acceptodds
Under review as a conference paper at ICLR 2027

From Least Action to Latent Forcing Tests for Root-Cause Attribution

Abstract

Root-cause attribution in latent dynamical systems asks which hidden process was perturbed and when. We study sparse additive forcing in latent stochastic differential equations using the classical innovation likelihood-ratio statistic from model-based fault detection and isolation. On an Euler–Maruyama grid with diagonal diffusion, the diffusion-normalised action evaluated at the profile-maximising latent forcing is exactly the per-atom generalised log-likelihood ratio. The oracle detection laws follow from standard Gaussian testing. Our focus is the identification, reconstruction and reporting conditions needed to use this statistic in learned latent coordinates. We derive a deterministic condition under which representation, drift and diffusion errors preserve the oracle attribution ranking. On synthetic latent-target benchmarks, the resulting one-pass estimator achieves comparable atom accuracy to an iterative solver while being substantially cheaper to evaluate. Physical-system simulations and real-data experiments also expose the importance of observability and the limits of transferring latent forcing attribution to observed-variable diagnosis. These results support latent forcing tests as an efficient and theoretically grounded approach to root-cause attribution when latent coordinates are meaningful and sufficiently observable.

Primary area:
learning on time series and dynamical systems
Keywords:
Root Cause Attribution, Time-series, Latent Dynamical Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QfN8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies root-cause attribution in partially observed latent dynamical systems, where the goal is to identify which latent process was perturbed and when. The proposed LARCA framework models incidents as sparse additive forcing in a latent SDE and shows that, under diagonal Gaussian transitions, the diffusion-normalized action evaluated at the profile-optimal forcing is exactly the per-atom GLRT statistic. This leads to a closed-form one-pass attribution procedure rather than iterative forcing optimization. The paper further analyzes when reconstruction, drift, and diffusion errors preserve the oracle ranking, and introduces selective reporting conditions for issuing named root-cause claims. Validation covers synthetic latent benchmarks, a physically grounded quadruple-tank simulation, real-world observed-variable benchmarks, and scaling experiments.

Critical strengths
  1. The action–likelihood connection gives the method a clean interpretation and removes the need for iterative optimization. On the latent benchmark, the one-pass estimator achieves essentially the same atom accuracy as the 550-step solver while reducing scoring time from about 984 ms to 0.43 ms. This is a concrete and practically meaningful simplification.
  2. The paper clearly separates coordinate identification, observability, ranking stability, and reporting reliability, and it explicitly evaluates important failure modes rather than only favorable settings. This substantially improves the credibility and interpretability of the empirical conclusions.
Critical weaknesses
  1. The core methodological novelty relative to classical innovation-based FDI/GLRT remains limited. In the known-state Gaussian setting, the singleton score is the classical Gaussian mean-shift / innovation GLRT. The paper’s main extension is to characterize the additional identification, approximation, and reporting conditions required in learned latent coordinates. While this synthesis is useful, the central inference mechanism itself is not fundamentally new.

  2. The paper highlights obtaining meaningful and observable latent coordinates as a key difficulty, but largely assumes rather than solves this problem. Named attribution requires external anchors, proxies, interventions, or structural constraints to fix the latent coordinate system. Moreover, the method performs substantially worse than AERCA on the observed-variable benchmarks. These results limit the current evidence that the framework addresses root-cause attribution beyond settings where the latent representation is already well anchored and observable.

Questions for authors
  1. Could the authors state more explicitly what they view as the main methodological contribution beyond classical innovation-based FDI/GLRT? In particular, given the same reconstructed state and nominal dynamics, which part of LARCA is fundamentally different in objective, inference procedure, or guarantee from a standard innovation GLRT?

  2. Once external anchors or calibrated proxies have fixed the physical meaning of the latent coordinates, what is the remaining distinction between the problem studied here and model-based fault isolation on reconstructed physical states? More specifically, which difficulties and contributions arise uniquely from the learned latent representation, rather than from the underlying innovation-based FDI framework?

Reviewer_5WaDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses root-cause localization in partially observed dynamics systems. The goal is to find which hidden state received an additive disturbance, and at which time step. A latent SDE is trained on fault-free data. At inference, the method encodes the incident and subtracts the learned drift. Each coordinate-time cell is scored by its diffusion-normalized squared residual, and the top cell is reported. The authors show that this score equals the classical per-atom GLR. In addition, the paper provides a deterministic ranking-stability condition and a split-calibrated bound on named-root errors. Experiments include a synthetic suite, a simulated plant with four tanks, and the SWaT and MSDS benchmarks. The results are comparable to those of the authors' iterative solver while requiring significantly less computational effort. However, for both real-world datasets, the results fall short of those achieved by AERCA.

