acceptodds
Withdrawn from ICLR 2027

From Covariance to Response: What Is Identifiable, and What Should We Measure?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Equilibrium measurements describe how variables fluctuate together, whereas a perturbation experiment asks how a change propagates from one variable to another. When can covariance and a list of decay rates answer this response question, and what should be measured when they cannot? We study stable Ornstein–Uhlenbeck networks with a known acyclic graph and unknown diagonal diffusion. For complete two-layer networks, we establish a sharp identification boundary using the rate sum and show how the full anonymous spectrum can resolve the remaining rate assignments. For a sparse family, an odd-cycle criterion identifies the missing information and guides targeted temporal measurements. Matching bounds on stars quantify the value of target self-lags; multiroot results connect response risk to graph geometry and amplitude calibration. On the star, with unknown independent sensor variances, tracking source and target attains the four-coordinate risk order. A local risk calculation predicts when temporal inversion improves on a prior-center prediction. All sixteen known-amplitude comparisons in independent synthetic experiments support the predicted directions, alongside matched-readout and likelihood-based comparisons. These results connect response identification to the choice and budget of temporal measurements.

Primary area:
learning theory
Keywords:
Identifiability, Stochastic dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FNnvOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies when equilibrium covariance information is sufficient/insufficient to determine transient intervention responses in stochastic dynamical networks. For a class of Ornstein-Uhlenbeck (OU) systems, the authors derive identifiability conditions, graph-theoretic characterizations, and minimax sampling rates that determine when responses either can or cannot be recovered.

The paper is primarily a contribution to a mathematical understanding of the specific problem and its representation as a known two-layer graph with diagonal diffusion. The mathematics certainly appear to be carefully carried out and correct. But, the problem is highly specialized, and I found it difficult to see a compelling connection to mainstream machine learning. The work reads more like a systems-theoretic identifiability analysis than a broadly impactful learning framework. It is unclear to me if this work will be of interest to most of the ICLR community.

Critical strengths
  1. The mathematics appear to be carefully and throughly carried out for this specific problem.
Critical weaknesses
  1. While the mathematical results appear to be technically sound and provide precise identifiability characterizations, I found it difficult to determine the broader significance of the problem being studied. From an uncertainty quantification (UQ) perspective, the observation that equilibrium covariance information alone may be insufficient to determine intervention response is largely unsurprising. The main contribution lies in the exact characterization of when identifiability does and does not hold. However, the paper's assumptions and application setting appear highly specialized, making it difficult for me to assess the likely impact of these results on the broader machine learning community.
Questions for authors

I do not have any clarifying questions for the authors.

Reviewer_Bdx3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the problem of determining the AA matrix of a model dXt=AXtdt+Q1/2dWtdX_t=AX_tdt+Q^{1/2}dW_t vased in the knowledge of the incidence graph of AA (i.e., which elements of AA have to be zero), the covariance of XtX_t and possibly the trace or the decay rates of AA. I was not able to decipher the rest of the statements.

Critical strengths

I have a hard time seeing any strengths.

Critical weaknesses

The first and main weakness is the presentation. The authors do not define carefully the objects they study, and as a result it is not clear what their mathematical statements mean. Just a few examples:

-- 138: 'Two-layer graphs establish the structural criteria; stars isolate the statistical value of each measurement, and multiroot families quantify the additional cost of graph geometry.' What is a two-layer graph, a star, which multiroot families are meant, etc. ?

-- 145: 'Start with a two-layer graph containing every root-to-leaf edge' Still not known what a two-layer graph is, what a root-to-leaf edge is, etc. -- 156: 'For leaf pairs i < j ': leaf pairs of which graph ?

-- 180-181: 'On a three-leaf star, covariance ratios recover relative rates and T supplies their scale. But O(ϵ)O(\epsilon) root–leaf covariances give informative residuals of only O(ϵ2)O(\epsilon^2), against O(N^[-1//2}) noise'

What is TT ? (the trace of AA ?), ϵ,N\epsilon, N ?

