acceptodds
Under review as a conference paper at ICLR 2027

Reliable Characterization of Local Model Complexity with Valid Localization

18 views

Abstract

Understanding training dynamics and learned models requires more than evaluating loss or predictive performance. Modern deep neural networks(DNNs) are often overparameterized, and solutions with similar performance can have different local structure, which motivates a principled measure of model complexity around a solution. Singular learning theory provides such a framework, and the local learning coefficient(LLC) quantifies local model complexity, including cases where parameter symmetries and redundancies make distinct parameter values represent the same predictive function. A scalable sampling-based approach estimates LLC around a target solution using Gaussian localization. Because the localization penalty is part of the sampled target, its strength can change the measured local model complexity. Weak localization can admit remote low-loss regions, whereas excessive localization can impose an artificial contraction scale and distort the target complexity. By resolving the loss and localizer in a common geometric representation, we derive model-agnostic localization conditions under which localization-based measurements asymptotically converge to the LLC of the target solution across a broad class of analytic models, without estimating model-specific singular geometry. We further establish classwise counterexamples showing that neither the lower nor upper localization requirement admits a uniform relaxation over this analytic model class. The guarantee is asymptotic. For finite computations, we separate the discrepancy between the population localized measurement and the target LLC from empirical and numerical errors, and derive exact diagnostics that quantify sensitivity to localization. Experiments on truth-known singular models and DNN training trajectories show that theory-guided localization yields reliable LLC measurements and reveals reproducible complexity changes during grokking and learning-rate transitions. These results establish when localization-based LLC measurements can be interpreted as properties of a target solution rather than consequences of the localization choice.

Primary area:
learning theory
Keywords:
local model complexity, local learning coefficient, singular learning theory, localization, estimation reliability, learning dynamics

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

The paper analyzes when the Gaussian localization penalty in the sampling-based LLC estimator of Lau et al. (2025) isolates a target solution without altering its local learning coefficient. This is formulated as a two-sided argument: the localization must be strong enough to suppress remote low-loss regions, yet weak enough not to impose its own contraction scale. Under a joint resolution of the loss and localizer, the paper shows that any localization strength that grows faster than logarithmically but slower than any polynomial in the inverse temperature results in asymptotic convergence to the target LLC without any model-specific knowledge. In addition, the paper separates the error into localization, empirical, and numerical errors and provides the exact response identities. Experimentally, the theoretical results are validated in a two-dimensional setting with a known LLC, and the LLC trajectories of an MNIST grokking MLP and a CIFAR-10 ResNet20 with learning-rate drops are analyzed.

Critical strengths

S1. LLC is a theoretically well-founded framework with practical applications. The localization strength in LLC estimation is currently chosen heuristically. Therefore, the clear, model-agnostic answer the paper provides is of practical significance. Theorem 2 gives a precise sufficient growth class. The counterexamples make the two-sidedness tight. This gives the main formal contribution of the paper.

S2. Numerically, a toy experiment validates the predicted regimes, and the DNN experiments on real data demonstrate localization sensitivity and reproducible qualitative patterns.

Critical weaknesses

W1. The guarantee is asymptotic in t=n/log⁡nt = n/\log n, but all DNN experiments are run at one fixed tt, which does not give a complete picture that aligns with the theoretical results. The finite-scale quantities are derived but are not computed or presented for the DNN experiments. Therefore, the empirical part of the headline claim that “theory-guided localization yields reliable LLC measurements” feels overstated; instead, what is shown corresponds more to qualitative trends that are robust to γ\gamma.

W2. The DNN experiments do not exhibit a case where validity changes a conclusion. The two emphasized signals (a late decline in MNIST and an increase followed by a plateau on CIFAR) appear under every tested γ\gamma. Similarly, the only known-truth validation is a 2-D toy experiment; a higher-dimensional model with a known LLC would test the finite-tt loss–localizer competition where it actually matters.

W3. Some of the illustrative plots (e.g., Fig. 2) are harder to understand, and the effort needed to understand them outweighs their contribution to understanding. Similarly, Fig. 1b uses variables that are not introduced beforehand, and it is not clear how the curves are created. In addition, the caption describes grokking, whereas the panel seems to display learning-rate phases.

