acceptodds
Under review as a conference paper at ICLR 2027

Least-Action Causal Discovery via SDEs for Stochastic Dynamical Systems

Abstract

We introduce the Principle of Least-Action Causal Identification of Dynamics (PLACID), a controlled-SDE testbed for learning latent dynamics and examining sparse causal drift structure from irregularly sampled time series. Under explicit regularity assumptions, Girsanov’s theorem relates path-space relative entropy to a quadratic steering cost, motivating a decomposition into reusable structural drift and data-dependent control. Theoretically, we give a conditional drift-consistency analysis and sufficient coverage, smoothness, and signal-margin conditions for oracle recovery, a counterexample shows that vanishing action and uniformly accurate drift can still retain a false edge. Empirically, within this controlled-SDE setting, Lorenz96 and Glycolysis experiments show drift-ranking signal, including for 50-variable Lorenz models, although fixed-threshold graph recovery remains poor in that high-dimensional setting. Checkpoint diagnostics show that similar prediction error can accompany different drift rankings. A matched Lorenz D10 comparison finds no resolved ranking benefit from control or stochastic diffusion over deterministic drift, a released-code NGM-GL comparator gives similar ranking with substantial checkpoint sensitivity. Ablations show that bounded control prevents constant-drift collapse in a zero-action-penalty stress test, while proximal sparsity improves Lorenz graph parsimony. Together, these results position PLACID as a testbed for examining how control regularization and sparse drift parameterization shape structure learning in stochastic dynamical systems.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Causality, Structure Discovery, Time-series, Stochastic Differential Equations

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

This paper introduces PLACID, a controlled stochastic differential equation framework for causal structure discovery from irregularly sampled dynamical systems. The method separates structural drift from data-dependent control, using control-energy regularization and group sparsity to encourage interpretable drift dependencies. The authors establish theoretical connections between path-space KL divergence, least action, and structural recovery, including conditions under which accurate drift estimation implies graph recovery.

The framework is evaluated on Lorenz–96 and Glycolysis systems, with comparisons against SCOTCH, Neural Graphical Models, and deterministic alternatives. The results show moderate causal graph recovery, although the experiments do not establish a consistent advantage from stochastic diffusion or control. Additional ablations examine controller collapse, sparsity, and the relationship between prediction accuracy and structural recovery.

Critical strengths
  • Interesting theoretical treatment of structural recovery. I find the distinction between drift accuracy and causal structure recovery important. In particular, the counterexample showing that vanishing action can coexist with incorrect graph structure is useful. The analysis also clarifies the additional assumptions required for structural recovery, including occupation coverage, smoothness, and sufficient interaction strength.

  • Careful separation of prediction and structure learning. The experiments explicitly distinguish forecast accuracy, causal graph ranking, and sparsity. I appreciate the inclusion of deterministic and no-control comparisons, alongside the controller-capacity ablations. These results expose important limitations of controlled SDE formulations rather than assuming that adding stochasticity necessarily improves causal discovery.

Critical weaknesses
  • The main methodological contribution is not supported by clear empirical improvements. The central motivation is that stochastic dynamics and least-action control can improve structural discovery. However, the matched Lorenz–96 experiments show essentially identical graph recovery between the full controlled SDE (AUROC 0.7456), the no-control SDE (0.7541), and the deterministic comparator (0.7450). The no-control model also achieves substantially lower prediction error on Glycolysis. These results raise the question of whether the additional control and stochastic components are necessary, particularly given their computational complexity.

  • Causal graph recovery remains relatively weak, particularly in higher dimensions. While the reported AUROC values suggest some structural ranking ability, these improvements do not translate into reliable graph recovery. For Lorenz–96 with 50 variables, directed SHD reaches approximately 343–366, whereas an empty graph would have an SHD of 200 under the same inclusive-edge convention. Furthermore, the headline AUROC includes self-edges, which makes the results appear stronger than the off-diagonal evaluation. For example, Lorenz D10 AUROC decreases from 0.7456 to 0.6608 when self-edges are excluded. I would like to see stronger graph recovery before considering the approach effective for causal discovery.

  • The theoretical guarantees are disconnected from the implemented optimization. The main consistency result assumes a fixed posterior joint law, known diffusion, and sufficient occupation coverage. In contrast, the implemented method jointly learns the encoder, diffusion, and structural drift using endpoint consistency and control-energy penalties. The theoretical recovery results also concern functional drift derivatives, whereas the experiments evaluate first-layer weight norms. Consequently, the analysis does not establish consistency or identifiability for the actual estimator. This substantially limits the theoretical justification for the proposed methodology.

  • The experimental comparisons and irregular-sampling evaluation are insufficient. Although SCOTCH and NGM are included, several reported comparisons use different training protocols, while the local SCOTCH evaluation is incomplete, particularly at higher dimensions. This makes it difficult to assess whether PLACID offers meaningful improvements over existing stochastic causal discovery methods. Furthermore, irregular sampling is a central motivation, but the experiments provide limited systematic analysis of how recovery changes with observation gaps, variable-specific missingness, or sampling frequency. More controlled comparisons would strengthen the empirical conclusions.

