [0,1.5][0,1.5], and the geodesic length is sampled uniformly from [0.1,4.0][0.1,4.0]. We uniformly sample N=200N=200 points along each geodesic.

For each kernel kk, denoting the canonical map as ϕ\phi, the RKHS curve length is approximated by the sum of consecutive RKHS distances:

LHk(N)=∑i=0N−2∥ϕ(xi)−ϕ(xi+1)∥=∑i=0N−2k(xi,xi)+k(xi+1,xi+1)−2k(xi,xi+1).L_{\mathcal H_k}^{(N)}=\sum_{i=0}^{N-2}\|\phi(x_i)-\phi(x_{i+1})\|= \sum_{i=0}^{N-2}\sqrt{k(x_i,x_i)+k(x_{i+1},x_{i+1})-2k(x_i,x_{i+1})}.

We compare it with the original hyperbolic geodesic length LH=dH(x0,xN−1)L_{\mathbb H}=d_{\mathbb H}(x_0,x_{N-1}) (dHd_{\mathbb H} is the geodesic distance of the Poincaré ball) using the absolute relative error

\frac{L_{\mathcal H_k}^{(N)}}{L_{\mathbb H}}-1 \right|.$$ We evaluate RIBin ($t=1$), RIInt, RIExp ($\lambda=1$), and RILog (with the origin as the reference point). The results are: | Kernel | Mean ratio | MAE | RMSE | Max error | |---|---:|---:|---:|---:| | RIBin | 0.999971 | 2.9e-5 | 3.8e-5 | 8.4e-5 | | RIInt | 0.999954 | 4.6e-5 | 6.0e-5 | 1.34e-4 | | RIExp | 0.999954 | 4.6e-5 | 6.0e-5 | 1.34e-4 | | RILog | 0.999989 | 1.1e-5 | 1.5e-5 | 3.4e-5 | Here, MAE, RMSE, and maximum error are computed from the absolute relative errors across the 500 geodesics. To examine the effect of numerical discretization, we further repeat the experiment with $N\in\{5,10,20,50,100,200,500\}$ points per geodesic, using the same set of 500 geodesics. The MAE consistently decreases as the sampling density increases. For example, for RIBin, it decreases from $6.03\times10^{-2}$ at $N=5$ to $2.9\times10^{-5}$ at $N=200$ and $4.6\times10^{-6}$ at $N=500$. The empirical convergence slopes, obtained by fitting the log-MAE against $\log(1/(N-1))$ for $N\geq20$, are approximately 2 for all four kernels. This is consistent with the second-order convergence of polygonal approximations to the arc length of a smooth curve. **4. Angle preservation.** We further evaluate angle preservation, another direct consequence of the preservation of the Riemannian metric. We generate 100 random non-degenerate geodesic triangles in a 3-dimensional Poincaré ball with $c=0.01$. At a selected vertex $p$ of each triangle, we consider two geodesic directions and construct short geodesic segments along them, controlled by a segment parameter $\epsilon$. Let $a_\epsilon$ and $b_\epsilon$ denote the endpoints of these segments. The original hyperbolic angle is computed using the hyperbolic law of cosines. $$\theta_{\mathbb H}=\arccos \left( \frac{ \cosh\left(\sqrt{c}\,d_{\mathbb H}(p,a_\epsilon)\right) \cosh\left(\sqrt{c}\,d_{\mathbb H}(p,b_\epsilon)\right)- \cosh\left(\sqrt{c}\,d_{\mathbb H}(a_\epsilon,b_\epsilon)\right) }{ \sinh\left(\sqrt{c}\,d_{\mathbb H}(p,a_\epsilon)\right) \sinh\left(\sqrt{c}\,d_{\mathbb H}(p,b_\epsilon)\right) } \right).$$ The corresponding angle between the RKHS feature-space displacement vectors is computed from the kernel-induced distances using the polarization identity: $$\theta_{\mathcal H_k}= \arccos \left( \frac{ d_k^2(p,a_\epsilon) +d_k^2(p,b_\epsilon) -d_k^2(a_\epsilon,b_\epsilon) }{ 2d_k(p,a_\epsilon)d_k(p,b_\epsilon) } \right),$$ where $d_k(x,y)=\sqrt{k(x,x)+k(y,y)-2k(x,y)}$. We report the absolute angular error in degrees. At $\epsilon=0.005$, the results are: | Kernel | Mean error (°) | Std (°) | Median error (°) | Max error (°) | |---|---:|---:|---:|---:| | RIBin | 0.003618 | 0.007003 | 0.000921 | 0.035390 | | RIInt | 0.005543 | 0.010988 | 0.001387 | 0.055865 | | RIExp | 0.005543 | 0.010989 | 0.001387 | 0.055870 | | RILog | 0.001707 | 0.003039 | 0.000438 | 0.014861 | The angular errors further decrease as $\epsilon$ becomes smaller. In particular, halving $\epsilon$ from 0.01 to 0.005 reduces the mean angular errors by approximately a factor of four, consistent with the finite-segment approximation effect. The original stress experiment evaluates the geometric consistency of learned representations after training, whereas the new curve-length and angle experiments directly evaluate geometric quantities induced by the canonical feature maps, independently of neural network training. The theoretical equality of the kernel-induced and original Riemannian metrics guarantees exact local metric preservation; the additional numerical experiments provide complementary empirical verification of its geometric consequences. We will revise Section 5.1 to clarify the distinct purposes of these evaluations and avoid interpreting representation-level stress as direct evidence of Riemannian isometry.

