acceptodds
Under review as a conference paper at ICLR 2027

A Theoretical and Experimental Analysis of the Robustness of Similarity-based Positional Encoding Under Rotation

Abstract

Transformer architectures heavily rely on positional encodings (PEs) to inject spatial structure; however, standard approaches remain highly sensitive to spatial rotations. In this paper, we present a unified theoretical and empirical investigation into the rotational robustness of similarity-based positional encoding (simPE). Theoretically, we prove that while simPE is not strictly rotation-invariant, it exhibits provable stability under simple Lipschitz regularity assumptions, yielding explicit Frobenius norm perturbation bounds that cleanly decouple architectural features from rotational magnitude. Empirically, we evaluate models trained on canonical orientations against increasing rotation angles across five diverse datasets, by comparing simPE with a comprehensive suite of baseline PEs. Across the tested datasets, simPE provides better performance (measured in terms of classification accuracy and F1-score) in small-to-moderate rotation regimes (), with respect to competing PE methods. Crucially, on a real clinical benchmark (Chest X-Ray), simPE outperforms all competing PEs across every evaluated angle (), and allows the corresponding ViT architecture to reach significant levels of accuracy. Our findings bridge theoretical stability guarantees with practical spatial robustness, establishing simPE as a highly effective encoding strategy for domains subject to routine geometric perturbations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Positional encoding, Rotational robustness, Vision Transformers, Lipschitz Stability, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UimQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes SimPE - a patch-similarity based positional encoding module to replace standard positional encoding based on patch global or relative positioning in vision transformers. The paper shows SimPE is robust to rotations using Lipschitz continuity bounds. The paper evaluates simPE rotation robustness against other PE variants in a controlled setting and demonstrates overall better robustness.

Critical strengths
  1. The proposed SimPE shows overall better rotation robustness in the controlled setting against the other evaluated PE, especially in Chest X-Ray benchmark.
Critical weaknesses
  1. The relation of the theoretical claims to the evaluated setting is unclear. The motivation of the method is to improve robustness to spatial rotation of 2D inputs, which is also the evaluation setting. The theoretical claims, however, are related to rotations of the embeddings in their latent Rd\mathbb{R}^d space, and it is unclear why spatial rotations of the inputs translates to such rotations in the embedding space.

  2. The empirical results do not show a consistent advantage for the proposed simPE. In some cases there is a large gap compared to other methods, for example vs. Sin at Arrows in 15 deg and FashionMNIST in 30 deg. Additionally, there are no reported confidence itervals.

  3. The method is not evaluated against data augmentations which is one of the most common approaches to improve robustness. The authors justify this as a controlled experimental setup, but this limits the ability to assess practical usability of the proposed method.

Questions for authors
  1. Can you explain how spatial rotations are translated to rotations of the latent space?
  2. Can you repeat the evaluation on multiple seeds and report confidence intervals (at least on a single benchmark)?
  3. Can you repeat the evaluation using rotation augmentations (at least on a single benchmark)?
3 comments
AuthorsOpenReview ↗

Rebuttal reviewer UimQ

We thank the reviewer and we acknowledge the important point issued by her/him concerning the mismatch between some theoretical results and experimentation. In particular, we point out that the main theoretical result (Theorem 1) applies also by directly considering the input space. The paper presents the result for any XX and YY in the feature/embedding space. If ff is the feature extractor, II an input image and IθI_{\theta} the rotated image through rotation RθR_{\theta}, then the theorem holds for X=f(I)X=f(I) and Y=f(Iθ)Y=f(I_{\theta}) by reconciling this part with experiments (it establishes a bound for the actual transformation used in the experiments starting from an image and the corresponding rotation). In addition, if we also assume that the feature extractor is Lipschitz (with constant LfL_f), then we directly obtain a corrected version of the theorem relating the result directly to the input space.

Let Δθ(I)=∥\mboxsimPE(f(Iθ))−\mboxsimPE(f(I))∥F\Delta_{\theta}(I)=\Vert\mbox{simPE}(f(I_{\theta}))-\mbox{simPE}(f(I))\Vert_F, then Δθ(I)≤2LΦLkLfN∥Iθ−I∥F\Delta_{\theta}(I)\leq 2 L_{\Phi} L_{k} L_f \sqrt{N} \Vert I_{\theta}-I \Vert_F We also acknowledge that Corollary 2 was not properly presented in the paper. It is a conditional corollary that, in order to be adapted to the input space, requires an additional assumption on the rotation and interpolation scheme.

