acceptodds
Under review as a conference paper at ICLR 2027

From Cross Mapping to Stable Dynamical Factors: Learning Aligned Dynamical Structure in Nonlinear Systems

Abstract

Learning aligned dynamical structure from paired nonlinear time series calls for a population correspondence alongside predictive skill. Cross-map prediction leaves this structural target unspecified. We define an aligned stable dynamical factor: a map that aligns paired states, commutes with dynamics, transports transmitted tangent directions, and controls metric-relative conditioning. Atlas GCCM–BCD and Neural GCCM estimate this relation with hard local and global differentiable surrogates, respectively. Matched-budget evidence isolates lower state and tangent incompatibility from structured losses, without uniform factor-error or categorical-direction gains; geometry and rank differences remain unresolved. The target uses latent dynamical manifolds (LDMs), combining reconstructed state, induced dynamics, a metric, and invariant measure. We prove reconstruction transport of an assumed factor and conditional regularized-target consistency for idealized globally smooth estimators under dependent sampling, plug-in consistency, risk separation, and approximate global optimization; hard-chart Atlas and no-factor consistency remain outside the result. An ancestry reading is external and conditional on ordered generative dynamics, component-local sensing, causal sufficiency, and absence of observational equivalence; otherwise the relation describes aligned dynamical information containment. Synthetic studies examine factor recovery, matched-independent negative controls, adverse feedback, synchronization, robustness, controlled observation-map dimension and conditioning, representation diagnostics, and targeted external baselines.

Primary area:
causal reasoning
Keywords:
causal reasoning, nonlinear dynamical systems, convergent cross mapping, dynamical systems, manifold learning, causal discovery, time series, 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_ahfJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends convergent cross mapping (CCM) from dynamical and geometric perspectives. While conventional CCM infers causal relationships through directional predictive skill, this paper introduces an aligned stable dynamical factor as its central concept. The main idea is that, when a smaller system (XX) drives a larger system (YY), the latter should not only predict the former but also consistently represent its dynamical evolution.

The proposed aligned stable-factor relation is formalized in Definition 4.3. It requires dynamics compatibility between two systems through a factor map Φy→x\Phi_{y \to x} satisfying this condition. Under appropriate submersion/constant-rank assumptions, the tangent space naturally decomposes into the kernel/vertical directions (which do not change the factor state xx to first order) and their orthogonal complement, the horizontal directions (which change xx). The authors focus on defining a stable dynamical factor map by constraining the horizontal mapping to be neither excessively contractive nor expansive (horizontal geometric stability).

The existence of such a stable factor map indicates directional dynamical information containment, which can be interpreted as causal ancestry under additional structural assumptions. As in conventional CCM, the framework operates on delay-reconstructed state spaces, with observationally indistinguishable latent states identified through observable quotients.

For implementation, the authors propose two estimators: Atlas GCCM-BCD and Neural GCCM. Their objectives incorporate required state alignment, dynamical compatibility, and additional geometric regularizers for the stable-factor relation. Several synthetic experiments are presented to evaluate factor recovery, directional identification, and robustness.

Critical strengths

The paper makes an interesting attempt to incorporate dynamical and geometric structures into CCM. Although related ideas have been explored in previous work (please see Q5), this paper provides a particularly detailed geometric formulation, such as the vertical-horizontal decomposition. I find this decomposition of the factor map particularly insightful, as it provides a geometric perspective on distinguishing dynamically preserved information from directions eliminated by the projection. This framework could serve as a foundation for further geometric investigations into the relationships between dynamical systems.

Critical weaknesses

W1. I believe the Introduction would benefit from a more reader-friendly and motivating explanation. Currently, the Introduction focuses more on describing what the authors have done than on explaining what problems they aim to solve, presenting their contributions through highly technical terminology. This makes it difficult for readers to understand the motivation behind the work. Even after carefully reading the paper and revisiting the Introduction, I still find it too condensed and abstract, particularly for material presented on the very first page. Similarly, the related work section mainly lists previous studies without clearly discussing what existing methods have achieved and what limitations remain. As a result, it is difficult to understand the specific gap this work aims to address and its main contributions.