-- 189: Theorem 2: "For pair-root models with m ≥ 3 leaves, positive rates and diffusion, and nonzero specified edge" What is a pair-root model ?

Perhaps by reading prior work on the use of SDEs to represent causal relationship some of these definitions could be decoded, but I do not think it is the job of the reader to do so.

Concerning related work: network identification has a long history in control theory and systems biology, with the underlying model often being a discrete-time version of the linear SDE mentioned in the paper. A few reference (very far from being an exhaustive list): J. Gonçalves and S. Warnick, “Necessary and sufficient conditions for dynamical structure reconstruction of LTI networks,” Automatic Control, IEEE Transactions on, vol. 53, no. 7, pp. 1670–1674, 2008. V. Chetty and S. Warnick, “Network semantics of dynamical systems,” in 2015 54th IEEE Conference on Decision and Control (CDC). IEEE, 2015, pp. 1557–1562. Y. Yuan, K. Glover, and J. Gonçalves, “On minimal realisations of dynamical structure functions,” Automatica, vol. 55, pp. 159–164, 2015. Y. Yuan, G.-B. Stan, S. Warnick, and J. Gonçalves, “Robust dynamical network structure reconstruction,” Automatica, vol. 47, no. 6, pp. 1230–1235, 2011, special Issue on Systems Biology. T. Nordling and E. Jacobsen, “On Sparsity as a Criterion in Recon- structing Biochemical Networks,” in 18th IFAC World Congress, 2011. H. H. M. Weerts, P. M. J. Van den Hof, and A. G. Dankers, “Identifiability of linear dynamic networks,” Automatica, vol. 89, pp. 247–258, 2018. P. M. J. Van den Hof, A. Dankers, P. S. C. Heuberger, and X. Bombois, “Identification of dynamic models in complex networks with predic- tion error methods—basic methods for consistent module estimates,” Automatica, vol. 49, no. 10, pp. 2994–3006, 2013. M. Dimovska and D. Materassi, “Granger-causality meets causal inference in graphical models: Learning networks via non-invasive observations,” in IEEE 56th Annual Conference on Decision and Control (CDC), 2017, pp. 5268–5273. D. Materassi and M. V. Salapaka, “On the problem of reconstructing an unknown topology via locality properties of the wiener filter,” IEEE Transactions on Automatic Control, vol. 57, no. 7, pp. 1765–1777, 2012. S. Shahrampour and V. M. Preciado, “Topology identification of directed dynamical networks via power spectral analysis,” IEEE Trans- actions on Automatic Control, vol. 60, no. 8, pp. 2260–2265, 2015. M. Nabi-Abdolyousefi and M. Mesbahi, “Network identification via node knockout,” IEEE Transactions on Automatic Control, vol. 57, no. 12, pp. 3214–3219, 2012. Kang, T., Moore, R., Li, Y., Sontag, E. D., & Bleris, L. (2015). "Discriminating direct and indirect connectivities in biological networks." Proceedings of the National Academy of Sciences (PNAS), 112(42), 12893-12898 Dasgupta, B., Vera-Licona, P., & Sontag, E. D. (2010). "Reverse engineering of molecular networks from a common combinatorial approach." In Algorithms in Computational Molecular Biology: Techniques, Approaches and Application'

All these references address a very similar problem.

Questions for authors

None, in my opinion the paper needs to be carefully rewritten to allow formulating questions.

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

The paper studies the identifiability and estimation of finite-horizon responses in stable Ornstein–Uhlenbeck networks from stationary covariance and limited temporal information. Assuming a known acyclic graph and unknown positive diagonal diffusion, the authors characterize when the covariance and the sum or unordered collection of decay rates determine a named source-to-target response.