W3 & Q2. Ablation Study on the Riemannian Isometry Condition

Response:

We thank the reviewer for this important question. We agree that it is necessary to determine whether the downstream improvements arise from the Riemannian isometry condition or merely from the distance transformations used by our kernels.

To address this, we conduct a controlled ablation study that removes the Riemannian isometry condition while keeping the distance-transformation functions unchanged. Specifically, for each proposed RI kernel, we construct a Euclidean counterpart (EN) by replacing the hyperbolic geodesic distance with the Euclidean distance. This allows us to isolate the contribution of the underlying Riemannian geometry.

Let x,y∈Bcnx,y\in\mathbb{B}_c^n. We define

\cosh\left(\sqrt{c}\|x-y\|_2\right).$$ The four EN counterparts are defined as follows: **ENBin:** \$k_{\mathrm{EN}}^{\mathrm{Bin}}(x,y) =\frac{1}{ct}r_{\mathrm{EN}}(x,y)^{-t}, \ t>0.\$ **ENExp:** \$k_{\mathrm{EN}}^{\mathrm{Exp}}(x,y) =\frac{1}{c\lambda} \exp\left[ -\lambda\left(1-r_{\mathrm{EN}}(x,y)^{-1}\right) \right], \ \lambda>0.\$ **ENInt:** \$k_{\mathrm{EN}}^{\mathrm{Int}}(x,y) =\frac{1}{c\left(1+\log r_{\mathrm{EN}}(x,y)\right)}. \$ **ENLog:** \$ \begin{aligned} k_{\mathrm{EN}}^{\mathrm{Log}}(x,y) =\frac{1}{c}\big[-\log r_{\mathrm{EN}}(x,y)+\log r_{\mathrm{EN}}(x_0,x) +\log r_{\mathrm{EN}}(x_0,y) \big], \end{aligned} \$ where $x_0$ is the reference point used in the corresponding RI construction. The RI and EN counterparts use the same scalar distance transformations. Specifically, define $$r_{\mathrm{RI}}(x,y)= \cosh\left(\sqrt{c}\,d_{\mathbb H}(x,y)\right).$$ Their squared kernel-induced distances can be written as $$d_{\mathrm{RI}}^2(x,y)= f(d_{\mathbb H}(x,y)), \qquad d_{\mathrm{EN}}^2(x,y)= f(\|x-y\|_2).$$ The corresponding scalar functions are: **Binomial:** \$f_{\mathrm{Bin}}(s) =\frac{2}{ct} \left[ 1-\cosh^{-t}(\sqrt{c}s) \right].\$ **Exponential:** \$f_{\mathrm{Exp}}(s) =\frac{2}{c\lambda} \left[ 1-\exp\left( -\lambda\left(1-\operatorname{sech}(\sqrt{c}s)\right) \right) \right].\$ **Integral:** \$f_{\mathrm{Int}}(s) =\frac{2}{c} \frac{\log\cosh(\sqrt{c}s)} {1+\log\cosh(\sqrt{c}s)}. \$ **Logarithmic:** \$f_{\mathrm{Log}}(s) =\frac{2}{c}\log\cosh(\sqrt{c}s). \$ Thus, each RI–EN pair uses exactly the same scalar function $f$ and corresponding parameters; only the input distance differs. We use $t=0.001$ for the binomial construction and $\lambda=0.001$ for the exponential construction, following our original experimental settings. All four EN constructions induce the same local tensor: $$g_{\mathrm{EN}}(x)=I_n.$$ In contrast, the Poincaré ball metric is $$g_{\mathbb H}(x)= \frac{4}{(1-c\|x\|_2^2)^2}I_n.$$ Consequently, $$g_{\mathrm{EN}}(x)\neq g_{\mathbb H}(x), \qquad \forall x\in\mathbb{B}_c^n.