acceptodds
Under review as a conference paper at ICLR 2027

Epistemic Representation under Predictive Compression: Separation, Excess Rate, and Collapse

Abstract

Can a compressed predictive representation preserve the information needed to evaluate a future measurement? We study settings in which an epistemic functional of a parameter-indexed observation experiment is recoverable from the complete predictive marginal before compression. We characterize residual sensing-context information at the predictive rate–distortion boundary. For a coordinate-weighted Gaussian model, we prove an all-encoder stability bound and derive the exact additional rate required for prescribed log-Fisher fidelity, including achievability. We then connect this mechanism to the native predictive KL: marginalizing an uncertain target suppresses predictive sensitivity to measurement quality while leaving log-Fisher sensitivity unchanged. For Gaussian predictive-family decoders, a global natural-KL theorem gives a strictly positive epistemic excess-rate bound at positive primary predictive rate. We test the mechanism with stochastic neural bottlenecks under the same natural-KL loss. A validation-locked candidate sweep followed by ten independent replication seeds per target-uncertainty setting yields 40/40 predictive-selected models satisfying predictive fidelity but 0/40 satisfying epistemic fidelity, whereas 40/40 epistemic-aware models satisfy both. The empirical-mixture mutual-information gap is 1.05–1.16 nats, while aggregate-prior mismatch is small, and binary sensing-action error falls from approximately chance (0.499) to 0.137–0.140. These results quantify how information present in a predictive distribution can remain costly to preserve under predictive compression.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Predictive Compression, Epistemic Representation, Rate–Distortion Theory, Fisher Information, Active Sensing

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

This paper asks whether a compressed representation optimized for prediction preserves information needed to evaluate future measurements. It considers a particularly interesting setting: the epistemic quantity is recoverable from the complete predictive distribution, so any loss arises during compression.

Critical strengths
  1. The central distinction is clear and meaningful. Separating information absent from the predictive marginal from information lost during compression sharpens the paper’s contribution. The Type-II assumption makes the result stronger than simply showing that prediction and inference require different information.
  2. The quantitative Gaussian analysis is convincing. Lemma 10 bounds conditional context information for every feasible encoder near the predictive optimum. The exact joint rate in Theorem 12, together with achievability, provides a complete characterization within the stated regime. The corresponding proofs appear internally consistent.
  3. The natural-KL analysis substantially strengthens the argument. A weighted coordinate distortion could build the desired mismatch into the objective. Section 6 instead explains how marginalizing target uncertainty creates the mismatch under a predictive likelihood loss. The separate global argument in Theorem 17 appropriately avoids treating a local quadratic approximation as globally exact.
  4. The empirical protocol is careful about selection and interpretation. Validation-only configuration locking, ten independent replication seeds, stochastic-latent evaluation, and separate rate diagnostics improve credibility. The authors also correctly acknowledge that probe failure does not establish unrestricted decoder impossibility, and that posterior-to-prior KL is not mutual information.
Critical weaknesses
  1. The incremental theoretical novelty needs sharper justification. The exact Gaussian solution relies on established reverse water-filling, Gaussian rate–distortion converses, and independent Gaussian test channels. The general separation criterion is useful, but its proof follows directly from comparing its two defining information quantities. The most substantive contributions appear to be the all-encoder stability bound and the global natural-KL bound. The paper would benefit from organizing its novelty claim around these results and explaining more precisely what their analysis adds beyond classical compression with multiple fidelity constraints.

  2. The likelihood-based theorem remains restricted in an important way. Theorem 17 permits arbitrary encoders but restricts predictive reconstructions to product Gaussians. Unrestricted forward-KL reconstruction generally permits mixture distributions, which can change the predictive rate–distortion boundary. The paper explicitly acknowledges this limitation, but the bound on I(V;YE+)I(V;Y_E^+) does not itself establish the same excess-rate theorem for unrestricted decoders. Extending the result, or quantifying the effect of this restriction, would strengthen its relevance to learned predictive models.