W2. The paper does not adequately discuss or compare several closely related CCM-based methods, particularly those incorporating dynamical and geometric structures (please see Q4 and Q5). Although the proposed geometric formulation is elaborate, it remains unclear whether these additional structures are necessary or provide substantial advantages over existing methods. Without appropriate comparisons and ablation studies, it is difficult to evaluate the practical significance of the proposed framework.

Questions for authors

Q1. It seems that the horizontal stability is one of the notable contributions of this paper; I have two questions regarding its practical significance:

Q1-1. Under the submersion assumption, the horizontal singular values are already positive and finite pointwise. What additional practical benefits arise from explicitly controlling these bounds? Specifically, could the authors provide empirical evidence that better horizontal conditioning of the factor map improves the identifiability of the dynamical relationship, or at least the accuracy of factor map reconstruction, as motivated by Theorem 6.1?

Q1-2. How can the appropriate lower and upper bounds (mm and MM in Eq. 16) be determined in practice? Are there theoretical or algorithmic procedures for estimating these bounds?


Q2. How does the choice of metric and horizontal connection affect the proposed framework? The horizontal subspace is defined as the gg-orthogonal complement of the vertical subspace, with the metric gg currently chosen as Euclidean in delay coordinates. How sensitive is the proposed method to this choice? More generally, any complementary subspace to the vertical distribution can define a horizontal distribution, corresponding to a choice of Ehresmann connection. Is there a particular theoretical reason to prefer the metric-orthogonal connection? Does this choice provide any specific advantage for identifying dynamical factor relations?


Q3. It seems that the tangent identity in Eq. 17 appears to be the differential counterpart of the dynamical compatibility condition in Definition 4.3. Thus, the tangent loss and dynamical compatibility loss appear to enforce closely related constraints at different implementation levels. Is there a particular reason for including both losses?


Q4. What practical advantages does the proposed framework provide over conventional CCM? I may be overlooking relevant results, as I am not an expert in CCM. However, the current experiments do not seem to clearly demonstrate the motivating case: where conventional CCM produces incorrect causal predictions due to its reliance on predictive skill, while the proposed framework successfully identifies the correct relationship owing to the additional geometric structure. Most experiments appear to focus on characterizing the behavior of the proposed methods on selected dynamical systems. Direct comparisons with conventional CCM and recent CCM-based methods would substantially strengthen the paper's main motivation.


Q5. Two particularly relevant papers [1, 2] appear to be missing from the discussion or experimental comparisons. First, Latent CCM [1] combines CCM with latent neural ODEs for causal discovery, making it closely related to the proposed framework.

More importantly, Tangent Space Causal Inference (TSCI) [2] shares a substantial part of the paper's central motivation: improving CCM by explicitly incorporating dynamical structure. In particular, TSCI uses a pushforward relation between vector fields, which can be understood as the infinitesimal form of flow semiconjugacy; the dynamical compatibility condition used in the present paper can be understood as the discrete version of flow semiconjugacy. Thus, both methods essentially exploit the same underlying dynamical compatibility principle, although the present paper additionally introduces vertical-horizontal structures and corresponding stability conditions.

Therefore, I believe a more detailed discussion of these connections, especially with TSCI, is essential to clarify the novelty of the proposed framework. A direct experimental comparison would also help establish whether the additional geometric constraints provide meaningful advantages beyond existing geometric methods.

[1] De Brouwer, E., Arany, A., Simm, J., & Moreau, Y. (2020). Latent convergent cross mapping. In International Conference on Learning Representations.

[2] Butler, K., Waxman, D., & Djurić, P. M. (2024). Tangent space causal inference: Leveraging vector fields for causal discovery in dynamical systems. Advances in Neural Information Processing Systems, 37, 120078-120102.


Minor comment: Author-defined acronyms, such as Atlas-GCCM-BCD, frequently appear without being defined. Please spell out each acronym upon its first occurrence.

1 comment
AuthorsOpenReview ↗

Response to Reviewer ahfJ

Thank you. We rewrote the motivation and now foreground two controlled diagnostic findings while directly testing the structural terms.

