acceptodds
Under review as a conference paper at ICLR 2027

Optimal t-kernel shape in t-SNE

Abstract

Neighbor-embedding methods have become an established tool for visualizing complex high-dimensional datasets in various scientific disciplines. The celebrated -SNE algorithm uses the Cauchy kernel to define embedding similarity between points, often strongly outperforming the original SNE that uses the Gaussian kernel. However, a theoretical justification and an empirical investigation of the optimal kernel shape have been lacking. Here we address this gap by studying the optimal kernel shape (-distribution degree of freedom) in -SNE. We use a simplified uniform ansatz to derive the optimal kernel shape and show that the tail thickness increases with intrinsic dimensionality. Implementing a fast kernel-shape optimization in both the Barnes–Hut and the FFT-based approximations of -SNE, we qualitatively confirm theoretical predictions on simulated datasets. Finally, we conduct a large-scale empirical investigation of the learned kernel shapes across various datasets, and show that optimizing the kernel can reduce t-SNE's tendency to over-fragment unstructured, low-dimensional data and can lead to more accurate visualizations of single-cell RNA-seq data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
visualization, dimensionality reduction, manifold learning, neighbor embeddings, t-SNE, single-cell 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 4 reviews, above 27% of the others.
Reviewer_ocuVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses a method for optimizing the parameters of the similarity kernel, which defines data similarity in the output space (the low-dimensional space) during t-SNE, and examines the relationship between these parameters and the underlying latent dimensionality. Specifically, the authors propose minimizing the KL divergence between the data distributions before and after embedding with respect to α\alpha, treating α\alpha (where the t-distribution degrees of freedom ν\nu is defined as ν=2α−1\nu = 2\alpha - 1) as a tunable parameter of the Cauchy similarity kernel. The study demonstrates that this optimal α\alpha is related to the dimensionality of the low-dimensional space under certain ideal circumstances, and validates, through real data experiments, that the optimized α\alpha generally holds this theoretical relationship across various latent dimensions. Furthermore, the experiments demonstrate that when the latent dimensionality is low, the theoretically expected optimal parameter α\alpha tends to decrease rapidly when noise is introduced into the k-NN graph. This empirically supports the intuition and practical observation that the heavy-tail nature of the t-distribution is effective when noise is present.

Critical strengths

To the best of the reviewer's knowledge, there are no prior studies that parameterize the similarity kernel used in t-SNE for embedding and discuss its optimality in relation to the dimensionality. The results obtained are reasonably coherent. The finding that the relationship found between latent dimensionality and the optimal parameter is largely consistent with experimental data is beneficial. Furthermore, the analysis of how the system behaves when noise is introduced to the k-NN graph was also convincing.

Critical weaknesses
  1. The most significant issue I find with this paper is the failure to properly formulate the problem, and the overall inaccurate description.

Firstly, it is unclear what is being optimized and how. The objective function L\mathcal{L} is supposedly minimized with respect to what? It is likely that the optimization is a joint, gradient-based process over both the embedded coordinates zi\mathbf{z}_i and the parameter α\alpha. However, this is not stated in the paper. Furthermore, the notation is inconsistent. Section 3.1 states that the input data has dimension DD and the output data has dimension dd, but from Section 5 onwards (at least), DD is treated as the intrinsic dimension. Did Proposition 4.3 not show that the optimal α\alpha depends on the embedding dimension dd? Additionally, the intrinsic dimension is used without being defined, and is immediately treated as a target for estimation using the Levina & Bickel method.

A minor point, but the use of KK to mean both the number of neighbors and the kernel is sloppy. Such negligence detracts from the overall readability and comprehensibility of the paper.

