acceptodds
Under review as a conference paper at ICLR 2027

Data-Driven Edge-Resolved Perturbation Propagation in Nonlinear Dynamical Networks

Abstract

Analysis of perturbation propagation in nonlinear networks has traditionally relied on explicit governing equations. When the governing equations are unknown, resolving source-specific perturbation responses into contributions from individual interactions and organizing them into source-connected propagation structures from observed trajectories remains challenging. To address this challenge, we introduce a data-driven method for finite-time perturbation analysis using observed trajectories and known network structure. A graph-constrained neural vector field is learned from the trajectories, and its time-varying Jacobian is evaluated along an observed unperturbed reference trajectory. The resulting Jacobian drives non-autonomous variational dynamics to infer source-specific first-order perturbation responses. We then apply variation of constants to decompose the response of each affected non-source node into time-integrated incoming-edge contributions and evaluate these contributions at the node’s inferred arrival time. These arrival-conditioned contributions are combined with arrival ordering to construct predecessor graphs and extract source-connected propagation paths. Across three nonlinear network systems, our method recovers state-dependent Jacobian structure, affected-node sets, propagation timing, and arrival-conditioned edge contributions. The resulting predecessor graphs closely agree with those constructed from the ground-truth Jacobian and retain most of the corresponding contribution mass.

Primary area:
learning on time series and dynamical systems
Keywords:
nonlinear dynamical networks, dynamical systems, perturbation propagation, time-varying Jacobian, variational equations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_vagROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper considers how a perturbation applied to one node propagates through a nonlinear network when the topology is known but the governing equations are unavailable. The authors fit a graph-constrained neural vector field to observed trajectories, evaluate its Jacobian along an unperturbed reference trajectory, and integrate the corresponding variational equations. They then use variation of constants to decompose each non-source node's response into signed contributions from its incoming edges. Evaluating these contributions at threshold-defined arrival times and imposing an arrival-order constraint produces predecessor graphs and source-connected paths.

The experiments cover three primary systems and five additional dynamics–topology combinations. Evaluation includes Jacobian recovery, perturbation responses, edge contributions, predecessor graphs, frozen-Jacobian comparisons, and nonlinear edge-removal tests. The neuronal example uses simulated dynamics on an empirical connectome.

Critical strengths
  1. The edge contributions have a clear mathematical definition. Equations (9)–(12), p. 5, separate the target's diagonal dynamics from its incoming inputs. Before predecessor filtering, the signed contributions sum exactly to the non-source node's variational response. The formulation preserves cancellation and accounts for the subsequent amplification or attenuation of each input. The authors also state clearly that an incoming-edge contribution includes the sender's upstream history. This makes the quantity well defined without implying a unique attribution to complete paths.

  2. The evaluation checks Jacobians rather than relying on rollout accuracy. This is an important choice because the downstream analysis depends directly on local derivatives. Table 5 reports structural-edge relative errors of 3.38%, 1.11%, and 3.97% on the three primary systems. Table 4 reports median edge-contribution relative errors of 8.00%, 1.16%, and 0.123%. The separate comparisons against analytical variational and nonlinear responses help distinguish learning error from linearization error. Appendix F adds five further pairings, including cases with less accurate graph recovery. The trajectory-level data split, clustered uncertainty estimates for Jacobian metrics, and source-level aggregation are also appropriate.

Critical weaknesses
  1. The arrival-order constraint in Eq. (14) can exclude genuine transmission routes. In a chain s → j → i, a subthreshold response at j can drive an above-threshold response at i, leaving the sole transmitting route absent from the predecessor graph.

  2. Contribution-mass recall in Eq. (29) uses an already arrival-filtered reference, while path agreement is conditional on jointly resolved targets. These metrics leave the filter's loss of incoming contribution mass and affected-target coverage unmeasured.

  3. The comparisons do not isolate the value of contribution-based edge analysis. Frozen-Jacobian baselines test time dependence, while the removal controls match edge count and approximate indegree without establishing comparable source–target reachability or route placement.

  4. Propagation and edge-attribution results use one designated test reference trajectory per system (Appendix A). The source-level bootstrap intervals do not establish stability across reference trajectories or independently trained models.

  5. The sensitivity of Jacobian and propagation recovery to observation noise and sampling density remains untested. Training uses finite-difference derivative targets (Eq. (5)), but the reliability of this step under less favorable observation conditions is unclear.

Questions for authors
  1. How should the graph be interpreted in the chain example?

