acceptodds
Under review as a conference paper at ICLR 2027

Latent Causal Recovery from Long-Lag Dependencies

Abstract

Inferring causal structure from observational time series is challenging in the best of conditions, and even more so when relevant variables are unobserved. Marginalizing latent variables typically induces complex long-lag dependencies among observed variables, where a lag denotes the number of time steps separating cause and effect, making the inverse problem of recovering latent structure ill-posed. To recover latent structure under such confounding, we introduce Causal Hidden Recovery of N-lagged Observed Structures (CHRONOS), a constraint-optimization method built on answer set programming. Given a lagged observed graph, CHRONOS searches for causal graphs with minimal latent complexity that are consistent with the observations. The method explicitly allows for multiple optimal solutions, yielding equivalence classes of latent structures rather than a single reconstruction. Through extensive simulations on synthetic causal graphs with varying graph density and numbers of latent variables, we evaluate when latent structure recovery is completed within the time budget, how many latent variables are required, and the limits of identifiability in the resulting reconstructions. We additionally apply CHRONOS to real traffic-sensor data (METR-LA). Our results characterize the regimes in which latent causal recovery from apparent long-range dependencies is tractable, and the identifiability limits that latent confounding imposes regardless of method.

Primary area:
causal reasoning
Keywords:
latent causal recovery, causal structure learning, latent confounding, time series, constraint optimization

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

The paper proposes CHRONOS, an answer set programming method for reconstructing latent graphs from observed lagged relations. It formalizes forward marginalization and searches for compatible unit-lag graphs, minimizing latent nodes and then edges while enumerating all optimal solutions. A relaxed formulation accommodates errors in the input relations. Validation includes synthetic experiments and an application to METR-LA traffic data using estimated lagged graphs.

Critical strengths

The ASP formulation explicitly accommodates multiple optimal latent structures and distinguishes proving optimality from completing their enumeration. The synthetic experiments systematically examine how graph size, density, and latent fraction affect runtime and solution ambiguity. This provides useful evidence about when exhaustive reconstruction is computationally feasible and how much ambiguity remains even under the minimality criterion.

Critical weaknesses

The paper does not establish the scientific value of recovering a minimal latent graph. Figure 1 shows that minimality can exclude the true structure, yet neither the theory nor the experiments explain what useful causal knowledge the resulting graph provides beyond the input lagged graph. Consequently, the paper demonstrates a graph optimization procedure without establishing its value for latent causal recovery, leaving the significance of its central contribution unsupported.

Questions for authors

What scientifically meaningful conclusions about the underlying causal system can be drawn from the minimal latent graphs returned by CHRONOS?

7 comments

We thank the reviewer for asking what scientific conclusions the reconstruction supports. CHRONOS makes explicit constraints on latent complexity and candidate mechanisms implied by the observed lag relations. We will clarify which conclusions follow from those relations and which additionally depend on the minimality criterion.

CHRONOS addresses a concrete question: what is the least hidden dynamical structure needed to explain the observed lag relations under the stated model? This differs from identifying the particular generating mechanism, and Figure 1 illustrates why the distinction matters.

With exact input, the optimal latent count establishes a lower bound on the number of latent variables required in our unit-lag representation. Any model in this class with fewer latent variables is incompatible with the observed relations. This is a constraint on possible explanations of the system, even when its actual structure remains unidentified. The bound concerns latent variables in the unit-lag representation, rather than necessarily distinct physical entities.

The returned graphs additionally provide explicit realizations of the joint lag constraints. They show how hidden mediators, shared latent influences, and latent feedback can collectively account for the observations. Their value is to turn an implicit space of possible explanations into concrete models that can be inspected and assessed against domain knowledge or additional measurements. These are candidate mechanisms; minimality alone does not establish that any particular one generated the data.

Enumeration makes the dependence on parsimony transparent. Features shared by all returned graphs are necessary within the optimal class, whereas features that vary expose alternatives unresolved by that criterion. A singleton class establishes uniqueness under the objective, not unconditional identification of the true system.

We will add a subsection distinguishing these three outputs: a lower bound on latent complexity, explicit minimal candidate mechanisms, and structural features conditional on minimality. The METR-LA application demonstrates construction of such hypotheses; it does not independently validate their physical interpretation.