$$ The EN counterparts therefore do not satisfy the Riemannian isometry condition. They provide non-isometric counterparts that retain the same scalar distance transformations while removing geometric preservation. We keep all other hyperparameters and experimental configurations identical between RI and EN, including the backbone, optimization settings, training protocol, and evaluation procedure. The curvature parameters are also matched to the corresponding RI experiments: $c=0.001$ for RIBin, RIExp, and RIInt, and $c=0.01$ for RILog. Results are reported as mean accuracy with 95% confidence intervals, and the performance gap is calculated as RI minus EN. **Table 1: Classification accuracy (%) on CUB (5-way 1-shot) with 95% confidence interval.** | Method | Bin | Exp | Int | Log | |:--|--:|--:|--:|--:| | RI (isometric) | 67.3 ± 0.2 | 67.3 ± 0.2 | 65.8 ± 0.2 | 68.6 ± 0.2 | | EN (non-isometric) | 61.8 ± 0.2 | 62.8 ± 0.2 | 63.2 ± 0.2 | 63.5 ± 0.3 | | RI − EN | **+5.5** | **+4.5** | **+2.6** | **+5.1** | **Table 2: Classification accuracy (%) on miniImageNet (5-way 1-shot) with 95% confidence interval.** | Method | Bin | Exp | Int | Log | |:--|--:|--:|--:|--:| | RI (isometric) | 55.2 ± 0.2 | 54.3 ± 0.2 | 54.5 ± 0.2 | 56.3 ± 0.2 | | EN (non-isometric) | 48.9 ± 0.2 | 49.3 ± 0.2 | 49.0 ± 0.2 | 47.9 ± 0.2 | | RI − EN | **+6.3** | **+5.0** | **+5.5** | **+8.4** | The results show that removing the Riemannian isometry condition consistently degrades downstream performance across all four kernel constructions on both datasets. Since the scalar distance transformations and all other experimental configurations remain unchanged, these differences cannot be attributed simply to different distance-transformation functions. Instead, they demonstrate the performance benefit of preserving the underlying hyperbolic geometry through the Riemannian isometry condition. ## Q3. Difference Between RILog and RIBin We thank the reviewer for this comment. As the same concern was raised by Reviewer p2Zq, we refer the reviewer to our response to Q2 in A2 for detailed analysis and additional experimental results.
Reviewer_Eec8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Riemannian isometric kernels on hyperbolic space as a special case of a general Riemannian isometric kernel theory on general manifold. The authors develop a family of kernels based on the Riemannian metric on the Poincare model. They prove that the kernel distance is a monotonically increasing function of geodesic distance. In the experiments, the paper shows that the introduced family of kernels enjoy small distortion and show good performance in zero-shot and few-shot learning, as well as graph node classification and semantic textual similarity evaluation, where RILog performs the best.

