acceptodds
Under review as a conference paper at ICLR 2027

A Finslerian Approach for Embedding Directed Data

Abstract

Many datasets carry an intrinsic directionality: citations point backward in time, cells differentiate along lineages, and traffic follows preferred routes. Spectral embedding methods, including most of their extensions to directed graphs, discard this information: they symmetrize the data and map it into a Euclidean space where asymmetry cannot be represented. We instead model directed data as sampled from a Finsler manifold, whose distance depends on the direction of travel, and study the kernel operator built from this asymmetric distance. Through a moment expansion of this operator, we show that its symmetric and anti-symmetric parts separate geometry from direction. As the bandwidth of the kernel vanishes, the symmetric part converges to a weighted Laplacian, recovering diffusion maps in the Riemannian case, while the anti-symmetric part converges to a first-order transport operator that encodes the directionality. We prove that the corresponding graph operators, built from finitely many samples, converge uniformly and almost surely to these limits. For Randers metrics, this vector field is explicit and yields an embedding algorithm recovering both the manifold structure, from the spectrum of the symmetric part, and the underlying drift. We illustrate the approach on synthetic directed graphs and point-clouds.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation learning, Embedding, Finsler Geometry, Randers metric, Directed graphs, Directed data

Reviews

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

This paper proposes a new framework for embedding directed data based on Finsler geometry, generalizing Riemannian geometry to direction dependent distances. The authors derive continuum limits for the symmetric and antisymmetric components of a kernel operator constructed from Finsler distances. They also provide consistency results for the corresponding discrete graph operators. Building on these results, they introduce a Randers approximation that allows reconstructing an approximate Finsler metric in an embedding space. The approach is demonstrated on synthetic manifolds and directed stochastic block models.

Critical strengths

Interesting theoretical perspective. Finsler geometry provides a natural framework for modeling direction-dependent distances, and its connection to directed graph embeddings is compelling.

Substantial theoretical development. The derivation of the continuum limits and the finite-sample convergence guarantees represents a meaningful mathematical contribution.

Critical weaknesses

No quantitative comparisons are provided against existing directed embedding methods or simpler drift estimators based on the antisymmetric adjacency matrix.

All experiments are synthetic, with little quantitative evaluation beyond drift recovery.

The theoretical results assume that graph weights arise from Finsler distances, which are computed using the ground-truth metric in the synthetic experiments. The relevance of these guarantees to arbitrary directed graphs is unclear.

Questions for authors

Can the authors compare against existing directed embedding methods and a simple drift estimator based on antisymmetric edge weights? Examples: [1,2]

How applicable are the theoretical guarantees to arbitrary directed graphs whose weights are not derived from Finsler distances?

Can the authors provide an experimental validation on non-synthetic data ? One possibility could be to use a predetermined label on the data and check whether clustering with your inferred distance aligns with it.

[1] Perrault-Joncas, Dominique, and Marina Meila. "Directed graph embedding: an algorithm based on continuous limits of laplacian-type operators." Advances in neural information processing systems 24 (2011). [2] Dagès, Thomas, et al. "Finsler multi-dimensional scaling: Manifold learning for asymmetric dimensionality reduction and embedding." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

1 comment
AuthorsOpenReview ↗

Answer to Reviewer rgcD

We thank the reviewer for the time spent on our submission.

Q1. We agree, and the revised version now includes a quantitative comparison (see Section 6). We compare with the method of Perrault-Joncas and Meila (2011), and with Finsler MDS (Dagès et al., 2025). An estimator that only uses the antisymmetric edge weights gives a direction, but no magnitude relative to a metric and no metric in the embedding space. The method of Perrault-Joncas and Meila is the natural version of such an estimator, so it serves as the requested drift baseline.

We measure the cosine between the estimated and the true drift and, for the methods that provide a metric, the relative error on the squared norm of the drift in the Binet-Legendre metric. On the Swiss roll, whose asymmetry is a difference of potentials, the three methods recover the direction. The picture changes on the ocean currents of GlobCurrent (see Q3), where the asymmetry is not a difference of potentials.

Perrault-Joncas and Meila recover the direction as well as we do, but they give no metric, and the magnitude of their field depends on the bandwidth and on the units of the embedding, so that recovering a metric is not straightforward. Finsler MDS estimates the strength of the drift with a smaller error than ours when its parameter ∥ω∥\lVert \omega \rVert is well chosen (0.01 to 0.02 against 0.08 to 0.11 on the Swiss roll), but how to choose it is left open. The error rises to 0.34 when ∥ω∥=0.5\lVert \omega \rVert = 0.5 is used for a drift of strength 0.9, and our method has no such parameter to tune. On GlobCurrent, the drift of Finsler MDS is the gradient of a potential and cannot represent circulation, so its cosine is low.