Thank you for the clarification. I understand the intended role of minimal latent structures, but my central concern about their scientific significance remains unresolved. As the authors acknowledge, reliable causal recovery and downstream practical benefits have not been independently validated. The response therefore does not establish how these outputs provide useful insights into the underlying causal system, and I maintain my original rating.

Thank you for your follow-up. We would like to address the remaining concern directly, since we believe the paper does provide evidence of the kind you are asking for.

Recovery against known ground truth: In our synthetic experiments, the true latent structure is known, so recovery can be checked directly. Across 3,624 cases, the number of latent variables recovered by CHRONOS equaled the true number in 86% of cases and never exceeded it, consistent with its role as a lower bound; 84% of cases had a single minimal explanation. We will state these results explicitly in the revision as validation of recovery, rather than only as properties of the search.

What the outputs tell us about the system: Recovery of latent structure from lagged relations is not identifiable in general (Figure 1). In such settings, as with the equivalence classes returned by PC or FCI, the scientifically meaningful conclusions are those shared by all admissible explanations. CHRONOS provides three: (i) a lower bound on how many hidden variables the system must have; (ii) the structural features common to all minimal explanations, which are established within the model class; and (iii) the features on which they differ, which identify exactly what additional measurements would be needed to distinguish them.

Scope of the claims: CHRONOS is designed for the setting the paper proves to be non-identifiable: it determines exactly what lagged observations imply about hidden structure, which is the strongest conclusion such data can support. We will make this framing, and the validation above, more prominent in the revision.

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

CHRONOS searches for parsimonious latent graphs that explain observed directed lag sets and bidirected lag-difference sets. It uses a unit-lag representation, allows multiple optimal latent explanations, and relaxes consistency through discrepancy penalties for uncertain input relations. Synthetic studies examine runtime, ambiguity, and lag perturbations. Traffic-matched estimator evaluation and METR-LA sensor subsets connect time-series estimation to a bounded latent-structure search.

Critical strengths
  1. The formulation takes non-identifiability seriously. Multiple minimal explanations are returned rather than one graph being presented as uniquely identified. The paper also distinguishes minimal latent complexity from recovery of the generating latent mechanism.

  2. The paper reports useful empirical boundaries. Completion and timeout behavior are described, and traffic-matched simulations use separate tuning and held-out graphs with frozen thresholds. The real-data section explicitly limits its output to candidate explanations of estimated predictive dependencies rather than independently validated causal relations.

Critical weaknesses
  1. The applicability conditions of the real-data pipeline remain insufficiently validated.

  2. Statistical robustness and computational generalization to more complex data are not yet established.

  3. The real-data study demonstrates bounded search feasibility, but not reliable causal recovery or practical advantage.

Questions for authors
  1. How can the applicability conditions be assessed on METR-LA?

  2. How does the full pipeline generalize to more complex data?

  3. What independently validates the real-data results?

2 comments

Thank you for the careful reading, and for recognizing that CHRONOS returns all minimal explanations rather than claiming a unique graph.

  1. Our checks support the estimator choice without establishing all causal assumptions. The traffic-matched simulations reproduce METR-LA’s median lag-1 autocorrelation (0.837) and cross-predictive gain (0.022). On METR-LA, the tested gradient-boosted predictor improved predictive R2R^2 by only about 0.01 over a linear model and added no cross-predictive gain, supporting linear estimation as a working approximation. Rank transformation reduces sensitivity to extreme marginal values, although the Gaussian simulations do not reproduce the observed heavy tails. Thresholds were tuned separately and frozen before held-out evaluation; no held-out score fell more than 0.01 below its tuning value. These checks assess predictive suitability and calibration. The causal interpretation additionally requires our graphical assumptions, including adequate sampling resolution, which these checks do not establish for five-minute measurements.

  2. We report both computational and statistical limits. All six real-data configurations reached proven optimality within minutes with at most two latents. Sparse sets completed in approximately one hour at N=16N=16 and thirteen hours at N=18N=18; denser or larger configurations and increased latent budgets exceeded 24 hours. On held-out N=12 simulations, directed estimation F1 declined from 0.789 to 0.571 as k increased from 0 to 3; bidirected F1 declined from 0.894 to 0.388 between k=1 and k=3. Thus, both estimation and search become harder with latent complexity. Applicability to more complex or nonlinear data requires further estimator validation as well as computational development.