  1. Practical value and recovery (Q1, Q4; Table 4, Appendices N.3 and O). The train-only local-affine adaptation attains clean shear/cover velocity cosines .9961/.9806 while full tangent-action errors remain .1217/.2644. These evaluations use the same test trajectories, but 1140 velocity-valid interior states versus 1152 clean-query states. In the clean a=.1 weak-direction experiment, Full's aggregate factor NRMSE is .0928, yet separately normalized weak-direction error is .5754; its fitted small-gain 5th percentile is .0194 against the true .1. Velocity alignment, aggregate error, and retained-direction recovery therefore require separate evaluation. Oracle diagnostics reveal these failures; they do not prove that our screen detects them or geometry regularization repairs them.

Two nonlinear factors, strong paired/kernel baselines, matched 8000-update training, and a disclosed post-hoc matched-16000 check show no consistent structural-loss advantage. Full and No-geometry nearly coincide. We do not establish that enforcing better horizontal conditioning improves identification or recovery.

The fixed-library CCM-style diagnostic attains reverse rho≈1 and NRMSE≈.002 at L=768 on the clean double cover although no global inverse factor exists (Appendix N.4). This illustrates prediction versus factor existence; it is neither a full CCM causal benchmark nor a demonstrated GCCM correction.

  1. TSCI and related methods (Q5; Related Work, Appendix O). We explicitly acknowledge that TSCI shares dynamical intertwining, Jacobian pushforward, and non-invertible cross-maps. Its default official-style forward cosines are .9748/.9339 clean and .9444/.9027 noisy; Neural scores are higher, but paired regression matches Full. Official-style TSCI uses the paired query target to estimate its Jacobian. The train-only adaptation above is separately labeled, not evidence that the official estimator fails at factor recovery. Latent CCM is discussed but was not experimentally run.

  2. Bounds and metric (Q1-Q2; Proposition M.1, Appendix A.3). A C1 submersion on a compact source with smooth source and target metrics already has uniform positive/finite horizontal bounds; their existence is not new. Our quantitative margin specifies tolerated derivative error. Training endpoints m0=.05 and M0=20 are heuristic soft barriers, not estimates of true bounds; we establish no calibrated procedure for selecting them. Given the source metric, the orthogonal lift is the unique minimum-norm preimage. Transporting both metrics preserves coordinate invariance. We establish neither an identification advantage of this connection nor robustness to changing the metrics.

  3. Two compatibility losses (Q3; Table 1). The tangent identity differentiates semiconjugacy, but finite value fitting need not recover derivatives. This motivates separate diagnostics; the ablations do not establish universal necessity or superiority of both penalties.

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

This paper introduces a new method for estimating the relationship between dynamics underlying two sets of observations, and provides a definition for the relationship between dynamics beyond prediction of one from the other. They attempt to estimate maps between the dynamics directionally, using a general non-parametric estimator and a neural estimator. They perform several experiments to evaluate their method, assessing recovery of known latent factors under various noise levels and feasibility of recovered latent factors.

Critical strengths

The main strength of this paper is that it provides both positive and negative results (such as experiment 6a), allowing for a balanced evaluation of their method.

Critical weaknesses

The main weakness of this paper is that it is very dense and challenging to untangle exactly what the authors did, both theoretically and experimentally:

  • In the main text, there is a large amount of notation that is not properly introduced or explained (such as ξ,dgx2(⋅),D,σ^,Ω(Φ))\xi,d^2_{g_x}(\cdot),D,\hat{\sigma},\Omega(\Phi)), making the motivation and problem setup very challenging to extract.
  • the justification for their objective is unclear: What the five terms are trying to regularize is not fully explained, nor is the tradeoff between them
  • the actual implementation and execution of their experiments is not self-contained; even with the extensive appendix it is challenging to untangle what the authors carried out. Experiments sweep over library sizes, but this term is never defined. Nor is the retained trajectory budget used in experiment 3. In some cases, such as experiment 3, it is unclear whether factor recovery is a positive or negative result.
Questions for authors

For the acceptance decision gates: what direction indicates acceptance, and how were the acceptance numbers chosen? Is performance sensitive to these thresholds?

How do the four terms loss terms relate to each other? How are they calculated in practice? How much do the results depend on their relative weighting?

What do the G and BCD stand for in your method's name?

3 comments
AuthorsOpenReview ↗

Response to Reviewer aFsG

Thank you. We revised the paper around concrete questions and made the execution protocol explicit.

