acceptodds
Under review as a conference paper at ICLR 2027

One Transition Score, Two Graphs: Untangling Dynamics and Synchronous Innovation Dependence

Abstract

Finite-interval sampling folds dynamical propagation and synchronous innovation dependence into the same transition, making the two mechanisms difficult to distinguish. We study how to jointly recover the dynamics and synchronous graphs from irregularly sampled observations. For additive-noise diffusions with correlated innovations, differentiating the conditional transition score with respect to the terminal increment and the starting state yields a curvature block and a cross block, respectively. The curvature block characterizes synchronous dependence and demixes the cross block to reveal dynamical sensitivity, allowing the two blocks to jointly identify both graphs. We construct systems with identical cross blocks but different two-layer structures, establishing that the cross block alone cannot identify the two graphs. We also derive support-recovery conditions that account for finite observation gaps and score-estimation error. CAST (Curvature and Sensitivity from Transitions) learns a terminal score conditioned on the actual observation gap directly from irregular transition pairs, without estimating trajectory derivatives or imputing onto a regular grid. Experiments on synthetic systems and external benchmarks validate the two-block readout and gap-conditioned learning. CAST outperforms published baselines for dynamical structure learning on three of four benchmark datasets.

Primary area:
causal reasoning
Keywords:
causal discovery, stochastic differential equations, irregularly sampled time series, score matching, identifiability

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

Finite-interval sampling obscures the distinction between dynamical propagation and synchronous dependencies, making joint structure recovery from irregularly sampled data challenging. This work introduces CAST, a framework that separates these mechanisms.

Critical strengths
  • The research direction appears interesting.
  • At first glance, the mathematics appears correct; however, I have not verified all the proofs.
Critical weaknesses
  • I believe the paper is not reproducible in its current form. It should include a description of the model architecture, as neither the main text nor the appendix makes it clear that a neural network is used. I inspected the accompanying source code and found that the model uses a simple MLP, but this should be explicitly described in the paper.
  • The paper presents numerous proofs and a novel methodology. However, even after reading it several times, I remained unsure about the research questions and the motivation behind the proposed approach. I suggest stating the research questions explicitly in the introduction and explaining how they motivate the methodology.
Questions for authors
  • The paper contains quite a few theorems and propositions. How do they relate to the experimental evaluation?
  • The paper states: “The directed dynamics graph describes how interactions propagate across time, whereas the undirected synchronous dependence graph encodes conditional dependence among innovations at the same instant. We refer to them as the dynamics and synchronous graphs, respectively.” How does this description relate to Figure 1? The figure appears to distinguish the two graphs only by whether their edges are directed or undirected, without clearly illustrating the temporal aspect.
  • Line 119: What specific conditions are meant by “standard regularity conditions”? Please state these conditions explicitly or provide an appropriate reference.
  • Equation 5: Please consider adjusting the formatting so that the brace labelled “demixing” does not appear between the numerator and the fraction bar.
  • Line 319: The evaluation includes several baselines. Please explain why these particular methods were selected.
  • In my view, placing the related work section after the experiments is unusual. Presenting it earlier could help readers understand the context and positioning of the proposed method.
4 comments
AuthorsOpenReview ↗

Response to reviewer Mxmh

Thank you for checking the code. For the main synthetic experiment, the score network is an MLP with three hidden layers, SiLU activations, and width max(192, 16D), where D is the state dimension. It takes the noised increment, the starting state, and two 32-dimensional sinusoidal embeddings for the log gap and noise level. It predicts the injected noise. We use a 100-step linear DDPM schedule with beta ranging from 0.0001 to 0.02. We train with Adam at a learning rate of 0.0003 for 6,000 updates, with batches of 256 and without weight decay or EMA.

Training uses 40,000 transition pairs from 100 segments at eight log-spaced gaps between 0.02 and 0.5. The readout uses 512 held-out states per gap at gaps of 0.076 and 0.2, with eight perturbation draws per state at each of four noise levels (alpha-bar of 0.988, 0.908, 0.769, and 0.610). The Gaussian control uses the same data, optimizer, and training budget, with an MLP of the same depth and width for the mean and a covariance head that depends only on the gap. We will add these settings to the paper.