  3. The study does not independently validate the inferred latent mechanisms. Road-network information provides a qualitative consistency check: it was used to select sensor sets but was not supplied to the estimator or inversion. Dense sets exhibit substantially more estimated bidirected relations than sparse sets (19/22/34 versus 3/7/7), consistent with shared unmeasured influences among nearby sensors. This pattern does not establish that the inferred latents represent upstream flow or exclude alternative explanations. The demonstrated practical capability is producing explicit, parsimonious candidate explanations from measured time series within the tested bounds; reliable causal recovery and downstream practical advantage remain to be independently evaluated.

We urge the reviewer to respond as soon as possible.

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

The paper focuses on recovering latent causal structure from a lagged graph over observed time series. Full system is written in a unit-lag form s.t. marginalizing latent variables (LVs) produces a compressed observed graph; with directed and bidirected edges carrying sets of lags, i.e., paths and forks through LVs. The paper sums up this "forward" marginalization in Thm1 -- which states that the result does not depend on the order in which latents are removed. They propose CHRONOS, an answer set programming (ASP) search that enumerates all unit-lag graphs with the fewest LVs (then fewest edges) that reproduce it exactly; given the observed lag-set graph. A constraint based version minimizes a weighted lag-mismatch cost plus a penalty on # LVs, to cover estimated I/p graphs. Simulations report runtimes, timeouts, equivalence-class sizes, and precision/recall under random lag flips; plus a traffic application which uses a rankVAR estimate of the observed graph as its I/p.

Critical strengths

Latent confounding in time series is common, enumerating all minimal latent explanations could be the right output for a non-identifiable inverse problem. ASP encodings make consistency and minimality explicit; forward marginalization is well defined.

Critical weaknesses
  • W1: My biggest concern is limited novelty compared to closest prior work, which was barely discussed. The proposed method builds on the RASL line (Plis et al., 2015; Hyttinen et al., 2016; Abavisani et al., 2023; 2025) -- This line uses compressed graphs with path directed and fork bidirected edges, inverts them with ASP, returns equiv. classes, and deals with noisy i/p graphs. CHRONOS instead swaps unobserved time steps for unobserved variables and a fixed undersampling rate for lag sets. Encoding, relaxed objective, and o/p follow the same approach. These works get only a little mention in related work, while undersampling is counted as a separate problem. A more comprehensive comparison (e.g., is undersampling a special case?) and experimental ones on overlapping scenarios are needed.

  • W2: I don't get to see the "nonparametric" claim. The contrast is with Gong et al. (2015) as parametric; the graph-level search is inherited from prior work. Cost of working at graph level is not discussed; for instance, discarding distributional info leaves large equiv. classes. Real-data pipeline is parametric anyway, no?

  • W3: Theory is somewhat straightforward. No identifiability or complexity result supports the abstract's claim about "identifiability limits ... regardless of method".

  • W4: Main simulations use the true G_O as I/p --> completed runs are correct by construction and only runtime is measured; robustness uses random lag flips, not estimated graphs. No baselines; small scale; on traffic data, no ground truth. Fewer recovered LVs than the truth (Fig 9) implies nonidentifiability, not recovery.

  • W5: Abstract and intro do not state the contribution clearly.

Questions for authors
  • What makes CHRONOS "nonparametric" beyond its predecessors, and what identifiability is lost by working only on the graph?
  • How accurate is the recovered latent structure end to end, when G_O is estimated from data rather than perturbed by random flips?
5 comments

We thank the reviewer for a detailed reading. We address each point below. Several concerns are answered by material already in the paper, which we will make more prominent.