Q2. The theoretical guarantees do not extend easily to arbitrary directed graphs. The method itself can be run on any directed graph, in the same way that the algorithm of Perrault-Joncas and Meila can be run on any adjacency matrix while their guarantees hold for graphs generated by their kernel. The Finsler assumption is a geometric model rather than a ground truth, and we do not claim that all directed graphs are Finsler. On the synthetic DiSBM, where there is no underlying manifold, the method still recovers an embedding with a plausible Finsler geometry. This question also arises for Riemannian manifolds and general undirected graphs. We have added a discussion of this point to the conclusion of the revised version (see Section 7).

Q3. We have added two experiments on real data in the revised version (see Section 6). In both, the data come with a real vector field but no directed distance. We therefore model the distances by Zermelo navigation in this field, which gives a Randers distance. The selected datasets are the following:

  • Ocean currents. Daily surface currents of the Agulhas region.
  • Single-cell RNA velocity. The scVelo pancreas and dentate gyrus datasets. The exact values are given in Table 2 of the revised version.

On the ocean currents, the eddies make the asymmetry of the distances different from a difference of potentials. Our drift and the one of Perrault-Joncas and Meila both recover the direction of the currents, and ours also estimate their strength, which PJM cannot do since their construction doesn't build a metric. Finsler MDS stays far from the currents, as its drift is the gradient of a potential, so that it cannot circulate along the eddies.

On the RNA velocity, differentiation is a flow along a tree, so that the asymmetry of the cells is almost curl-free and the comparison is closer. Our drift and the one of Perrault-Joncas and Meila both recover the direction of the velocity well. Ours is above theirs when we remove the density normalization (θ=0\theta = 0) and below it at θ=1\theta = 1, where the dense regions have a weak drift whose direction is degraded by the normalization. Finsler MDS depends strongly on its parameter ∥ω∥\lVert \omega \rVert and stays below the best gradient drift, so that the presence of curl does not explain its gap on these data. The cells have no ground truth for the strength, so that we only compare the direction.

We hope that these answers address the concerns of the reviewer, and we would be happy to provide further details on any point.

Reviewer_tMZ2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a Finslerian approach for embedding directed data by modeling asymmetric relationships through direction dependent distances on a Finsler manifold. The main ingredients are constructing an asymmetric kernel from Finsler distances (Sec 2.2), decomposing the normalized kernel into symmetric and antisymmetric operators (Eq 8), and using these operators to recover geometric structure and directional drift (Eq 5 and 6). The authors then construct a moment Randers approximation of the underlying Finsler metric (Def 5), which is incorporated into a directed embedding algorithm (Algo 1).

Mathematically, the central mechanism is the separation of diffusion and directional transport through the first two moments of the Finsler unit ball. The paper establishes three main theoretical results. Theorem 1 derives a moment expansion of the Finsler kernel operator, while Theorem 2 shows that the symmetric and antisymmetric operators converge to a weighted Laplacian associated with the Binet Legendre metric and a first order transport operator. Theorem 4 establishes uniform almost sure convergence of the corresponding discrete graph operators. Experiments on synthetic Swiss roll, Matsumoto manifold, and directed stochastic block model datasets illustrate the recovery of geometry and directionality.

Critical strengths

(1) Although related approaches already exist, I find the Finslerian geometric formulation interesting and potentially fruitful. In particular, I see several promising directions for extending these frameworks for the future machine learning literature.

(2) The theoretical analysis is a strong aspect of the manuscript. The characterization of the limiting operators in Theorem 2 and the finite sample convergence result in Theorem 4 are mathematically interesting and provide a principled foundation for the proposed construction.

Critical weaknesses

(1) My main concern is the limited empirical validation. The experiments are restricted to synthetic datasets, without substantial comparisons against existing methods. While the theoretical results are interesting, they do not establish whether the proposed framework provides practical advantages for directed data embedding. Given the paper's machin learning motivation, I believe a more comprehensive experimental evaluation is still necessary.

(2) I am not entirely convinced that the Finslerian formulation provides additional expressive power over existing approaches. I am not sure what additional geometric information the proposed approach can recover compared with existing methods.

(3) The authors should also position their contribution more clearly relative to recent work on Finsler geometry in machine learning. In particular, it would be helpful to clarify how the proposed framework differs from existing Finsler based embedding and graph learning methods, and what additional theoretical or practical capabilities it provides.