  3. The experiments do not isolate the empirical cost of epistemic fidelity. In Table 1, epistemic-aware models achieve predictive distortion around 0.046–0.049, whereas predictive-selected models achieve approximately 0.079–0.087. Both satisfy the threshold, but the former use additional rate while delivering substantially better prediction. Consequently, the reported information gap cannot be attributed solely to epistemic preservation. The authors acknowledge this; a comparison at matched predictive distortion, using denser operating-point searches or constrained optimization, would make the empirical evidence more informative.

  4. The empirical scope is narrow relative to the motivation. All executed neural experiments use the same static, two-variable Gaussian construction. Varying target uncertainty tests the proposed mechanism, but does not establish robustness to correlated sources, nonlinear observation models, higher-dimensional contexts, or sequential learning. The binary sensing task also explicitly makes the optimal action depend on the sign of the context with a fixed utility margin. This is a useful controlled demonstration, but provides limited evidence about naturally arising sensing decisions.

  5. Decoder and reproducibility controls could be stronger. The epistemic results rely on a prescribed neural probe. Stronger probe families or an independent conditional-mean estimate would help distinguish information loss from readout limitations. Appendix E specifies many implementation details, but does not clearly provide training/validation/test sample counts, maximum training duration, or the complete probe-training protocol. Appendix G reproduces analytical checks rather than the neural study. These omissions limit independent reproduction.

Questions for authors

Please see the weakness part.

1 comment
AuthorsOpenReview ↗

Response on Near-Boundary Stability, Empirical Confounds, and Reproducibility

We thank the reviewer for the detailed technical assessment and for recognizing the all-encoder stability result and the global natural-KL analysis. We have sharpened the theoretical contribution and directly tested three major empirical alternative explanations.

1. Theoretical contribution beyond classical rate-distortion

Our novelty claim is not reverse water-filling or the mere existence of competing distortion criteria, but quantitative epistemic-information rigidity near the predictive rate-distortion boundary.

For every admissible stochastic encoder satisfying DP∗(Z)≤DD_P^*(Z)\le D and I(H;Z)≤RP(D)+ϵI(H;Z)\le R_P(D)+\epsilon, we establish

\qquad D_E^*(Z)\ge c^2\sigma_V^2e^{-2\epsilon/\kappa_\lambda}.$$ These results apply to all stochastic encoders under the stated Gaussian assumptions, not merely Gaussian test channels. The exact joint rate-distortion theorem further provides achievability and quantifies the required additional rate. A second substantive contribution is the global natural-KL analysis: marginalizing target uncertainty suppresses predictive sensitivity to sensing quality under a likelihood-based objective. Its finite-$M$ excess-rate bounds assume product-Gaussian reconstructions; we do not claim an unrestricted finite-$M$ decoder theorem. ### 2. Robustness to mixture decoders We trained 2-, 4-, and 8-component GMM predictive decoders (18 models). All satisfy $D_P\le0.1$; all nine predictive-only models fail $D_E\le0.16$, whereas all nine epistemic-aware models satisfy both thresholds. Mean epistemic MSEs are 0.635–0.648 versus 0.128–0.129. This directly tests whether the observed separation depends on unimodal Gaussian prediction. ### 3. Matched predictive fidelity We addressed the original Table 1 confound using a validation-selected three-seed comparison with the original single-Gaussian decoder at $M=32$. | Metric | Predictive-only | Epistemic-aware | |---|---:|---:| | $D_P$ | 0.08482 | 0.08359 | | $D_E$ | 0.63820 | 0.13644 | | Action error | 0.4981 | 0.1460 | | Variational rate | 1.6115 | 2.4136 | **Predictive distortion differs by only 0.00123, while epistemic MSE differs by 0.50176.** The separation therefore persists at matched predictive fidelity. An additional matched 8-GMM experiment provides consistent evidence. The different variational rates are consistent with the capacity mechanism but are not identified with the unrestricted theoretical $\Delta R_E$. ### 4. Stronger epistemic readouts We froze the matched-$D_P$ encoders and trained 64×2, 128×3, and 256×3 MLP probes, plus a nonparametric kNN conditional-mean estimator. Across three seeds, predictive-only epistemic MSE remains 0.6346–0.6367, near the zero-context benchmark of 0.64, whereas epistemic-aware MSE is 0.1348–0.1420. Thus, the separation persists under stronger readouts, although these estimates do not certify Bayes-optimal epistemic distortion. ### 5. Reproducibility and scope The additional GMM controls use independently generated validation/test sets of 5,000/6,000 samples per seed. Stronger-probe experiments use 35,000/6,000/12,000 training/validation/test samples. These controls use 650 optimization steps, distinct from the original early-stopping protocol. The revised manuscript documents the executed procedures and per-seed results. Neural predictive KL uses five-node numerical quadrature, whereas the theory uses exact forward KL. We distinguish variational KL, mutual information, and unrestricted excess rate. Our experiments isolate the mechanism within a controlled Gaussian sensing source; they do not establish generality across non-Gaussian sources, nonlinear observations, higher-dimensional contexts, or sequential policies. ### Conclusion **The principal theoretical and empirical concerns have now received direct, substantive responses.** The all-encoder rigidity and native-KL results clarify the scientific contribution, while matched-$D_P$, mixture-decoder, and stronger-probe controls address the principal alternative explanations. We respectfully request reconsideration of the initial assessment based on the additional evidence and clarified scope.
Reviewer_iJ7SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how compressing information for prediction can affect a system’s ability to recognize what it does not know. Its main finding is that a compressed representation can preserve enough information to make accurate predictions while losing information needed to assess uncertainty or distinguish between different possible explanations. The authors develop a framework to measure this trade-off and identify conditions under which predictive accuracy and uncertainty awareness can or cannot be preserved together. They also provide a mathematical example showing how much additional information may be needed to retain uncertainty-related details, and use a neural-network experiment to illustrate the problem in practice. Overall, the paper highlights that a representation that is good at making predictions is not necessarily good at representing uncertainty, and that preserving both may require additional information.