W4. The main theorem’s assumptions are not connected adequately to the empirical DNN checkpoints. The theory defines KK around a population minimizer and requires analytic resolved geometry. The experiments instead freeze intermediate training checkpoints in ReLU networks. Neither the minimizer assumption nor the required analytic setting is established. As in W1, this makes the connection between the numerical results and the theory weaker.


If the discrepancies between the theory and the simulations and experiments can be resolved, I would be willing to increase my score.

Questions for authors

Q1. How do the suggested γ\gamma values from the paper compare with those used in prior LLC studies, and which fall outside the sufficient guarantees of the paper?

Q2. At fixed t≈145t\approx145, what finite-scale criterion justifies choosing γ=log⁡tlog⁡log⁡t=7.98\gamma=\log t\log\log t=7.98 as the primary localization strength? Since Clog⁡tlog⁡log⁡tC\log t\log\log t is asymptotically admissible for every fixed C>0C>0, how is this particular choice calibrated or shown to preserve the target LLC at the experimental scale?

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer sHUA (1/2)

We appreciate the reviewer's detailed feedback, in particular, the one on the “discrepancies between the theory and the simulations and experiments.” We strengthened the connection through known-truth deep linear networks (DLNs) at several temperatures and explicitly separated those tests from exploratory DNN measurements.

Higher-dimensional known-truth evaluation (W1/W2).

To connect theory and experiments more strongly, we added the truth-known DLN experiments. Fig. 5 evaluates 20 DLNs with analytic LLCs: ten approximately 100K and ten 1M parameters (Aoyagi, 2024). By changing tt for population loss and nn for empir ical loss of DLNs, we confirmed dependence on tt or nn instead of multi-layer perceptron for grokking and ResNet20 for multi-stage linear perceptron. With γ=log⁡tlog⁡log⁡t\gamma=\log t\log\log t, median population errors stay below 2% over t∈{104,106,108}t \in \{10^4,10^6,10^8\}, while empirical errors reach 3.87% and 3.01% at the largest n=108n=10^{8}. Differences from fixed(γ=1\gamma=1)/linear(γ=t\gamma=t) localizations show that γ\gamma materially affects accuracy. These are known-truth finite-chain comparisons and we confirmed the validity of the asymptotic localization theory.

Connection between theory and experiments (W4).

To make the relationship between theory and experiments, as we explained it in the first item, we added the truth-known DLN experiments and evaluated how well we can confirm the asymptotic localization theory, in addition to the existing truth-known 2D toy model. By contrast, there are gaps in non-analytically of ReLU, training dynamics with checkpoints (not necessarily local minimum). Therefore, we cast the truth-known 2D toy model and 100k/1M DLNs for confirming the validity of our localization theory. We retain multi-stage learning rates in the main text because ResNet20 training while we increased the number of training data 1000 (submission) to 50000(rebuttal). By contrast, we moved the Grokking experiment to Appendix F because it does not have a strong relationship between asymptotic-tt valid localization theory and learning dynamics where a local minimum is not always guaranteed in training.

Revision of headline claims in Abstract (W1).

In Abstract, we changed the headline claim from “theory-guided localization yields reliable LLC measurements” to "Our theory establishes conditions for asymptotically recovering the target LLC, while the experiments distinguish quantitative accuracy against known ground truth from finite-scale sensitivity in DNN training." .

Finite-tt theory and diagnostics (W1/W2/W4).

Theoretically, Theorem 1 is asymptotic. Proposition D.1 gives a finite-tt partition-ratio bound, while Proposition E.5 rules out universal calibration. Empirically, Appendix D derives exact response identities. In the multi-stage learning rate experiment, Fig. 6 show that the 48k–52k change is 73.43±1.2073.43 \pm 1.20 for γ=(log⁡t)2\gamma=(\log t)^2 versus 67.54±11.8967.54 \pm 11.89 for γ=1\gamma=1, preserving the training-stage signal with less seed variability. Please note that we changed the number of training data from 1000 (submission) to 50000(rebuttal) because original ResNet20 experiments were conducted in the latter experimental setting (He et al., 2016). In Appendix F, we also provide Table 3 for computing sample-moment plug-ins across four checkpoints and four strengths, as suggested kindly by the reviewer.