At the image level we will assume the following regularity assumption: ∥Iθ−I∥F≤cI∣θ∣\|I_{\theta}-I\|_F \leq c_I |\theta| The assumption is satisfied by rotations and interpolation schemes that produce continuous changes in pixel values as the rotation angle changes. In particular, since we deal with discrete digital images (finite real-valued pixel matrices),

given an image II, its rotated version IθI_{\theta} follows a specific interpolation scheme of an underlying continuous rotation process from angle t=0t=0 to angle t=θt=\theta, and its rotatory derivative is ∂∂tIt\frac{\partial}{\partial t} I_t. Smooth changes can be modeled by the property of local Lipschitz continuity of the rotational derivative that allows us to conclude (Rademacher's theorem) that Iθ−I=∫0θ∂∂tIt\mboxdt.I_{\theta}-I=\int_0^\theta \frac{\partial}{\partial t} I_t \mbox{ }dt. Now consider a discrete image II and the rotated image IθI_\theta obtained using a specified interpolation scheme.If the interpolated scheme produces a locally Lipschitz continuous rotational derivative (i.e., interpolated pixel values vary continuously, with a uniformly bounded rate of change almost everywhere, as a function of the rotation angle), and we further assume that rotational derivative is essentially uniformly bounded in the norm by a constant cIc_I (i.e. the rotation rate has a specific finite limit that applies almost everywhere), then we can finally prove that

\Vert\int_0^\theta \frac{\partial}{\partial t} I_t \mbox{ } dt \Vert_F \leq \vert \int_0^\theta \Vert \frac{\partial}{\partial t} I_t \Vert_F \mbox{ } dt \vert \leq c_I |\theta|.$$ The required assumptions are not valid for any rotation, but they are valid for several rotation of interest such as bilinear, bicubic, or suitable band-limited interpolation. In particular, in the experiments we used bilinear interpolation having rotational derivative defined almost everywhere (failing to be differentiable only along grid boundaries} and satisfying the required conditions of the corollary. In the revised verson, we will provide the details of the proof together with the required references justifying any assumption made. As a corollary of the corrected formulation of Theorem 1 we finally get the corrected formulation of the angle bound (Corollary 2): $$\Delta_\theta(I)\leq 2 L_{\Phi} L_{k} L_f \sqrt{N} c_I |\theta|$$ This reconciles theory and experiments. Concerning statistical reliability, we will add in revised version the required means and confidence intervals for a better comparison. Concerning rotation agumentation , this has been deliberately avoided since the main goal of the experiments are to test the robustness to rotations, thus it would make a little sense to add rotations in the training set that would inform the model about the rotation process. We believe that a sound approach is to have ONLY unrotated images in the training, by testing robustness by rotating images in the test set. This is particularly relevant in the tested clinical application where rotations are errors in the image capturing and should not be used tro train the model.

I thank the authors for seriously engaging with my comments and their detailed response. Unfortunately, some of my concerns remain. Regarding the theoretical results, the authors’ clarification suggests that the claimed stability under input rotations follows from a general Lipschitz continuity argument. Rotation enters this argument only through the distance between the original and rotated images; no rotation-specific property of SimPE is used. Under the same assumptions on image rotation, analogous results could therefore be obtained for any alternative positional encoding that is Lipschitz continuous in its input. Consequently, the revised result does not identify a rotation-robustness property particular to SimPE or explain its advantage over competing positional encodings. Second, my concerns regarding evaluation with data augmentation were not sufficiently addressed. While I appreciate the authors’ approach of evaluating the method in a controlled setting, I believe a complementary evaluation with rotation augmentation could have clarified the practical benefits of the proposed method. Due to the above, I maintain my original score.