Questions for authors

(1) Could you clarify the precise distinction between their approach and Perrault-Joncas and Meila (NeurIPS 2011)? That work already separates diffusion and directional transport, and recovers a vector field through an embedding. The present manuscript appears to follow a similar operator based mechanism. Is the main novelty the derivation from Finsler distances and the resulting geometric interpretation, or does the proposed framework offer additional theoretical or representational capabilities?

(2) Can you provide an example where the Finslerian formulation recovers geometric or directional information that existing manifold based approaches cannot? Such an example would considerably strengthen the motivation for the proposed framework.

1 comment
AuthorsOpenReview ↗

Answer to Reviewer tMZ2

We thank the reviewer for the time spent on our submission. Our two answers also cover the three weaknesses. The positioning with respect to recent Finsler work in machine learning (W3) is addressed in Q1, while the limited empirical validation (W1) and the additional expressive power of the formulation (W2) are addressed in Q2.

Q1 (W3). The main distinction is that Perrault-Joncas and Meila (2011) do not consider Finsler geometry. Their analysis assumes a hand-crafted kernel, made of a Euclidean symmetric part and a small antisymmetric part carrying a vector field, whereas we start from an arbitrary Finsler metric. Although we follow an analogous analysis by studying a transport operator, the Finsler perspective leads to three differences.

  • Scope of the guarantees. Their kernel resembles a Randers metric, and extending their analysis to a general Finsler metric is not immediate. Their algorithm can run on any adjacency matrix, yet the guarantees are stated for graphs generated by this kernel, in which the vector field enters the construction. Our guarantees hold for graphs built from any Finsler distance and only require the distances.
  • Geometry of the embedding space. In their setting, the embedding space is Euclidean and carries a vector field, which does not necessarily define a Finsler metric. Without a metric, the magnitude of the recovered drift has no clear geometric meaning. In our setting, the embedding space is equipped with a moment-Randers metric, and we prove that the metric estimated there is exactly related to the moment-Randers approximation of the original Finsler metric.
  • Interpretation of the limits. Since our input is a Finsler metric, the symmetric and antisymmetric parts of our limit operators have a clear interpretation, namely the Laplacian of the Binet-Legendre metric and a drift, which is not available in their setting.

In short, our framework differs both in its starting point (Finsler distances) and in what it provides, namely a metric in the embedding space and limit operators with a geometric interpretation. We have made these distinctions clearer in the revised version, together with a comparison with other relevant works (see Section 1).

Q2 (W1, W2). We agree that the paper remains primarily theoretical, and the revised version includes a more quantitative evaluation (see Section 6). We compare with Finsler MDS (Dagès et al., 2025, 2026) and with the method of Perrault-Joncas and Meila (2011), and we add two real datasets, the ocean currents of GlobCurrent and the RNA velocity of single cells. When the asymmetry is a difference of potentials, as on the Swiss roll and on the Matsumoto surface, the three methods recover the direction of the drift.

The proposed approach brings more when the asymmetry circulates. The embedding space of Finsler MDS has a single constant drift ω\omega, in which d(x,y)−d(y,x)=2ω⊤(y−x)d(x, y) - d(y, x) = 2 \omega^\top (y - x). This identity follows directly from the definition of the canonical Randers space (Dagès et al., 2026, page 6, and Dagès et al., 2025), so the asymmetry sums to zero around every closed cycle, whatever the embedding dimension and the value of ∥ω∥\lVert \omega \rVert. On the cyclic directed SBM, Finsler MDS recovers the communities, but the cosine between its drift and the cycle is about 00, while ours is above 0.990.99 as soon as ∣p−q∣≥0.1\lvert p - q \rvert \geq 0.1. On the ocean currents, where the eddies make the asymmetry far from a difference of potentials, our approach recovers the true currents, while Finsler MDS struggles with this task.

We hope that these answers address the concerns of the reviewer, and we would be happy to provide further details on any point.

Reviewer_iaM8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors tackle problems in learning with directional data by embedding the data in a Finsler manifold where distances depend on the direction of travel. The study the respective graph operators and prove convergence. Furthermore they show experiments on synthetic directed graphs and point clouds.

Critical strengths

I have trouble identifying clear strengths because the authors fail to clarify how their work differs from prior work.