AuthorsOpenReview ↗

Author Response to Reviewer sHUA (2/2)

Readability in figures (W3)

We moved Fig.2 in the submitted version to Appendix C for further reference in the rebuttal version. Instead, we add Fig.2 and Fig.3 for readers to understand intuitively the two localization failures and the core idea of our theory for expressing loss, localizer, and weighted volume as local powers with the same chart. Fig. 1(b) now summarizes the theoretical main result, rather than foregrounding learning dynamics because the main contribution of this paper lies in asymptotic valid localization theory rather than empirical findings in learning dynamics with finite-tt.

Choice of γ\gamma in previous work and finite-tt prefactors (Q1/Q2).

Published fixed strengths (γ=1\gamma=1 in Lau et al., 2025; γ=200\gamma=200 in Wang et al., 2025) differ in objective, temperature, and sampled subspace. Please note that even if it is outside Eq. (4)'s sufficient asymptotic class, it does not invalidate finite-tt measurements. At t=1000/log⁡1000≃145t = 1000/\log{1000} \simeq 145, γ≃7.98\gamma \simeq 7.98 with unit prefactor for the multi-stage learning rate and the grokking experiments was illustrative, not calibrated. Section 4 quadratic identity explains the absence of a universal finite-tt prefactor.

References

  1. Aoyagi, M. (2024). Consideration on the learning efficiency of multiple-layered neural networks with linear units. Neural Networks, 172, 106132.
  2. Lau, E., Furman, Z., Wang, G., Murfet, D., & Wei, S. (2025). The Local Learning Coefficient: A Singularity-Aware Complexity Measure. AISTATS, 244–252.
  3. Wang, G., Hoogland, J., van Wingerden, S., Furman, Z., & Murfet, D. (2025). Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient. ICLR 2025.
AuthorsOpenReview ↗

Official Comment by Authors

With the discussion phase approaching its end, we would sincerely appreciate it if you could take a moment to review the additional explanations presented in our rebuttal and share your thoughts. We trust that our responses have helped resolve the issues you raised, and we would be thankful if you could reconsider your initial evaluation in light of these clarifications. If any aspects of our work or responses remain unclear, or if you have further recommendations, we would welcome your input. We remain fully committed to addressing any outstanding concerns and incorporating further refinements into the manuscript as needed.

We are truly grateful for the time and effort you have devoted to reviewing our submission and for your valuable insights. We would greatly appreciate hearing your perspective before the discussion period closes.

Best regards, The Authors

Reviewer_2NwTOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The local learning coefficient (LLC), introduced by Lau et al. 2025, is a model complexity measure deriving from singular learning theory (SLT). Its estimator has become a useful tool for studying how networks learn: for example, detecting developmental stages in transformers (Hoogland et al., 2025) and tracking grokking (Cullen et al., 2026). It also underpins newer methods for interpretability and data attribution (Baker et al., 2026, Kreer et al., 2026). All of this relies on the estimator’s locality in parameter space, enforced by a Gaussian confining prior with variance 1/γ1/\gamma centered at the target parameter. Yet this mechanism was only justified heuristically, and its effect on the estimated quantity was never rigorously analyzed.

This paper provides theoretical support for the use of the Gaussian confining prior in LLC estimation. The core idea is to use a localization strength γ\gamma which changes with the effective inverse temperature tt: precisely, γt\gamma_t must increase faster than log⁡t\log t, but slower than every power tat^a. Within this regime, the authors prove that (a population version of) the localized estimator converges to the true LLC for a broad class of models, and show that the aforementioned conditions on γt\gamma_t cannot be relaxed in general. Experiments on a toy model with known LLC and on deep neural network training runs (grokking, learning-rate transitions) illustrate the predicted failure modes and the insensitivity of qualitative LLC behavior to the specific localization strength.

