Source-linked AI summary
Edge Accuracy Is Not Enough: Why Dynamics-Learned Structure Fails to Transfer to Inverse Problems
Nicholas Tan Jerome, Fangnian Wang
TL;DR
Inverse problems often have scarce or costly labels, motivating the reuse of relational structure learned from abundant forward simulations. This paper analyzes and tests that transfer strategy, including a Jaccard-based transferability heuristic. Dynamics-learned structure fails to transfer across the evaluated domains despite meeting an edge-error condition, and the heuristic separates transfer outcomes in the four evaluated pairs.
Problem
Inverse-task labels are often expensive or dangerous to collect, motivating the question of whether structure learned from cheaper forward simulations can constrain inverse models.
Source: §1 · p. 1Method
The paper proves an edge-error robustness condition, tests dynamics-learned structure across domains and estimators, and evaluates a Jaccard test against a partially observed target-task graph.
Source: §1 · p. 2, §5.4 · p. 7Results
Takeaways & Limitations
The edge-accuracy condition controls estimation error but not approximation error when forward and inverse tasks have different optimal structures.
Source: §1 · p. 2, §7 · p. 7Both evaluated domains are simulated, so real-world sensor-deployment validation remains needed to test whether the conclusions hold outside simulation.
Source: §6 · p. 7
Abstract
from arXiv · showhide
A natural strategy for inverse problems with scarce labelled data is to transfer relational structure learned from abundant forward-simulation data. We show this strategy fails systematically, even when it satisfies the standard theoretical justification for why structure should help. We prove that approximate structure provides estimation-error benefits whenever the edge error satisfies $\Delta < n^2 - kn$, reducing sample complexity from $O(n^2)$ to $O(kn+\Delta)$. Structure learned via Neural Relational Inference (NRI) from dynamics prediction satisfies this condition, yet on a source-localisation task across 180 CFD-simulated hydrogen-leak scenarios and 180 acoustic scenarios, it degrades performance by 116% and 201% relative to a flexible, task-optimised attention baseline, while a physics-based prior (Green's function) degrades by only 69-72%. Four independent lines of evidence show this is not a tuning failure: NRI improves only 0.5% when given 18x more training data (versus 16.6% for the task-optimised baseline, $p<0.001$); performance is insensitive to the NRI edge threshold across a wide range; the dynamics-learned graph overlaps the task-optimal graph on only 6% of edges; and two further dynamics-derived structure estimators (correlation- and mutual-information-based) show no measurable benefit over a structure-free baseline, with the correlation-based estimator performing markedly worse. We formalise this gap as a statement about approximation error that the edge-accuracy condition cannot control, and we provide a lightweight transferability test (Jaccard similarity against a partially-observed target-task graph) that separates successful from failed transfer in all four domain/structure pairs we evaluate, using under an hour of computation and 15-20% of target-domain data; we present this as a heuristic calibrated on few cases, not a validated general threshold.
1 INTRODUCTION
The paper tests whether relational structure learned from abundant forward dynamics can help inverse problems with scarce labels. It finds that edge accuracy alone does not ensure transfer when the forward and inverse tasks favor different structures.
1 INTRODUCTION
- The paper evaluates transfer on source localisation, using hydrogen-leak detection as its primary case study.The setting concerns sensor placement in underground parking facilities; forward CFD simulation takes hours per scenario.Source: §1 · p. 1
- Dynamics-learned structure fails to transfer when its learning objective is misaligned with the inverse task, despite satisfying the theoretical edge-error condition.The condition controls estimation error, but not approximation error when the tasks have different optimal structures.Source: §1 · p. 2
- Theorem 3 gives an edge-error condition under which approximate structure retains sample-complexity benefits: Δ < n^2 − kn.Known sparse structure reduces sample complexity from O(n^2) to O(kn), while the approximate-structure benefit degrades gracefully under the stated condition.Source: §1 · p. 2
- Across two physically distinct domains and three dynamics-derived structure estimators, dynamics-learned structure fails to transfer; a Jaccard-overlap test separates successful from failed transfer in all four evaluated domain/structure pairs.The test compares candidate structure with a partially observed target-task graph, using a fraction of target data and under an hour of compute. The J ≥0.30 threshold is an initial calibration, not a validated general rule.Source: §1 · p. 2
2 RELATED WORK
Prior work studies physics-based and learned approaches to source localisation, and uses forward-dynamics graphs in several scientific domains. Whether such graphs transfer to inverse tasks with different objectives remains undercharacterised.