We respectfully disagree witrh the reviewer conclusions. It is not trivially true that "under the same assumptions on image rotation, analogous results could therefore be obtained for any alternative positional encoding that is Lipschitz continuous in its input". This shoud be formally proved as done by us for the specific formulation of SimPE. We are not aware of any formal proof analogouos to ours for any other PE. Moreover, we are not claiming that only simPE is robust, we are claimimg that, given the specific theoretical properties (formally proved), simPE is robust and experiments suggets that it is more robust than competing approaches. For the augmentation argument, it is honestly true that a complementary evaluation with rotation augmentation would be biased, since adding rotation in the tranining set wil defiinitely inform the model about the rotation process. On the contrary, we have set an experimental framework where the prediction capabilities of the model are challanged in the hardest way.

Reviewer_rU8POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates rotational robustness of similarity-based positional encoding (SimPE) for Vision Transformers, combining theory and experiments. Contributions: (1) theoretical analysis formulating SimPE as a pairwise similarity mapping composed with a learned linear projection, yielding a Frobenius-norm perturbation bound under boundedness and Lipschitz assumptions and a bound proportional to rotation angle; (2) a distinction between invariance and stability, arguing SimPE need not be strictly invariant but may change in a controlled way under small perturbations, desirable when orientation remains class-relevant; (3) empirical comparison against learned absolute PE, sinusoidal encoding, RoPE, PEG, and no-PE under a shared ViT, across Arrow, Digits, Shapes, FashionMNIST, and Chest X-Ray, trained on unrotated images and tested at increasing angles.

Critical strengths

A systematic, controlled empirical comparison of positional encodings under rotation. A shared ViT backbone and a protocol training exclusively on unrotated images help isolate PE effects on generalization to unseen rotations; synthetic datasets, a standard benchmark, and a clinical dataset provide reasonably diverse evaluation. This matters because robustness to spatial transformations is practically important, and the comparisons yield useful observations independently of the theory.

Critical weaknesses

(a) A fundamental mismatch between the theoretical rotation model and the actual spatial image rotations undermines the main theoretical contribution. The analysis studies X↦RXX\mapsto RX with R∈SO(d)R\in SO(d), while experiments rotate images before patch extraction and embedding. These are mathematically different operations. With ff the patch embedding and Rθ\mathcal R_\theta a spatial rotation, the paper does not establish that

f(RθI)=Rθf(I).f(\mathcal R_\theta I)=R_\theta f(I).

Without such a relationship, the feature-space Lipschitz bound does not establish stability under spatial image rotations. More importantly, cosine similarity is exactly invariant under simultaneous orthogonal transformations:

(Rxi)⊤(Rxj)∥Rxi∥∥Rxj∥=xi⊤xj∥xi∥∥xj∥.\frac{(Rx_i)^\top(Rx_j)} {\|Rx_i\|\|Rx_j\|} = \frac{x_i^\top x_j}{\|x_i\|\|x_j\|}.

Therefore SimPE itself is exactly invariant under the transformations analyzed in the theorem, assuming nonzero feature vectors and fixed projection weights, and the angle-dependent upper bound does not explain the observed sensitivity to image rotations. Secondary concerns: cosine similarity needs a positive lower bound on feature norms for a uniform Lipschitz constant, which bounded pixel intensities do not guarantee; and Appendix B infers that a change in the similarity matrix necessarily produces a change after multiplication by WW, which is invalid without additional assumptions on WW. This is critical because the primary theoretical contribution is intended to explain SimPE's robustness to image rotations; the current analysis establishes general continuity but does not characterize the experimentally studied transformations.

(b) Experimental evidence is insufficient to substantiate the claimed robustness advantages reliably or to validate the proposed mechanism. First, results lack standard deviations, confidence intervals, or repeated-training statistics, which matters because several improvements are small—e.g., on Chest X-Ray without rotation, 86.70% versus 86.52% for PEG, a difference of 0.18 points—leaving robustness to initialization and training variability unclear. Second, the experiments report accuracy and F1 but do not directly measure representation sensitivity to rotation, e.g.,