Critical strengths

This paper patches a significant theoretical gap in the foundations of LLC estimation: the justification behind the localization method. This has posed an obstacle to further theoretical characterization of LLC estimation, and in principle a soundness issue for the method itself; furthermore, I have seen several researchers attempt unsuccessfully to fix this issue. The results are model agnostic, which is important, given that LLC estimation is applied across a wide range of machine learning models and architectures.

The authors’ proof technique is nontrivial, and requires jointly performing resolution of singularities on the loss function and the localizing prior. They also show that the upper and lower bounds for γt\gamma_t are tight in general.

Beyond the main theorem, the paper offers useful conceptual clarifications. It shows that a numerically stable LLC estimate is not necessarily a correct one, since excessive localization can produce a stable but wrong regime. It also separates localization error from empirical and sampling error, which clarifies what any given LLC measurement is actually measuring. Both points can inform how the community interprets LLC estimates going forward.

Critical weaknesses

Although LLC estimation is an increasingly popular tool, I expect the direct readership of this paper to be primarily researchers working on the estimator’s foundations, which is a fairly limited audience. For practitioners, the experiments suggest that qualitative conclusions are already fairly robust to γ\gamma. This limits the paper’s immediate impact outside the singular learning theory community, though I believe it will be valuable to that group.

Given that the paper’s value lies mainly in its foundational role, its contribution would be substantially strengthened and less incremental if it fully established the theoretical foundations here. The original Lau et al. 2025 paper did not prove consistency of the estimator; given that there is an emerging research subcommunity built around variations and extensions of this estimator, establishing a rigorous foundation for the tool would be significant. The work in this paper bridges a significant gap towards establishing this consistency, but does not complete it. Were the authors to fully establish the consistency of the LLC estimator, I would consider raising my score.

Separately, I worry about the assumption that the true parameter w∗w^* be a global minimizer of the loss. In practice, LLC estimation is applied throughout the entire training trajectory, not only at the global minima. Of course, the meaning of the LLC estimator at non-local-minima is a preexisting theoretical problem I do not expect the authors to solve, but I am uncomfortable that the results do not even appear to apply at local minima which are not global minima.

Finally, I am unsure to what extent the experiments actually support the theoretical results in the paper. The admissible values of γ\gamma are only determined asymptotically, not at a finite tt; the experimental results do not appear to noticeably distinguish between admissible and inadmissible regimes as far as I can tell; and the theory’s assumptions are violated in many experiments (e.g. ReLU is not real-analytic, ResNet20 has batch norm, etc).

Questions for authors
  • Can one relax the assumption that w∗w^* is a global minimizer of the loss, in any way?
  • What are the main obstacles to extending Theorem 2 to the empirical target KnK_n, which (as far as I understand it) would yield consistency of the actual estimator?
2 comments
AuthorsOpenReview ↗

Author Response to Reviewer 2NwT

We appreciate the reviewer's careful assessment, especially the request to “fully establish the consistency of the LLC estimator.” We now give consistency results for actual finite-chain estimators under explicit statistical and sampling assumptions, not just for an ideal population Gibbs observable.

Estimator consistency (W2/Q2).

The main obstacles to extending the population result to the actual LLC estimator are twofold. First, the empirical excess loss KnK_{n} is random and is not necessarily nonnegative or minimized at the population anchor w∗w^{\ast}; its fluctuations may be amplified by the inverse temperature tnt_{n}. Therefore, applying Theorem 1 directly to KnK_{n} does not establish population LLC consistency. Second, convergence of the empirical Gibbs expectation does not guarantee convergence of a finite-chain estimator.

We address both obstacles. Proposition 5 controls empirical fluctuations near low loss and excludes spurious low-loss contributions elsewhere. For fixed-dimensional analytic categorical networks, Corollary E.13 establishes empirical Gibbs consistency with discrepancy Op(tn/n)O_{p}(\sqrt{t_{n}/n}) when tn/n→0t_{n}/n \to 0, without convexity or full Hessian rank. Proposition E.25 and Corollary E.26 then establish consistency of an actual finite-chain independence Metropolis–Hastings estimator, albeit with exponential worst-case cost. Propositions E.21–E.23 and Corollary E.24 additionally establish finite-update SGLD consistency for specified nonconvex quartic likelihoods, including Op((log⁡n)−1/2)O_{p}((\log{n})^{-1/2}) error under the explicit schedules of Propositions E.22–E.23.