  2. How much does the arrival filter discard?

  3. Does the intervention advantage survive route-aware controls?

  4. What observation and numerical settings are required for the reported results?

9 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Thank you for raising this important issue. In practical observations, finite temporal sampling and measurement noise make it difficult to distinguish very weak initial responses from numerical fluctuations or observational noise. We therefore introduce a detection threshold to define detectable arrival times and use the arrival-order condition to construct the predecessor graph. We acknowledge that this filtering involves a trade-off between suppressing noise-induced responses and retaining weak but potentially relevant interactions. To evaluate this trade-off, we added threshold and noise sensitivity analyses. Following your suggestion, we also revised the calculation of the incoming contribution mass retention ratio, using the total absolute contribution mass of all structural incoming edges as the reference. Under the lower calibrated noise level, increasing α\alpha from 0.005 to 0.01 reduces the false-positive rate to below 0.01% in all three systems under both white and temporally correlated noise. Further increasing α\alpha to 0.02 reduces the number of responses detected on clean trajectories. We also evaluated the proportion of total absolute incoming-edge contribution mass retained by the predecessor graphs. At the adopted threshold of α=0.01\alpha=0.01, the median retention rates are 94.56%, 87.28%, and 99.95% for Mutualistic–SW, Neuronal–Brain, and SIS–Gamma, respectively. These results support the adopted threshold as a reasonable balance between noise suppression and contribution retention. The threshold sensitivity and noise analyses are reported in Appendix G (pp. 28–30), Tables 18–19 (p. 29; p. 30). The contribution mass retention metric is defined in Equation (30) (p. 16), and its results are reported in Table 7 (p. 19).

This threshold-based filtering does not prevent edge contributions from being calculated, as contributions are computed for all structural incoming edges before the predecessor graph is constructed. For the chain example s→j→is\rightarrow j\rightarrow i, the edge j→ij\rightarrow i may have a nonzero accumulated contribution even when node jj remains below the detection threshold at the time node ii reaches its threshold. Before predecessor filtering, we calculate the accumulated contribution at the target's arrival time for every structural incoming edge of node ii:

=\int_{t_0}^{\tau_i^{\mathrm{arr}}} G_i(\tau_i^{\mathrm{arr}},u)J_{ij}(u)\delta_j(u)\,\mathrm{d}u.$$ Therefore, the contribution from node $j$ is still included in the full incoming-edge decomposition, provided that its perturbation response generates a nonzero interaction input during the integration interval. For a non-source node with zero initial perturbation, the signed contributions from all structural incoming edges still sum to its first-order variational response at its arrival time. The potential loss occurs during the subsequent predecessor-graph filtering. Specifically, the graph requires the sender to satisfy the arrival-order condition $\tau_j^{\mathrm{arr}}\leq\tau_i^{\mathrm{arr}}+\varepsilon_t$. Consequently, an edge with a nonzero accumulated contribution may still be excluded from the final graph if its sender has not reached the detection threshold. Based on this distinction, we have further clarified the intended interpretation of the predecessor graph. It is not intended to recover all actual perturbation transmission relations in the underlying dynamical system. Instead, it organizes structural incoming interactions with nonzero arrival-conditioned contributions according to detectable arrival order and traces source-connected paths through the retained interactions. We have clarified this interpretation and its scope in the revised Conclusion and Discussion section to avoid interpreting the predecessor graph as a complete reconstruction of all physical perturbation transmission relations (Section 4, p. 10).
AuthorsOpenReview ↗

Response to Weakness 2 and Question 2: Quantifying Information Loss from Arrival Filtering (Part1)

Thank you for pointing out the limitation of our original evaluation metric in quantifying the information loss introduced by arrival-order filtering. We agree that the original contribution-mass recall used the analytically filtered predecessor edges as its reference set, and therefore could not fully characterize the contribution mass discarded during filtering. To address this concern, we redefine the incoming contribution-mass retention ratio by using all structural incoming edges as the reference, rather than the already filtered analytical predecessor set.