The introduction will ask whether both graphs can be recovered from one transition model, how curvature separates dynamical sensitivity from innovation precision, and when a learned score improves on direct Gaussian estimation. Table 2 compares the raw cross block, diagonal scaling, and full demixing, with dynamics AUPRC of 0.793, 0.854, and 0.967. The appendix checks the finite-gap expansion on the population readout. The irregular-sampling experiment compares direct transition learning with impute-then-differentiate controls. The strong-noise DblMass experiment uses larger readout gaps, as motivated by Proposition C.7. The main tables report empirical recovery accuracy. Corollary 4.1 gives the conditions for exact recovery. We will add a table showing how the theoretical results relate to these experiments.

We will add a time axis to Figure 1. The dynamics arrows will connect successive times. The synchronous edges will connect innovation components at the same time. The figure will also show the observed transition over a gap, which combines both effects.

For the limits in Eq. (2), one sufficient condition is a continuously differentiable, globally Lipschitz drift and a constant positive-definite innovation covariance. The main synthetic drift includes quadratic terms. Its acyclic ordering and linear self-decay allow moment bounds to be obtained recursively along the ordering. In the simulator, each coordinate is clipped to the interval from -8 to 8 after each integration step, as described in the appendix. We will make the distinction between the continuous-time assumptions and the numerical protocol explicit.

PCMCI+ and VARLiNGAM represent conditional-independence testing and non-Gaussian structural VAR, respectively. DYNOTEARS provides a score-based dynamic Bayesian network comparison. NGM and CaDyT cover continuous-time neural ODE modeling and derivative-based discovery. The matched Gaussian tests whether the learned score improves on direct Gaussian transition estimation. We will explain these choices, move related work before the method, and move the demixing brace outside the fraction in Eq. (5).

I want to thank you for the detailed reply. However, please take note that ICLR allows updating the submission with explicitly marked corrections. I remain unconvinced until explicit changes are made in the text.

The revised PDF is uploaded, with changes in blue. Each item of our earlier reply is now in the text.

Section 4.2 now describes the score network and its training. The network is a three-hidden-layer MLP with SiLU activations. Tables 8 and 9 in Appendix E.2 give the training and readout settings for CAST and the matched Gaussian, including the settings that vary by benchmark.

The introduction now starts with the difficulty of separating the two graphs when each observed transition mixes their effects. This motivates the two-block readout. The introduction then explains why we compare the learned score with direct estimates of the transition mean and a Gaussian transition model.

Table 6 in Appendix E explains how the theoretical results relate to the experiments. For example, Proposition 4.1 shows that at a fixed gap, the cross block alone cannot identify both graphs. The readout ablation measures what curvature adds in finite samples. The fraction of pairs with only a synchronous edge that are incorrectly reported as dynamics edges falls from 10% with the raw cross block to 2% with full demixing.

Figure 1 now has a time axis. Dynamics edges run across a short time step, and synchronous edges connect innovations at the same instant. Section 3 states the regularity conditions, and Appendix A proves the two limits in Eq. (2) and checks the conditions for our generator. We moved the demixing brace in Eq. (5) so it covers only the demixing product. Section 6 explains which family each baseline represents, and related work has moved to Section 2.

I want to thank you for addressing my remarks. I need some time to reevaluate the manuscript. However, I am grateful to the authors for the work done.

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

The authors proposed a score-based method for learning graphical structure from irregularly sampled time series. The method introduced a DAG describing dynamical interactions and an undirected graph describing conditional dependence among synchronous innovations. It trains a conditional denoising model on observed transitions and extracts both graphs from derivatives of the learned transition score. Its central operation uses score curvature to remove noise weighting from the cross derivative, yielding an estimate of dynamical sensitivity.

Critical strengths
  1. The paper addresses a meaningful problem in continuous-time causal structure learning.
  2. The curvature-demixing interpretation is mathematically sound under the stated linear Gaussian assumptions.
Critical weaknesses
  1. The fundamental methodological contribution is insufficiently established. The two target graphs are determined by quantities already studied in stochastic system identification. The central curvature-demixing identity follows from implicit differentiation and the chain rule for conditional location families. The paper does not establish why learning a conditional score provides an advantage over estimating the conditional drift and innovation covariance directly.