These results provide a rigorous foundation for LLC estimation in nontrivial singular models, while efficient consistency guarantees for general trained DNNs remain open.

Nonglobal minima (W3/Q1).

We agree with your concern: "I worry about the assumption that the true parameter be a global minimizer of the loss." In full-space sampling, a distant lower-loss region can dominate when γt=o(t)\gamma_{t}=o(t) (Proposition C.7). Therefore, we also show Proposition 4/Eq. (5) by restricting sampling to a neighborhood with a unique local minimum. As Section 4 states, “No global-minimum assumption on LL outside VlocV_{\mathrm{loc}} is required” (p. 8). There, the lower localization-growth condition γt/log⁡t→∞\gamma_{t}/\log{t} \to \infty also becomes unnecessary. This extension covers strict nonglobal local minima, but not arbitrary nonstationary training checkpoints.

Finite-tt relevance (W1/W4).

Theoretically, Theorem 1 and Proposition 3 establish asymptotic validity and classwise sharpness, not finite-tt accuracy. Proposition D.1 supplies a partition-ratio-based finite-tt bound. Proposition E.5 rules out universal calibration. Fig. 4 checks exact Gibbs values and step-size effects for an analytic model.

By contrast, experimentally, Fig. 5 tests 20 known-truth deep linear networks(DLNs) over several tt/nn, with below 2% median population errors for γ=log⁡tlog⁡log⁡t\gamma=\log t\log\log t in both 100K/1M families. Unlike the original training-dynamics experiments, these known-truth comparisons directly quantify estimation errors and reveal substantial accuracy differences across localization strengths. There, flagged fixed-strength γ=1\gamma=1 runs remain reported. Fig. 6 shows multi-stage learning-rate dynamics in training for ResNet20. In comparison with the submitted version, we increased the number of training data from 1000(submission) to 50000(rebuttal). By contrast, we moved grokking to Appendix F because it does not have a strong relationship between asymptotic-tt valid localization theory and learning dynamics where a local minimum is not always guaranteed in training. As the reviewer pointed out, the analytic-minimizer assumptions are not established for the trained DNN checkpoints; in particular, ReLU networks do not satisfy the real-analyticity assumption.

These model-specific results are intended to support further theoretical work toward LLC estimators whose reliability can be established rather than assumed.

Readership and readability (W1).

We reorganized the manuscript, in particular, in Section3 and 4, for readers unfamiliar with singular learning theory (SLT), using intuitive examples and diagrams before the formal results. Beyond SLT specialists, these results are relevant to ML researchers using LLC measurements to study training dynamics and learned representations, because the known-truth experiments show that localization choices can substantially affect quantitative accuracy even when qualitative trends appear stable.

References

  1. Mengersen, K. L., & Tweedie, R. L. (1996). Rates of Convergence of the Hastings and Metropolis Algorithms. Annals of Statistics, 24(1), 101–121.
AuthorsOpenReview ↗

Official Comment by Authors

With the discussion phase approaching its end, we would sincerely appreciate it if you could take a moment to review the additional explanations presented in our rebuttal and share your thoughts. We trust that our responses have helped resolve the issues you raised, and we would be thankful if you could reconsider your initial evaluation in light of these clarifications. If any aspects of our work or responses remain unclear, or if you have further recommendations, we would welcome your input. We remain fully committed to addressing any outstanding concerns and incorporating further refinements into the manuscript as needed.

We are truly grateful for the time and effort you have devoted to reviewing our submission and for your valuable insights. We would greatly appreciate hearing your perspective before the discussion period closes.

Best regards, The Authors

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

The paper studies the validity of Gaussian localization for estimating local learning coefficients (LLCs). It derives asymptotic conditions on localization strength, provides counterexamples illustrating their necessity for uniform guarantees, and presents experiments on synthetic models and neural networks.

Critical strengths

The research question is relevant.

The claimed theoretical results seem strong and to provide useful conditions for localization.