W1 The reviewer’s main novelty concern rests on a conflation of two distinct inverse problems. RASL corrects for unobserved time steps; CHRONOS reconstructs unobserved variables. A system can exhibit both, requiring both problems to be solved. These are complementary inference tasks, and RASL does not perform the latent-variable reconstruction evaluated here. CHRONOS therefore cannot be obtained simply by replacing RASL’s sampling rate with lag sets. RASL searches over connections among a fixed set of measured variables, whereas CHRONOS must infer additional latent variables and their connections while jointly explaining the observed directed and bidirected lag sets. Its output is a set of minimal latent expansions, a reconstruction target absent from RASL’s formulation. The shared use of compressed graphs, ASP, and constraint relaxation does not make the methods interchangeable or establish the claimed lack of novelty. We will expand Related Work to make their relationship and complementary scope explicit; this clarification does not support treating RASL as a competing recovery baseline.

W2/Q1 The question conflates the scope of our assumptions with the source of our contribution. We do not claim that CHRONOS is “more nonparametric” than earlier graph-level methods. As explained in W1, RASL solves a different inverse problem; CHRONOS contributes a method for reconstructing minimal latent-variable expansions. “Nonparametric” specifies that this inversion requires no functional form or noise distribution, subject to our stated graphical assumptions. Its applicability therefore extends across data-generating mechanisms sharing the same lag-set graph. Additional distributional information and assumptions can sometimes distinguish structures that GOG_O alone cannot, but the resulting identification gains depend on their validity. CHRONOS explicitly reports the ambiguity remaining at the graph level by enumerating minimal compatible structures. In the real-data application, rank-VAR estimates GOG_O, while inversion remains independent of that estimator’s linear model. We will clarify that “nonparametric” applies to inversion, not the entire demonstrated pipeline.

W3 Figure 1 provides a constructive nonidentifiability example: two distinct latent graphs, with different latent counts, induce exactly the same observed lag-set graph GOG_O. Any method receiving only GOG_O therefore receives identical input for these two generating systems and cannot uniquely recover the true structure or latent count in both cases. This limitation follows from the information available, independently of the reconstruction algorithm. We will state this argument explicitly and replace the abstract’s unqualified “regardless of method” with “for reconstruction from the observed lag-set graph alone.” Additional distributional information or structural assumptions may resolve ambiguities unaddressable by this input. The other theoretical results establish the reconstruction framework: Theorem 1 proves order independence of marginalization; Lemma 1 establishes existence of (potentially infinite) latent realizations; and Corollary 1 justifies the unit-lag representation at the level of observed lag sets. We do not present these results as sufficient conditions for unique recovery. Nor do we claim a formal computational complexity characterization: our tractability results are empirical, reporting runtimes and timeout rates across graph sizes, densities, and latent fractions (N, h, e).

Continued in the next comment due to the portal's character limit.

W4/Q2 Exact GOG_O inputs deliberately isolate the computational difficulty of inversion from estimation error. CHRONOS seeks minimal explanations of observed lag relations while the generating GVG_V need not be minimal. A reconstruction with fewer latents that preserves those relations is therefore a successful, more parsimonious explanation, while also demonstrating nonidentifiability. Edgewise accuracy against the particular generating graph would conflate these two issues. We report the computational limits explicitly: all oracle instances complete at N=12N=12, while timeouts become common at N=36N=36. The evaluation also goes beyond oracle inputs and random flips. Section 4.3 and Appendix A.8 evaluate estimated graphs from traffic-matched simulated time series, using separate tuning and held-out sets with frozen thresholds. Table 1 reports observed-graph estimation accuracy, Appendix A.8 reports downstream inversion completion. These establish estimation performance and pipeline feasibility, although they do not measure the final explanations’ agreement with the ground-truth lag graph. Controlled flips serve the separate purpose of isolating sensitivity to input errors. The traffic application illustrates the resulting explanations without claiming ground-truth validation. RASL is not a competing reconstruction baseline for the reasons given in W1.

W5 Our contribution is CHRONOS, a constraint-based method that reconstructs explicit latent-variable explanations from an observed lag-set graph. Within specified latent-count and lag-horizon bounds, it enumerates compatible structures minimizing latent count and then edge count; a relaxed formulation accommodates noisy observations. The marginalization results establish the mathematical basis for this reconstruction, and the experiments characterize its tractability and solution ambiguity. These contributions are stated in the introduction’s final paragraph; we will foreground the reconstruction target and output more directly in both the abstract and introduction.

We urge the reviewer to respond as soon as possible.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.