The revised retention ratio is defined as

\qquad M_i^{\mathrm{pred}}=\sum_{j\in\mathcal R_i} \lvert C_{j\rightarrow i}^{\mathrm{arr}}\rvert, \qquad M_i^{\mathrm{all}}=\sum_{\substack{j\ne i\\ A_{ij}\ne0}} \lvert C_{j\rightarrow i}^{\mathrm{arr}}\rvert.$$ Here, ($M_i^{\mathrm{pred}}$) denotes the absolute contribution mass retained by the predecessor graph, while ($M_i^{\mathrm{all}}$) denotes the total absolute contribution mass from all structural incoming edges. This revised metric directly quantifies the fraction of incoming contribution mass preserved in the final predecessor graph. Under the setting of $\alpha=0.01$, the median contribution-mass retention ratios are 94.56%, 87.28%, and 99.95% for Mutualistic–SW, Neuronal–Brain, and SIS–Gamma, respectively. These results indicate that the predecessor graphs retain the majority of the total incoming contribution mass across all three systems. The revised definition is provided in Equation (30), Appendix D.1 (p. 16), and the results are reported in Table 7 (p. 19). Furthermore, we recognize that contribution-mass retention alone does not fully characterize the coverage of affected targets by the inferred predecessor graph. The original path-agreement evaluation was restricted to targets for which both the learned and analytical analyses successfully identified corresponding paths. Consequently, it did not account for affected targets that were missed or could not be connected to the perturbation source through the inferred predecessor graph. To address this limitation, we introduce an additional metric, affected-target coverage, defined as $$C_{\mathrm{aff}}(s,m;\mathcal G) =\frac{\lvert\mathcal A_{s,m}^{\mathrm{NL}} \cap\operatorname{Reach}_{\mathcal G}(s)\rvert} {\lvert\mathcal A_{s,m}^{\mathrm{NL}}\rvert}.$$ where $\mathcal A_{s,m}^{\mathrm{NL}}$ denotes the set of non-source nodes whose ground-truth nonlinear responses reach the same node-specific RMS detection thresholds at one or more sampled times for source–trajectory pair $(s,m)$, and $\operatorname{Reach}_{\mathcal G}(s)$ denotes the set of non-source nodes reachable from perturbation source $s$ through directed paths in the retained predecessor graph $\mathcal G$. This metric quantifies the proportion of nonlinear affected targets covered by source-connected paths in the inferred graph. It accounts for both targets missed by the learned variational response and detected targets that cannot be connected to the perturbation source through the inferred predecessor graph. Coverage is defined for conditions with at least one nonlinear affected target.
AuthorsOpenReview ↗

Response to Weakness 2 and Question 2: Quantifying Information Loss from Arrival Filtering (Part 2)

Using the same 20 fixed perturbation sources on each of the 10 test trajectories and α=0.01\alpha=0.01, we evaluated affected-target coverage across all three systems. The results are summarized below:

Source-connected coverage of nonlinear affected targets. The confidence intervals apply to condition medians. Pooled coverage combines affected target instances across valid conditions.

SystemJacobianValid<br>pairsMedian [IQR]<br>(%)95.00% CI<br>(%)Pooled<br>(%)Covered /<br>affected
Mutualistic–SWLearned20098.72 [95.83, 100.00][98.57, 98.86]95.8932,390 / 33,777
Analytical20098.15 [96.17, 100.00][97.71, 98.51]98.1733,160 / 33,777
Neuronal–BrainLearned19499.11 [98.26, 99.76][98.87, 99.39]98.7758,952 / 59,688
Analytical19499.64 [98.88, 100.00][99.54, 99.71]99.2259,220 / 59,688
SIS–GammaLearned195100.00 [100.00, 100.00][100.00, 100.00]99.171,189 / 1,199
Analytical195100.00 [100.00, 100.00][100.00, 100.00]100.001,199 / 1,199

These results demonstrate that the inferred predecessor graphs cover the vast majority of targets whose nonlinear responses reach the specified detection thresholds across all three systems. Incoming contribution mass retention and affected-target coverage assess the retained incoming contribution mass and the source-connected coverage of nonlinear affected targets, respectively.

The definition of affected-target coverage is provided in Equation (34), Appendix D.6 (p. 21). Its evaluation protocol, aggregation procedure, and detailed results are presented in Appendix D.6 (pp. 21–22), with the general trajectory-level bootstrap uncertainty estimation described in Appendix D.3 (p. 17). The main-text results are reported in Section 3.4 (p. 9), while the detailed statistics for learned and analytical Jacobians are presented in Table 10 (p. 22).

Reviewer_yspBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a data-driven method for inferring perturbation propagation in nonlinear dynamical networks when governing equations needs to be estimated. The data is given observed trajectories where each variable is a node on a graph. A graph-constrained neural dynamics model is learned the data, its time varying Jacobian is evaluated along a reference trajectory, and the resulting nonautonomous variational system is integrated to obtain first order responses. Variation of constants then decomposes each nonsource node's response into time integrated incoming edge contributions, which are combined with arrival ordering to construct predecessor graphs and source connected paths. The framework is evaluated on 3 primary and some additional network dynamics combinations. The work is technically sound of straitforward computation from perturbation analysis.

Critical strengths

This paper offers a way to studying the perturbation from obserational networked dynamic data. The presentation and mathematical derivation is very clear. The decomposition of the response of node i into a sum of edge contributions C_{j->i}(t) is exact under the variational approximation, and the edge-aligned structure of GCND ensures off-diagonal Jacobian entries correspond to structural interactions, which is necessary for the downstream attribution. The derived arrival-conditioned predecessor graph is interesting.