∥SimPE⁡(f(RθI))−SimPE⁡(f(I))∥F\left\| \operatorname{SimPE}(f(\mathcal R_\theta I)) - \operatorname{SimPE}(f(I)) \right\|_F

as a function of angle, which is needed to determine whether results reflect the proposed mechanism. Third, SimPE's advantage is not uniform: at 30° on FashionMNIST it achieves 23.84% versus 42.00% for sinusoidal PE, and on Digits RoPE substantially outperforms SimPE at several larger angles, including 90° (29.00% versus 18.70%). These crossovers are not fatal given the emphasis on small rotations but require careful interpretation. Finally, the paper does not sufficiently isolate the effects of the similarity function and learned projection, despite identifying them as central to the stability bound. Without reliability estimates and representation-level measurements, the results show promising performance differences but do not establish that these arise from the proposed robustness mechanism.

Questions for authors

(1) Theoretical formulation of rotation. How does R∈SO(d)R\in SO(d) acting on token embeddings relate to the two-dimensional spatial image rotations used in experiments? Can the authors establish a bound for

∥SimPE⁡(f(RθI))−SimPE⁡(f(I))∥F\left\| \operatorname{SimPE}(f(\mathcal R_\theta I)) - \operatorname{SimPE}(f(I)) \right\|_F

that explicitly accounts for Rθ\mathcal R_\theta and ff? This is important because cosine similarity is exactly invariant under the shared feature-space rotations currently analyzed; a rigorous clarification or corrected result would substantially improve the soundness assessment.

(2) Statistical reliability. Are results based on a single run or averages over multiple seeds? Can the authors provide means, standard deviations or confidence intervals for the main comparisons, especially on Chest X-Ray, and assess whether differences from the strongest baselines are statistically significant across independently trained models?

(3) Relationship between theoretical stability and observed performance. Can the authors directly measure perturbation magnitude of positional representations under spatial image rotations and compare it across SimPE and baselines? Does SimPE exhibit smaller representation-level perturbations in the same angular regimes where it achieves higher accuracy? This would help determine whether improvements are attributable to the proposed stability mechanism rather than other architectural or optimization effects.

1 comment

We acknowledge the important point issued by the reviewer concerning the mismatch between some theoretical results and experimentation. In particular, we point out that the main theoretical result (Theorem 1) applies also by directly considering the input space. The paper presents the result for any XX and YY in the feature/embedding space. If ff is the feature extractor, II an input image and IθI_{\theta} the rotated image through rotation RθR_{\theta}, then the theorem holds for X=f(I)X=f(I) and Y=f(Iθ)Y=f(I_{\theta}) by reconciling this part with experiments (it establishes a bound for the actual transformation used in the experiments starting from an image and the corresponding rotation). In addition, if we also assume that the feature extractor is Lipschitz (with constant LfL_f), then we directly obtain a corrected version of the theorem relating the result directly to the input space.

Let Δθ(I)=∥\mboxsimPE(f(Iθ))−\mboxsimPE(f(I))∥F\Delta_{\theta}(I)=\Vert\mbox{simPE}(f(I_{\theta}))-\mbox{simPE}(f(I))\Vert_F, then Δθ(I)≤2LΦLkLfN∥Iθ−I∥F\Delta_{\theta}(I)\leq 2 L_{\Phi} L_{k} L_f \sqrt{N} \Vert I_{\theta}-I \Vert_F We also acknowledge that Corollary 2 was not properly presented in the paper. It is a conditional corollary that, in order to be adapted to the input space, requires an additional assumption on the rotation and interpolation scheme.

At the image level we will assume the following regularity assumption: ∥Iθ−I∥F≤cI∣θ∣\|I_{\theta}-I\|_F \leq c_I |\theta| The assumption is satisfied by rotations and interpolation schemes that produce continuous changes in pixel values as the rotation angle changes. In particular, since we deal with discrete digital images (finite real-valued pixel matrices),

given an image II, its rotated version IθI_{\theta} follows a specific interpolation scheme of an underlying continuous rotation process from angle t=0t=0 to angle t=θt=\theta, and its rotatory derivative is ∂∂tIt\frac{\partial}{\partial t} I_t. Smooth changes can be modeled by the property of local Lipschitz continuity of the rotational derivative that allows us to conclude (Rademacher's theorem) that Iθ−I=∫0θ∂∂tIt\mboxdt.I_{\theta}-I=\int_0^\theta \frac{\partial}{\partial t} I_t \mbox{ }dt. Now consider a discrete image II and the rotated image IθI_\theta obtained using a specified interpolation scheme.If the interpolated scheme produces a locally Lipschitz continuous rotational derivative (i.e., interpolated pixel values vary continuously, with a uniformly bounded rate of change almost everywhere, as a function of the rotation angle), and we further assume that rotational derivative is essentially uniformly bounded in the norm by a constant cIc_I (i.e. the rotation rate has a specific finite limit that applies almost everywhere), then we can finally prove that

