acceptodds
Under review as a conference paper at ICLR 2027

Whitening Improves Robustness to Spurious Correlations in Linear Probes

Abstract

Deep neural networks tend to rely on simple features that may be spurious and thus fail to generalize. We study this problem in the setting of linear probes, where a (generalized) linear model is fitted on the representations of a (pretrained) model. We use the connection of these models to the max-margin classifier, and show they favor directions associated with large eigenvalues of the covariance matrix. Whitening removes this preference by equalizing the eigenvalues of the covariance matrix. This observation motivates whitening as a preprocessing step that can reduce reliance on spurious correlations without requiring prior knowledge of their presence or labeled data. We examine the effect of whitening on a synthetic data-generating process and standard spurious correlation benchmarks, and find that it improves robustness. We also find that whitening can improve robustness when added to existing approaches.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Spurious correlations, Linear probes, Whitening, Robustness

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_jvojOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the problem of spurious correlations in linear probes, where a classifier may rely on simple but irrelevant features instead of the actual features needed for classification. The main observation is that linear classifiers tend to favor feature directions associated with large eigenvalues of the representation covariance matrix, and these high-variance directions may correspond to spurious features. To address this, the authors apply whitening to the learned representations before training the linear probe. Whitening equalizes the eigenvalues and removes the classifier's preference for high-variance directions. The authors provide theoretical analysis of this behavior and evaluate the approach on synthetic data and several spurious-correlation benchmarks, including Waterbirds, CelebA, and MultiNLI. The experiments show that whitening can substantially improve worst-group accuracy in some settings and can also complement existing approaches for handling spurious correlations.

Critical strengths

The main observation of the paper is interesting: linear probes can favor feature directions associated with large eigenvalues of the covariance matrix, which can increase their reliance on simple or spurious features. Based on this observation, the paper proposes a simple solution using whitening to remove this preference. I also find it interesting that such a relatively simple preprocessing step can provide substantial improvements on some spurious-correlation benchmarks without requiring prior knowledge of the spurious features or additional labels.

Critical weaknesses

My main concern is that whitening changes the geometry of the representation by equalizing the eigenvalues of the covariance matrix. While this can help when directions with large eigenvalues correspond to spurious features, there is no guarantee that this is always the case. High-variance directions may contain useful information for the actual task, and suppressing their relative importance could therefore hurt performance on data where spurious correlations are absent or do not align with these directions. The weaker results on MultiNLI already suggest this limitation. I would like to see a more extensive evaluation on standard datasets without intentionally introduced spurious correlations to better understand when whitening helps and when it can hurt.

Questions for authors

How does whitening affect performance on standard datasets where there are no known or intentionally introduced spurious correlations? Could equalizing the eigenvalues hurt performance by reducing the importance of genuinely useful high-variance features?

Is there a way to determine when large-eigenvalue directions are likely to represent spurious features versus useful features before applying whitening?

The results on MultiNLI suggest that whitening is not always beneficial. Can the authors provide more insight into what properties of a representation or dataset determine whether whitening will help?

Have the authors considered partial or adaptive whitening, rather than forcing all directions toward equal variance, to preserve useful representation geometry while reducing the dominance of unusually large eigenvalues?

3 comments

We thank the reviewer for their thoughtful feedback, as well as their suggestions for improvement, which we found very valuable. We also thank them for commenting that it is `interesting that such a relatively simple preprocessing step can provide substantial improvements on some spurious-correlation benchmarks'. Below, we address the feedback point-by-point.

My main concern is that whitening changes the geometry of the representation by equalizing the eigenvalues of the covariance matrix. While this can help when directions with large eigenvalues correspond to spurious features, there is no guarantee that this is always the case.

We thank the reviewer for highlighting this limitation of our work. We agree that there is no guarantee that directions associated with large eigenvalues correspond to spurious features. We explicitly acknowledge this limitation in the paper (lines 394–397 and 525–529), but agree this deserves more attention. We have thus updated our introduction (lines 52-55 in the new manuscript) to highlight this.

