acceptodds
Under review as a conference paper at ICLR 2027

CHIRP: Recovering Network Topology from Coupling-Response Profiles

Abstract

Synchronized trajectories can arise from a direct link, an indirect path, or a shared source. Oscillator experiments often sweep coupling strength only to describe the approach to synchronization. We instead write each phase-difference coefficient as a dose-independent term plus a term that changes with the known dose, and rank candidate edges by the second term. A single dose leaves the two terms collinear; varying the dose separates them. We bound the coefficient error and the condition for recovering the edge set. Choices were fixed on five physical networks and applied once to fifteen held-out graphs of 28 chaotic electronic oscillators. The locked estimator ranks edges above each declared baseline. Cutting the raw traces to a 3 s budget before filtering still gives higher average precision than a single-dose model under that same raw-time budget. Otsu's rule selects a graph from the ranking without using the known number of edges. In synthetic trials, average precision falls in a confounded scenario that adds a constant-amplitude common drive together with other disturbances, and falls again when the dose response is nonlinear. These results suggest that a calibrated intervention can make topology easier to identify than an observational fit when the intervention is structurally relevant and approximately affine over the fitted range.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
network topology inference, Kuramoto, coupled oscillators, phase dynamics, coupling-response analysis, structural identifiability

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

The paper proposes CHIRP for recovering network structure from experiments performed at several known coupling strengths. Instead of relying on how strongly two nodes appear related in a single experiment, CHIRP asks how their estimated interaction changes as the coupling strength is varied. The authors provide theoretical justification for why multiple intervention levels can help distinguish structural interactions from nuisance effects, and validate the method on networks of electronic oscillators.

Critical strengths

The central idea of this paper is simple and intuitive: controlled interventions can reveal network structure more reliably than observational relationships alone. The experimental evaluation of the paper is careful designed and executed, with major choices fixed before testing on held-out networks, and the reported performance is very strong.

I also appreciate that the paper includes meaningful controls and negative results. In particular, the paper reveals that performance can degrade substantially when the intervention is confounded or when the response to intervention becomes nonlinear. This helps clarify when the method should and should not work.

Critical weaknesses

For me, the main limitation is how general are the main results of the paper. The method is developed and tested for a rather specific oscillator model with a fixed phase-based interaction form. The strong experimental results therefore do not establish that the approach will work equally well for more general dynamical systems.

The experimental validation also seems narrow to me: all physical experiments use the same family of 28-node oscillator networks and a single type of scalar intervention. The paper does not address directed networks, hidden variables, or substantially different intervention mechanisms.

Finally, I am somewhat uncertain whether the paper fits for ICLR. The main contribution seems closer to experimental design, system identification, and network reconstruction than to a broadly applicable machine-learning methodology. The underlying learning method is relatively simple, while the strongest novelty comes from how the experiments are designed and interpreted.

Questions for authors

See the identified weaknesses. I would like to see how the authors respond to them.

1 comment

Thank you for the positive reading of the frozen protocol and of the negative results. All of the new analyses described below were done after the reviews came in. The paper marks them as post-review, and none of them touches the frozen protocol or its endpoint (R6–R20, mean AP 0.9980). The frozen estimator still has the same hash. Several answers below refer to the new synthetic checks (Section 7, Table 3; Appendix D). Each one starts from the matched model and changes a single assumption: dense static coupling, edges that do not respond to the dose, directed edges, a phase-lagged coupling function, or hidden nodes. There are 15 networks per scenario, 90 in total. Code, configs, unit tests and hashes are in the reproduction record (Run 5).

Comment: "The method is developed and tested for a rather specific oscillator model with a fixed phase-based interaction form."

-> The sine feature was fixed in advance for this benchmark. The method itself does not need it, since any coupling basis can be interacted with the dose. Section 3 now describes a (sin⁡,cos⁡)(\sin, \cos) variant that scores each pair by the norm of its two slopes. In a new phase-lag scenario with coupling sin⁡(Δϕ−1.2)\sin(\Delta\phi - 1.2), the sine-only fit falls to AP 0.450 and the (sin⁡,cos⁡)(\sin, \cos) basis brings it back to 1.000 (Table 3). A (sin⁡,cos⁡)(\sin, \cos) single-dose fit also reaches 1.000 in that scenario, so a phase lag is handled by the basis. The slope helps with nuisance terms, which is a separate issue. We do not claim that this covers general dynamical systems. The affine dose response and a known interaction basis are still assumptions.