\Vert\int_0^\theta \frac{\partial}{\partial t} I_t \mbox{ } dt \Vert_F \leq \vert \int_0^\theta \Vert \frac{\partial}{\partial t} I_t \Vert_F \mbox{ } dt \vert \leq c_I |\theta|.$$ The required assumptions are not valid for any rotation, but they are valid for several rotation of interest such as bilinear, bicubic, or suitable band-limited interpolation. In particular, in the experiments we used bilinear interpolation having rotational derivative defined almost everywhere (failing to be differentiable only along grid boundaries} and satisfying the required conditions of the corollary. In the revised verson, we will provide the details of the proof together with the required references justifying any assumption made. As a corollary of the corrected formulation of Theorem 1 we finally get the corrected formulation of the angle bound (Corollary 2): $$\Delta_\theta(I)\leq 2 L_{\Phi} L_{k} L_f \sqrt{N} c_I |\theta|$$ We appreciate the observation that bounded feature norms alone does not guarantee a uniform Lipschitz constant for cos sim. We will make the nondegeneracy assumption explicit ($0<\lambda \leq \Vert x \Vert \leq M$), by also noting that in practical implementations (e.g. PyTorch) cos sim is numerically stabilized against zero-norm tokens. We also thank the reviewer for pointing out an imprecision in appendix B. We can restate the claim as "there exist an image $I$ and a projection $W$ such that $simPE(R_θ I) \neq simPE(I)$". Defining $\Delta S =S(f(R_\theta I)) - S(f(I))$, its $(k, p)$ entry is therefore strictly positive: $$\Delta S_{kp} = 1 - \kappa(x_b, x_p) > 0.$$Now choose the projection matrix $W$ such that its first column is the $p$-th standard basis vector $e_p$ (i.e., $W_{p1} = 1$ and $W_{j1} = 0$ for all $j \neq p$). Since $simPE(X) = S(X)W$, the $(k, 1)$ entry of $simPE(R_\theta I) - simPE(I)$ evaluates to $$(simPE(R_\theta I)-simPE(I))_{k1}=\sum^N_{j=1} \Delta S_{kj}W_{j1}=\Delta S_{kp}>0$$ The claim follows given that $f$ is non degenerate (does not map patches such that $x_p=c x_b$ for $c>0$) otherwise $1-\kappa(x_b,x_p)=0$ Concerning statistical reliability, we will add in revised version the required means and confidence intervals for better comparison.
Reviewer_MwxFOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper theoretically analyzes the promising properties of the recently proposed similarity-based positional encoding (SimPE) method, which enables vision transformers to produce consistent predictions under rotational variations. The existing SimPE method lacks theoretical investigation. **(main contribution) This paper presents a theoretical analysis of the SimPE method, explaining why it performs well under rotational variations. Empirical results demonstrate that the SimPE method performs well under rotational variations across five datasets (e.g., chest X-rays).

Critical strengths

(S1) This paper theoretically establishes the Lipschitz continuity of SimPE under mild regularity assumptions, providing a theoretical foundation for its robustness to rotational perturbations.

Critical weaknesses

(W1) Low accuracy of SimPE. Despite its theoretical guarantees, SimPE performs similarly to or worse than the baselines in some experimental settings, raising questions about the extent to which the theoretical analysis translates into empirical improvements.

(W2) The experimental scale is limited. The experiments are mainly conducted on four-layer ViT architectures trained on a relatively small number of samples. This reviewer suggests that the authors conduct additional experiments at a larger scale to further validate the effectiveness of the proposed method.

Questions for authors

See major weaknesses (W1, W2).