Critical strengths
  1. This paper proposes a simple but interesting and previously underexplored idea: instead of keeping the distances, the proposed idea maintains the Riemannian metric, which naturally leads to the specific family of hyperbolic kernels. The authors also clearly differentiate their contribution to previously proposed hyperbolic kernels.
  2. This paper presents a comprehensive set of experiments, ranging over different types of datasets.
  3. The paper is generally well-written. It starts with a clear motivation, continues with a clear introduction of the relevant mathematics, and then the central idea and the experiments. It is enjoyable to read the paper.
Critical weaknesses
  1. It looks to me that the idea of section 4.2 could widely apply to many manifolds that are used in machine learning. It is unclear why the authors choose only to focus on hyperbolic space, and only the Poincare model of hyperbolic space. It seems that the only thing that matters for the development of the hyperbolic kernels from section 4.2 to section 4.3 is the specific formula of hijh_{ij} in Poincare model. And the result in section 4.4 seem not to rely on the negative curvature of hyperbolic space. The authors could either explain why the method is expected to work especially well in hyperbolic space, or extend the discussion to wider families of kernels, which I guess is not painful.
  2. It is good that tables 2 and 3 contain both means and confidence intervals. However, the confidence intervals are missing from tables 1, 4, 5. For instance, for table 1, the authors mention “train hyperbolic prototypical networks” as “geo”, so if I understand correctly, “geo” itself is random, so confidence intervals can be added.
  3. Also, it is unclear to me why the “hyperbolic prototypical networks”, as “geo”, could serve as the measure for computing the stress.
Questions for authors

See "weaknesses".

4 comments

We sincerely thank the reviewer for the careful reading of our paper and the encouraging assessment of its motivation, theoretical development, and experimental evaluation. We are glad that the reviewer finds the proposed Riemannian isometric kernel framework interesting and appreciates its potential for learning on non-Euclidean spaces. We appreciate the constructive comments and address each concern below.

Q1. Scope of the framework and the focus on hyperbolic space

A1: We agree with the reviewer that the theoretical framework developed in Section 4.2 is not intrinsically restricted to hyperbolic manifolds. Indeed, a central motivation of our work is to establish a general framework for constructing Riemannian isometric kernels on Riemannian manifolds.

We choose hyperbolic space as our initial focus because of its suitability for representing hierarchical and tree-structured data. Its negative curvature induces exponential volume growth, allowing such structures to be embedded with lower distortion than in Euclidean space [1] [2]. This motivates its widespread use in machine learning, including NLP, computer vision, and graph learning [3] [4] [5]. We view this study as a first step toward extending the framework to other Riemannian manifolds. We view this work as an initial step toward exploring Riemannian isometric kernels in a broader range of geometric spaces. Extending the proposed framework to other Riemannian manifolds is an interesting direction for future research, which we intend to investigate.

Different models of hyperbolic space are mathematically equivalent representations of the same underlying geometry; thus, choosing the Poincaré ball does not fundamentally restrict our framework. We adopt it for its bounded coordinate domain and the convenient expressions it provides for the Riemannian metric and geodesic distance, facilitating kernel derivation and implementation.

We also clarify that, although the results in Section 4.4 do not explicitly rely on the negative curvature of hyperbolic space, they are derived based on the specific analytical forms and properties of the kernels constructed in Section 4.3. Therefore, this conclusion is not independent of curvature, but is applicable to hyperbolic kernels with different curvatures. Extending these results to other manifolds would require constructing suitable kernels and establishing the corresponding properties under the geometric and analytical conditions of each manifold.