  1. Issues with the theoretical section

Firstly, Theorem 1 lacks novelty; it merely states a relatively self-evident fact. Again, the inaccuracy of the description is a major problem. In the proof of Proposition 4.2, the argument—"perplexity calibration ensures 2A/D2A/D is finite; the mapping $2A/D \mapsto \operatorname{Perp}(p_m)$ is strictly monotone and injective for any DD, therefore 2A/D2A/D converges"—is incorrect or at least lacks sufficient argument. The final conclusion ("therefore") is unfounded. Injectivity alone does not guarantee the convergence of the inverse function cDc_D for each DD. Uniform convergence plus uniqueness is required. Furthermore, the meaning of "average of pi∣jp_{i|j} values" must be clarified: is it an average over the population or over the observed data? While it would be acceptable if it were over the population, the argument in this proposition is somewhat inaccurate if it applies to empirical data.

Questions for authors
  • How was the optimization performed (i.e., what was optimized and how)?
  • Is the usage of DD and dd consistent?
  • What is the definition of the Intrinsic Dimension used here?
2 comments

Dear reviewer, thank you for your comments. We appreciate your critical reading; we have updated the paper and we believe we could address all your concerns.

Firstly, it is unclear what is being optimized and how. The objective function is supposedly minimized with respect to what?

Thank you. We added an explicit statement to Section 3.2:

In what follows we consider the loss $\mathcal L(\mathbf z_i, \alpha)$ to be a function of the embedding coordinates $\mathbf z_i$ and the parameter $\alpha$, and perform gradient-descent optimization with respect to both.

Furthermore, the notation is inconsistent. Section 3.1 states that the input data has dimension D [...], but from Section 5 onwards (at least), D is treated as the intrinsic dimension.

Thank you. We renamed the ambient dimensionality in Section 3.1 to D_\text{amb}, and added the following clarification to the beginning of Section 4:

Throughout the paper, $D$ denotes the \emph{intrinsic dimensionality} of the input data, that is the dimension of the manifold on which the data approximately lie. As $t$-SNE's affinities only depend on distances to the nearest neighbors, and any smooth $D$-dimensional manifold looks locally like $\mathbb R^D$, the ambient dimension $D_\text{amb}$ plays no role in our analysis.

Did Proposition 4.3 not show that the optimal alpha depends on the embedding dimension d?

Yes indeed. To make it more explicit, we now write \hat\alpha^*(D,d) after Proposition 4.3 when we talk about finding the optimal alpha numerically.

Additionally, the intrinsic dimension is used without being defined, and is immediately treated as a target for estimation using the Levina & Bickel method.

Indeed. We now added the definition to the beginning of Section 4, see above.

A minor point, but the use of K to mean both the number of neighbors and the kernel is sloppy.

Thank you. We changed the kernel notation to TT (this only affects Section 3.2).

Theorem 1 lacks novelty; it merely states a relatively self-evident fact.

Fair enough, and we did not want to claim novelty here. We renamed this statement to Proposition. The subsequent content of Section 4 is the actual substance.

In the proof of Proposition 4.2, the argument [...] is incorrect or at least lacks sufficient argument.

We thank the reviewer for spotting this gap. Injectivity alone indeed does not imply convergence. We now prove convergence in a short lemma in the Appendix (Lemma~A.1): since the perplexity is strictly decreasing in 2A/D2A/D and continuous in 2/D2/D, the calibrated value 2A/D2A/D is squeezed into any neighborhood of β\beta for large DD.

Furthermore, the meaning of "average of p_{i|j} values" must be clarified

Thank you. pmp_m denotes the population average (expected value); we now state this explicitly, and clarify that the approximation consists of replacing the random neighbor distances by their typical values.

We hope that our edits make the manuscript more precise and address your concerns! As you appreciated the novelty and convincingness of our results, we hope that you will be willing to revisit your score.

Thank you for reply to my questions and explanation to the point I enumerated in the "weakness". The notational problems are revolved. Proposition 4.2 seems to become accurate. Overall, the readability and accuracy of the paper have been improved.

This work has certain amount of novelty, and the coherence between theory and experiments.

Considering that the improved readability and soundness, in addition to the contribution itself, and taking into account the upcoming inter-reviwer discussion/cross-check, I will consider the score again.