1 comment
AuthorsOpenReview ↗

Rebuttal Reviewer MwxF

We thank the reviewer for the useful observations and comments. Here are our responses:

W1:

  1. Medical imaging is the primary application domain motivating SimPE. On such real clinical benchmark (Chest X-Ray, Table 2), SimPE outperforms every single competing PE (Sinusoidal, Learned, ROPE, PEG, and no-PE) across every evaluated angle without exception. In this setting, competing methods degrade substantially under rotation while SimPE is not.
  2. There is a consistent superiority in small-to-moderate rotations (<= 15 degrees) across al datasets.
  3. Concerning the alignment between theory and empirical findings, presented theoretical bounds (Theorem 1, Corollary 2) establish that the perturbation in the encoding is bounded linearly by the rotation angle. We actually have provided an adjustment to the theoretical results showing that under simple additional assumptions we can restate the results in terms of images in the input space. In particular, if the feature extractor is Lipschitz (and standard patch-wise convolution is), theorem 1 can be restated as follows (ff is the feature extractor, II the input original image and IθI_\theta the rotated input image):: ∥∣simPE(f(Iθ))−simPE(f(I)∥∣≤2LΦLkLfN∥∣Iθ−I∥∣\||simPE(f(I_\theta))-simPE(f(I)\||\leq 2 L_\Phi L_k L_f \sqrt{N} \|| I_\theta - I \|| where LfL_f is the Lipschitz constant relative to the extractor. Moreover, if the rotation is smooth (the interpolated image depends differentiably on the rotation angle) and the rotational derivative is bounded in norm (the image's pixels can only change at a controlled finite speed), then also corollary 2 can be restated in terms of images in the input space: ∥∣\mboxsimPE(f(Iθ)−\mboxsimPE(f(I)∥∣≤2LΦLkLfNcI∣θ∣\|| \mbox{simPE}(f(I_{\theta})-\mbox{simPE}(f(I) \||\leq 2 L_{\Phi} L_{k} L_f \sqrt{N} c_I |\theta| with cIc_I is the bound on the norm. Several usuful rotations satisfy such assumption (usually with some exceptions at the image boundary pixels), and in particular bilinear interpolation, which is the one used in the experiments. We can now say that there is no gap between theory and experiments, and the empirical results confirm that perturbations scale smoothly with the rotation angle, providing controlled robustness precisely where needed. We will modify the paper in agreement with the above observations.

W2:

  1. As summarized in Table 1, our study relies on standard, large-scale benchmarks which are not so limited for the required analysis.
  2. The choice of a 4-layer ViT backbone was deliberate and guided by sound experimental methodology: (a) deeper networks introduce significant implicit regularization, high parameter redundancy, and deep feature averaging that obscure the isolated effects of early-stage positional encodings, which is our target of study; (b) In domain-specific tasks such as clinical X-ray diagnosis, scaling to a deeper ViT architecture leads to severe parameter overcapacity and memorization (see https://doi.org/10.64898/2025.12.06.25341759). A 4-layer ViT may then provide a more appropriate capacity-to-data ratio. Moreover, theoretical results are independent from the architectural choices: in particular, considering theorem 1 and corollary 2 the derived Lipschitz constants and the angle stability bound depend on the similarity function and projection weights, holding independently of the number of subsequent self-attention layers in the downstream backbone.

Comments on the paper

1 comment

We thank all the reviewers for their constructive and important comments. There were 2 main issues raised by the reviewers:

  1. mismatch between the theoretical results and the experimentation framework
  2. statistical reliability of the results

We have solved 1) by restating the results in terms of the input space, in order to reconcile them with the experimental framework; the properties are now proved as requested by the reviewers.

To address 2) we have computed the requested confidence intervals for the experiments and we will report them in the updated version of the paper to show the statistical reliability of the results.

We want to highlight that the main claim of the paper is not to prove that SimPE is the only PE robust under rotation as one might implicitly infer from a reviewer’s comment, but to show: (a) theoretically, that is robust wrt to some specific properties (e.g., angle rotation); (b) experimentally, that it is more robust that competing approaches.

We hope this will clarify the claim and the merit of the approch.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.