Critical strengths
  1. The idea of studying compressed representation can preserve enough information to make accurate predictions while losing information needed to assess uncertainty or distinguish between different possible explanations is interesting.

  2. The paper highlights a subtle but meaningful limitation of compressed representations: preserving predictive accuracy does not necessarily preserve information about uncertainty.

Critical weaknesses

1.The paper is difficult to follow because mathematical symbols and concepts are introduced before they are properly defined. For example, in Section 1, quantities such as DP(Z)D_P(Z) and DPreqD_P^{\mathrm{req}} are used before being formally introduced in Section 3. The notation should also be more clearly explained. For instance, Section 3 should explicitly define $d_P(\cdot,\cdot)$ as a measure of discrepancy between the true predictive distribution KK and its reconstruction K^\hat K, rather than leaving its meaning implicit.

  1. The application of the framework to the Gaussian case seems is less convincing for the following reasons:

a) The exact separation and excess-rate formulas rely on a specialized Gaussian model. It remains unclear whether similar results hold for non-Gaussian distributions and nonlinear observation models.

b) The separation is influenced by the weighted squared-error objective, which assigns relatively little importance to sensing quality in certain parameter regimes. Robustness to alternative predictive losses should be investigated.

c) The natural-KL result is established for a specified Gaussian decoder family, while the practical benefits are demonstrated for a particular sensing decision problem. Broader decoder classes and downstream tasks should be explored.

  1. The experiments focus on a specific Gaussian-based sensing setup and a finite family of neural bottleneck models. Testing on non-Gaussian data, different architectures, and more realistic sensing tasks would strengthen the evidence.

  2. The reported information-rate gap is estimated from learned models and does not equal the theoretical excess rate for unrestricted encoders. The experiments support the proposed mechanism, but further work is needed to quantify the fundamental rate cost and establish robustness across settings.

Questions for authors

Please see all the weaknesses above, additionally:

Could you clarify how general the proposed Gaussian separation result is? In particular, do you expect the separation phenomenon and excess-rate formula to hold for non-Gaussian distributions or nonlinear observation models, or are they specific to the assumptions of the current Gaussian setting?

3 comments
AuthorsOpenReview ↗

Response on Theoretical Scope, Natural-KL Mechanism, and Empirical Generality

We thank the reviewer for emphasizing clarity of presentation, theoretical scope, and empirical generality. We have revised the manuscript and conducted additional experiments directly addressing the predictive-loss and decoder-flexibility concerns.