Comment: "All physical experiments use the same family of 28-node oscillator networks and a single type of scalar intervention. The paper does not address directed networks, hidden variables, or substantially different intervention mechanisms."

-> We agree that the physical evidence covers one device family and one scalar intervention. Section 8 says so, and we do not claim more. We were not able to add a second physical system with a calibrated coupling dose in the time available for the revision. We did add simulations for the two cases the reviewer mentions, both in Table 3 with intervals in Table 10. For directed networks, each edge is oriented one way (42 of 756 ordered pairs), and the unsymmetrized slopes recover the directed graph at AP 1.000. For hidden variables, 4 of the 28 nodes are unobserved, and CHIRP keeps AP 0.960 on the observed subgraph, where a single dose gets 0.393. Section 8 says that these are simulations and not physical results, and that disturbances which move with the dose have not been tested.

Comment: "I am somewhat uncertain whether the paper fits for ICLR. The main contribution seems closer to experimental design, system identification, and network reconstruction …"

-> We think the paper is within ICLR's scope of structure learning and causal discovery. Learning graphs from interventions is an active topic in ML (Hauser & Bühlmann 2012; Mooij et al. 2020; Brouillard et al. 2020), and so is relational inference for interacting dynamical systems (Kipf et al. 2018; Löwe et al. 2022). The revised introduction and related work place CHIRP between these two lines. Interventional structure learning usually intervenes on node variables in static data, and relational inference usually learns from observation only. CHIRP applies a continuous, calibrated intervention to the interactions themselves and reads the graph from how they respond. The learning machinery is deliberately simple, and the paper says so. Our contribution is the choice of what to estimate, along with its identifiability, its limits (Corollary 2, Table 3) and a frozen held-out evaluation.

Reviewer_TU6pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduces a method (i.e., CHIRP) for recovering network topology from oscillator dynamics. The propose architecture separates pairwise interaction coefficients into dose-independent and dose-dependent components, and uses the estimated response slopes to identify network connections. The papers give theoretical analyses of identifiability and estimation error, and evaluate the method on physical oscillator networks and synthetic data.

Critical strengths
  1. The work investigates how to recover the underlying network topology from the dynamical trajectories of oscillators, which is an interesting and practically relevant problem.

  2. Using dynamical responses under different coupling strengths to identify network connections is an interesting research topic.

Critical weaknesses
  1. CHIRP mainly constructs [Xe,deXe][X_e, d_e X_e] and then uses Ridge Regression to estimate the coefficients. This seems very close to existing studies, e.g., parameter-dependent regression, which makes me feel that the novelty is limited.

  2. Proposition 1 mainly consider the full-rank condition of the design matrix, Proposition 2 utilizes a standard variance bound, and Proposition 3 is a conventional Ridge Regression error analysis. In my view, these results do not provide new theoretical insights.

Questions for authors
  1. One result in Section 6 is somewhat confusing to me. After removing the dose-independent block, the model still achieves a mean AP of 0.9980, exactly the same as the full CHIRP model.

  2. CHIRP determines whether an edge exists based only on bijb_{ij}. If the coupling strength of a true physical connection does not change under a intervention, its bijb_{ij} may be zero, and the method would fail to identify that edge. I am curious whether CHIRP recovers the complete network topology or only intervention-responsive edges.

  3. The authors use 95% network-bootstrap intervals in most experiments. I am not arguing against this approach, but it does not seem very common in machine learning papers. I would like to see more motivation for using this strategy?

3 comments

Thank you for the careful reading and for the direct questions about novelty and theory. All of the new analyses described below were done after the reviews came in. The paper marks them as post-review, and none of them touches the frozen physical protocol or its endpoint (R6–R20, mean AP 0.9980). The frozen estimator still has the same hash. Several answers below refer to the new synthetic checks (Section 7, Table 3; Appendix D). Each one starts from the matched model and changes a single assumption: dense static coupling, edges that do not respond to the dose, directed edges, a phase-lagged coupling function, or hidden nodes. There are 15 networks per scenario, 90 in total. Code, configs, unit tests and hashes are in the reproduction record (Run 5).