Our paper establishes that the max-margin classifier has a preference for features associated with large eigenvalues of the covariance matrix, and that whitening removes this preference. This holds regardless of whether these features correspond to spurious or core features. However, whether whitening improves robustness depends on the extent to which these high-variance directions correspond to spurious features. When they contain useful information for classification, we agree that whitening may hurt performance.

As discussed in Sections 2 and 3, a substantial body of previous work relates the reliance of DNNs on spurious correlations to simplicity bias [1–6]. This provides some motivation for expecting spurious features to be associated with high-variance directions, and thus that at least in some cases (as shown in the paper) that whitening can improve robustness to spurious correlations.

[1] Shah et al. (2020) https://proceedings.neurips.cc/paper/2020/hash/6cfe0e6127fa25df2a0ef2ae1067d915-Abstract.html

[2] Pezeshki et al. (2021) https://proceedings.neurips.cc/paper_files/paper/2021/file/0987b8b338d6c90bbedd8631bc499221-Paper.pdf

[3] Teney et al. (2022) https://openaccess.thecvf.com/content/CVPR2022/html/Teney_Evading_the_Simplicity_Bias_Training_a_Diverse_Set_of_Models_CVPR_2022_paper.html

[4] Vasudeva et al. (2024) https://arxiv.org/abs/2310.06161

[5] Yang et al. (2024) https://proceedings.mlr.press/v238/yang24c.html

[6] Qiu et al. (2024) https://proceedings.mlr.press/v235/qiu24a.html

The results on MultiNLI suggest that whitening is not always beneficial. Can the authors provide more insight into what properties of a representation or dataset determine whether whitening will help?

We believe that two properties are particularly relevant.

First, the original classifier should exhibit a substantial preference for simple features, as defined in Definition 1. This is investigated empirically in Figure 4. If the classifier does not exhibit such a preference, whitening likely has little effect.

Second, and crucially, the directions favored by this preference should align with spurious features. In this case, removing the preference can improve robustness.

Figure 4 suggests that the first property holds for MultiNLI, but this does not translate into a clear improvement in worst-group accuracy. One possible explanation we provide in the paper is that the high-variance directions in MultiNLI are less strongly aligned with spurious features than in the other benchmarks. We emphasize, however, that we have not directly measured this alignment.

High-variance directions may contain useful information for the actual task, and suppressing their relative importance could therefore hurt performance on data where spurious correlations are absent or do not align with these directions. The weaker results on MultiNLI already suggest this limitation.

How does whitening affect performance on standard datasets where there are no known or intentionally introduced spurious correlations? Could equalizing the eigenvalues hurt performance by reducing the importance of genuinely useful high-variance features?

I would like to see a more extensive evaluation on standard datasets without intentionally introduced spurious correlations to better understand when whitening helps and when it can hurt.

We agree that this is a possibility, and have added several experiments to the paper to investigate this. First, we have added an experiment where we create a version of Waterbirds where the spurious correlation is absent - e.g. there is no correlation between the background type and bird type. In this experiment, whitening does not lead to a statistically significant reduction in performance.

Second, we have added an experiment on CIFAR-10, where we compare re-training the final layer with no preprocessing, standardization, and whitening, from a finetuned Resnet50. This is a standard benchmark without an intentionally introduced spurious correlations. We again observe no statistically significant reduction in performance from whitening.

Third, we vary the scale of the spurious feature in the synthetic experiment and observe an asymmetry. Whitening substantially improves performance when the spurious feature has a large scale, while having little effect when its scale is equal to or smaller than that of the core feature. This provides further insight into the circumstances under which whitening can help, and why otherwise it does not degrade performance (see the new paragraph near line 275 for details).

These three experiments show no signs that whitening reduces performance in these datasets. While we cannot guarantee that the concern of the reviewer never is an issue, they at least provide empirical evidence that it is a limited risk.

Is there a way to determine when large-eigenvalue directions are likely to represent spurious features versus useful features before applying whitening?