Reviewer_4312OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This study establishes a theoretical connection between the intrinsic manifold dimension of high-dimensional data and the optimal kernel shape—parameterized by the degrees of freedom of a tt-distribution—within the tt-SNE framework. Although derived under the simplifying assumption of a locally uniform distribution, the theoretical result aligns remarkably well with real-world datasets in numerical evaluations. Furthermore, the necessity of parameter optimization is demonstrated using noisy synthetic datasets generated from a Swiss roll manifold.

Critical strengths

The theoretical analysis provided by the authors is intriguing, as it aligns surprisingly well with real-world data despite relying on simplifying assumptions. This finding is crucial for elucidating the key performance differences between t-SNE and standard SNE with a Gaussian kernel.

Critical weaknesses
  1. Several relevant studies are missing from the literature review, as detailed in the "Questions for Authors" section.

  2. Section 6 offers an interesting analysis; however, its connection to the primary theoretical findings is unclear. If a connection exists, the authors should provide a more explicit and thorough discussion.

Questions for authors
  1. The following two related papers should be cited, and their relevant results should be discussed, because both papers examine the relationship between the intrinsic dimension and the optimal kernel shape, while proposing extended methods by replacing a Gaussian kernel in the original dimension with another kernel.

[1] De Bodt, Mulders, Verleysen, Lee (2018)「Perplexity-free t-SNE and twice Student tt-SNE」ESANN 2018, pp.123–128

[2] Schubert & Gertz (2017)「Intrinsic t-Stochastic Neighbor Embedding for Visualization and Outlier Detection — A Remedy Against the Curse of Dimensionality?」SISAP 2017, pp.188–203(doi:10.1007/978-3-319-68474-1_13)

  1. Figure 5 shows a consistent gap between theoretical predictions and empirical results. Further discussion regarding this point, including the potential causes/origins of this gap, is required.

  2. Is there a connection between the main theoretical result and the analysis in Section 6? If so, please clarify how they are related.

1 comment

Dear reviewer, thanks a lot for your comments and the positive evaluation! We really appreciate that you liked our theoretical analysis that indeed ``aligns surprisingly well with the real-world data despite relying on simplifying assumptions''!

We have updated the paper and we believe we could address all your points.

The following two related papers should be cited, and their relevant results should be discussed, because both papers examine the relationship between the intrinsic dimension and the optimal kernel shape.

Thanks a lot for mentioning these papers! We added the following paragraph to Related Work:

Two other papers connect similarity kernels to intrinsic dimensionality, but both change the the \emph{high-dimensional} affinities while leaving the Cauchy embedding kernel unchanged. \citet{debodt2018perplexity} (tt-SNE) replaced the Gaussian high-dimensional affinities with $t$-distributed ones, setting the degree of freedom from an estimated intrinsic dimensionality. \citet{schubert2017intrinsic} (it-SNE) applied a per-point transform to high-dimensional distances based on estimated intrinsic dimensionality. In contrast, we leave the high-dimensional affinities unchanged (at Gaussian or even uniform on the $k$NN graph) and \emph{optimize} the shape of the embedding kernel.

Figure 5 shows a consistent gap between theoretical predictions and empirical results. Further discussion regarding this point, including the potential causes/origins of this gap, is required.

Our theoretical predictions rely on strong assumptions, so we find some difference between the actually observed α⋆\alpha^\star values and our idealized predictions not surprising. What is more important, in our opinion, is that the predictions hold up qualitatively: α⋆\alpha^\star decreases and eventually plateaus with increasing intrinsic dimensionality. We have now clarified in the text that we focus on qualitative agreement here.

Section 6 offers an interesting analysis; however, its connection to the primary theoretical findings is unclear. If a connection exists, the authors should provide a more explicit and thorough discussion.

We added the following clarification to the beginning of Section 6:

While our theory does not explicitly consider low-dimensional $k$NN graphs with added noise, we expect them to behave similarly to higher-dimensional $k$NN graphs in the sense that the graph structure cannot be fully represented in 2D and some edges (here, spurious edges) need to be `torn'.

In response to another reviewer, we also included additional clarification regarding the strength of attractive forces as a function of α\alpha to Section 3.1. Together, we hope that these edits provide a clearer connection to Section 6.

We hope that our revision addresses your concerns!