Comment: "CHIRP mainly constructs and then uses Ridge Regression to estimate the coefficients. This seems very close to existing studies, e.g., parameter-dependent regression … novelty is limited."

-> We agree that the estimator is a varying-coefficient (parameter-dependent) ridge regression. The paper now says so and cites Hastie & Tibshirani (1993) and the LPV literature. What we think is new is what the regression is used to estimate. When a known dose acts on the interactions, we read the slope bb as the graph and leave the dose-independent part cc dense and discard it. The new experiments show that this choice decides the result. Under dense static coupling, the same ridge machinery gets AP 0.363 without the cc block and 0.614 at a single dose, while the joint slope keeps 0.987 (Table 3). A parameter-dependent regression that does not separate cc from bb, or that reads structure from the level at one setting, would not get this result.

Comment: "Proposition 1 mainly consider the full-rank condition … Proposition 2 utilizes a standard variance bound, and Proposition 3 is a conventional Ridge Regression error analysis … these results do not provide new theoretical insights."

-> We accept this. Section 4 now says openly that the proofs are short least-squares arguments and are not new regression theory. What they do is state which property of the experiment each step depends on: dose variance for identifiability, endpoint spread for efficiency, and a slope margin for support recovery. We also added Corollary 2, which we hope is closer to the insight the reviewer was asking for. With exact coefficients, a single-dose fit ranks pairs by ∣c+d b∣|c + d\,b|. A non-edge whose dose-independent coupling is as large as an edge's coupling at that dose can therefore outrank a true edge, and a dense nuisance defeats the single-dose fit at any sample size. The support condition for the slope does not involve cc at all. The new dense-static-coupling scenario tests exactly this contrast (single dose 0.614, CHIRP 0.987).

Comment: "After removing the dose-independent block, the model still achieves a mean AP of 0.9980, exactly the same as the full CHIRP model."

-> The two means only agree to four decimals (0.99797 and 0.99799). Graph by graph, the fit without the block is higher on 7 graphs, lower on 4, and tied on 4 where both reach AP 1. The largest paired difference is 0.0046, and a two-sided Wilcoxon test gives p=0.96p = 0.96. The reason is physical. On R6–R20 the fitted dose-independent coupling is small. At d=0.05d = 0.05 the median ∣c^∣|\hat{c}| on non-edges is 0.70, while the median dose term on edges is 13.1, a ratio of 0.053. With a nuisance this small, the dose block can explain the coupling by itself. The block starts to matter when the nuisance is large. In the new dense-static-coupling scenario, removing it drops AP from 0.987 to 0.363, because the slope then has to absorb the static term. Section 6 and Appendix D.1 now report these numbers.

Comment: "If the coupling strength of a true physical connection does not change under an intervention, its bb may be zero, and the method would fail to identify that edge. … whether CHIRP recovers the complete network topology or only intervention-responsive edges."

-> The reviewer is right that CHIRP recovers the edges that respond to the intervention. The introduction, the method section and the discussion now all say this. In the physical benchmark the dose is the global coupling strength, so every edge responds. To look at the general case we added a partial-response scenario. One third of the edges couple at every dose with the strength the other edges have at the top dose, but they do not respond to the dose. The slope ranks the responsive edges at AP 0.975 and misses most of the fixed ones (AP 0.882 on all edges). The fixed edges are still visible in the same fit, since ∣c^∣|\hat{c}| ranks them against non-edges at AP 0.983. We note the caveat in the paper: c^\hat{c} also picks up nuisance, so it only identifies fixed edges when the nuisance is small.

Comment: "The authors use 95% network-bootstrap intervals … not very common in machine learning papers. … more motivation?"

-> We want the result to generalize over graphs, so the graph is the unit we resample. Windows from one recording are strongly dependent, and an interval computed over rows would be much too narrow. Resampling whole graphs is a cluster bootstrap (Efron & Tibshirani, 1993; Field & Welsh, 2007), and with 15 graphs it avoids having to assume normality. Section 6 now explains this. As a check, Appendix D.1 (Table 9) gives a Student-t interval over the 15 graph means for every method. For CHIRP it is [0.9967, 0.9993], and the bootstrap interval is [0.9968, 0.9990]. The conclusions are the same under either choice.