Critical weaknesses

My main concern of this method is the interpretation of perturbation arrival times and the thresold value related arrival-conditioned predecessor. The proposed defines perturbation arrival times eq8 through a relative detection threshold and later uses these times to build predecessor graphs. While the underlying variational dynamics and so on are mathematically correct, I am not sure how could the predecessor graph from eq14 be useful. In continuous-time dynamical networks, perturbations generally propagate to dynamically connected nodes immediately, although their magnitudes may be extremely small. Eq8 gives a threshold crossing time, not the earliest time at which the physical perturbation influences node regardless what the value of alpha (>=0) is. One can construct examples that by turning some variables, you can easily change the order according to eq14 that is against the network connection (as directed graph). This raises a question about the validity of using threshold crossing order to infer propagation pathways. The method appears to construct an operational, threshold-dependent predecessor graph rather than necessarily recovering the actual dynamical transmission structure. Since this is data driven, what should the practioners read from the result? Maybe a sensitivity analysis with respect to alpha is could be useful?

Another one is that the method considers first-order variational responses, which can vanish even when the nonlinear response is nonzero at higher orders. The resulting affected-node sets should therefore be interpreted as first-order, threshold-dependent classifications rather than complete characterisations of nonlinear perturbation influence.

Questions for authors

What is the intuition of the tolerance in the ordering of the arrival time of adjacent nodes in eq14?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1: Arrival-Time Threshold and Predecessor Graph (Part1)

Interpretation of the Predecessor Graph and the Role of the Detection Threshold

We sincerely thank you for pointing out this important distinction. We fully agree that threshold-crossing times do not represent the earliest physical influence of a perturbation. In continuous-time dynamical networks, a perturbation can begin to influence dynamically connected nodes almost immediately, while the initial responses may be arbitrarily small. Therefore, the arrival time defined in our work should be interpreted as the first time at which a node's response becomes detectable under a prescribed threshold, rather than the onset of its physical response. The predecessor graph is consequently not intended to identify the earliest physical transmission of a perturbation or to reconstruct all actual propagation paths. Instead, it organizes arrival-conditioned edge contributions according to the ordering of detectable node responses, providing a contribution-informed representation of the incoming interactions associated with detectable responses and their source-connected candidate routes. We acknowledge that the previous manuscript may not have sufficiently distinguished detectable arrival from the onset of physical influence, potentially leading to this misunderstanding. To address this concern, we have revised the definition of the detection threshold and arrival time in Section 2.5 (p. 5), explicitly clarifying the threshold-based nature of the arrival time. We have also expanded the corresponding discussion in the revised manuscript, including the concluding discussion (Section 4, p. 10), to further clarify the intended interpretation and scope of the predecessor graph.

We also acknowledge that introducing a detection threshold inevitably excludes some weak responses and their associated contributions. Nevertheless, such a criterion is important when considering practical observations, where finite temporal sampling and measurement noise make arbitrarily small initial responses difficult to distinguish from numerical fluctuations or observational noise. The threshold therefore involves a trade-off between detecting weak responses, avoiding false positives, and retaining relevant edge-contribution information. To examine this trade-off, we conducted additional sensitivity experiments under different detection thresholds and observation noise levels (Appendix G, pp. 28–30, Tables 18–19). Under the lower noise level κref=0.10\kappa_{\mathrm{ref}}=0.10, with both white noise and temporally correlated AR(1) noise, setting α=0.01\alpha=0.01 reduces the false-positive rate to below 0.01% in all three systems. At higher noise levels, false positives remain more pronounced, while increasing the threshold further reduces false detections at the cost of missing some genuine responses, illustrating the dependence of threshold selection on observation conditions. We also evaluated the incoming contribution mass retained by the inferred predecessor graphs. At α=0.01\alpha=0.01, the median retention rates are 94.56%, 87.28%, and 99.95% for Mutualistic–SW, Neuronal–Brain, and SIS–Gamma, respectively. These results indicate that the selected threshold retains most of the incoming contribution mass while providing effective suppression of noise-induced false detections under the tested lower-noise conditions. Taken together, the noise sensitivity and contribution retention results support α=0.01\alpha=0.01 as a reasonable common operating threshold for the three main systems.

AuthorsOpenReview ↗

Response to Weakness 1: Arrival-Time Threshold and Predecessor Graph (Part2)

Revision of the Arrival-Time Detection Threshold

We sincerely appreciate you for pointing out that the original arrival-time detection criterion could be sensitive to differences in variable scales. A common absolute threshold can produce different detection results across nodes with different state magnitudes. This can also change their inferred arrival order.