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

This paper addresses a long-standing issue in the fields of dimensionality reduction and data visualization: how to appropriately select the shape of the kernel function used in neighborhood embedding methods. Widely used techniques such as SNE and t-SNE characterize sample similarities in a low-dimensional embedding space via kernel functions, subsequently minimizing the Kullback–Leibler (KL) divergence between the similarity distributions in the high-dimensional and low-dimensional spaces. However, there remains a lack of robust theoretical grounding or standardized criteria for choosing the shape of these kernel functions and their associated parameters.

To resolve this, the authors propose an approach that adaptively learns the shape parameters of the kernel function during the t-SNE optimization process. They provide two efficient implementations of this method, based on the Barnes–Hut algorithm and the Fast Fourier Transform (FFT), respectively. Furthermore, the paper analyzes the relationship between optimal kernel parameters and the intrinsic dimensionality of the data through simplified assumptions, and validates the approach using both synthetic and various real-world datasets—including single-cell transcriptomic data. Overall, the experimental results demonstrate that this method can improve the visualization performance of traditional t-SNE, particularly by better preserving continuous data structures and reducing excessive fragmentation.

Critical strengths

This paper proposes a straightforward and intuitive scheme for optimizing kernel parameters. Specifically, the authors derive the gradient of the Kullback–Leibler (KL) divergence with respect to the shape parameters of the kernel function. By jointly optimizing these parameters alongside the low-dimensional embedding coordinates, the kernel shape can be determined adaptively via gradient descent, eliminating the need for pre-specified, fixed parameters. Furthermore, to address the high computational complexity involved in gradient calculation, the authors develop accelerated implementations based on the Barnes–Hut algorithm and the Fast Fourier Transform (FFT), enabling the method to be applied efficiently to large-scale data visualization tasks.

Critical weaknesses

1.The methodological innovation requires further clarification. Kobak et al. (2019) have already investigated the use of kernels with different tail shapes in t-SNE and their impact on visualization results. Building on this, the primary methodological contribution of the current paper appears to be the derivation of gradients for optimizing kernel parameters, along with an efficient computational implementation. While these improvements offer practical utility, the extent of their innovation at a methodological level warrants further discussion. The authors are encouraged to articulate more clearly the core technical contributions of this work relative to existing literature. 2.The practical significance of the research problem requires more robust validation. Although t-SNE is a widely used visualization algorithm, other non-linear dimensionality reduction methods—such as UMAP—are also extensively employed. This paper focuses on the optimization of kernel parameters within t-SNE, yet it does not sufficiently demonstrate the unique advantages this optimization offers compared to other existing visualization techniques. Consequently, the authors are advised to conduct more comprehensive experimental comparisons to further validate the necessity of the proposed method in practical data analysis, as well as its value relative to other mainstream dimensionality reduction approaches. 3.There is still room to enhance the comprehensiveness of the experimental evaluation. Although the authors conducted large-scale experiments across datasets from various domains, the focus remained primarily on the performance of t-SNE under different kernel parameter settings; there is a lack of sufficient quantitative comparison against other competitive dimensionality reduction and visualization methods. Furthermore, the current evaluation relies heavily on Kullback-Leibler (KL) divergence and qualitative visual results. It is recommended to incorporate additional quantitative metrics—such as those reflecting the preservation of local neighborhoods, the maintenance of global structure, and the stability of the embedding—to provide a more thorough assessment of the method's practical advantages.

Questions for authors

The article notes that standard t-SNE tends to produce excessive fragmentation when applied to data with low intrinsic dimensionality, whereas SNE may lead to significant overlap between distinct data clusters when handling high-dimensional data. The authors attribute these phenomena to the tail characteristics of the kernel functions; however, the precise mechanisms underlying this relationship remain insufficiently explained.

Specifically, could the authors elaborate on how the shape of the kernel function influences the balance between attractive and repulsive forces within the low-dimensional embedding, ultimately resulting in the aforementioned visualization outcomes? Furthermore, to what extent are these phenomena determined by the kernel function’s shape, as opposed to being significantly influenced by parameters such as perplexity, initialization methods, or other optimization hyperparameters? It is recommended that the authors further clarify the interplay among these factors and consider conducting targeted ablation studies to isolate and verify the specific role played by the kernel function’s shape.