  • Limited evidence of generalization and causal identifiability. The experiments primarily use simulated systems with known coordinates and fixed underlying graph structures. The learned latent coordinates are aligned using validation simulator states, which would generally be unavailable for an unknown physical system. Additionally, the exploratory DREAM3 experiments achieve AUROC close to random guessing, approximately 0.49–0.50. It therefore remains unclear whether the framework can recover meaningful causal relationships in more realistic, partially observed systems with unknown latent representations.

Questions for authors

See above.

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

This paper introduces PLACID, a machine learning framework designed to uncover cause-and-effect relationships between variables in complex systems governed by random fluctuations (stochastic dynamical systems), even when data is noisy or recorded at irregular intervals.

The core approach revolves around five main ideas:

  1. Modeling the underlying mechanics: A neural network learns the governing equations of the system (the drift term of a stochastic differential equation), predicting how each component changes based on the state of the others.

  2. Adding a corrective "controller": A secondary neural network applies temporary nudges to keep predictions aligned with actual real-world measurements. These corrections are strictly penalized using a principle of least effort: ideally, the primary model should account for the data on its own without needing constant steering.

  3. Pruning non-essential connections: By applying sparsity penalties (encouraging the model to rely on as few variables as possible), the algorithm eliminates weak connections, leaving behind a network diagram that represents direct causal influences.

  4. Analyzing theoretical limits: The authors mathematically prove that even a model that reproduces trajectories almost perfectly can still map out an incorrect dependency network. They lay out specific conditions—such as comprehensive data coverage and smooth underlying behavior—required for the true network to be reliably recovered.

  5. Testing on simulated benchmarks: The framework is tested on standard synthetic problems, including atmospheric simulations (the Lorenz-96 model) and cellular chemistry (a glycolysis model). While it identifies some relationships, the resulting network maps remain flawed, especially in larger systems. Crucially, adding the corrective controller and noise terms does not show a clear performance advantage over simpler baseline models.

Critical strengths

The primary strength of the work is its rigorous mathematical look at the disconnect between learning system trajectories and accurately uncovering their underlying network structure. Proving that an algorithm can simulate data near-flawlessly while still drawing the wrong causal map is an important insight, and the authors do well to formalize the conditions needed to prevent this failure. The experimental evaluation is also refreshingly honest and systematic, offering helpful comparisons against simpler baselines and an unusually candid discussion of negative results.

Critical weaknesses

The main drawback is that the fundamental methodology feels largely incremental. Core elements—such as sparse identification of dynamical systems, penalizing control inputs, and probability-based path adjustments—are already well explored in existing methods like Neural Stochastic Differential Equations and SCOTCH.

• No clear real-world performance edge: In benchmarks, the full model often fails to outperform stripped-down baselines that omit the controller and random noise, calling into question whether this extra complexity is justified.

• Limited scope of evaluation: The empirical tests rely on a small set of synthetic simulations, offering little evidence of how the method performs on messy real-world data or against established system-identification tools.

Questions for authors
  1. Distinction from established methods: Could you clarify the primary methodological advance over existing frameworks like SCOTCH, standard neural differential equation models, and sparse identification algorithms like SINDy? Splitting system dynamics from an energy-penalized corrective controller is already a recognized technique—what unique capability or theoretical insight sets PLACID apart?

  2. Justifying the controller module: The benchmarks show that omitting the corrective controller does not hurt performance, and the simpler setup occasionally performs better. Under what specific conditions do you expect this controller to provide a clear advantage? Why is it necessary for discovering causal graphs rather than just smoothing out predicted curves?

  3. Reconciling theory with implementation: Your theoretical recovery proofs assume an idealized mathematical loss function evaluated across entire trajectory paths, whereas the implemented program optimizes endpoint prediction errors and penalty weights. What justifies the assumption that your practical implementation will yield the theoretical guarantees, and what additional conditions are required to bridge that gap?

  4. Distinguishing correlation from causation: The algorithm extracts dependencies directly from the mathematical formulas it learns. Under what assumptions can these be interpreted as genuine cause-and-effect relationships rather than mere statistical associations? In particular, how would unmeasured variables, coordinate transformations, or noise that changes with system state impact these conclusions?

  5. Ranking versus binary graph accuracy: The system ranks potential links reasonably well but performs poorly when making concrete binary choices (link vs. no link), especially in the 50-variable Lorenz simulation. How do you explain this discrepancy? Is it primarily an issue of setting the wrong cutoff thresholds, an inability to fit the correct mathematical forms, or a broader limitation in how the dynamics are learned?

Reviewer_xB21OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PLACID, a controlled latent-SDE framework for causal discovery from irregularly sampled time series. It decomposes the learned dynamics into a structural drift and a data-dependent controller, regularizing control energy through a Girsanov-inspired least-action objective and encouraging sparse drift dependencies using adaptive group sparsity.