We thank the reviewer for this suggestion. This is also something we considered ourselves. Ideally, one could measure the extent to which spurious features (e.g. backgrounds) align with high eigenvalues, and then compare this to the extent core features (e.g. bird characteristics) align with high eigenvalues. However, we do not deem this viable for two reasons. First, it is very difficult to unambiguously identify separate directions in the feature space associated with the spurious and core feature. While attempts at this have been made (see [7]), these methods do not guarantee directions solely associated with either of the features. On top of that, there is also no guarantee that there exists a unique direction representing either feature. Second, even if we knew how to determine the directions representing spurious and core features, it is impractical to estimate these directions; this likely would require pre-existing knowledge of the spurious correlation as well as the availability of group labels. As such, we believe that when group labels are available for the validation set, it is best to use worst-group accuracy to measure whether whitening improves robustness. Otherwise, class-balanced accuracy is the best alternative, although it can miss performance on (small) subgroups. A careful examination as to how one can select the best preprocessing approach (none, standardization or whitening) beforehand is outside the scope of the paper.

[7] Holstege et al. https://arxiv.org/abs/2310.11991

Have the authors considered partial or adaptive whitening, rather than forcing all directions toward equal variance, to preserve useful representation geometry while reducing the dominance of unusually large eigenvalues?

We agree that partial whitening is an interesting approach, and has the potential to address some limitations of the paper.

Our analysis focuses on full whitening, as this allows us to isolate the effect of removing the max-margin classifier's preference for high-variance directions. A systematic investigation of partial whitening is beyond the scope of this paper, as partial whitening brings its own challenges and complexities - mainly, how one can determine the optimal amount of whitening beforehand. We have added partial whitening as a suggestion for future work in Section 6.

Reviewer_jvojOpenReview ↗

I am okay with acceptance

Considering the response and other reviews. I am okay to accept.

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

This paper first frames the problem of deep neural networks typically preferring / training towards simpler features, as defined by features associated to large covariances in the training data. This tendency creates a bias known as simplicity bias, hypothesized to be correlated to spurious features (such as image backgrounds, etc.). The paper is well framed in terms of existing methods of addressing this bias and uses straighforward mathematical justification (Sect. 3) to show that whitening training features (i.e., removing linear relationships / correlations between features) directly eliminates the tendency of models to fit towards simple features and avoids the resulting bias. Models and training are all based on Linear probes (linear classification of deep features). A toy example on synthetic data (Sect. 4) shows that whitening indeed works as intended and removes sensitivity to the scales of the spurious features (and seems to be tolerant to noise with the use of the L&W nonlinear shrinkage estimator). Empirical evaluation on Waterbirds and CelebA (using ImageNet-pretrained ResNet-50), and on MultiNLI (using BERT) demonstrates whitening dramatically improves the accuracy (especially worst-group accuracy) for the two image datasets (Waterbirds and CelebA), whereas the change is insignificant on the MultiNLI experiment. Whitening also does better than Standardization, and in Fig. 4 it is shown that whitening results in far less simplicity. The insignificance on MultiNLI may be a result of the feature chosen to be considered spurious (presence of a negation word in the hypothesis). An acknowledged limitation is that this method assumes that any simple features (with large eigenvalues) are spurious and undesired, which may not always be true.

Critical strengths
  1. Excellent mathematical explanation of problem and justification of solution! Furthermore, the synthetic data and practical experiments nicely demonstrate both the indented behaviour as well as cases where the whitening approach may not be suitable.
  2. Well laid-out paper, easy to read and understand.
Critical weaknesses
  1. There are a few cases of notation issues surrounding Eq. 16, which makes that subsection difficult to follow. See my clarifying questions below.
Questions for authors

Clarification:

  1. I'm a little unsure about notation surrounding Eq. 16. What is xax_a? I don't think it is defined. I assume this is meant to be xcx_c? I think there may be a few minor notation issues there. Please double-check notation related to aa, ss, and cc. Clarify what dd and qq are. I assume dd is the dimension of the feature vector xx, and qq is the number of noise features?
  2. What is y^\hat{y} in Fig. 1? Is this meant to be aa?
  3. For Fig. 2, does large qq not imply small noise, based on Eq. 16, where qq is seen in the denominator of the covariance of ϵ\epsilon ? I do understand (with moderate confidence) that qq is the number of noise features, and therefore there are many noise dimensions with large qq.
  4. In Fig. 2, what is the "Worst Group"? I don't think you defined this term beforehand. I think you define it well enough on line 334, later. Is "Equal-group Accuracy" (Fig. 3) the same as "Overall Accuracy" (Fig. 2)?
  5. I'm confused by line 471 "Notably, in our results whitening only worsens the linear probe in a comparison.". Does whitening not generally improve the linear probe? Did you mean standardization worsens the linear probe?
4 comments

We thank the reviewer for their feedback and for their positive remarks about the paper. In addition, we thank them for their remark that the paper has an `Excellent mathematical explanation of problem and justification of solution'. We address the feedback point-by-point below.