3 comments

Dear reviewer, thanks a lot for your comments. We have updated the paper and we believe we could address all your points.

The methodological innovation requires further clarification.

In Kobak et al. (2019), the tail shape is treated as a freely adjustable hyperparameter (similar to perplexity), that the user can change if they want to focus on what the authors call "finer cluster structure". In contrast, our approach here is to optimize this parameter. In this sense our work is very different from Kobak et al. (2019). We have now clarified it in the Related Work section.

Methodologically, our innovation was indeed to implement alpha learning into the Barnes--Hut and FFT approximations. This was never done before. This is precisely what allowed us to perform the large-scale empirical study of 600+ datasets.

Also, please note our theoretical analysis that we believe to be novel.

The practical significance of the research problem requires more robust validation. Although t-SNE is a widely used visualization algorithm, other non-linear dimensionality reduction methods—such as UMAP—are also extensively employed.

Thank you for raising this point. We now include two new supplementary figures showing UMAP embeddings on the datasets used throughout the paper (Fig. S4) and UMAP's behavior in the presence of noise (Fig. S5). We also report two quality measures (kNN recall and Gap Index; see below) for a quantitative comparison. Our analysis shows that on the Swiss Roll, UMAP also suffers from overfragmentation (Fig S4), achieves substantially worse kNN recall, and is considerably more sensitive to noisy kNN graphs (Fig S5).

Although the authors conducted large-scale experiments across datasets from various domains, the focus remained primarily on the performance of t-SNE under different kernel parameter settings; there is a lack of sufficient quantitative comparison against other competitive dimensionality reduction and visualization methods.

We have now added an explicit comparison to UMAP, see above. We also extended the discussion of UMAP in Related Work. t-SNE is known to be by far the best method in terms of local structure preservation (as measured e.g. by kNN recall; see review by de Bodt & Diaz-Papkovich et al., arXiv 2025).

It is recommended to incorporate additional quantitative metrics—such as those reflecting the preservation of local neighborhoods [...]

Thank you for this suggestion. We used the kNN recall (with k=100), a widely used measure of local structure preservation, and the gap index (GI), a recently proposed measure to quantify distortions in low-dimensional embeddings (Ros et al. 2026, Measuring Distortion in the Empty Regions of Dimensionality Reduction Scatterplots with the Gap Index). GI is particularly sensitive to the extent with which the empty regions in the embedding also exist in the high-dimensional space and is therefore well suited to quantify overfragmentation.

We now report both measures in Figs. 1 and 6, as well as in the newly added supplementary figures. Optimizing α\alpha improves the measures (esp. the GI) and t-SNE with optimized α\alpha outperforms UMAP.

Specifically, could the authors elaborate on how the shape of the kernel function influences the balance between attractive and repulsive forces within the low-dimensional embedding, ultimately resulting in the aforementioned visualization outcomes?

Thank you for this question. The problem lies not in the balance between the attractive and repulsive forces, but in how the attractive force depends on the embedding distance. In SNE, the attractive force between the two points that share a kNN edge scales as F(d)∝dF(d) \propto d. This strong attractive force leads to a good embedding of Swiss roll (Figure 1), but is very sensitive to noise (Section 6) leading to what the 2008 t-SNE paper called "crowding problem" on datasets like MNIST. In t-SNE, the attractive force scales as F(d)∝1/(1+d2)F(d) \propto 1/(1+d^2) which goes to zero for large dd. This allows some neighbors to drift away, which is however suboptimal on datasets like Swiss roll.

We have now added this to the end of Section 3.1 and also to the Discussion.

Furthermore, to what extent are these phenomena determined by the kernel function’s shape, as opposed to being significantly influenced by parameters such as perplexity, initialization methods, or other optimization hyperparameters?

In all our experiments comparing different α\alpha values we use the same perplexity, same initialization, and same optimization engine and hyperparameters. All differences therefore must be due to the α\alpha alone.