Critical weaknesses
  1. The idea of embedding in Finsler spaces is far from new. This was done by many of the works cited in this paper, including Lopez et al. 21, Weber et al. 24, Dages et al. 25 and 26 and many more. Even the specific choice of Finsler metrics used in this paper, namely -Randers metric and Matsumoto metric have been used in the works of Dages et al.
  2. Moreover, operator limits and Legendre transforms for linking directed parts to Finsler have been explored by Roddenberry and Baraniuk (https://arxiv.org/abs/2606.17185), a work that is not mentioned in this paper.
  3. The prior work is either not mentioned or not properly cited. For example, the proposed Finsler manifold learning task was actually introduced in Dages et al. "Harnessing data asymmetry: Manifold learning in the Finsler world". In that sense, the contribution claimed by the authors: "This work proposes a principled approach for representing directed data, grounded in Finsler geometry: we model the data as sampled from a Finsler manifold, so that the asymmetry comes from an arbitrary under- lying geometry rather than from a hand-crafted kernel, and we recover this geometry, through a Randers approximation of the metric, in the embedding space (see Figure 1)" has zero novelty. It is in fact identical with what Dages et al. did.
  4. Also, most experiments of recovering asymmetries in this fashion seem identical with the one done by Dages et al. Specifically, the examples of the swiss roll and of the mountain were already presented in the papers of Dages et al. What is the point of reinventing all these ideas and claiming them as your own?
  5. These incorrect contribution claims are then obfuscated by pages and pages of mathematical derivations that are likely designed to intimidate the reader.
  6. The experiments are quite limited. There is only one quantitative result (table 1) and it lacks comparison with baseline methods such as Finsler MDS, Finsler tsne, Finsler umap, no downstream results, no stability analysis...
  7. The authors claim a Matsumoto embedding, yet their embedding is purely Randers. In the mountain example, the inputs are Matsumoto, yet the embedding is again Randers only. In that sense, the claims are flawed as well.
Questions for authors

What exactly is your novelty compared to works by Weber et al. 24, Dages et al. 25, Dages et al. 26 and Roddenberry & Baraniuk? I fail to see any and I believe you significantly overclaimed what the novelty was (if any).

2 comments
AuthorsOpenReview ↗

Answer to Reviewer iaM8 (1/2)

We thank the reviewer for the time spent on our submission. The assessment rests on the single claim that our contribution is identical to that of Dagès et al. (2025, 2026), yet no specific result of those works is identified as corresponding to ours. We disagree, and we answer in concrete terms. In short, our framework takes an arbitrary Finsler metric as input and outputs a general Randers metric, whereas both of those works embed in the canonical Randers space with a single constant drift, which they list among their own limitations.

W1. We do not claim to have introduced the use of Finsler geometry in machine learning, and the background section of the paper cites several relevant works, including all but one of those the reviewer points at and asks for head-to-head comparisons with.

Both papers pointed out by the reviewer (Dagès et al., 2025, 2026) note as a limitation that they use the canonical Randers metric, which has a fixed drift, and not a more refined extension.

  • Dagès et al. (2025) write that "all the asymmetry is encoded in a single dimension" and suggest that other Finsler metrics might be preferable to produce disentangled asymmetric embeddings (page 8, p. 25849 in the CVPR proceedings).
  • Dagès et al. (2026) write that "We also provide a single asymmetry direction, yet applications might prefer disentangling more asymmetry directions, thus requiring other Finsler metrics." (page 13).

Regarding Matsumoto metrics, the reviewer claims that they have already been used in those works. This is not accurate. In Dagès et al. (2025), they appear only once by name in the main text (page 4, p. 25844 in the CVPR proceedings), where the paper notes that the (α,β)(\alpha,\beta) family induces Randers, Matsumoto or Kropina metrics before restricting to Randers. In Dagès et al. (2026), they are cited without being named (page 6, refs [79], [80]), only as examples of parametric Finsler metric families. Beyond citing them, neither work instantiates Matsumoto metrics in practice, so the direction taken by our work is not covered by those papers.

W2. Roddenberry and Baraniuk indeed study the convergence of a Laplacian operator, but this Laplacian differs from ours, and the symmetric and antisymmetric parts of their operator are not investigated. We view the two works as complementary and have cited that paper in the introduction of the revised version (page 2).

W3. We do not claim to have introduced the Finsler manifold learning task, and we acknowledge that the writing may have been unclear at some points. The sentence quoted by the reviewer contrasts our approach with the work of Perrault-Joncas and Meila (2011), which assumes a kernel made of a Euclidean symmetric part and a small antisymmetric part carrying a vector field, and does not take Finsler geometry into account.

We have reworked the introduction to make our contributions and their relation to prior work clearer, and we summarize them below.

  • We study the decomposition of the transport operators induced by directed graphs built from Finsler distances, without assuming a specific form for the input Finsler metric.
  • We prove that the symmetric part of these operators converges to a weighted Laplacian of the underlying Binet-Legendre metric, while the antisymmetric part converges to a drift.
  • We show that, together, they define a moment-Randers metric, which matches the first two moments of the unit ball of the input metric and admits the same limit operators, up to the reference measure.
  • We show that this moment-Randers metric is well suited to embed data in a Finslerian embedding space, notably by proposing admissibility criteria, so that the metric in the embedding is valid.
  • We prove that the metric estimated in the embedding space is exactly related to the moment-Randers metric.