1. Definitions and presentation

We have revised the Introduction to define KY(H,a)K_Y(H,a), E(H,a)=ϕY(KY(H,a))E(H,a)=\phi_Y(K_Y(H,a)), DP∗(Z)D_P^*(Z), and DE∗(Z)D_E^*(Z) before their first substantive use.

Section 3 explicitly defines

\mathbb E\left[ d_P\left(K_Y(H,A),\widehat K(Z,A)\right) \right].$$ This clarifies that predictive distortion measures the discrepancy between the true predictive distribution and its reconstruction from the compressed representation. ### 2. Weighted distortion versus natural predictive KL The separation is not based solely on a prescribed weighted squared-error objective. Section 6 independently analyzes forward predictive KL. Its local sensitivity to sensing context scales as $$\left(\frac{n(v)}{M+n(v)}\right)^2,$$ while squared log-Fisher sensitivity remains proportional to $c^2$. Increasing target uncertainty can therefore suppress predictive sensitivity to sensing quality without eliminating its epistemic relevance. Crucially, the finite-$M$ theorem uses a separate global expected-KL argument, not an extrapolation of the local quadratic approximation. ### 3. Generality of the theoretical results The general separation criterion and distortion-compatible collapse condition do not require Gaussian sources. However, the exact excess-rate formula and quantitative all-encoder stability bound are established under the specified Gaussian assumptions. **We do not claim that these closed forms extend unchanged to arbitrary non-Gaussian sources or nonlinear observation models.** The finite-$M$ natural-KL excess-rate bounds assume product-Gaussian predictive reconstructions, whereas asymptotic context suppression is established under unrestricted Bayes decoding. ### 4. Decoder robustness and matched predictive fidelity We trained 2-, 4-, and 8-component Gaussian-mixture predictive decoders (18 models). All satisfy $D_P\le0.1$; all nine predictive-only models fail $D_E\le0.16$, whereas all nine epistemic-aware models satisfy it. We also performed a matched-$D_P$ 8-GMM experiment at $M=32$: | Metric | Predictive-only | Epistemic-aware | |---|---:|---:| | $D_P$ | 0.07936 | 0.08076 | | $D_E$ | 0.6474 | 0.1300 | Thus, the epistemic gap persists despite nearly identical predictive fidelity. These controls establish robustness across the tested predictive decoders, not across arbitrary source distributions, encoder architectures, or downstream tasks. Non-Gaussian sources, nonlinear observations, and realistic sequential sensing remain beyond the present empirical scope. ### 5. Interpretation of empirical rates We explicitly distinguish variational rate from mutual information: $$R_{\mathrm{KL}} = I(H;Z)+D_{\mathrm{KL}}(q(z)\Vert p_0(z)).$$ Accordingly, learned-model rate differences are not estimates of the unrestricted theoretical excess rate $\Delta R_E$. The experiments support the predicted mechanism within the controlled model; the fundamental rate statements follow from the specified information-theoretic assumptions. ### Conclusion The revised definitions, independent natural-KL analysis, mixture-decoder experiments, and matched-fidelity control directly address the principal conceptual and empirical concerns. The assumptions and remaining generality limitations are now explicitly distinguished. **We respectfully ask the reviewer to reconsider the assessment based on these substantive clarifications and additional evidence.**

Thank you for clarifying the definitions in the paper. I now have a good picture of the paper's ideas, however, I will still keep my rating.

Thank you for confirming that the revised presentation has clarified the paper's ideas.

Beyond improving the definitions, our rebuttal provides additional theoretical analysis, 18 mixture-decoder experiments, and a matched-predictive-fidelity control addressing several of your principal concerns.

Could you please clarify which substantive concerns remain unresolved and continue to justify the original Weak Reject recommendation?

This would help both the authors and the Area Chair understand how the revised manuscript and additional evidence have been taken into account in your final assessment.

Reviewer_Kd8hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether predictive compression preserves information needed to evaluate future measurements. It considers settings where the full predictive distribution contains this information. It derives a general separation criterion and exact excess rates for a weighted Gaussian model. It also proves bounds under predictive KL with Gaussian-family decoders. Controlled neural experiments illustrate the effect on epistemic fidelity and sensing decisions.