The theoretical contributions include an oracle drift-consistency result, a counterexample demonstrating that vanishing action does not imply correct graph recovery, and sufficient coverage, smoothness, and signal-strength conditions for recovering the functional drift graph. Experiments on Lorenz96 and Glycolysis investigate graph recovery, controller capacity, proximal sparsity, and comparisons with deterministic and stochastic alternatives.

Critical strengths
  1. The paper provides a useful distinction between recovering a drift function and identifying its dependency graph. Proposition 4.5 constructs a sequence of drifts with uniformly vanishing approximation error and action but a persistent false edge. Theorem A.1 complements this observation by deriving a derivative-error bound of order under additional coverage and smoothness conditions. These results clarify why trajectory fitting alone cannot guarantee graph recovery.

  2. The controller-capacity stress test shows that an unregularized controller can explain trajectories while the structural drift collapses. The matched comparisons also separate the roles of control, stochastic diffusion, and proximal sparsity.

Critical weaknesses
  1. Theorem 4.3 assumes a fixed exact posterior joint law whose path marginal equals the true law, together with known common diffusion and fixed latent coordinates. Under these assumptions, minimizing the expected conditional KL is equivalent, up to a constant, to minimizing KL(P⋆∥Pθ)KL(P^\star\|P_\theta). Thus, the oracle consistency result largely follows from the nonnegativity of KL divergence. However, PLACID jointly learns the encoder, drift, diffusion, and controller using an endpoint-consistency objective rather than the oracle path KL. Furthermore, the method restarts inference at each observation interval, so these fitted intervals do not necessarily define a consistent global posterior path law. The conditions of Theorem 4.3 are therefore not established for the actual estimator. Exact conditioning on noiseless continuous-valued observations also generally produces singular conditional path laws, which further limits direct applicability to the benchmarks. More importantly, the graph is estimated from first-layer group norms Sij=∥Wj,:,i(1)∥2S_{ij}=\|W_{j,:,i}^{(1)}\|_2, whereas Theorem A.1 concerns functional drift derivatives. Proposition 4.6 only establishes that zero input groups exclude functional edges, not that nonzero groups correspond to true edges. The learned coordinate representation introduces another ambiguity, and the experimental permutation alignment uses simulator states unavailable in general applications. Consequently, the theoretical results do not establish consistency or identifiability for the graph actually returned by PLACID.

  2. The experiments do not establish a clear advantage from introducing a controller or stochastic diffusion. In the matched Lorenz D10 comparison (Table 2), the controlled SDE achieves AUROC 0.74560.7456, compared with 0.75410.7541 for the no-control SDE and 0.74500.7450 for the deterministic comparator. The differences are not statistically resolved. On Glycolysis, removing control also substantially improves forecast MSE.

Although Lorenz D50 produces reasonable AUROC, binary graph recovery remains poor. Table 7 reports SHD values of 366.2366.2 and 343.4343.4, whereas the empty graph has SHD 200200 for the 50-variable Lorenz system, including self-edges. Thus, the proposed thresholding procedure performs worse than the trivial empty-graph baseline in these settings.

The comparison with SCOTCH is also not conducted under a fully matched protocol, limiting conclusions about relative performance. Additionally, the DREAM3 experiments obtain approximately random off-diagonal AUROC, providing little evidence of transfer beyond the main simulators.

Overall, the experiments provide useful diagnostics but do not convincingly establish the practical value of the proposed controlled-SDE architecture over simpler alternatives.

Questions for authors
  1. Can the authors provide either a theoretical connection between the implemented endpoint-consistency objective and the oracle action in Theorem 4.3, or empirical evidence that reducing the implemented action leads to improved functional drift accuracy and graph recovery?

  2. How sensitive are the reported edge rankings to latent-coordinate transformations? For example, under an invertible diagonal rescaling z′=Dzz'=Dz, the same structural drift can be reparameterized with first-layer weights Wj′=WjD−1W'_j=W_jD^{-1}, changing raw input-group norms without changing functional dependencies. Can the authors evaluate an identity encoder/decoder or fixed observed-coordinate version to isolate this ambiguity?

  3. Given that the controlled SDE does not outperform the no-control or deterministic alternatives in the matched Lorenz experiment, what evidence supports using the more complex PLACID formulation? Can the authors provide stronger matched baseline comparisons and demonstrate improved binary graph recovery on Lorenz D50, where the current SHD exceeds that of the empty graph?

Comments on the paper

1 comment
Area_Chair_nVVNOpenReview ↗

Please join the author-reviewer discussion

Dear all reviewers,

Our author and reviewer discussion will end soon in 18 November (AOE). If you have not yet joined the discussion, now it is the time to start it. We highly appreciate your time on discussing and exchanging thoughts to keep our ICLR standards. Many thanks!

Bests,
Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.