  2. The theoretical guarantees do not adequately support the implemented structural recovery procedure. At finite observation gaps, the dynamics readout contains indirect-path contributions, while the main synchronous readout uses uncorrected curvature from artificially perturbed increments, whose precision support need not match that of the underlying innovations. These concerns are compounded by synthetic experiments using additive drift functions and disjoint correlated innovation pairs, which substantially simplify the recovery task. The reported results therefore provide limited evidence for reliable recovery of the claimed structures in more general systems.

Questions for authors

How were the structural-readout gaps selected? Were they selected independently of graph-recovery performance?

2 comments
AuthorsOpenReview ↗

Response to Reviewer SkUy

The drift Jacobian and innovation precision are classical targets, and the demixing identity follows from implicit differentiation. Proposition 3.1 gives a fixed-gap counterexample in which two OU systems have the same cross block but different dynamics and synchronous graphs. Curvature distinguishes the two systems. In the readout ablation in Table 2, dynamics AUPRC rises from 0.793 with the raw cross block to 0.967 with full demixing. Theorem C.1 shows how two-sided differencing can introduce a synchronous innovation term into drift regression that persists as the step shrinks in the OU reference model.

The matched Gaussian directly estimates the conditional mean and a gap-dependent covariance from the same transitions, with the same training budget and a comparable network. Its score Jacobians give the two readout blocks in closed form. The comparison in our general response shows where the learned score helps empirically and where the Gaussian performs better. We will state this distinction in the introduction and cite the connection to stochastic system identification.

In the linear reference, the leading indirect-path term in Eq. (8) scales linearly with the observation gap. Proposition C.7 shows that the bias from starting-state measurement noise can grow inversely with the gap. The recovery conditions require edge margins large enough to separate true edges from the combined bias and estimation error. We will add a non-additive Lorenz-96-type system and hub and random sparse innovation precision graphs on fresh seeds.

We selected the main synthetic readout gaps using dynamics AUPRC on separate pilot seeds 0 to 9. The pilot compared gaps of 0.02, 0.075, 0.2, and 0.5. The two intermediate gaps performed best, with AUPRC across the four gaps ranging from 0.984 to 1.000. We rounded 0.075 to 0.076 on the simulation grid and fixed 0.076 and 0.2 before generating data for the final main-table seeds.

An earlier run on seeds 100 to 109 drew training pairs from one short trajectory and was retained as a stress test. A readout-budget debugging run exposed seeds 200 to 202, so we retired the entire block from 200 to 209. The main table uses seeds 210 to 219, chosen before generating their data. CAST's model, optimization, gap, and readout settings stayed fixed across these seed replacements. We will document the selection and report results at each candidate gap.

The revised PDF is uploaded, with changes in blue.

The revision adds direct estimators of the conditional mean of the standardized increment on the same trajectories as the main experiment (Appendix E.9, Table 13). The controls were fixed before scoring the final seeds. The MLP mean, matched Gaussian, and CAST use the same data, hidden width, and training budget.

At D = 5, dynamics top-k F1 is 0.920 for the MLP mean, 0.930 for the Gaussian, and 0.960 for CAST. At D = 10, the corresponding values are 0.660, 0.705, and 0.850. At D = 15, they are 0.490, 0.567, and 0.733. At D = 10 and D = 15, the paired bootstrap intervals for CAST's gains over both controls exclude zero. Ridge regression reaches 0.720 and 0.647 at those two dimensions, but recovers fewer than half of the quadratic edges. Table 3 gives the Gaussian comparison across benchmarks, including NetSim, where the Gaussian is more accurate.

Section 2 now relates the matched Gaussian to Gaussian transition quasi-likelihoods used in stochastic system identification. Proposition D.3 and Corollary D.4 address the uncorrected synchronous readout used in the experiments.