We believe these clarifications will better highlight both the generality of the proposed framework and the motivation for choosing hyperbolic space and the Poincaré ball model as our initial setting. And we will further extend the discussion to explain how the framework may be applied to other Riemannian manifolds in our future works.

Q2. Confidence intervals in Tables 1, 4, and 5

A2: We thank the reviewer for pointing out the missing confidence intervals in Tables 1, 4, and 5. We would like to clarify that, for few-shot learning experiments, the confidence intervals reported in our paper are obtained by evaluating the trained models over multiple randomly sampled episodes, following the standard few-shot evaluation protocol. Therefore, these intervals characterize the variability across evaluation episodes, rather than the variability caused by different training random seeds.

We would also like to clarify that the selection of Geo is not arbitrary or random. Geo refers to the conventional hyperbolic Prototypical Network trained using the standard hyperbolic geodesic distance, which serves as a natural geometric reference in our experiments. We provide a more detailed explanation of this choice in our response A3 to Q3.

To assess the variability across training seeds, we have additionally conducted multi-seed experiments for Tables 1, 4, and 5. The results for the geometric stress experiment are reported below.

Table 1:Geometric stress of embeddings (mean ± standard deviation across 5 different random seeds) relative to the geodesic-distance reference.

CosineCHLCHPolyCHRBFCHLapAHLAHPolyAHRBFAHLapAHRadRIBinRIExpRIIntRILog
Stress0.2996 ± 0.00010.6978 ± 0.00010.6999 ± 0.00010.6961 ± 0.00010.6833 ± 0.00010.5482 ± 0.00010.5505 ± 0.00020.1622 ± 0.00010.1656 ± 0.00010.2791 ± 0.00010.0399 ± 0.00010.0397 ± 0.00020.0404 ± 0.00010.0393 ± 0.0001

We will include these results, together with the corresponding multi-seed results for Tables 4 and 5, in the revised manuscript.

Q3. The use of Geo as the reference for computing stress

A3: We thank the reviewer for raising this question. We would like to clarify that the choice of Geo is motivated by its underlying geometry and is not arbitrary.

Specifically, Geo refers to the conventional Hyperbolic Prototypical Network, which uses the hyperbolic geodesic distance as its distance metric for classification. The hyperbolic geodesic distance is a natural and canonical choice for measuring distances in hyperbolic space, as it is directly determined by the underlying Riemannian geometry. Therefore, Geo provides a natural reference for evaluating the geometric properties of our proposed kernels.

In particular, the stress evaluation aims to quantify how closely the pairwise distance structures induced by our proposed kernels agree with those obtained using the standard hyperbolic geodesic distance. Thus, Geo serves as a reference for conventional hyperbolic geometry.

We acknowledge that the learned representations of Geo may vary across random seeds due to stochastic training. However, this does not affect the rationale for selecting Geo as the reference: its defining distance metric is the standard hyperbolic geodesic distance. We have provided multi-seeds results in the response A2, and we will add them in the revised manuscript.

References

[1] Maximilian Nickel and Douwe Kiela. Poincare embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems, 2017.

[2] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In Graph Drawing, 2012.

[3] Octavian Ganea, Gary Becigneul, and Thomas Hofmann. Hyperbolic neural networks. In Advances in Neural Information Processing Systems, 2018.

[4] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. In Advances in Neural Information Processing Systems, 2019.

[5] Ines Chami, Zhitao Ying, Christopher Re, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In Advances in Neural Information Processing Systems, 2019.

Reviewer_Eec8OpenReview ↗

Thanks for the rebuttal

Thanks for the rebuttal, which addresses my questions. I maintain the positive rating.

AuthorsOpenReview ↗

Thanks for the feedback

Thank you for your time and for the positive assessment. We greatly appreciate your insightful feedback.