Critical strengths
  1. The paper isolates compression loss from predictive insufficiency. The relevant information is recoverable before compression. Theorems 11–12 then quantify its loss near the predictive optimum. The bounds apply to all encoders under the stated assumptions.
  2. The natural-KL result goes beyond a chosen quadratic distortion. Theorem 17 uses a global KL bound. It does not rely on extending a local approximation beyond its valid range. This strengthens the connection to predictive learning.
Critical weaknesses
  1. The evidence remains confined to a restricted Gaussian setting. Theorem 17 assumes product-Gaussian predictive decoders. The neural experiments use the same source model and decoder family. They confirm the predicted behavior within this setting. However, they provide little evidence that it persists with more flexible predictive distributions. This limits the broader implications for predictive representation learning.
  2. The neural results do not isolate the rate needed for epistemic fidelity. Table 1 reports predictive distortion of about 0.046–0.049 for epistemic-aware models, versus 0.079–0.087 for predictive-selected models. Both satisfy the predictive threshold, so the feasibility comparison remains valid. However, some additional rate may support better prediction. Comparisons near the same predictive distortion would clarify how much of the gap is attributable to the epistemic requirement.
Questions for authors
  1. Does the observed separation persist with mixture predictive decoders on the same benchmark?
  2. Does the rate gap remain when models are compared near the same predictive distortion? Can the existing candidate pool support this comparison?
3 comments
AuthorsOpenReview ↗

Response on Mixture-Decoder Robustness and Matched Predictive Fidelity

We thank the reviewer for recognizing the significance of compression-induced epistemic separation and the global natural-KL analysis. We conducted two targeted experiments addressing the reviewer's questions.

1. Robustness to mixture predictive decoders

We trained learned 2-, 4-, and 8-component Gaussian-mixture predictive decoders using the same synthetic sensing source and 64-unit network width, with three seeds per configuration (18 models).

DecoderPred. DPD_PEpist.-aware DPD_PPred. DED_EEpist.-aware DED_E
2-GMM0.04020.04130.63520.1287
4-GMM0.03970.04190.63610.1294
8-GMM0.03940.04150.64760.1285

All 18 models satisfy DP≤0.1D_P\le0.1. All nine predictive-only models fail DE≤0.16D_E\le0.16, whereas all nine epistemic-aware models satisfy both thresholds.

The separation therefore persists across the tested mixture-decoder families, rather than depending on unimodal Gaussian prediction. We do not interpret this as unrestricted-decoder optimality.

2. Matched predictive fidelity and rate gap

We performed a dedicated three-seed matched-DPD_P experiment at M=32M=32 using an 8-GMM predictive decoder.

MetricPredictive-onlyEpistemic-aware
DPD_P0.079360.08076
DED_E0.64740.1300
Action error0.49870.1446
Variational rate (nats)1.6282.421

Predictive distortion differs by only 0.00140, while epistemic MSE differs by 0.5174 and the variational-rate gap remains approximately 0.793 nats. The epistemic-aware model achieves slightly worse predictive distortion, so improved predictive fidelity does not explain its epistemic advantage.

Regarding the existing candidate pool, its original selection criterion targeted threshold feasibility and rate minimization, not matched predictive distortion. We therefore conducted a dedicated matched-fidelity comparison rather than treating the original selected models as matched. We do not claim an exhaustive matching audit of the original pool.

The observed rate difference is consistent with additional capacity for epistemic preservation, but it is not identified with the unrestricted theoretical excess rate.

Conclusion

Both requested experimental controls have now been performed. The separation persists under more expressive predictive decoders and at closely matched predictive fidelity, together with an observed additional variational-rate cost.

These results directly address the reviewer's central concerns. We respectfully ask the reviewer to reconsider the assessment in light of the new evidence.

Thank the authors for the additional experiments. The mixture-decoder results address my concern that the separation might depend on a single-Gaussian decoder. The matched-fidelity comparisons also show that the epistemic gap persists at nearly equal predictive distortion. The observed rate gap is specific to the trained models, and the finite-MM bound in Theorem 17 still assumes product-Gaussian reconstructions. The rebuttal strengthens the empirical case, and I maintain my rating.

