acceptodds
Under review as a conference paper at ICLR 2027

Rotated Manifold Optimization for Low-Rank Adaptation

Abstract

We propose a novel optimizer for low-rank adaptation (LoRA) that explicitly incorporates the gauge symmetry of low-rank factorization. Our optimizer extends recent matrix optimizers for full-parameter training to the manifold of fixed-rank matrices by interpreting them as normalization under a rotated basis. We show how rotation and normalization can be integrated with the fixed-rank manifold efficiently. Our optimizer converges faster to lower held-out loss and achieves better or comparable downstream performance on both supervised finetuning and reinforcement learning tasks.

Primary area:
optimization
Keywords:
Manifold optimization, LoRA

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_McShOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Rotated Manifold Optimization (RoM), an optimizer for LoRA that operates on the rank-r Riemannian manifold. RoM applies row and column normalization in a coordinate frame defined by the left singular vectors of the tangent momentum, followed by tangent-space projection and retraction. By exploiting the low-rank structure of tangent vectors, the method avoids materializing full weight-sized matrices. Experiments cover supervised finetuning and reinforcement learning, demonstrating consistently lower validation loss and more pronounced downstream improvements in the RL setting.

Critical strengths
  • The paper evaluates RoM on mathematical reasoning and code generation under SFT, as well as mathematical reasoning under GRPO. The SFT experiments cover multiple model families and sizes, compare several LoRA optimizers.
  • The factorized implementations of gradient computation, projection of the previous momentum onto the current tangent space, rotation, and retraction preserve the memory efficiency of LoRA. Reducing the required SVDs to small matrices makes the proposed method has practical utility.
  • RoM achieves lower validation loss across the experiments.
Critical weaknesses
  • Section 3.2 cites prior work on rotated optimization and multi-norm normalization, and explains how the low-rank structure makes the left singular frame efficient to compute. However, the paper provides limited justification for preferring this frame over alternative orthogonal bases or a right-sided formulation, and for adding column normalization beyond row normalization.
  • The relationship between the normalization objective and its implementation requires clarification. Equation 16 constrains both the maximum row norm and the maximum column norm to equal one, whereas Algorithm 3 alternately normalizes every row and every column. For a reduced matrix C∈R2r×nC\in\mathbb{R}^{2r\times n} with n>2rn>2r, ending with unit-norm columns gives ∣∣C∣∣F2=n||C||_F^2=n, while a maximum row norm of one would imply ∣∣C∣∣F2≤2r||C||_F^2\leq 2r. Thus, the returned matrix cannot satisfy both stated constraints in this setting. The paper should clarify the dimension-dependent scaling or the precise sense in which Algorithm 3 approximates Equation 16.
  • The RL comparison includes only Adam and uses a single model and training recipe.
Questions for authors

Section 3.2 motivates the design through prior work on rotated optimization and multi-norm normalization. However, the specific advantages of using the left singular basis and adding column normalization remain insufficiently explained. Could the authors clarify these choices and provide controlled comparisons with alternative bases and a row-only normalization variant?

6 comments

We thank Reviewer McSh for their time and valuable feedback. We are glad that they appreciated the breadth of our evaluation across SFT, RL, and model families, and that they found our factorized implementation practical. We respond to the raised concerns as follows.

W1 & Q. Section 3.2 motivates the design through prior work on rotated optimization and multi-norm normalization. However, the specific advantages of using the left singular basis and adding column normalization remain insufficiently explained. Could the authors clarify these choices and provide controlled comparisons with alternative bases and a row-only normalization variant?

Rotation basis: The use of left singular vectors is actually a natural choice motivated by existing matrix optimizers. Consider Muon as an example, for a gradient GG with its SVD decomposition: G=USV⊤G = USV^{\top}, Muon normalizes GG as Muon(G)=UV⊤\text{Muon}(G) = UV^{\top}. Under the rotation-normalization framework, Muon can be exactly recovered if we choose UU as the rotation basis and row-normalization as the normalization function applied to the rotated representation U⊤GU^{\top}G. Specifically, given the orthogonality of UU and VV, we have: Urow-normalize(U⊤G)=Urow-normalize(SV⊤)=UV⊤=Muon(G)U \text{row-normalize}(U^{\top}G) = U \text{row-normalize}(SV^{\top}) = UV^{\top} = \text{Muon}(G)

Earlier matrix optimizers like SOAP/Shampoo consider eigenvectors of GG⊤GG^{\top} for preconditioners, which are exactly the left singular vectors of GG. Therefore, choosing the singular vectors as the rotation basis is a natural choice extending the design of existing matrix optimizers.