We have revised the detection criterion using node-specific RMS scales. For a perturbation applied at source node s, the threshold at node i is defined as:

ηi∣s=α∣ϵ∣qiqs,\eta_{i\mid s}=\alpha|\epsilon|\frac{q_i}{q_s},

where qiq_i is the time-weighted RMS of the unperturbed reference state at node i:

\int_{t_0}^{T}\bar{x}_j^2(t)\,\mathrm{d}t\right)^{1/2}.$$ Here, $\epsilon$ is the initial perturbation amplitude at the source, and $\alpha$ is a dimensionless detection ratio. The corresponding arrival time is: $$\tau_i^{\mathrm{arr}}=\inf\{t:|\delta_i(t)|\geq\eta_{i\mid s},\ t\in[t_0,T]\}.$$ Equivalently, the detection condition can be written as: $$\frac{|\delta_i(t)|}{q_i}\geq\alpha\frac{|\epsilon|}{q_s}.$$ This formulation makes arrival-time detection invariant to positive rescaling of individual node states. Under a rescaling $x'_j=c_jx_j$, the reference RMS and perturbation response become $q'_j=c_jq_j$ and $\delta'_i(t)=c_i\delta_i(t)$, respectively. The source perturbation amplitude also becomes $\epsilon'=c_s\epsilon$. Both ratios in the detection condition therefore remain unchanged. The detected arrival times and their ordering are consequently preserved when the measurement units of individual nodes change. The revised RMS threshold and its scale-invariance property are presented in Section 2.5 (Equations (8)–(9), p. 5) and Appendix G (p. 28). Appendix G further explains the rationale for the node-specific threshold and evaluates its sensitivity to different values of $\alpha$ (pp. 28–30, Tables 18–19).
AuthorsOpenReview ↗

Response to Weakness 2: First-Order Jacobian-Based Perturbation Responses

Thank you for raising this important point. We agree that first-order variational dynamics describe local perturbation behavior around the reference trajectory and do not necessarily capture all higher-order nonlinear effects. However, for sufficiently small perturbations, the Jacobian provides an accurate approximation of the first-order response, while higher-order nonlinear effects are expected to be relatively small. To examine this, we conducted experiments comparing first-order variational responses with nonlinear responses across different perturbation amplitudes. The results show that, at smaller perturbation amplitudes, the first-order responses obtained from the Jacobian capture the affected nodes and their arrival ordering well. As the perturbation amplitude increases, the discrepancy between the variational and nonlinear responses becomes larger. This behavior is consistent with the local nature of the first-order approximation. This is also why we use r=1.00%r=1.00\% as the common small-perturbation setting for the subsequent edge-contribution and propagation-structure analyses (Section 3.3, pp. 7–9; Figure 3, p. 8; Table 2, p. 8).

To further clarify the scope of the proposed method, we have added this limitation to the Conclusion of the revised manuscript. We clarify that the framework relies on a first-order linearization along the reference trajectory and is therefore best suited to small perturbations. Higher-order nonlinear effects may reduce the accuracy of inferred responses at larger perturbation amplitudes. Extending edge-resolved perturbation analysis to larger perturbation amplitudes is left for future work (Section 4, p. 10).

Reviewer_sG72OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper studies how perturbations propagate through nonlinear dynamical networks when the governing equations are unknown but the network structure is available. The authors train a graph-constrained neural vector field on observed trajectories, calculate its time-varying Jacobian, and integrate the corresponding variational equations to estimate source-specific perturbation responses. They then apply the standard variation-of-constants formula to decompose each node's response into contributions from incoming edges. These contributions, together with threshold-based arrival ordering, are used to construct predecessor graphs and source-connected paths.

Experiments on three primary synthetic dynamical systems show strong agreement with analytical-Jacobian calculations and substantial response suppression following targeted edge removal.

Critical strengths
  • The paper is technically coherent.
  • The authors provided useful evidence that the numerical pipeline works as intended.
Critical weaknesses
  1. Insufficient scientific justification for the central contribution. The paper combines established techniques, including neural vector-field learning, variational dynamics and variation of constants, without establishing what new understanding of perturbation propagation their combination provides. More importantly, the authors interpret threshold-ordered predecessor graphs as propagation pathways without justifying this interpretation. Even with a perfect Jacobian, the inferred graph need not represent the actual propagation structure. The paper does not adequately confront this fundamental issue or establish how its construction advances existing work on transient network responses and path contributions.
  2. The experiments do not distinguish the proposed mechanism from simpler explanations. Much of the evaluation measures agreement between quantities computed using learned and analytical Jacobians through the same procedure. High agreement establishes numerical fidelity but does not validate the physical interpretation of those quantities.
Questions for authors
  • What justifies interpreting the inferred predecessor graphs as actual perturbation propagation pathways? I reckon a node can transmit a substantial perturbation downstream without itself crossing the threshold?
4 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 and Question (Part1)

We greatly appreciate your thoughtful comments and the opportunity to clarify the scientific contribution and interpretation of our method. We further clarify the scientific contribution of this work, the interpretation of the predecessor graph, and the rationale for its construction.