Follow-up on Addressed Concerns and Final Assessment

We thank the reviewer for recognizing that both requested experimental controls have addressed the original concerns regarding decoder flexibility and matched predictive fidelity.

We agree that the empirical rate gap is model-specific and that Theorem 17's finite-MM bounds retain the stated product-Gaussian reconstruction assumption. These limitations are explicitly distinguished from our general separation criterion and asymptotic result under unrestricted Bayes decoding.

Given that both principal experimental concerns have now been addressed, we respectfully ask whether the remaining scope limitations warrant retaining the original rating, or whether the strengthened theoretical and empirical evidence supports a more favorable assessment.

We appreciate your careful and constructive evaluation.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Rebuttal Summary and Request for Reassessment

Dear Area Chair,

We respectfully wish to emphasize that we have addressed the principal technical and empirical concerns raised by all three reviewers through additional controlled experiments, rigorous theoretical arguments, and substantial clarification of the manuscript. Importantly, our responses go beyond explanatory revisions. We have conducted three targeted experimental investigations that directly test the alternative explanations raised during review:

  1. Matched predictive fidelity: At nearly identical predictive distortions (0.08482 versus 0.08359), the epistemic distortion remains sharply separated (0.63820 versus 0.13644). This demonstrates that the observed epistemic gap cannot be attributed simply to the predictive-fidelity mismatch in the original comparison.

  2. More expressive predictive decoders. Across learned 2-, 4-, and 8-component Gaussian-mixture decoders, all nine predictive-only models satisfy the predictive-fidelity requirement but fail the epistemic-fidelity requirement, whereas all nine epistemic-aware models satisfy both. Thus, the observed separation persists beyond the original single-Gaussian decoder.

  3. Stronger epistemic readouts. Increasing probe capacity up to three-layer, 256-unit networks and introducing a nonparametric conditional-mean estimator does not eliminate the observed separation. This substantially weakens the explanation that the result arises from insufficient decoder capacity.

These findings are supported by our theoretical characterization. In particular, the all-encoder stability result quantitatively bounds the epistemic information retained by any stochastic representation near the primary predictive rate–distortion optimum. Together with the exact Gaussian excess-rate characterization and the global analysis under native predictive KL, this establishes a rigorous mathematical foundation for the reported phenomenon.

We have also clarified the assumptions and scope of each result, distinguishing general information-theoretic statements, Gaussian-specific theorems, and finite-model empirical evidence. We believe the combined theoretical and experimental evidence now directly addresses the major concerns underlying the initial reviews. In particular, the three principal empirical alternative explanations have been explicitly tested rather than left unresolved.

We therefore respectfully ask that the paper be assessed on the basis of the complete evidence presented in the manuscript and rebuttal, rather than solely on the concerns reflected in the initial ratings. We believe the work makes a substantive contribution by establishing when and why predictive compression can suppress information required to evaluate future measurements, even when that information is recoverable from the complete predictive distribution.

Thank you for your careful consideration.

Sincerely,

The Authors

AuthorsOpenReview ↗

Request for Reassessment of Reviewer iJ7S's Recommendation

Concern Regarding Reviewer iJ7S's Post-Rebuttal Assessment

Dear Area Chair,

Reviewer iJ7S acknowledges that our clarifications have resolved the difficulties in understanding the paper, yet maintains the original Weak Reject rating without identifying any remaining substantive objections.

Our rebuttal goes well beyond presentation improvements. It provides independent natural-KL analysis, theoretical results under unrestricted Bayes decoding, 18 additional Gaussian-mixture decoder experiments, and a matched-predictive-fidelity experiment addressing several of the reviewer's principal concerns.

Our concern is that the unchanged rating, in the absence of a reasoned post-rebuttal assessment, may no longer accurately reflect the scientific merits of the revised manuscript or the extent to which the original concerns have been addressed.

We respectfully request that the Area Chair independently evaluate the revised manuscript and the substantive evidence provided during rebuttal, rather than interpreting the unchanged rating as evidence that the original concerns remain unresolved.

Thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.