  1. Motivation and notation (Introduction, Figure 1, Tables 1-2). The map (u1,u2,v) -> (u1,epsilon*u2) now distinguishes legitimate removal of v from poor preservation of u2. GCCM means Geometry-aware Convergent Cross Mapping; BCD means block-coordinate descent. X->G Y is evaluated by fitting Y->X. Table 1 explains paired fit, dynamics commutation, tangent commutation, retained-direction conditioning, and coefficient regularization. Dynamics constrains values and tangent commutation constrains derivatives; reductions and weights are stated because identical coefficients need not imply identical contributions.

  2. Reproducible execution (Algorithm 1; Appendices G, K, L, N, O). Library size is the selected sample prefix; retained trajectory count is reported separately. Experiment 3's L=2000 eight-orbit prefix contains only two orbits, and its outcome is explicitly mixed. Effective training/validation counts, preprocessing, and checkpoint selection are now distinguished. The original matched-budget ablation uses 72 scheduler cycles and 288 sampled updates; these are not complete dataset passes. Historical RSB-v2 Pair-only has no L2, whereas the matched ablation and new Pair+L2 controls retain the shared penalty. Raw-state discussion experiments are explicitly separated from the original delay-space standardization.

The nonlinear study uses 2304/1152/1152 training/validation/test pairs from disjoint trajectories and scores every validation state. Histories, selected/resume checkpoints, and per-query outputs are archived. Table 4 defines factor/full-tangent NRMSE locally. The matched comparisons show no consistent structural-loss gain; geometry-weight sensitivity also does not establish general weight robustness.

  1. Gates and sensitivity (Table 2; Appendices G.2, N.4, N.6). Acceptance requires all gates: dimension feasibility, first-coordinate RMSE <=.8 and correlation >=.35, vector state RMS <=.8, tangent residual <=.9, and positive rank LCB. The four non-rank tolerances are frozen heuristics; the rank-rule construction was selected on separate calibration controls using the joint screen and historical per-coordinate state metric. On the original holdout controls, the current vector-state replay gives 56/60 positive and 0/90 null passes, versus historical 60/60 and 0/90. Local sensitivity preserves the current counts, but circle clean-forward BCD acceptance changes 0/2/10 at tangent cutoffs .81/.90/.99. We keep .90 fixed and do not claim generally calibrated screening.

Thank you for your response and your revisions! I appreciate the effort that you have put towards improving this paper. The name is now explicitly defined in the paper, and there is an explicit table documenting what the terms in the loss are meant to regularize/penalize. There is also now an extensive appendix documenting their experiments. However, reading through the revisions, it is still very challenging to understand. Much of the notation I mentioned in my comment is still undefined or unexplained, in addition to some acronyms (RSB-v2). Some additional points that remain difficult to untangle:

  • How the terms optimized relate to the loss terms. There is no explicit link between Φ\Phi, B, σ\sigma, and MM.
  • An intuition for picking loss weights. The authors state that that equal weights do not imply equal contributions or minimizers; the authors in their response restate this fact but do not provide evaluation of the sensitivity of their model to weights other than λgeo\lambda_{geo}
  • Rationale for their cutoff choices. Again, the authors state that the cutoffs are operational, not calibrated significance levels, and I appreciate their transparency -- but this is worrying from a reproducibility standpoint. How can future users implement and assess this method if they don't have a good way to select these cutoffs?

The even more extensive appendix further accentuates these problems, it is far too much to ask any reader to go through. For future submissions of this work, I would recommend making the intuition behind the method a heavier focus in the main text: why did you make the design choices that you made, why do they work, and how they relate to prior work could all be much clearer. Similarly, the experiments should be aimed at showing the intuition and assumptions behind the model hold; think about the minimal set of experiments needed to show that your method works as you intend and expand from there. In this current state it's unclear why most of the experiments were performed, what they contribute towards the assessment of your method, and how the results support your method.

AuthorsOpenReview ↗

Clarification of implementation, weights, and cutoffs

Thank you for rereading the revision. We agree that adding appendix detail does not by itself fix the main-text explanation. Our previous response also did not resolve your questions about selecting weights and cutoffs.