Scientific Contribution

The central goal of this work is to identify which incoming interactions contribute to the formation of a target node's transient perturbation response and quantify their respective contributions when the governing equations of the system are unknown.

The main contribution of our work is a finite-time, edge-level perturbation contribution analysis method for unknown nonlinear network dynamics. The method decomposes the perturbation response of a target node at the first time it reaches a detectable level into signed, time-integrated contributions from its structural incoming edges. These contributions reflect the upstream perturbation history carried by the sender node and the time-varying local interactions. They also account for the amplification or attenuation of inputs at different times by the receiver node's own dynamics. For a non-source node with zero initial perturbation, the sum of contributions from all structural incoming edges exactly reconstructs its first-order variational response at the arrival time (Section 2.6, p. 5). This decomposition allows us to distinguish the roles of different incoming interactions. It identifies which interactions enhance the target response, which interactions counteract it, and how much cumulative influence each interaction produces. The mathematical definition of edge contributions is provided in Section 2.6, Equations (13)–(14) (p. 5).

Compared with existing studies, this work provides a complementary analytical perspective. Relative to previous research on transient network responses and path contributions, we further focus on time-varying local interactions along nonlinear reference trajectories. We decompose the finite-time response of a target node into signed accumulated contributions from different incoming edges and quantitatively characterize how individual interactions enhance or counteract the target response. Our work goes beyond identifying which nodes are affected by a perturbation or which network relationships are involved in generating a response. It further quantifies the role of individual local interactions in forming the transient response of a specific target node at a specific response time.

AuthorsOpenReview ↗

Response to Weaknesses 1 and Question(Part2)

Interpretation of the Predecessor Graph and Rationale for the Threshold (Response to Your Question)

We would like to clarify the purpose of constructing the predecessor graph. We regard the predecessor graph as a structured representation for interpreting and organizing perturbation interactions, rather than a unique or complete reconstruction of the true propagation paths. The predecessor graph provides an interpretable graph for organizing interactions. Specifically, it organizes the incoming interactions associated with a target response and their source-connected relationships according to the finite-time edge contributions and the detectable arrival order of node responses. This allows us to identify and display important interactions and their candidate connecting paths that contribute to the target perturbation response.

Regarding the reviewer's concern that the predecessor graph may not represent the true propagation structure even when the Jacobian is perfectly accurate, we agree with this observation. One reason is that, in continuous-time dynamical systems, perturbation responses can begin immediately after the perturbation is applied, while their initial magnitudes may be arbitrarily close to zero. In practical numerical computations and observations, finite sampling resolution, numerical errors, and measurement noise can make these extremely weak responses difficult to distinguish from numerical fluctuations or observation noise. If the earliest nonzero response were directly used as the arrival time, noise or small numerical fluctuations could be identified as effective responses. This could affect the inferred arrival order and the stability of the predecessor graph. Therefore, we introduce a response detection threshold and define arrival as the first time a node's perturbation response reaches a detectable level. We use a node-specific RMS threshold for this purpose (Equation (8), p. 5).

We also recognize that introducing a threshold inevitably leads to some information loss. Some weak responses may not reach the detection criterion. The threshold must balance noise suppression against the retention of meaningful perturbation responses. To address this issue, we added the rationale for threshold selection and a noise sensitivity analysis in the revised manuscript. We also quantitatively evaluate the retained edge-contribution mass by calculating the ratio of the absolute contribution mass retained by the predecessor graph to the total absolute contribution mass of all structural incoming edges (Appendix G, pp. 28–30, Tables 18–19; Equation (30), p. 16; Table 7, p. 19).

With α=0.01\alpha=0.01, the median contribution mass retention rates in Mutualistic–SW, Neuronal–Brain, and SIS–Gamma are 94.56%, 87.28%, and 99.95%, respectively. This metric is calculated for each valid trajectory–source pair as the ratio of the absolute incoming contribution mass retained by the predecessor graph to the total absolute contribution mass from all structural incoming edges. These results show that, although the threshold and subsequent predecessor selection inevitably discard some weaker interaction contributions, the final predecessor graphs retain most of the incoming absolute contribution mass in all three systems. Together with the additional threshold and noise sensitivity analyses, these results support the use of α=0.01\alpha=0.01 as the primary experimental setting. This threshold provides a reasonable balance between reducing noise interference and retaining the contributions of major interactions.

It is important to note that edge-level accumulated contributions are calculated for all structural incoming edges of a target node before the predecessor graph is constructed. They are not excluded in advance simply because the sender node has not reached the detection threshold. The threshold determines the arrival time of the target node and therefore the upper integration limit of its edge contributions. Only after this calculation are predecessor edges selected according to their contributions and the arrival-order condition. Thus, the threshold may affect the numerical values of edge contributions and the final predecessor graph, but it does not directly exclude incoming contributions from subthreshold sender nodes during the contribution calculation (Equation (15), p. 6).