There are a few cases of notation issues surrounding Eq. 16, which makes that subsection difficult to follow. See my clarifying questions below. I'm a little unsure about notation surrounding Eq. 16. What is xax_a ? I don't think it is defined. I assume this is meant to be xcx_c? I think there may be a few minor notation issues there. Please double-check notation related to a,ca,c, and ss.

We thank the reviewer for allowing us to clarify the notation in Eq. 16. xsx_s is the spurious feature, which is obtained by multiplying the random variable xax_a with γ\gamma, giving xs=γxax_s = \gamma x_a. In hindsight, we agree that this notation is somewhat confusing. We have updated the Eq. 16 as follows:

x_c \mid y \sim \mathcal{N}(y,\sigma_y^2), \quad x_s \mid a \sim \mathcal{N}(\gamma a,\gamma^2\sigma_a^2), \quad \mathbf{\epsilon} \sim \mathcal{N}\left( \boldsymbol{0}, \frac{\sigma_\epsilon^2}{q}\boldsymbol{I}_q \right). \end{aligned}$$ and have removed the notation $x_a$ from the paper. > Clarify what $d$ and $q$ are. I assume $d$ is the dimension of the feature vector, and $q$ is the number of noise features? This is correct. In order to make this clearer, we have updated the text in the paper to read as follows: *“Here, $d$ denotes the number of features, and $q=d-2$ denotes the number of independent noise features. The first two coordinates, $x_c$ and $x_s$, are the ‘core’ and ‘spurious’ features, respectively.”* > What is $\hat{y}$ in Fig. 1? Is this meant to be $a$? Here, $\hat{y} = \mathrm{sign}(\boldsymbol{x}^{\top}\boldsymbol{w})$ is the prediction of the model. We understand that this might be confused for $a$ because of the quadrant. We have updated the caption of Figure 1 to clarify this. > For Fig. 2, does large $q$ not imply small noise, based on Eq. 16, where $q$ is seen in the denominator of the covariance of $\epsilon$? I do understand (with moderate confidence) that $q$ is the number of noise features, and therefore there are many noise dimensions with large $q$. Correct. We scale the variance of $\mathbf{\epsilon}$ with $q$ to isolate the total level of noise from the number of dimensions. Because we divide by $q$, we have $\mathbb{E}\lVert\mathbf{\epsilon}\rVert_2^2 = \sigma_{\epsilon}^2$. Otherwise, we would have **$\mathbb{E}\lVert\mathbf{\epsilon}\rVert_2^2 = q\sigma_{\epsilon}^2$.** We have updated the writing in the manuscript to make this clearer: *“Scaling the variance of each noise coordinate by $1/q$ ensures that the expected squared norm of the noise features remains constant, $\mathbb{E}\lVert\mathbf{\epsilon}\rVert_2^2 = \sigma_{\epsilon}^2$.”* > In Fig. 2, what is the "Worst Group"? I don't think you defined this term beforehand. I think you define it well enough on line 334, later. We thank the reviewer for pointing this out. We have clarified the definition of worst-group accuracy in the caption of Figure 2, where it first appears. We keep the more detailed discussion of the evaluation metrics in Section 5. > Is "Equal-group Accuracy" (Fig. 3) the same as "Overall Accuracy" (Fig. 2)? In this case, yes. The 'overall accuracy' is simply the accuracy over all samples. In Fig. 2, because each group has the same number of samples, this is the same as equal-group accuracy. We have updated the y-label in Fig. 2 and swapped the two plots for consistency with the remaining figures. > I'm confused by line 471 "Notably, in our results whitening only worsens the linear probe in a comparison.". Does whitening not generally improve the linear probe? Did you mean standardization worsens the linear probe? We thank the reviewer for highlighting this typo. We have rewritten this sentence to *“Notably, in our results whitening worsens the linear probe in terms of worst-group accuracy in only one comparison (when whitening is combined with NT and applied to MultiNLI)."* Generally, whitening either improves the linear probe compared to both no preprocessing and standardization, or does not alter performance. Only in one case (when combining whitening with NeuronTune to MultiNLI) does applying whitening lead to a statistically significant decrease in the worst-group accuracy (when compared to no preprocessing).