Which side to rotate: In the paper, we apply rotation to the left side of the adapter WW (thus using left singular vectors) for ease of illustration of the idea. In implementation, we rotate the side with the smaller dimension for computational efficiency. If WW is square, we rotate the side that faces the input (exactly the left side in this paper's notation)

As an ablation study of this design choice, we experimented with rotating the output side of each square adapter (using singular vectors of the output side) and compared the final held-out loss (×10−3\times 10^{-3}) with our original design. Experiments were based on math SFT, using Llama 3B and 8B with rank 16. All hyperparameters were the same as those for our main results.

Llama 3BLlama 8B
Rotate input side182.69$\pm$0.21157.60$\pm$0.06
Rotate output side182.60$\pm$0.17157.53$\pm$0.10

The difference is on the same order as seed variance and negligible compared with difference between optimizers

Row-only normalization variant: From the above Muon example, we can see that Muon is equivalent to row-normalization in the singular-vector frame. Plugging the row normalization as the function hh into RoM's update rule (Eq 17), we can get that the row-only normalization variant of RoM is equivalent to applying Muon first and then projecting the result onto the tangent space of the fixed-rank manifold, which is exactly Riemannion. Therefore, RoM and Riemannion have already formed a controlled comparison regarding the additional column normalization, and RoM is consistently better than Riemannion.

We have added the RL results for Riemannion, and the results further validate the advantage of RoM's normalization

MATH-500GSM8KOlympiad
Adam62.9$\pm$0.382.7$\pm$0.425.9$\pm$0.8
Riemannion62.1$\pm$0.983.3$\pm$0.226.9$\pm$0.1
RoM (ours)64.9$\pm$0.684.2$\pm$0.429.0$\pm$0.8

W2. The relationship between the normalization objective and its implementation requires clarification.

We acknowledge that the previous normalization objective Eq (16) was not accurate. We missed the constants in the norm constraints. The corrected normalization objective is:

h(G)=argmaxH∈Rm×n⟨G,H⟩s.t.maxi∣∣Hi,:∣∣2n=1andmaxj∣∣H:,j∣∣2m=1h(G) = \text{argmax}_{H \in \mathbb{R}^{m \times n}} \langle G, H \rangle \quad \text{s.t.} \quad \frac{\text{max}_i ||H_{i, :}||_2}{\color{red}\sqrt{n}} =1 \quad \text{and} \quad \frac{\text{max}_j ||H_{:, j}||_2}{\color{red}\sqrt{m}} = 1

To achieve this, Algorithm 3 operates directly on the rotated representation C∈R2r×nC \in \mathbb{R}^{2r \times n}, but is equivalent to the classical Sinkhorn algorithm [1] applied to the element-wise square of CC, and converges to a fixed point whose rows have the same norm and columns have the same norm [2]. In each iteration:

Ci,:←nCi,: / ∣∣Ci,:∣∣2fori=1,…,2rC_{i,:} \leftarrow {\color{red}\sqrt{n}} C_{i,:} \ / \ ||C_{i, :}||_2 \quad \text{for} \quad i=1,\dots,2r C:,j←2rC:,j / ∣∣C:,j∣∣2forj=1,…,nC_{:,j} \leftarrow {\color{red}\sqrt{2r}} C_{:,j} \ / \ ||C_{:,j}||_2 \quad \text{for} \quad j=1,\dots,n

In our implementation of Algorithm 3 (and the initial version of the submission), we didn't multiply by the two constants because:

  • Column normalization is invariant to scalars, so n\sqrt{n} has no effect
  • 2r\sqrt{2r} depends only on the rank rr and is the same for all LoRA layers, so we absorbed it into the learning rate, and the learning rate was swept in all experiments

Therefore, whether or not the two constants are explicitly multiplied in Algorithm 3 won't affect our results or findings

Updates in revised pdf (marked in red): We have corrected the constants to make the normalization objective and algorithm consistent. We have moved Algorithm 3 to the main body (now it's Algorithm 1), and have added sentences explaining its connection to the Sinkhorn algorithm and why explicitly multiplying by these constants is not necessary in practice.