The functional relevance of the predecessor graph is not established solely through the numerical accuracy of the recovered Jacobian or edge contributions. We further evaluate its functional relevance through interventions in the ground-truth nonlinear systems. The corresponding experimental design and results are presented in our response to Weakness 2.

AuthorsOpenReview ↗

Response to Weaknesses 2 (Part1)

Thank you so much for pointing out this important issue. We agree that the agreement between the learned and analytical Jacobians mainly validates the numerical accuracy of local dynamics and edge-contribution recovery. This agreement alone cannot sufficiently establish the physical interpretation of the inferred results. We therefore added the following analyses based on interventions in the ground-truth nonlinear dynamical systems to further evaluate the functional relevance of the identified predecessor edges.

Original Ancestry-Removal Experiment

The original manuscript already included an ancestry-removal experiment in the ground-truth nonlinear dynamical systems. We compared the effects of removing inferred source-connected predecessor edges with those of matched random edge removals on the target node's perturbation response. We measured the proportion of cases in which the target response failed to reach the detection threshold within the analysis window. The arrival-elimination rates are reported in Appendix D.4 (p. 20), with the intervention contrasts in Table 8 (p. 20).

Ancestry-removal effects on detectable arrival. Rates summarize target conditions.

SystemTargeted removalMatched random removal
Mutualistic–SW83.33% (50/60)0.00% (0/180)
Neuronal–Brain95.00% (57/60)1.67% (3/180)
SIS–Gamma100.00% (49/49)0.00% (0/147)

In all three systems, targeted removal of the inferred predecessor ancestry edges prevents the target node from reaching the detection threshold more frequently than matched random removal. These results provide initial evidence that the inferred source-connected predecessor structures have functional relevance.

AuthorsOpenReview ↗

Response to Weaknesses 2 (Part2)

Fixed-Budget Intervention Experiment

To further rule out differences in the number of intervened edges and their distribution across receiving nodes, we added a fixed-budget intervention experiment to the revised manuscript. This experiment provides a more direct test of whether interactions selected by contribution ranking have stronger functional effects (Section 3.4, p. 10; Appendix D.5, pp. 20–21; Table 9, p. 21).

We hold the unperturbed reference trajectory fixed and intervene on selected interaction edges in the ground-truth nonlinear system. We set the sender's perturbation component to zero on each selected edge while retaining the interaction's dependence on the receiver node's perturbed state. The random controls exactly match the total number of intervened edges and the number of intervened edges at each receiving node.

We compare the normalized changes in the target response produced by the two intervention strategies. We calculate the paired difference ΔIi\Delta I_i between contribution-ranked selection and matched random selection. A positive ΔIi\Delta I_i indicates that contribution-ranked selection produces a larger change in the target response than matched random selection.

The median paired differences and trajectory-clustered 95.00% confidence intervals across the eight system–budget comparisons are reported below.

Fixed-budget intervention effects. Values summarize the same paired conditions within each system and budget.

KKnnn1/n2/n3n_1/n_2/n_3Learned$I_i$Random$I_i$ΔIi\Delta I_i [95.00% CI]
Mutualistic–SW
51820/0/1820.57480.00500.5564 [0.4041, 0.6559]
101820/0/1820.99490.01800.9279 [0.8876, 0.9438]
201440/0/1440.99910.09900.8853 [0.8729, 0.8954]
Neuronal–Brain
51640/0/1640.45440.02510.4120 [0.3736, 0.4368]
101640/0/1640.70630.04180.6285 [0.6106, 0.6591]
201640/0/1640.91000.07020.7794 [0.7646, 0.7937]
SIS–Gamma
119475/34/851.00000.00660.9918 [0.9904, 0.9935]
2708/11/510.99950.01500.9801 [0.9743, 0.9852]

All eight 95.00% confidence intervals lie entirely above zero. This shows that, under the same intervention budget and with the receiving-node distribution strictly matched, interactions selected by contribution ranking consistently produce larger changes in the target's nonlinear response than matched random interactions.

These experiments directly evaluate the functional effects of the identified interactions in the ground-truth nonlinear dynamical systems. They do not merely verify the agreement between two Jacobian calculations. The additional fixed-budget intervention further controls the number of intervened edges and their distribution across receiving nodes. The results therefore provide a more direct assessment of the selection value offered by contribution ranking. These findings further support the practical value of edge-contribution ranking and the resulting predecessor graph for identifying and organizing important perturbation interactions. We do not claim that the predecessor graph recovers all true propagation paths, but emphasize its ability to identify dominant interactions with substantial functional effects. The corresponding experimental protocol and results have been added to Section 3.4 (p. 10) and Appendix D.5 (pp. 20–21), including Table 9 (p. 21).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response Summary and Updated Manuscript for All Reviewers

We sincerely thank all reviewers for their careful evaluation and constructive feedback. In particular, we greatly appreciate:

  • Reviewer sG72 for recognizing the technical coherence and numerical evaluation of our framework, and for valuable comments on its scientific contribution and the interpretation of predecessor graphs.
  • Reviewer yspB for recognizing the clarity and mathematical soundness of the formulation, and for helpful comments on arrival-time detection, threshold selection, and the first-order approximation.
  • Reviewer vagR for recognizing the well-defined edge contributions and Jacobian-based evaluation, and for constructive suggestions on evaluation metrics, intervention controls, and robustness analyses.