We selected the readout gaps using dynamics AUPRC on separate pilot seeds and fixed them before generating the final main-table data. Appendix E.3 gives the pilot results for each candidate gap and documents the seed blocks. In the irregular-sampling experiment, dynamics AUPRC ranges from 0.986 to 0.994 across fourteen individual readout gaps.

Reviewer_sf5jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAST to jointly recover directed dynamics and synchronous innovation-dependence graphs from irregularly sampled observations. The method learns a gap-conditioned transition score and uses curvature and cross-derivative blocks to identify both structures. Theoretical analysis establishes identification properties and conditional support-recovery guarantees. Experiments on synthetic systems and external benchmarks demonstrate competitive dynamical structure recovery. However, the theoretical assumptions and limited validation of synchronous dependencies weaken the generality of the conclusions.

Critical strengths
  1. The paper addresses the important problem of distinguishing dynamical interactions from synchronous innovation dependencies under irregular sampling. The two-block formulation provides an interesting theoretical perspective.

  2. CAST avoids explicit trajectory-derivative estimation and regular-grid imputation. The experiments include multiple baselines and meaningful ablations demonstrating the benefits of curvature-based demixing for dynamical graph recovery.

Critical weaknesses
  1. The theoretical guarantees rely on accurate score Jacobians and finite-gap assumptions that are not directly verified for the learned model. Moreover, the support-recovery guarantee concerns analytically corrected precision, whereas the main synchronous experiments use uncorrected curvature rankings. This creates a gap between theory and implementation.

  2. The validation of synchronous graph recovery is insufficiently comprehensive. The main synthetic experiments use disjoint correlated innovation pairs, where covariance and precision supports coincide. External benchmarks lack independently established synchronous-graph ground truth or use constructed precision structures. Consequently, the generalization of joint two-layer recovery remains uncertain.

  3. The advantage over the matched Gaussian baseline is inconsistent. The Gaussian model outperforms CAST on NetSim and achieves better synchronous recovery on Causal Dynamics-CI. Furthermore, top-k evaluation assumes knowledge of the true edge count.

Questions for authors
  1. How do the theoretical support-recovery guarantees apply to the uncorrected curvature readout used in the main experiments?

  2. Can the authors provide stronger validation on complex synchronous precision structures beyond disjoint correlated pairs?

  3. Under what conditions does CAST consistently outperform the simpler matched Gaussian baseline?

  4. How does performance change when the true number of graph edges is unknown and must be estimated?

2 comments
AuthorsOpenReview ↗

Response to reviewer sf5j

The support guarantee requires bounds on the learned score Jacobians and on population finite-gap bias. We monitor the conditioning of the pooled curvature, which affects both readouts. The learned Jacobian errors and the edge-margin conditions remain unverified. The next version will use the corrected precision readout analyzed by the guarantee.

In the three-node chains of Appendix E.4, the covariance contains an extra edge between the endpoints of each chain. The uncorrected readout gives AUPRC of 0.691 against precision support and 0.598 against covariance support. The paired difference is 0.094, with a seed-bootstrap 95% interval of 0.039 to 0.141. On CausalDynamics-CI with constructed chain precision, CAST gives a synchronous top-k F1 of 0.908 and the Gaussian gives 0.967. Both experiments use block structures with constructed ground truth. We will add hub and random sparse precision graphs on fresh seeds.

We have yet to establish when CAST will outperform the matched Gaussian. On CausalDynamics-CI, the matched-null Mardia diagnostic detected no departure from the Gaussian reference at the tested gaps. This is consistent with the Gaussian's strong synchronous results. Why CAST still leads on dynamics in the same benchmark remains open.

AUPRC evaluates the ranking without choosing a threshold or using the true edge count to select edges. Table 10 includes it for the synchronous graph. Our top-k F1 uses the true edge count. Corollary 4.1 establishes that a separating threshold exists under the margin conditions. We will add F1 with a threshold selected from the data to evaluate recovery when the edge count is unknown.

The revised PDF is uploaded, with changes in blue.

The revision keeps the uncorrected synchronous readout as primary. Table 5 compares it with the corrected readout on the same trained models. Proposition D.3 characterizes the population bias from artificial noise, and Corollary D.4 gives a support-recovery condition that depends on the curvature-estimation error and the edge margin.