  1. The implementation link. The reconstructed state spaces are denoted by My,Mx\mathcal M_y,\mathcal M_x, and Φ\Phi maps a source state to its paired target state. Within an Atlas chart, Φ^(q)=cx+Ux[BUy⊤(q−cy)+b],\widehat\Phi(q)=c_x+U_x[B U_y^\top(q-c_y)+b], where cic_i are chart centers, UiU_i are local PCA bases, and ξi=Ui⊤(q−ci)\xi_i=U_i^\top(q-c_i) are local coordinates. Atlas optimizes B,bB,b; Neural instead optimizes Φθ\Phi_\theta and computes B=Ux⊤DΦθUyB=U_x^\top D\Phi_\theta U_y. Thus BB represents the map's derivative in tangent coordinates, and σj(B)\sigma_j(B) are its local gains. Dynamics loss compares Φ^(qy,t+1)\widehat\Phi(q_{y,t+1}) with F^x(Φ^(qy,t))\widehat F_x(\widehat\Phi(q_{y,t})); tangent loss compares Bt+1AyB_{t+1}A_y with AxBtA_xB_t, where AiA_i advances a local tangent perturbation. Geometry loss penalizes gains outside [m0,M0][m_0,M_0]. Plain MM denotes a population upper gain bound, whereas M0=20M_0=20 is a chosen soft-barrier endpoint; neither is another optimized map. The squared distance dgx2d_{g_x}^2 measures target-state error under the chosen metric, and Ω\Omega penalizes fitted coefficients. RSB-v2 means Reconstructed Shared Benchmark v2, our frozen shared-data benchmark, not another estimator.

  2. Weights. The intended trade-off is paired-data fit versus time-evolution, derivative, and conditioning preferences; errors in estimated dynamics/tangents can make these preferences harmful. The three structural coefficients of 0.1 and the coefficient penalty of 10−410^{-4} were fixed heuristic settings, not validated optimal choices. Our component ablations compare zero with the fixed nonzero weight; only the geometry coefficient received a multi-value nonzero sweep. These do not establish sensitivity to the remaining weights, and reweighting an evaluation score on fitted maps does not fill that gap.

  3. Cutoffs. Their purpose was to reject poor prediction, incompatibility, or degeneracy, but the four non-rank values were heuristic; we have no derivation or validation-based selection record for them. Separate calibration selected the rank-rule construction, not those four values, and used the historical state-error normalization. The concern about transfer to new data is therefore valid: fixing a rule makes our reported decisions reproducible but does not justify its use elsewhere. The fitted map and continuous diagnostics can be reported separately; categorical screening on new data requires additional, independently validated calibration that this paper has not established.

The smallest experimental argument is: nonlinear factors test actual map/derivative recovery against strong paired regression; velocity-versus-full-tangent comparisons test what trajectory alignment misses; weak-direction tests show what aggregate error misses. These support separate diagnostics, but do not demonstrate a consistent recovery benefit from the added losses. We acknowledge that this rationale needed to be clearer in the main text.

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

The paper studies how to recover directed dynamical structure from paired nonlinear time series. It introduces an aligned stable dynamical factor: a potentially dimension-reducing map satisfying paired-state alignment, dynamical compatibility, and conditioning bounds on horizontal tangent directions. Two estimators are proposed, Atlas GCCM–BCD and Neural GCCM. The theoretical results establish transport of an assumed factor through valid delay reconstructions and conditional consistency for an idealized regularized estimator. Synthetic experiments investigate factor recovery, negative controls, synchronization, noise, and observation-map conditioning.

Critical strengths

A well-defined structural target beyond cross-map prediction. The paper formulates aligned stable dynamical factors through paired-state alignment, dynamical compatibility, and conditioning on horizontal tangent directions. The horizontal formulation accommodates dimension reduction without requiring preservation of response-specific directions. This is the strongest consideration supporting acceptance: it provides an interpretable mathematical object for studying dynamical information containment and clarifies the additional assumptions needed for an ancestry interpretation.