Guided by this feedback, we have revised the manuscript and added the analyses and experiments summarized below.

Revision / Additional ExperimentReviewer(s)Location in Manuscript
Clarified the scientific contribution, interpretation of predecessor graphs, and scope of the first-order approximationsG72, yspB, vagRSections 2.6–2.7 and 4
Introduced node-specific RMS thresholds and added threshold-sensitivity and false-positive analysessG72, yspB, vagRSection 2.5; Appendix G
Explained the arrival-order tolerance and added a numerical-tolerance sensitivity analysisyspBAppendix C.2
Revised incoming contribution mass retention and added source-connected coverage of nonlinear affected targetssG72, vagRSection 3.4; Appendices D.1 and D.6; Tables 7 and 10
Added fixed-budget nonlinear interventions with a fixed reference trajectory and matched random controlssG72, vagRSection 3.4; Appendix D.5; Table 9
Expanded the main propagation evaluations to 10 test trajectories per system, using the same 20 fixed sources on each trajectory and trajectory-level bootstrap uncertainty estimatesvagRSections 3.1–3.4; Appendices D.2–D.3
Added observation-noise experiments at 5.00% derivative-label noise with three training seeds per system, reduced-sampling-density experiments, and numerical-resolution checksvagRAppendices H.1–H.3; Tables 20–21

We believe these revisions strengthen the presentation and empirical support of our work and address the key concerns raised by the reviewers. We hope the revised manuscript conveys the contribution and value of our framework more clearly.

AuthorsOpenReview ↗

Summary of Major Contributions and Experimental Validation

Dear AC, SAC, and PCs,

We sincerely appreciate your time and effort in handling our submission. To facilitate the overall assessment of our work, we provide a concise summary of its core contributions and empirical validation.

1. Core Contributions

Our work addresses the problem of inferring source-specific, edge-resolved perturbation propagation in nonlinear dynamical networks when the governing equations are unknown but network structure and observed trajectories are available.

The main contributions are:

  • Data-Driven Perturbation Response Inference. We combine learned trajectory-dependent local dynamics with variational analysis to infer responses to small perturbations, affected-node sets, and detectable arrival times along unperturbed reference trajectories.

  • Edge-Resolved Accumulated Contribution Decomposition. We decompose each affected non-source node's response into signed, time-integrated contributions from individual structural incoming edges. This quantifies how individual interactions contribute to the transient response while accounting for upstream perturbation history and subsequent amplification or attenuation.

  • Contribution-Informed Propagation Structure. We organize arrival-conditioned edge contributions according to detectable arrival order to construct predecessor graphs and extract source-connected candidate propagation paths. The predecessor graphs provide an interpretable representation of the dominant interactions contributing to target responses.

2. Main Experimental Validation

Validation TargetExperimentKey Findings
Local-dynamics and Jacobian recoveryTrajectory Rollout and Jacobian FidelityThe learned dynamics reproduce held-out trajectories and preserve edge-aligned local sensitivities consistent with analytical Jacobians.
Small-perturbation response inferenceFinite-Time Perturbation Amplitude SweepAt the selected small-perturbation setting, learned variational responses closely reproduce nonlinear affected-node sets and their arrival order.
Importance of time-varying JacobiansFixed-Jacobian ComparisonsFreezing the Jacobian at perturbation onset or near steady state reduces response and predecessor-graph agreement.
Edge-resolved contribution accuracyDirect Edge-Contribution AgreementLearned signed, time-integrated edge contributions closely agree with their analytical references.
Predecessor-graph recovery and contribution retentionPredecessor-Graph Agreement and Incoming Contribution Mass RetentionThe inferred predecessor graphs closely match analytical-Jacobian references and achieve high incoming contribution mass retention.
Functional relevance of contribution-ranked predecessor edgesFixed-Budget Edge InterventionsContribution-ranked edge selections produce larger changes in target responses than matched random selections.

We believe these contributions establish a connection between data-driven local dynamical recovery and edge-resolved perturbation propagation analysis, providing a systematic approach to studying how perturbations develop through individual interactions in nonlinear networks with unknown governing dynamics.

Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.