Reviewer_p2ZqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies kernel functions in hyperbolic space. It introduces a new family of kernels, namely Riemannian isometric kernels, whose canonical feature maps define Riemannian isometries between the input mainfolds and their embeddings in a reproducing kernel Hilbert space (RKHS). The authors built four concrete Riemannian isometric kernels: RILog, RIBin, RIExp and RIint.

In experiments, the authors evaluated the proposed Riemannian isometric kernels on geometric-stress experiment and four downstream tasks: generalized zero-shot learning, few-shot learning, graph node classification and semantic textual similarity, comparing to no-kernel baseline and other hyperbolic kernels . Analysis shows that the proposed Riemannian isometric kernels achieves lower geometric distortion, and outperform existing hyperbolic kernels.

Critical strengths
  • The paper formulates an explicit condition under which a kernel’s canonical feature map preserves the Riemannian metric, and provides concrete positive-definite kernels satisfying it. The formulation is clean and theoretically correct. The paper provided solid proofs for their claims.

  • The paper presented broad empirical evaluations, which includes geometric distortion analysis on CIFAR-FS, zero-shot and few-shot learnings on image recognition, graph node classification and textual similarity.

Critical weaknesses

The contribution of the paper is modest and the practical benefit of this is unclear. In a contrastive pipeline, using the proposed kernels only changes how distances are transformed into training scores. At inference time, every RI kernel ranks candidates exactly as plain hyperbolic distance does. In addition, most of the gains over the strongest baseline kernel (AHRad) are single digits, yet the results on RILog, RIBin which should be similar in the setting, differs by a large margin. Those experiments can be more convincing if multi-seed results and avg/variance of the results are reported.

Questions for authors

Can you clarify on which baseline numbers were copied from prior work and which were re-run? It seems that there are results that looks lower than the original paper, i.e. the hyperbolic ProtoNet baseline on CUB and mini-ImageNet.

3 comments

We sincerely thank the reviewer for the careful reading of our manuscript and for the constructive feedback. We try our best to address your questions as follows.


Q1. Identical Ranking at Inference Time

Reviewer concern: In the contrastive learning pipeline, the kernels only affect training scores. At inference time, the ranking induced by each RI kernel is identical to that induced by the ordinary hyperbolic distance.

A1: We agree with the reviewer’s observation. For a fixed embedding and a fixed set of prototypes, the distances induced by our proposed kernels are strictly monotonic transformations of the hyperbolic geodesic distance. Therefore, they preserve pairwise distance rankings.

We would like to clarify that this property does not imply that different kernels produce identical models after training. The distinction is between that each kernel induces an RKHSs, a complete inner-product space which serves as the feature embedding space. Therefore, different kernels lead to different gradients and optimization trajectories, learning classification boundaries of distinct structure .

More specifically, each positive-definite kernel kk induces a canonical feature map ϕk:M→Hk\phi_k:\mathcal{M}\rightarrow\mathcal{H}_k, where Hk\mathcal H_k is the associated RKHS, such that

k(x,y)=⟨ϕk(x),ϕk(y)⟩Hk,dk(x,y)=∥∣ϕ(x)−ϕ(y)∥∣Hk.k(x,y)=\langle \phi_k(x),\phi_k(y) \rangle_{\mathcal{H}_k}, \quad d_k(x,y)=\||\phi(x)-\phi(y)\|| _{\mathcal H_k}.

Thus, although different RI kernels preserve the same distance ranking for a fixed representation, they induce different feature maps and RKHS representations. During training, the loss is computed from kernel-induced similarities or distances, and its gradients are backpropagated through the corresponding feature map to the feature extractor. Consequently, different kernels can induce different gradient signals and optimization trajectories, leading to different learned representations.

To directly address this concern, we made a controlled experiment in which models trained with different kernels are evaluated using the same inference metric. Specifically, we compared:

  • Hyperbolic-distance-trained representations (Geo);
  • RIBin-trained representations;
  • RILog-trained representations.