Critical weaknesses

Mathematical exposition is not sufficiently self-contained for a non-expert in singular learning theory (i.e. most of ICLR audience). The preliminaries and the results are difficult to follow without substantial prior knowledge in singular learning theory. Central concepts, particularly the LLC and its connection to the proposed estimator, are insufficiently defined or motivated in the main text. Important assumptions are introduced only later or relegated to the appendix.

The authors should consider either submitting to another more specialized venue or improve the writing to allow ICLR audience to read the paper without difficulty.

I would be willing to change my score if during the revision period the authors are able to perform a significant rewrite that would make the paper easier to follow.

Questions for authors

Could the authors provide a more self-contained introduction to the necessary preliminaries and to their results, including their precise definition of LLCs and connection to the proposed estimator?

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer ChLn

We appreciate the reviewer's constructive feedback, particularly the request for a “significant rewrite that would make the paper easier to follow.”

Self-contained explanation and enhanced readability (W1).

We reorganized the manuscript, in particular, in Section3 and 4, for readers unfamiliar with singular learning theory (SLT), using intuitive examples and diagrams before the formal results.

Accessible formulation and estimator link (W1/Q1).

Section 3 proceeds from population/empirical losses to weighted low-loss volume, contrasting K(u)=u2K(u)=u^2 and K(u)=u4K(u)=u^4 (LLCs 1/21/2 and 1/41/4). As the text explains, “LLC measures the rate at which this volume shrinks” (p. 4). Definition 1 and Eq. (1) then connect local geometry to the sampling-based Gibbs observable. Eq. (2) distinguishes localization, finite-data, and sampling errors instead of equating finite-tt measurements with LLC.

Explicit assumptions and visual explanation (W1).

Fig. 1 introduces the two localization failures and highlights the main contribution of this paper for asymptotic requirements. Fig. 2 demonstrates the failures analytically. Fig. 3 explains joint coordinates for loss and localizer. Assumption 1 and Theorem 1 (Eq. (4)) now appear in Section 4, with the resolution argument in Appendix B (Proposition B.3, Corollary B.4). Global optimality, target isolation, and integrability are stated explicitly; Proposition 3 establishes classwise boundary sharpness. Figs. 4–5 test known-LLC accuracy, while Fig. 6 is labeled a finite-chain DNN illustration.

Importance to ICLR (W2).

LLC-based methods already inform analysis of learning stages, attention specialization, and learned network structure (Lau et al., AISTATS 2025; Wang et al., ICLR 2025; Gordon et al., ICML 2026), along with work on neural parameter symmetries (Lim et al., NeurIPS 2024). As Section 2 states, we derive “validity conditions for the Gaussian localization” (p. 3) underlying these measurements. This is a foundational reliability question for ML research, which makes ICLR an appropriate venue.

References

  1. Lau, E., Furman, Z., Wang, G., Murfet, D., & Wei, S. (2025). The Local Learning Coefficient: A Singularity-Aware Complexity Measure. AISTATS 2025, 244–252.
  2. Wang, G., Hoogland, J., van Wingerden, S., Furman, Z., & Murfet, D. (2025). Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient. ICLR.
  3. Gordon, A., Baker, G., Wang, G., Snell, W., van Wingerden, S., & Murfet, D. (2026). Towards Spectroscopy: Susceptibility Clusters in Language Models. ICML, 36247–36314.
  4. Lim, D., Putterman, T. M., Walters, R., Maron, H., & Jegelka, S. (2024). The Empirical Impact of Neural Parameter Symmetries, or Lack Thereof. NeurIPS, 28322–28358.
AuthorsOpenReview ↗

Official Comment by Authors

With the discussion phase approaching its end, we would sincerely appreciate it if you could take a moment to review the additional explanations presented in our rebuttal and share your thoughts. We trust that our responses have helped resolve the issues you raised, and we would be thankful if you could reconsider your initial evaluation in light of these clarifications. If any aspects of our work or responses remain unclear, or if you have further recommendations, we would welcome your input. We remain fully committed to addressing any outstanding concerns and incorporating further refinements into the manuscript as needed.

We are truly grateful for the time and effort you have devoted to reviewing our submission and for your valuable insights. We would greatly appreciate hearing your perspective before the discussion period closes.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.