Thank you for your clarifications! I am happy with these changes.

One further minor edit request: Relating to notation and Fig. 1, I think it would be a nice touch to include that "y^=sign(x⊤w)\hat{y} = \mathrm{sign}(x^\top w) is the prediction of the model", if space allows.

Thank you!

We thank the reviewer for this acknowledgement and are happy to hear they are satisfied with the changes. We will implement their suggestion for the caption of Fig. 1 in the final version of the manuscript.

Thank you for agreeing to that change. Placing the equation in the caption would be good, or in the main body text might be better if it would have broader applicability throughout the rest of the paper. I'll accept either.

Thank you once again for your receptiveness to revision.

Reviewer_ReapOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whitening representations before linear probing by investigating whitening's ability to prevent spurious correlations from affecting out-of-distribution generalization.

The authors first define the notion of simplicity and show that max-margin classifiers have a simplicity bias, i.e., they favour features that are associated with large eigenvalues. They further argue that this is undesirable in the context of spurious correlations.

The paper proceeds by showing that whitening can remove this simplicity bias and that, consequently, the model then selects (separating) directions based on their margins (in the whitened space). It also discusses the connections with standardization and the max-margin formulation, and suggests an estimator for the covariance matrix.

Experimentally:

  1. The authors demonstrate on a synthetic dataset that whitening leads to significantly higher overall and worst-group accuracies against the baseline and against standardization. They then present Theorem 1, which shows that whitening lowers sensitivity to the scale of the spurious features.
  2. They also demonstrate the method's benefit on real-world vision and language datasets, where results suggest better robustness than standardization or no pre-processing. They further show that the approach improves robustness even when combined with various pre-processing approaches.
Critical strengths
  • The paper is exceptionally well-written and easy to follow: it defines all relevant notions, provides intuition when required, clearly states the limitations of its scope and results, etc. Beyond its contributions, it is a strong reference for explaining the role of spurious correlations.
Critical weaknesses

None (the paper has weaknesses, but none are critical)

Questions for authors
  • Could you clarify how the work compares to the publications "Kernel-Whitening: Overcome Dataset Bias with Isotropic Sentence Embedding" (Gao et al., 2022) and "Mitigating Spurious Features in Contrastive Learning with Spectral Regularization" (Ghanooni et al., NeurIPS 2025)? Please note that the latter is not exactly "semi-supervised"; a more accurate description could look something like "self-supervised pretrained, with group-label-based tuning and probe balancing".
  • (minor) It would be helpful to the reader if supplementary material were referenced more often in the paper's main content (e.g., the additional experimental results). Also, the dataset descriptions (in the main text) lack some important details (like the number of classes in each dataset or the fact that the CelebA results are from an ImageNet pre-trained model).

The reviewer is willing to raise its suggested rating to "Clear accept" if concerns are properly addressed (and unless other reviewers point out an important weakness that changes the reviewer's assessment).

2 comments

We thank the reviewer for their kind words and feedback. We also thank them for their remark that the paper is `exceptionally well-written and easy to follow'. Below, we address the feedback point-by-point.