[1] Concerning nonnegative matrices and doubly stochastic matrices. Sinkhorn & Knopp, 1967
[2] Gradient Multi-Normalization for Efficient LLM Training. Scetbon et al. NeurIPS 2025

W3. The RL comparison includes only Adam and uses a single model and training recipe.

We have added the RL results for Riemannion to Figure 2 and Table 4 (numbers are in our first comment). We tuned its learning rate and used the same experimental setup. RoM improved performance faster than Riemannion and also achieved better final performance

We hope our response and revision can address the reviewer’s concerns. We would be grateful if the reviewer could consider revising their score. We are happy to answer any further questions during the discussion period.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer McSh,

Thank you once again for your time and feedback! As the discussion period nears its end, we would greatly appreciate it if you could take a moment to check our rebuttal and revision. If you have any further questions or concerns, we would be more than happy to address them!

Thanks

Reviewer_XB87OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Rotated Manifold Optimization (RoM), a LoRA optimizer that directly optimizes the fixed-rank adaptation matrix. It combines factorized Riemannian gradients, momentum transport, a conditioning frame derived from tangent-momentum singular vectors, alternating row and column normalization, and truncated-SVD retraction. Thin factorizations avoid materializing full-sized updates. Experiments cover mathematical SFT with Llama and Gemma models, code SFT with Llama-3.1-8B, and GRPO finetuning of Qwen2.5-3B. RoM achieves lower held-out loss across the evaluated SFT settings, while downstream improvements are mostly modest. RL results show clearer accuracy gains over Adam, with reported end-to-end per-step training time of 1.08 times Adam’s.

Critical strengths
  1. The paper provides a concrete factorized implementation of fixed-rank manifold optimization. Equations (11)–(14) and Algorithms 1–2 describe Riemannian gradient computation, retraction, rotated updates, and momentum transport using thin factors, avoiding explicit construction of full-sized gradient and update matrices.

  2. The experiments provide consistent evidence of improved optimization. Across the mathematical and code SFT settings in Tables 1–3, RoM achieves lower validation loss than all evaluated baselines, including Riemannion, LoRA-RITE, and BaLoRA. Figure 4(a) further demonstrates the additional benefit of rotation over normalization alone in the tested math SFT setting, supporting the contribution of this design component.

Critical weaknesses
  1. The normalization objective and implementation appear inconsistent. Equation (16) requires both the maximum row norm and maximum column norm to equal one, whereas Algorithm 3 ends by normalizing each nonzero column to unit norm. For a 2r-by-n output with all columns nonzero and n > 2r, the squared Frobenius norm is therefore n, implying a maximum row norm of at least sqrt(n/(2r)) > 1 and violating the row constraint in Eq. (16). The authors should specify the exact normalization objective and explain how Algorithm 3 relates to it, so that the mathematical formulation matches the implemented procedure.

  2. The RL evaluation provides limited evidence of RoM’s advantage relative to other geometry-aware optimizers. Section 4.3 compares only with Adam, citing the absence of established RL recipes for the other baselines, although Riemannion is identified as the closest method and is evaluated in SFT. Because the clearest downstream gains occur in RL, this omission limits the assessment of RoM’s additional benefit over the closest competing approach. A comparison with Riemannion under the same GRPO setup, or a concrete explanation of the obstacles to adapting it, would help assess the comparative value of these gains.

Questions for authors
  1. How does Eq. (16) relate to the normalization implemented in Algorithm 3? When all n columns are nonzero and n > 2r, the final column normalization gives squared Frobenius norm n and maximum row norm at least sqrt(n/(2r)) > 1, violating the row constraint in Eq. (16). Please clarify the intended objective and the exact normalization procedure used to obtain the reported results.