Critical weaknesses

There are two critical weakness:

  1. The theoretical guarantees do not cover the principal implemented estimator. Theorem 6.2 explicitly excludes Atlas GCCM–BCD because its hard chart switching does not satisfy the globally smooth estimator assumptions. For Neural GCCM, the required consistency and optimization conditions remain unverified, and the implemented surrogate differs from the displayed theoretical objective. Theorem 6.1 transports an already-assumed factor through reconstruction coordinates rather than establishing recoverability. Although these limitations are acknowledged, they leave the main algorithm without a demonstrated connection to the recovery guarantee. This materially weakens the paper’s theoretical contribution.

  2. The experiments do not sufficiently establish the added value of structural constraints for factor recovery. The matched-budget comparison demonstrates lower state and tangent incompatibility, but does not establish improvements in factor error, geometric conditioning, rank, or categorical decisions. Lower compatibility residuals alone do not show that the true factor or its derivative is recovered more accurately. Moreover, Experiment 1 rejects the reverse direction through dimensional infeasibility before fitting, limiting its evidence for learned directional discrimination. These findings leave uncertain whether the additional machinery delivers a substantial recovery advantage over well-tuned paired regression, which is central to assessing the method’s significance.

Questions for authors
  1. What guarantee or quantitative error analysis directly applies to the implemented Atlas estimator?

  2. Do the structural losses improve ground-truth factor or Jacobian recovery under matched budgets and adequately tuned baselines?

  3. How much directional discrimination remains when both directions pass the dimension-feasibility check?

  4. How reliably do the screening rules distinguish stable factors from near-degenerate mappings across observation representations?

1 comment
AuthorsOpenReview ↗

Response to Reviewer oEzN

Thank you. We directly tested the recovery question and added analysis for the implemented Atlas.

  1. Quantitative Atlas connection (Proposition M.1). For one hard-chart cell, after aligning tangent spaces and metrics, derivative error epsilon_J < m* guarantees sigma_min(B_hat) >= m* - epsilon_J and condition number at most (M* + epsilon_J)/(m* - epsilon_J). The preceding decomposition identifies coefficient, local approximation, and coordinate errors. This conditional cell-interior result does not obtain epsilon_J from BCD descent, a statistical rate, seam continuity, or exact target-manifold membership of ambient predictions. The smooth-reference theorem remains separate: the ambient Neural MLP is not constrained to its target-manifold/hard-bound class, and its additional statistical and optimization premises remain unverified.

  2. Ground-truth recovery and strong controls (Table 4; Appendices N.2 and O). We evaluated actual Atlas factors and derivatives. BCD has mixed effects: on the noisy circle, factor NRMSE improves .04804 to .03352, while derivative NRMSE worsens .14105 to .18703; clean recovery also worsens. These within-Atlas comparisons avoid conflating its 582 training states with Neural's 768.

Two nonlinear T3-to-T2 factors add independent whole-trajectory splits, five seeds per noise condition, 64 validation-only learning-rate calibrations, and 80 matched 8000-update MLP fits. Pair+L2 keeps Full's weight penalty; unregularized Pair and kernel regression provide additional controls. Full shows no consistent advantage: clean paired intervals cross zero; noisy-shear factor/full-action errors are .01190/.03216 versus Pair+L2's .00984/.02681. Full and No-geometry nearly coincide. Validation-triggered paired continuations and a separately disclosed post-hoc matched-16000 check leave this conclusion unchanged. We retain all learning curves without claiming convergence.

  1. Direction and screening (Appendices N.2, N.4, N.6). When both circle directions are dimension-feasible, Atlas BCD reverse acceptance is 0/10, but clean forward acceptance is only 2/10. Changing only the tangent cutoff from .81 to .90 to .99 gives 0/2/10 clean-forward BCD acceptances; the official .90 is unchanged. On the original controls, the current vector-state replay gives 56/60 positive and 0/90 null joint passes, although all 150 pass the zero-margin rank gate alone. We have not validated near-degenerate true factors under this frozen Atlas joint screen.

These checks strengthen the recovery evidence and identify diagnostic limits; they do not establish an additional recovery advantage from the structural losses.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up on revised manuscript and responses

Thank you for considering our responses. We have uploaded the revised manuscript and complete supplementary evidence. The main updates are summarized in our individual replies, with the recovery comparisons in Table 4 and Appendices N–O and the conditional Atlas analysis in Proposition M.1. If any decision-critical concern remains unresolved, we would appreciate knowing which point needs further clarification before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.