Could you clarify how the work compares to the publications "Kernel-Whitening: Overcome Dataset Bias with Isotropic Sentence Embedding" (Gao et al., 2022)

We thank the reviewer for the opportunity to clarify the relationship between our work and that of Gao et al. [1]. The two approaches differ in three key respects. First, Gao et al. [1] focus on mitigating dataset bias in sentence embeddings, whereas our work considers last-layer representations across different modalities. Second, Gao et al. [1] apply whitening during fine-tuning, whereas we apply whitening before fitting a linear probe. Third, Gao et al. [1] propose a kernel-based approach to address nonlinear feature dependencies, whereas we do not. We have updated our writing in the paper to reflect these differences as follows:

“Whitening-based normalization layers have been used to stabilize and speed up the training of DNNs (Huang et al., 2018; 2019). The work of Gao et al. (2022) uses kernel whitening of sentence embeddings during fine-tuning to improve robustness.”

and "Mitigating Spurious Features in Contrastive Learning with Spectral Regularization" (Ghanooni et al., NeurIPS 2025)? Please note that the latter is not exactly "semi-supervised"; a more accurate description could look something like "self-supervised pretrained, with group-label-based tuning and probe balancing"

We thank the reviewer for pointing out that the description of Ghanooni et al. [2] in relation to our paper could be more precise. We have updated the description in the paper to the following:

“A similar finding was reported by Ghanooni et al. (2025), who show that truncating and flattening the spectrum of frozen self-supervised features improves the robustness of linear probes, and propose regularizing pretraining toward a uniform eigenspectrum while avoiding full whitening. They select the regularization strength using group labels and train the probe on group-balanced data. We prove that full whitening, without truncation or retraining, removes the max-margin classifier’s preference for high-variance directions, and show that it improves robustness without group labels.”

Please note that we also cite Ghanooni et al. [2] below Equation 4 in the manuscript, where we point out that with gradient descent learning is dominated by covariance eigenvectors with large eigenvalues.

It would be helpful to the reader if supplementary material were referenced more often in the paper's main content (e.g., the additional experimental results)

As the reviewer suggests, the main text already contains references to the supplementary experimental results. To make the results easier to locate, we have added references to the corresponding figures in the appendix. We have also added references to the numerical experiments confirming our theoretical results in Section 4.

Also, the dataset descriptions (in the main text) lack some important details (like the number of classes in each dataset or the fact that the CelebA results are from an ImageNet pre-trained model).

We thank the reviewer for this suggestion. We have clarified the number of classes for each dataset. Regarding the pretrained models, we specify in lines 342-343 of Section 5 that we fine-tune an ImageNet-pretrained ResNet-50 for both Waterbirds and CelebA. We thus will keep this description.

Thank you for the clarifications and edits!

The authors addressed all my concerns, and the paper clearly states the assumptions under which the approach should help (i.e., when directions associated with large eigenvalues correspond to spurious features), as discussed with Reviewer jvoj. I therefore raise my recommendation to "Clear accept".

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to all reviewers

We are grateful to the reviewers for their attentive consideration of the paper. We would like to also thank the reviewers for acknowledging that the paper is well-written, provides interesting insights supported by theory and experiments, and is clear about its limitations.

We also thank the reviewers for their constructive feedback, which we have used to improve the paper. Based on the reviews, we have made three key changes to the manuscript.

  1. We have added two sets of experiments. The first investigates whether applying whitening to the last layer worsens performance on a dataset without a spurious correlation or on a standard benchmark (CIFAR-10). The second varies the scale of the spurious feature in the synthetic experiment to examine whether whitening can harm performance when that feature has low variance. These experiments intend to address the feedback of Reviewer Jvoj.
  2. To answer the questions of Reviewer Reap, we have sought to clarify the difference between our approach and existing work in the related work section.
  3. Per the feedback of Reviewer uxrP, we have updated our notation in Section 4, as well as figure 2.

In addition to these three changes, we have made several small changes in the text per the suggestions of Reviewers Reap and uxrP.

Finally, we noticed a typo in the earlier version in Equation 14 - we apologize for this and have updated the manuscript accordingly.

We have uploaded an updated version of the manuscript where all changes are temporarily colored orange in order to highlight the changes and make it easier for reviewers to check if their concerns have been addressed.

AuthorsOpenReview ↗

Thank you to all reviewers

As we are approaching the end of the discussion phase, we would like to thank the reviewers again for their careful feedback. We believe that thanks to their comments, the paper has substantially improved. We are also happy to hear that reviewers uxrP and Reap appreciated our answers and changes to the manuscript. Furthermore, we hope our answers and changes to the manuscript have alleviated the concerns of reviewer jvoj.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.