  2. When the tangent momentum has rank below 2r, how is the conditioning frame constructed, and how are zero or near-zero row and column norms handled in Algorithm 3? Please describe the handling used in the experiments, including any rank thresholds, frame truncation or completion rules, or denominator safeguards, if applicable.

  3. Section 4.3 notes that the other baselines lack established RL recipes. Could the authors clarify whether adapting Riemannion, identified in Appendix B as the closest related method, presents specific technical difficulties? If any comparison is already available, it would help contextualize RoM’s gains over Adam.

  4. The submission provides pseudocode but no runnable implementation or code repository link. Could the authors make an anonymous implementation of RoM available? Alternatively, providing the implementation details underlying the normalization and numerical handling discussed above would help resolve the remaining ambiguity about the optimizer used in the experiments.

3 comments

We thank Reviewer XB87 for their time and detailed review. We are glad that they found our factorized implementation of fixed-rank manifold optimization concrete, and that our experiments provide consistent evidence of improved optimization. We respond to the raised concerns as follows.

W1 & Q1: How does Eq. (16) relate to the normalization implemented in Algorithm 3? Please clarify the intended objective and the exact normalization procedure used to obtain the reported results

We acknowledge that the previous normalization objective Eq (16) was not accurate. We missed the constants in the norm constraints. The corrected normalization objective is:

h(G)=argmaxH∈Rm×n⟨G,H⟩s.t.maxi∣∣Hi,:∣∣2n=1andmaxj∣∣H:,j∣∣2m=1h(G) = \text{argmax}_{H \in \mathbb{R}^{m \times n}} \langle G, H \rangle \quad \text{s.t.} \quad \frac{\text{max}_i ||H_{i, :}||_2}{\color{red}\sqrt{n}} =1 \quad \text{and} \quad \frac{\text{max}_j ||H_{:, j}||_2}{\color{red}\sqrt{m}} = 1

To achieve this, Algorithm 3 operates directly on the rotated representation C∈R2r×nC \in \mathbb{R}^{2r \times n}, but is equivalent to the classical Sinkhorn algorithm [1] applied to the element-wise square of CC, and converges to a fixed point whose rows have the same norm and columns have the same norm [2]. In each iteration:

Ci,:←nCi,: / ∣∣Ci,:∣∣2fori=1,…,2rC_{i,:} \leftarrow {\color{red}\sqrt{n}} C_{i,:} \ / \ ||C_{i, :}||_2 \quad \text{for} \quad i=1,\dots,2r C:,j←2rC:,j / ∣∣C:,j∣∣2forj=1,…,nC_{:,j} \leftarrow {\color{red}\sqrt{2r}} C_{:,j} \ / \ ||C_{:,j}||_2 \quad \text{for} \quad j=1,\dots,n

In our implementation of Algorithm 3 (and the initial version of the submission), we didn't multiply by the two constants because:

  • Column normalization is invariant to scalars, so n\sqrt{n} has no effect
  • 2r\sqrt{2r} depends only on the rank rr and is the same for all LoRA layers, so we absorbed it into the learning rate, and the learning rate was swept in all experiments

Therefore, whether or not the two constants are explicitly multiplied in Algorithm 3 won't affect our results or findings

Updates in revised pdf (marked in red): We have corrected the constants to make the normalization objective and algorithm consistent. We have moved Algorithm 3 to the main body (now it's Algorithm 1), and have added sentences explaining its connection to the Sinkhorn algorithm and why explicitly multiplying by these constants is not necessary in practice.

[1] Concerning nonnegative matrices and doubly stochastic matrices. Sinkhorn & Knopp, 1967
[2] Gradient Multi-Normalization for Efficient LLM Training. Scetbon et al. NeurIPS 2025

Q2. When the tangent momentum has rank below 2r, how is the conditioning frame constructed, and how are zero or near-zero row and column norms handled in Algorithm 3?