Reviewer_TU6pOpenReview ↗

Response to the Authors

Thank you for the authors' responses. They have addressed some of my concerns. However, I still find the novelty of the work limited.

Given the standards of ICLR, I will maintain my original rating.

Reviewer_us51OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper formulates network topology recovery as an estimation of an intervention response rather than selection from a single observation fit. It derives a rank condition for dose variation and provides finite-noise support bounds. The paper conducted experiments on the public benchmark that contains 20 independently wired networks of 28 chaotic electronic Rossler oscillators. However, the physical evidence is narrow. A high average precision does not mean the coupling magnitudes are accurate.

Critical strengths

None.

Critical weaknesses

The paper is not well written and it is difficult to follow the main ideas and results of the paper. For example, in Section 5, there is no detailed discussion of the main results (Table 1).

Questions for authors

My main suggestion to the authors is to improve the writing and presentation of the paper.

1 comment

Thank you for reading the paper closely. All of the new analyses described below were done after the reviews came in. The paper marks them as post-review, and none of them touches the frozen physical protocol or its endpoint (R6–R20, mean AP 0.9980). The frozen estimator still has the same hash. To make room for the new discussion and stay within nine pages, we moved the proof of Proposition 4 and the synthetic-stress bar chart to the appendix and cut sentences that repeated results. Code, configs, unit tests and hashes for the new analyses are in the reproduction record (Run 5).

Comment: "The paper is not well written and it is difficult to follow the main ideas and results of the paper. For example, in Section 5, there is no detailed discussion of the main results (Table 1)."

-> We agree that the results discussion was too thin. (Table 1 is in Section 6; Section 5 describes the protocol.) We changed Section 6 and the introduction as follows.

Section 6 has a new paragraph, "Reading Table 1". It sorts the rows into four groups, and each group answers one question. CHIRP at 18 s, at 3 s and with one repeat shows how much data the slope needs. The separate-fit and two-stage slopes on the same dose path test whether the joint fit matters. The single-dose and zero-dose fits test whether the slope does better than the level. The four observational scores (the dose-independent coefficient c, phase locking, voltage correlation and partial correlation) show how far observation alone gets.

The same paragraph then points to the three comparisons the claim rests on. (i) A single regression gives both blocks, yet the slope ranks edges at AP 0.9980 while the dose-independent coefficient ranks them at 0.2034, close to the random reference of 0.111. So the structural information is in the slope. (ii) With the same 3 s raw-time budget, CHIRP reaches 0.9350 and a single dose reaches 0.8814, so the gain does not come only from having more data. (iii) None of the observational scores goes above 0.3407.

The introduction now separates the standard machinery (a varying-coefficient ridge regression) from the idea we are testing, which is to read the dose slope as the graph and leave the nuisance dense. It also says which edges the method recovers.

Comment: "The physical evidence is narrow."

-> We agree, and Section 8 says this. Two things are new. The synthetic checks in Table 3 vary one assumption at a time, and the discussion makes clear that they are simulations and do not count as physical evidence. On the physical data we added a mechanism check: the fitted dose-independent block is about 5% of the dose term on an edge (Appendix D.1). That explains why the benchmark is favorable to the method, and it also tells the reader when it would stop being favorable.

Comment: "A high average precision does not mean the coupling magnitudes are accurate."

-> We agree. The sentence comes from our own limitations section, and we now put numbers on it. In the matched simulation the true slopes are known and every network reaches AP 1. Even so, over all ordered pairs the directed slope estimates have a mean relative L2 error of 0.19 (0.15 on edges). Their median ratio to the truth is 0.94, which fits ridge shrinkage, and their correlation with the true edge weights is only 0.64. Directed support is exactly separated on 10 of the 15 networks (Appendix D.2, Table 11). For these reasons the paper claims topology and makes no claim about coupling magnitudes. The physical benchmark only gives binary adjacency, so magnitudes cannot be checked on it.

Comments on the paper

1 comment
Area_Chair_bnk1OpenReview ↗

Please join the author-reviewer discussion

Dear all reviewers,

Our author and reviewer discussion will end soon in 18 November (AOE). If you have not yet joined the discussion, now it is the time to start it. We highly appreciate your time on discussing and exchanging thoughts to keep our ICLR standards. Many thanks!

Bests,
Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.