The three-node precision chains extend the disjoint-pair design. The endpoints of each chain are connected in the covariance graph and have no edge in the target precision graph. Their mean absolute partial-correlation score is 0.047, close to 0.046 for pairs in different components (Appendix D). These chains still use a constructed block structure.

On the main synthetic system, dynamics top-k F1 at D = 15 is 0.490 for MLP mean regression, 0.567 for the matched Gaussian, and 0.733 for CAST. At D = 10 and D = 15, the paired intervals for CAST's gains over both controls exclude zero (Appendix E.9). The comparison varies across settings. On CausalDynamics-CI, CAST is more accurate on the dynamics graph, while the Gaussian is more accurate on the synchronous graph. The Gaussian also leads on NetSim (Table 3).

AUPRC evaluates the edge ranking without using the true edge count to select edges. Table 13 reports it for the dynamics layer of the main synthetic system. We did not measure F1 with a data-selected threshold in this revision.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

Thank you for the reviews. The main synchronous recovery tables use uncorrected curvature. The support-recovery guarantee applies to the analytically corrected precision readout and requires the finite-gap and score-Jacobian errors to be small relative to the edge margins. In the Gaussian reference, terminal perturbation rescales the transition covariance and adds a diagonal term before inversion. A covariance made up of disjoint pairs retains its precision support under this transformation. The main synthetic design therefore gives limited evidence about how the two readouts differ on more general graphs.

We compared the saved readouts from the three-node precision-chain experiment in Appendix E.4. Across the 40 configurations with nonempty precision graphs, the corrected and uncorrected readouts gave precision-support AUPRCs of 0.694 and 0.691, respectively. Both gave a top-k F1 of 0.658. We will use the corrected precision as the primary synchronous readout in the next version and keep the uncorrected result as a comparison.

The matched Gaussian is a direct transition model with a closed-form two-block readout. For dynamics, CAST gives a top-k F1 of 0.733 versus 0.567 at dimension 15 on the main synthetic system. Its AUPRC is 0.612 versus 0.451 on CausalDynamics-CI and 0.763 versus 0.662 on DblMass under strong measurement noise. In the multiplicative-noise study, CAST's dynamics AUPRC ranges from 0.900 to 0.928, while the Gaussian ranges from 0.692 to 0.749. On NetSim, the Gaussian leads with 0.853 versus 0.786. For synchronous recovery, the methods tie in the main synthetic table. CAST leads at dimensions 30 and 50 in the scaling study, with top-k F1 of 1.000 versus 0.920 and 0.996 versus 0.760. The Gaussian leads on both the paired and chain settings of CausalDynamics-CI. We will put these comparisons together in the main text.

Our reproducibility statement incorrectly said the appendix listed the model hyperparameters. We give the main synthetic settings in our reply to Reviewer Mxmh. We will add the architecture and training details to the manuscript, state the research questions in the introduction, move related work ahead of the method, and redraw Figure 1 with a time axis.

We have uploaded a revised PDF with changes marked in blue. Related work is now Section 2. Proposition 3.1 and Corollary 4.1 in the original submission are now Proposition 4.1 and Corollary 5.1.

In our earlier response, we said we would make the corrected precision the primary synchronous readout. After comparing both readouts from the same trained models, we kept the uncorrected readout as primary and report the corrected results alongside it in Table 5. The two readouts agree through D = 10. At D = 15, the corrected readout improves top-k F1 from 0.986 to 1.000. At D = 30 and D = 50, the uncorrected readout gives 1.000 and 0.996, compared with 0.793 and 0.432 after correction. Both give a top-k F1 of 0.658 on the precision chains.

Proposition D.3 explains how artificial noise affects the population readout. It preserves the precision support for disjoint pairs and introduces no edges between different connected components. Corollary D.4 gives a support-recovery condition that accounts for both this bias and curvature-estimation error. The analytic correction removes the perturbation bias, but its two matrix inversions can amplify curvature-estimation error.

The revision does not include the previously proposed experiments on hub and random sparse precision graphs, a non-additive Lorenz-96-type system, or F1 with a data-selected threshold.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.