In theory, it's true that the tangent momentum MM can have rank < 2r2r. In that case, we complete the frame FF using arbitrary orthonormal vectors that are orthogonal to MM's singular vectors (it's the default behavior of PyTorch's QR function). This can make the rotated representation, i.e., F⊤MF^{\top}M, have zero or near-zero rows. We address this by introducing a ratio τ\tau (by default 1e-4), and we don't normalize rows whose norms are below τ×M\tau \times M's max row norm. We also apply the same τ\tau ratio to column normalization.

In our experiments, we tried different orders of magnitude for τ\tau, from 1e-4 to 1e-10, and found that it didn't affect the results, indicating that rank-deficiency barely happens in realistic training settings

We use an ϵ\epsilon = 1e-12 to safeguard the denominator when doing normalization

W2 & Q3. Section 4.3 notes that the other baselines lack established RL recipes. Could the authors clarify whether adapting Riemannion, identified in Appendix B as the closest related method, presents specific technical difficulties? If any comparison is already available, it would help contextualize RoM’s gains over Adam.

We have added the RL results for Riemannion to Figure 2 and Table 4. We tuned its learning rate and used the same experimental setup. RoM improved performance faster than Riemannion and also achieved better final performance

MATH-500GSM8KOlympiad
Adam62.9$\pm$0.382.7$\pm$0.425.9$\pm$0.8
Riemannion62.1$\pm$0.983.3$\pm$0.226.9$\pm$0.1
RoM (ours)64.9$\pm$0.684.2$\pm$0.429.0$\pm$0.8

Q4. The submission provides pseudocode but no runnable implementation or code repository link. Could the authors make an anonymous implementation of RoM available?

https://anonymous.4open.science/r/RoM-5118/

Here we provide the code implementation for RoM, along with a runnable training pipeline on the math SFT task that reproduces the reported numbers, and an Adam comparison

We hope our response and revision can address the reviewer’s concerns. We would be grateful if the reviewer could consider revising their score. We are happy to answer any further questions or clarify any remaining points during the discussion period.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer XB87,

Thank you once again for your time and feedback! As the discussion period is approaching its end, we would greatly appreciate it if you could take a moment to check our rebuttal and revision. If you have any further questions or concerns, we would be more than happy to address them!

Thanks

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer XB87,

The discussion period will end within one day. We would be grateful for your feedback on our response and revisions, and we are still available for any further questions. Thank you for your time!

Reviewer_f2xmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes RoM to optimize the LoRA update on the fixed-rank manifold, which rotates the momentum into its singular basis before normalization. Experiments cover math and code SFT on Llama/Gemma and GRPO on Qwen2.5 against five LoRA optimizers.

Critical strengths
  1. The paper is well motivated to keep rotation, normalization, projection, and retraction entirely in thin factors.
  2. The paper's insight on reading recent matrix optimizers as normalization in a rotated basis is interesting.
Critical weaknesses
  1. The paper's SFT downstream gains are within seed variations, which needs further justification for why this is a good indicator to support the framework.
  2. The paper's visual presentation on (Pseudo-)Algorithm 1 and Figures 1/2/4 are below expectations. While the presented framework has good insights, the presentation needs extra effort to improve. Algorithm 1 has 6 lines of comments, which are almost equal to the number of actual operations in the algorithm. Figures 1/2/4 occupy overwhelmingly large spaces to show few line plots, which are not informative. Clearly conveying the framework would be beneficial for readers to understand the ideas.
Questions for authors

N.A.

3 comments

We thank Reviewer f2xm for their time and feedback. We are glad that they found our factorized implementation well motivated and our interpretation of matrix optimizers interesting. We respond to the raised concerns as follows.

W1. The paper's SFT downstream gains are within seed variations, which needs further justification for why this is a good indicator to support the framework.

We respectfully note that this holds for some but not all comparisons. On MATH, which has longer solutions and a larger test set than GSM8K (5K vs 1.3K test problems), we observe a clearer separation between optimizers: RoM is the only optimizer that belongs to the best-performing group in every math SFT setting (3B, 8B, 27B), while each baseline is separated below RoM in at least one setting.