Critical strengths
  • The paper is well-written and easy to follow.
  • The proposed score needs no optimizer, gradient or temporal mask. Once the latent path and drift are evaluated, building and scanning the action amp costs O(Kd)O(Kd). On the synthetic suite, it matches the iterative solver's atom accuracy with the reduced runtime. The scale study extends this to 4096 states and about 16.8 million candidate cells. Latency grows almost linearly with the number of cells and stays near 3 s at the largest size. Since every prefix map is a slice of the full map, rescanning after each step needs no extra multiplicity correction. Therefore, exhaustive state-time search becomes practical for large monitored systems.
  • Drift subtraction removes motion and nominal model already explains. A downstream state that merely responds to an upstream fault keeps a zero-mean residual. In the downstream-symptom regime, the method finds the entry atom in 92.2% of incidents. Inverse-variance weighting then places all coordinates on a common statistical scale. This is important when sensors differ in noise level. Under the oracle model, the null law gives a principled scan threshold. Therefore, the detector is interpretable and calibrated without tuning the metric itself.
Critical weaknesses
  • The scoring rule is the Gaussian mean-shift GLR. The paper states this itself. The intended contribution thus lies in identification, ranking stability and selective reporting. Identification is assumed through external anchors, not established. Theorem 1 is a triangle-inequality bound that depends on true states, so it cannot be checked at deployment. Theorem 2 is standard split calibration. Therefore, the paper applies an existing detector to a learned model without a new mechanism that makes this combination succeed.
  • Learned latent coordinates are meant for systems whose states are not directly measured. That is exactly where the method falls short. On real data, it underperforms AERCA by a wide margin and is near chance on SWaT. In the tank plant, a known-physics EKF does better, especially with weak proxies. On synthetic data, accuracy collapses once observability weakens. All favorable results come from simulations the authors designed, without classical FDI baselines. The claim of an efficient and theoretically grounded approach is thus unsupported in any practical setting.
Questions for authors
  • Is there any setting, synthetic or physical, where the learned pipeline outperforms a filter-based innovation GLR given the same information? A convincing positive answer would raise my score.
  • Theorem 1 requires bounds on true-state reconstruction error. Can any part of its condition be checked at deployment without true states?
Reviewer_cHbQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LARCA, a method which attributes an incident to a latent coordinate and time at which extra additive forcing entered a nominal stochastic dynamical system. For diagonal Euler-Maruyama transitions, profiling a singleton forcing amplitude gives precisely the classical Gaussian mean-shift GLRT. With a focus on identifying conditions for this to be applicable, a derived one-pass estimator attains competitive atom accuracy to a more expensive iterative solver, subtracting fitted drift, normalizing excess increments, etc in a single pass. Experiments are carried out over synthetic latent-target regimes, hydraulic simulations, real observed-variable diagnosis, and a large synthetic scale study.

Critical strengths
  1. I found the core statistical argument was carefully scoped, technically consistent, and is presented with transparent analyses. There is explicit recognition of the GLRT as established, the constrained singleton statistic for correlated diffusion is provided, detection from exact support and top-1 recovery is separated, and a pseudo-GLRT on estimated states from an oracle chi-squared law is distinguished. Theorems are clearly presented with explicit assumptions, and Theorem 1, for instance, nicely tracks increment reconstruction error, drift error, and variance error separately. These distinctions are valuable guidance, and I thought the presentation was strong in this regard.
Critical weaknesses
  1. The main analysis doesn't provide a method for overcoming the hard latent-inference problem, leaving the methodological advance over classical innovation testing fairly limited. The additional contribution beyond the established GLRT is principally a sufficient perturbation analysis and reporting framework. Positionally, it seems the paper needs a stronger case that this advances latent root attribution, rather than describing conditions under which a standard statistic can be applied. This is primarily a concern about the overall novelty impact (not the validity of the presented theorems).
  2. The positive evidence doesn't seem to establish a clear practical advantage in the settings where latent root attribution is most needed. In the hydraulic simulation, upper states are recovered from calibrated local flow proxies. This supplies physically meaningful coordinates but is easier than identifying hidden forcing from indirect, distributed measurements. For the prespecified 20 cm3 endpoint, ordinary Euclidean forcing residuals already match or slightly exceed the normalized method (Table 32), so this experiment seems to validate reconstruction/semantics more than demonstrating a new attribution benefit.
Questions for authors
  1. Can any nontrivial ranking-stability condition be evaluated without true incident latents and/or is there evidence connecting observable reconstruction diagnostics to oracle agreement? If not, what concrete modeling or reporting decision does Theorem 1 enable beyond its synthetic verification? This would help contextualize the practical significance of the main analytical result.
  2. Is there an independent physical or real-data setting in which the deployed one-pass method achieves useful accuracy and nontrivial accepted coverage with the confidence family and threshold fixed before validation? Existing SWaT/MSDS results don't seem to support this. A positive existing validation result, or possibly a narrower statement of the paper's contribution as a conditional framework, could bolster the core impact of the work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.