We hope that our revision addresses your concerns and you will be willing to revisit your score!

Dear reviewer, there are only some hours left until the end of the discussion period. We would really appreciate your response! Let us know if you have any remaining concerns.

Thank you for the detailed rebuttal and the additional experiments. I appreciate the authors' efforts to address my concerns, particularly the newly added comparisons with UMAP, the quantitative evaluations using kNN recall and the Gap Index, and the clarification of how kernel shape affects the attractive forces in the embedding. Although the additional experiments provide evidence of practical improvements, I remain uncertain whether these gains are sufficiently substantial and generalizable to justify the contribution, particularly given the simplifying assumptions underlying the theoretical analysis. Overall, I acknowledge the improvements made during the rebuttal, but I do not think they fully resolve my concerns about the novelty and broader significance of the work. Therefore, I would prefer to maintain my original rating of 2 (Weak Reject).

Reviewer_myB1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the optimal shape of the t-kernel in t-SNE, that is, the degree of freedom of the t-distribution used for embedding similarity. The authors turn this degree of freedom α into a learnable parameter, derive a closed form for ∂L/∂α, and implement fast optimization of it under both the Barnes-Hut and the FFT approximations. On the theory side they work from a simplified uniform model and derive an optimal α̂* = βd/2 that falls as intrinsic dimensionality rises. On the empirical side they measure the learned α across 633 real datasets, and add two single-cell examples plus an experiment that injects noise into the kNN graph.

Critical strengths

S1. The question is well chosen and has genuinely been left open. The original justification for the Cauchy kernel in t-SNE was heuristic from the start, and van der Maaten himself noted in 2009 that the same argument applies equally well to any degree of freedom. S2. I checked the theory independently, and the formulas, all four reported values of β and the α̂* curve reproduce. S3. The numerical implementation is better than what I usually see in this area. The tests compare the analytic gradient against finite differences and against an exact computation, so they were not written merely to make the code run. S4. The learned α does behave like a property of the data. I ran my own control with random and PCA initialization, and the two agree closely. S5. The authors state plainly in the discussion that for many real datasets the optimal degree of freedom is close to the default, which amounts to conceding that the improvement is not universal.

Critical weaknesses

W1. The largest selling point of the paper is the study over 633 datasets, yet the text reports no correlation coefficient, no goodness of fit and no error bars, so the claim that α* is inversely related to intrinsic dimensionality rests on eyeballing Fig. 5. From the supplementary material it looks like the authors did compute how many datasets fall between the theoretical prediction and the measured curve, and that number simply never made it into the paper. Runs whose learned α came out very low are also excluded altogether, while the text only mentions excluding simulated datasets and near duplicates and never says how many runs were dropped. If these numbers are added I would have no remaining objection on this point, and I would be willing to raise my score accordingly. W2. Both the text and the Fig. 5 caption say the real datasets were downsampled to 5 000 points, but the supplementary material actually uses 100 000. This matters, because the learned α depends on the sample size. Please state the number of points actually used and add a sensitivity analysis of α with respect to it. W3. In Fig. 6 the judgment that the embedding became more accurate after learning α rests entirely on visual inspection plus appeals to prior literature, with no computable metric. The human testis data was subsampled to 20 000 cells, while the core argument of Section 6 is precisely that kNN graph noise pulls the optimal α toward 1, and subsampling changes that noise level. No sensitivity analysis over cell number is given. Both examples are also continuous developmental series, which is exactly the regime where the theory already predicts a larger α, so validating a prediction on the special case it was built for carries limited force. Adding one quantitative measure would let me accept a raised score here as well.

Questions for authors

Q1. Please address the weaknesses above as thoroughly as you can, and indicate which of them you can resolve during the rebuttal period and which would have to wait for a revision. Q2. The theoretical prediction is independent of sample size, while the measured α is not. On what basis then does Fig. 5 compare the theoretical curve, the hypercube curve and the real datasets in one plot? Q3. typo error: l.60 "yields adequate embeddings datasets" -> "yields adequate embeddings for datasets" l.185-186 "This partial derivative can be can expanded" -> "This partial derivative can be expanded" l.412 "the embedding combined togther all Sst types" -> "the embedding combined together all Sst types"

2 comments

Dear reviewer, thanks a lot for your comments. We really appreciated your positive remarks and your verifications. We have updated the paper and we believe we could address all your points.