More importantly, we argue that the held-out loss is the appropriate primary metric for comparing optimizers. An optimizer minimizes the training objective, while downstream accuracy additionally depends on the mismatch between token-level cross-entropy and exact-match accuracy, and on distribution shift. On this primary metric, RoM's gain is significant compared to seed variation in every SFT setting. This was also recognized by the other reviewers: Reviewer XB87 noted that "The experiments provide consistent evidence of improved optimization", and Reviewer McSh noted that "RoM achieves lower validation loss across the experiments". Consistent with this view, RoM's downstream gain is the clearest in RL, where the training objective is closely aligned with the downstream evaluation metric: RoM outperforms both Adam and Riemannion on all three downstream datasets, with non-overlapping intervals (Table 4).

Beyond the final metrics, two more signals support RoM's optimization advantage. First, Figures 1 and 2 show that RoM's gain is persistent over the training trajectory in both SFT and RL rather than appearing only at the final checkpoint, indicating faster convergence. Second, the learning rate sensitivity analysis (Section 4.4) shows that RoM is robust to learning rate selection and degrades less than baselines.

Summing them up, we argue that our experimental results provide strong evidence to support RoM's advantage

W2. The paper's visual presentation on (Pseudo-)Algorithm 1 and Figures 1/2/4 are below expectations. While the presented framework has good insights, the presentation needs extra effort to improve. Algorithm 1 has 6 lines of comments, which are almost equal to the number of actual operations in the algorithm. Figures 1/2/4 occupy overwhelmingly large spaces to show few line plots, which are not informative. Clearly conveying the framework would be beneficial for readers to understand the ideas.

We respectfully disagree that Algorithm 1 and Figures 1/2/4 are uninformative, though we agree they can be presented more compactly. The comments in Algorithm 1 separate different sub-steps of one RoM optimization step, and we believe a combination of text and math makes the algorithm more understandable than pure math notations. The figures show that RoM's gain persists over the training trajectory rather than appearing only at the final checkpoint

Updates in revised pdf: We have reduced the number of comments in the original "Algorithm 1" to 4, while still separating different sub-steps (in the revised pdf it's Algorithm 2). We have merged Figure 1 and Figure 4 into a single row (the new Figure 1 in the revised pdf). We used the saved space to expand the explanation of the normalization function

We hope our response and revision can address the reviewer’s concerns. We would be grateful if the reviewer could take these clarifications into account and consider revising their score. We are happy to answer any further questions during the discussion period

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer f2xm,

Thank you once again for your time and feedback! As the discussion period nears its end, we would greatly appreciate it if you could take a moment to check our rebuttal and revision. If you have any further questions or concerns, we would be more than happy to address them!

Thanks

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer f2xm,

The discussion period will end within one day. We would be grateful for your feedback on our response and revisions, and we are still available for any further questions. Thank you for your time!

Comments on the paper

1 comment
AuthorsOpenReview ↗

A summary of our response and revisions

Here we summarize our response and revisions:

  • We have corrected the missing constants in the norm constraints of our initial submission, which addresses the concerns raised by Reviewer XB87 and Reviewer McSh, and we are glad to see that Reviewer McSh has acknowledged that their concern has been resolved.
  • We have added the RL results for Riemannion, as requested by Reviewer XB87 and Reviewer McSh, and the results show that our proposed approach is consistently better than Riemannion
  • We have clarified how we address potential numerical issues in our implementation. As per the request from Reviewer XB87, we have provided the code implementation of our approach that reproduces the reported numbers: https://anonymous.4open.science/r/RoM-5118/
  • In our response to Reviewer f2xm, we have clarified why our experimental results show a convincing optimization advantage of our proposed approach, as also recognized by Reviewer XB87 and Reviewer McSh in their initial reviews.
  • We have made the algorithm and figure illustrations more compact, corresponding to Reviewer f2xm's concern

We thank all reviewers once again for their time and valuable feedback. We believe that we have adequately addressed their concerns through our latest responses and revisions. We are pleased that Reviewer McSh has acknowledged that their concern has been resolved and subsequently revised their score. Unfortunately, we have not yet received responses from Reviewer XB87 and Reviewer f2xm regarding our rebuttal; we sincerely hope they will have an opportunity to review our responses and revisions. We would greatly appreciate their reconsideration of their initial assessments and scores in light of our clarifications and revisions

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.