To the best of our knowledge, the cited works do not provide any result on the validity of the metric in the embedding space, nor a theoretical analysis of the convergence of the antisymmetric part of the operator to a drift. The moment-Randers metric and the embedding in a general Randers space are also new. We have clarified these contributions in the revised version of the paper.

AuthorsOpenReview ↗

Answer to Reviewer iaM8 (2/2)

W4. We reproduce some experimental setups from the literature on purpose, since this highlights what is specific to our method in illustrative settings that the community may find familiar. These examples serve as proofs of concept. The Swiss roll is a standard example in manifold learning, and adding a drift to it is a natural choice.

Likewise, we do not see why setups close to those of Dagès et al. (2025, 2026) would be an issue. Since those works are important for Finsler manifold learning, as the reviewer also states, comparing within their example settings to showcase the differences is natural. Besides, neither of the two papers works with a "mountain" example. The only related experiment we found uses the US cities latitude and longitude dataset, which is indirectly biased by the altitude (Dagès et al., 2026). Beyond this, we do not see how their experiments relate to ours.

W5. Complete proofs are the minimum expected from a mainly theoretical paper, and we provide a primer on Finsler geometry in the appendix to make the paper accessible to readers unfamiliar with the topic. Moreover, the proofs come with recaps, sketches and comments that guide the reader, and the derivations are explained as they go. Regarding the intentions, we deny that the writing style was part of a plan to intimidate or trick the reader.

W6. We agree that the experiments are limited, as also noted by Reviewers tMZ2 and rgcD. Since the main focus of the paper is the theoretical analysis, the experiments were designed as proofs of concept. In response to the comments of all reviewers, we have added a more quantitative evaluation in the revised version (see Section 6). In particular, we include a comparison with existing directed embedding methods, namely Finsler MDS (Dagès et al., 2025) and the approach of Perrault-Joncas and Meila (2011).

W7. Our framework takes any Finsler metric as input, and the Randers metric is the linearization used in the embedding space. In the mountain example, the data are generated from a Matsumoto metric and the embedding is Randers by construction. The purpose of this experiment is precisely to show that the linearization performs well even though the underlying geometry is not Randers. The same holds for the directed stochastic block model, whose underlying geometry is not Randers either, and not even Finslerian. We do not claim in the paper to obtain a Matsumoto metric in the embedding space.

Q. We briefly answer the question of the reviewer, at the cost of some repetition.

  • Digraph from Finsler distances. We construct a directed graph from a kernel built on Finsler distances. To the best of our knowledge, the cited works neither propose this construction nor study its properties.
  • General framework. Our framework takes any Finsler metric as input, and we prove that the symmetric and antisymmetric parts of the resulting operator converge to the Laplacian of the Binet-Legendre metric and to a drift, respectively. In contrast, the cited works (namely Dagès et al.) focus on the canonical Randers metric, which restricts the asymmetry to a single constant drift. Moreover, our construction gives a mathematical characterization of the metric obtained in the embedding space, which is not provided in the cited works. We also show that the moment-Randers metric matches the first two moments of the unit ball of the original Finsler metric, which is, to the best of our knowledge, a new result.
  • Relation to Roddenberry and Baraniuk. Their Laplacian differs from ours, and the symmetric and antisymmetric parts of their operator are not investigated (see W2). We view the two works as complementary and cite theirs in the revised version.

We hope that these clarifications address the concerns of the reviewer regarding the contribution and novelty of our work. We would be happy to provide more details on any specific similarity that the reviewer identifies between our work and prior works.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment on the revisions

We would like to thank the reviewers for their thoughtful comments. We summarize below the main changes made to the paper, which are highlighted in purple in the revised PDF.

Additional experiments. We have added two baselines, namely the method of Perrault-Joncas and Meila (2011) and Finsler MDS (Dagès et al., 2025), and three real-life datasets, namely the ocean currents of the Agulhas region and two RNA velocity datasets. Overall, the results show that our method recovers the direction of the drift on par with Perrault-Joncas and Meila, who provide no metric, and that it also estimates the strength of the drift without a parameter to tune. Finsler MDS is competitive when the asymmetry is a difference of potentials, but its embedding cannot represent circulation, as shown on a cyclic directed SBM and on the ocean currents.

Revised related work. We have revised the related work to make our contributions and their position relative to prior work clearer, and we now cite Roddenberry and Baraniuk.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.