The largest selling point of the paper is the study over 633 datasets, yet the text reports no correlation coefficient

Thank you for the suggestion. We now report Spearman's correlation coefficients in Fig 5. Correlations are -0.76 in (a) and -0.84 (b), both very significant given n > 600 (p << 0.0001).

Runs whose learned α came out very low are also excluded altogether

We thank the reviewer for bringing this to our attention. Reviewing our exclusion criteria revealed a bug in the code, and we found several datasets containing duplicate rows that resulted in degenerate learned values of alpha. Fixing the bug fixed all degenerate runs, and the results now contain far fewer outliers than our initial submission. We have therefore removed the exclusion criteria entirely. While some of the runs are still not shown in the plot due to our choice of x- and y-limits (2 datasets in 2D and 5 datasets in 3D), these were not excluded when calculating the correlation.

Both the text and the Fig. 5 caption say the real datasets were downsampled to 5 000 points, but the supplementary material actually uses 100 000.

Thank you for catching this error. When conducting the experiments, we had initially subsampled to 5,000 points, but in the final figure, we actually used 100,000 points. We have now corrected all mentions of this to 100,000 points.

add a sensitivity analysis of α with respect to it. [...] The theoretical prediction is independent of sample size, while the measured α is not. On what basis then does Fig. 5 compare the theoretical curve, the hypercube curve and the real datasets in one plot?

We now investigated the effect of sample size on the learned degree of freedom in Fig. S3. We used points sampled from the 5-dimensional hypercube with sample sizes ranging from 1,000 and 100,000. We found only a very weak dependence of the alpha-hat on the sample size, suggesting that there is no problem with Figure 5 (where real datasets have different sample sizes).

In Fig. 6 the judgment that the embedding became more accurate after learning α rests entirely on visual inspection

Agreed. We now include two complementary measures of embedding distortion to quantify embedding quality. We used the kNN recall (with k=100), a widely used measure of local structure preservation, and the gap index (GI), a recently proposed measure to quantify distortions in low-dimensional embeddings (Ros et al. 2026, Measuring Distortion in the Empty Regions of Dimensionality Reduction Scatterplots with the Gap Index). GI is particularly sensitive to the extent with which the empty regions in the embedding also exist in the high-dimensional space and is therefore well suited to quantify overfragmentation. In Fig. 6 the GI drops from 0.74 to 0.60 in the human testis dataset and from 0.72 to 0.56 in the mouse cortex dataset (lower is better) when learning the α\alpha. This example, along with those in Fig. 1, show that our approach produces more faithful embeddings.

We now discuss these results in the last paragraph of Section 5 and include a description of both metrics in the appendix.

Typos

Thanks a lot, we fixed all these (and also some other) typos!

As you wrote that you would be willing to raise your score if the issues are addressed, we hope that you will consider them fixed!

Dear reviewer, there are only some hours left until the end of the discussion period. We would really appreciate your response, especially because you wrote that you would be willing to raise your score if the issues are addressed. We hope you will consider them addressed!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal summary

We were happy to see that the reviewers agreed that the question we study is important and has been open since the original t-SNE paper, and that our theory "aligns surprisingly well with real-world data". In response to the raised issues, we did the following improvements:

  • Added quantitative evaluation (kNN recall and gap index) supporting our conclusions;
  • Added comparison to UMAP;
  • Added quantification (Spearman correlations) to the empirical study;
  • Confirmed that the sample size does not strongly affect the tail shape;
  • Clarified notation, definitions, and problem statement throughout the paper;
  • Expanded the incomplete step in the proof of Prop. 4.2;
  • Expanded discussion of related work.

We believe that these changes strengthened the paper and addressed all of the raised concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.