All representations were evaluated using the same inference metric at inference time. We ran a fully crossed 3×3 experiment on CUB for 1-shot. All experiments used the same Conv-4 backbone and experimental settings, with the curvature fixed at c=0.1c=0.1. The only difference among the training configurations was the training metric. The results are here:

Table 1: Training kernel and inference metric cross-evaluation on CUB for 1-shot with 95% confidence interval.

Training kernel \ Inference metricGeoRILogRIBin
Geo0.582 ± 0.0020.582 ± 0.0020.582 ± 0.002
RILog0.666 ± 0.0020.666 ± 0.0020.666 ± 0.002
RIBin0.652 ± 0.0020.652 ± 0.0020.652 ± 0.002

The results reveal two important findings. First, for a specific trained representation, changing the inference metric produces virtually identical predictions (with prediction agreement of at least 99.985%). This is consistent with the monotonicity property discussed above. Second, the performance differs substantially across the three training kernels, even when the inference metric is held fixed.

These results indicate that the performance differences primarily originate from the effect of different kernels on representation learning during training, rather than from their inference-time distance transformations.

Q2. Experimental Variance and the Difference Between RILog and RIBin

Reviewer concern: The improvements over AHRad are relatively small in some settings. Moreover, RILog and RIBin exhibit unexpectedly different results despite their apparent similarity. More random seeds and variance estimates are needed.

A2: We thank the reviewer for raising these important concerns. We agree that the relatively small improvements over the strongest baseline, and we agree that the results should be evaluated across independent training runs to assess their stability. We have therefore conducted additional experiments using 5 independent random seeds and report the mean and standard deviation of the GZSL metrics.

Table 2. GZSL results (mean ± standard deviation across 5 independent random seeds, %).

DatasetMetricBaselineAHRadRIBinRILog
AWA1U29.8 ± 0.765.1 ± 0.958.8 ± 1.271.8 ± 1.0
S76.4 ± 0.483.6 ± 0.583.5 ± 0.476.4 ± 0.4
H42.8 ± 0.573.2 ± 0.769.0 ± 0.874.1 ± 0.7
AWA2U25.6 ± 0.869.1 ± 1.150.2 ± 0.974.0 ± 1.2
S76.3 ± 0.484.1 ± 0.386.5 ± 0.578.8 ± 0.5
H38.3 ± 0.675.8 ± 0.863.5 ± 0.776.3 ± 0.8
CUBU18.5 ± 0.649.2 ± 0.951.1 ± 1.049.3 ± 1.0
S44.7 ± 0.553.9 ± 0.550.2 ± 0.455.8 ± 0.5
H26.4 ± 0.551.3 ± 0.750.6 ± 0.652.3 ± 0.7

We also appreciate the reviewer's observation that the squared kernel-induced distances of RIBin converge to those of RILog as tt approaches zero. However, we would like to clarify that the closeness of kernel-induced distances does not necessarily imply that the corresponding RKHSs have similar function-space structures. We illustrate this distinction with a simple example.

Consider the following two positive-definite kernels on [−1,1][-1,1]:

\qquad k_2(x,y)=xy+\varepsilon e^{-(x-y)^2}, \qquad \varepsilon>0.$$ Their squared kernel-induced distances are $$d_{k_1}^2(x,y)=(x-y)^2, \quad d_{k_2}^2(x,y) =(x-y)^2+2\varepsilon(1-e^{-(x-y)^2}).$$ Consequently, $$\sup_{x,y\in[-1,1]} \left|d_{k_2}^2(x,y)-d_{k_1}^2(x,y)\right| \leq 2\varepsilon,$$ which tends to zero as $\varepsilon\to0^+$. Nevertheless, the corresponding RKHSs have fundamentally different dimensionalities. Specifically, $$\mathcal H_{k_1}=\operatorname{span}{x}, \qquad \dim\mathcal H_{k_1}=1.$$ By the sum-of-kernels property, $$\mathcal H_{k_1}+\mathcal H_{\varepsilon k_G}, \qquad k_G(x,y)=e^{-(x-y)^2}.$$ Since the Gaussian kernel induces an infinite-dimensional RKHS, we have $$\dim\mathcal H_{k_2}=\infty.$$ Thus, even when two kernels induce arbitrarily close squared distances, their RKHSs can differ substantially in dimensionality and function-space structure. This example illustrates that the closeness of kernel-induced distances alone is insufficient to characterize the relationship between the corresponding RKHSs. Moreover, for RILog and RIBin, their functional forms are also different. In particular, $$\lim_{t\to0}[k^{\mathrm{log}}(x,y)-k^{\mathrm{bin}}_t(x,y)]=\infty.$$ This explains to some extent why they exhibit such different results. More directly relevant to our experiments, different kernel-induced distance functions can produce different embeddings, which is consistent with our controlled cross-evaluation experiment in response A1, where models trained with RILog and RIBin achieve different performance even when evaluated using the same inference metric. This demonstrates that the choice of training kernel can affect the learned representations, rather than merely changing the inference-time distance calculation. Overall, the multi-seed results establish that the performance difference between RILog and RIBin is reproducible across independent training runs. We will incorporate these results and the above clarification into the revised manuscript to avoid conflating the mathematical convergence of their kernel distances with the empirical behavior of independently trained models.

Q3. Baseline Provenance

Reviewer concern: It is unclear which baseline results were copied from previous work and which were reproduced by the authors. The reported Hyperbolic ProtoNet results on CUB and mini-ImageNet also appear lower than those in the original paper.

A3: Thank you for pointing out this ambiguity. We agree that the provenance of the reported baseline results should be clearly documented to ensure transparency and reproducibility.

Regarding the Hyperbolic ProtoNet results on CUB and mini-ImageNet, the results reported in our original manuscript were taken from [1], rather than independently reproduced by us. We acknowledge that this distinction was not made sufficiently clear in the original manuscript.

To address the reviewer's concern, we compare the results reported in [1] with those in the original Hyperbolic ProtoNet paper [2], alongside the results of our proposed RILog method. The comparison is presented below.

Table 3. Comparison of Hyperbolic ProtoNet results reported in [1] and [2], and the results of our proposed RILog method.

DatasetSettingHyperbolic ProtoNet [1]Hyperbolic ProtoNet [2]RILog
CUB1-shot59.6 ± 0.264.0 ± 0.268.6 ± 0.2
CUB5-shot78.3 ± 0.282.5 ± 0.185.1 ± 0.2
mini-ImageNet1-shot52.7 ± 0.254.4 ± 0.256.3 ± 0.2
mini-ImageNet5-shot71.1 ± 0.272.7 ± 0.273.8 ± 0.2

As shown in the table, the Hyperbolic ProtoNet results reported in [2] are consistently higher than those reported in [1]. Nevertheless, RILog outperforms both sets of baseline results across all four evaluation settings. In particular, even when compared against the stronger results reported in the original Hyperbolic ProtoNet paper [2], RILog maintains a consistent performance advantage. Therefore, this discrepancy does not change the relative performance comparison or the conclusions drawn from our experiments.

For completeness, we also clarify that the GZSL and STS baseline results were adopted from [2]. We will explicitly document the provenance of all baseline results in the revised manuscript, distinguishing results taken from prior work from those independently reproduced by us.

We thank the reviewer again for highlighting this issue, which has helped us improve the transparency of our experimental comparisons.

References

[1] Leping Si, Meimei Yang, Hui Xue, Shipeng Zhu, and Pengfei Fang. Adaptive hyperbolic kernels: Modulated embedding in de branges-rovnyak spaces. In Proceedings of the AAAI Conference on Artificial Intelligence, 2026.

[2] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.