- 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
-
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.
-
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
-
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.
-
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.
-
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.
-
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.
-
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
-
How should the graph be interpreted in the chain example?
-
How much does the arrival filter discard?
-
Does the intervention advantage survive route-aware controls?
-
What observation and numerical settings are required for the reported results?
9 comments
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.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 , 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.
| System | Jacobian | Valid<br>pairs | Median [IQR]<br>(%) | 95.00% CI<br>(%) | Pooled<br>(%) | Covered /<br>affected |
|---|---|---|---|---|---|---|
| Mutualistic–SW | Learned | 200 | 98.72 [95.83, 100.00] | [98.57, 98.86] | 95.89 | 32,390 / 33,777 |
| Analytical | 200 | 98.15 [96.17, 100.00] | [97.71, 98.51] | 98.17 | 33,160 / 33,777 | |
| Neuronal–Brain | Learned | 194 | 99.11 [98.26, 99.76] | [98.87, 99.39] | 98.77 | 58,952 / 59,688 |
| Analytical | 194 | 99.64 [98.88, 100.00] | [99.54, 99.71] | 99.22 | 59,220 / 59,688 | |
| SIS–Gamma | Learned | 195 | 100.00 [100.00, 100.00] | [100.00, 100.00] | 99.17 | 1,189 / 1,199 |
| Analytical | 195 | 100.00 [100.00, 100.00] | [100.00, 100.00] | 100.00 | 1,199 / 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).
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 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 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, 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→i, the edge j→i may have a nonzero accumulated contribution even when node j remains below the detection threshold at the time node i reaches its threshold. Before predecessor filtering, we calculate the accumulated contribution at the target's arrival time for every structural incoming edge of node i:
=\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).