For complete two-layer networks, they establish a sharp root–leaf identification boundary. For a sparse pair-root family, they derive an odd-cycle criterion and characterize additional temporal measurements sufficient for identification. They further establish minimax estimation bounds for star and multiroot models, quantifying the statistical advantages of oriented temporal pairs, target self-lags, and source–target tracking under measurement noise. Synthetic experiments validate the predicted sample-complexity differences and finite-budget measurement trade-offs.

Critical strengths
  1. The paper provides theoretical results connecting graph structure to the identifiability of dynamical responses. Theorem 1 establishes the generic identification boundary for complete two-layer networks when covariance and the rate sum are known. Theorem 2 extends the analysis to a sparse family using the bipartite components and odd cycles of an associated graph.

  2. The paper goes beyond population identifiability by quantifying the sample cost of different observation schemes. For example, Theorem 5 establishes a minimax risk decomposition. Theorems 6–7 further demonstrate the statistical value of target self-lags and show that tracking only source and target can attain the full-coordinate minimax risk order under independent noise.

Critical weaknesses
  1. The main identification results assume a known acyclic graph, positive diagonal diffusion, and specific two-layer structures. The sharp minimax results rely on even more specialized model classes, including stars or pair-root graphs, constrained rate ranges, and particular covariance parameterizations. In several risk comparisons, the estimator additionally receives substantial information about the underlying system, such as the exact rate sum, stationary variances, known graph structure, or root–leaf covariance amplitudes. Although the authors provide meaningful extensions to unknown common amplitude and sensor noise, these do not address uncertainty about the graph itself or general correlated diffusion. Three further restrictions limit the scope:
  • Every quantitative separation is a weak-coupling result. The star classes require ϵ≤0.01\epsilon\le0.01, the multiroot classes ϵ≤c0/m\epsilon\le c_0/m, and risk is normalized by ϵ\epsilon. The central ϵ−4\epsilon^{-4} versus ϵ−2\epsilon^{-2} gap comes from static residuals being O(ϵ2)O(\epsilon^2) in this regime. Nothing is shown for moderate coupling, where responses are large enough to matter in practice.
  • The identification results take the exact trace TT or the exact anonymous spectrum RR as given. Obtaining these exactly is itself a temporal-estimation problem, and the paper does not analyze how errors in them propagate, apart from the noisy-trace case in Theorem 22.
  • The star results use three leaves, a trace fixed exactly at a+∑idi=10a+\sum_i d_i=10, and known ϵ\epsilon.

Consequently, it remains unclear how the proposed measurement-selection principles extend to more general stochastic dynamical systems, where the graph is unknown and multiple sources of structural ambiguity coexist. This limits the applicability of the theoretical results.

  1. All experiments use simulated linear stochastic systems with known response functions. The positive findings are encouraging, but the experiments largely validate the predicted behavior within the same structured model families used in the analysis. In particular, the greatest finite-budget improvements rely on known graph structure and carefully specified prior information. The matched readout experiments improve the comparison, but the accounting does not include the potential experimental costs of preparing stationary samples or acquiring the supplied rate information.

The reported gains are also small in practical terms, and the budgets needed are large. Furthermore, the experiments do not evaluate the proposed measurement principles on real dynamical data or under misspecified graph and diffusion assumptions. Such experiments would help determine whether the theoretical advantages translate into practically useful measurement-design strategies.

Questions for authors
  1. To what extent can the odd-cycle identification criterion and targeted-completion results be extended beyond the specific pair-root construction? In particular, is there an analogous rank or graph-theoretic characterization for arbitrary known DAGs?

  2. How sensitive are the proposed measurement recommendations to incorrect graph assumptions or correlated process noise?

  3. The theoretical comparisons count scalar observations but do not account for obtaining exact rate summaries or for the additional preparation and synchronization costs of temporal measurements. Could the authors clarify under what practical experimental-cost regimes the proposed measurement strategies retain their advantage over static sampling?

  4. Do the measurement recommendations persist at moderate coupling, where static residuals are no longer O(ϵ2)O(\epsilon^2)?

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

Related papers

Discussion (0)

Sign in to comment.