2 RELATED WORK
- Physics-based source localisation combines a forward transport or dispersion model with sparse observations, while learned approaches map sensor readings to locations.The paper compares these traditions through a Green’s Function baseline and a Learned Attention baseline, respectively.Source: §2 · p. 2
- Prior scientific-ML work learns interaction graphs from forward dynamics, but transfer conditions are rarely tested on downstream tasks with different objectives.Earlier studies often assume transfer succeeds or evaluate on tasks similar to the structure-learning objective.Source: §2 · p. 2
- NRI learns latent interaction graphs from observed dynamics, but prior extensions do not test transfer to tasks with different objectives.The paper selects NRI as a sharp test because its edge error is low enough to satisfy Theorem 3’s condition.Source: §2 · p. 2
- The theoretical analysis applies structured-hypothesis-class Rademacher complexity bounds to attention-based localisers and examines what edge-level error controls.The cited foundations are Bartlett & Mendelson (2002), Wainwright (2019), and Mohri et al. (2018).Source: §2 · p. 2
3 THEORETICAL ANALYSIS
Structure masks reduce the estimation cost of attention-based localisation, and approximate masks retain this benefit under bounded edge error. However, edge-error guarantees do not control approximation error when learned and inverse-task structures differ.
3 THEORETICAL ANALYSIS
- Known sparsity k yields ε-optimal excess risk with m = Ω(kn/ε^2) samples, versus Ω(n^2/ε^2) without structure.The improvement is Θ(n/k).Source: §3 · p. 3
- Under the restricted-eigenvalue assumption, LASSO recovers the true k-sparse linear-dynamics graph from O(k log(n/δ)) samples, while any recovery algorithm requires Ω(k log n).Source: §3 · p. 3
- With edge error ∆, the masked class dimension is at most kn + ∆ and sample complexity is O((kn + ∆)/ε^2); structure helps when ∆ < n^2 − kn.Source: §3 · p. 3
- When the optimal localiser lies outside the structured class, the bound adds an approximation-error term it does not control; edge-count closeness does not ensure task-relevant graph closeness.Source: §3 · pp. 3–4
- The authors conjecture that approximation error can dominate under objective misalignment, but do not prove this generally and instead test alternatives empirically.Their tests address insufficient data, threshold tuning, sparsity, and architecture-specific explanations.Source: §3 · p. 4
4 METHODS
The methods compare four source localisers using shared sensor inputs and outputs. They contrast unstructured models, a physics-based prior, and an NRI-derived graph mask, with additional dynamics-derived estimators used to check architecture dependence.
4 METHODS
- All four localisers share sensor readings and positions as inputs, a 3D source-location output, and 64 query points.Source: §4 · p. 4
- The MLP is the unmasked full-class baseline; “no structure” means no sparsity mask, not restricted model capacity.Source: §4 · p. 4
- Learned Attention can express any attention pattern and remains in the unstructured class, while Green’s Function is fixed, sparse, and grounded in transport physics.Source: §4 · p. 4
- NRI-Graph thresholds edge probabilities learned from 60-timestep concentration sequences using dynamics reconstruction loss plus an edge-sparsity penalty; τ = 0.3 is selected by held-out dynamics edge F1.Source: §4 · p. 4
- NRI-Graph receives the sharpest scrutiny because it can meet the edge-error condition while potentially retaining approximation error; correlation and mutual-information estimators check architecture independence.Source: §4 · p. 4
5 EXPERIMENTS
Across CFD hydrogen dispersion and acoustic propagation, the experiments compare task-optimised, physics-based, and dynamics-derived structures, then test whether NRI’s deficit reflects tuning, scaling, or task mismatch.
5.1 STRUCTURE COMPARISON FAILS TO FAVOUR DYNAMICS-LEARNED STRUCTURE
- Learned Attention achieves the best RMSE in both domains, beating Green’s Function by 42% and NRI-Graph by 54% in CFD, with acoustic errors of 0.95 m versus 2.86 m for NRI.The same ranking is preserved across the two domains governed by different physics.Source: Table 1 · p. 5, §5.1 · p. 5
5.2 FOUR LINES OF EVIDENCE AGAINST A TUNING EXPLANATION
- With 18× more scenarios, NRI-Graph improves only 0.5%, while Learned Attention improves 16.6% (p < 0.001), consistent with approximation rather than estimation error.The data-scaling experiment uses CFD scenarios.Source: Table 2 · p. 5, §5.2 · p. 5
- Across thresholds τ=0.2–0.5, NRI-Graph remains at 11.49 ± 0.11 m, while its task-optimal-graph overlap is J≈0.06 in CFD and J≈0.19 in acoustic localisation.The learned structure therefore differs substantially from the proxy graph for the localisation task.Source: §5.2 · p. 5
- Correlation-Graph underperforms the unstructured MLP, while Mutual-Information-Graph is statistically indistinguishable from it, so neither additional dynamics-derived estimator measurably helps.The comparison is in CFD; the table reports both non-NRI baselines.Source: Table 4 · p. 6, §5.2 · p. 6
5.3 SENSOR-COUNT SCALING CONFIRMS STRUCTURE ENABLES SCALABILITY
- As sensor count rises, all three structured methods improve while the unstructured MLP degrades; NRI-Graph improves but never closes the Learned Attention gap.This supports that NRI-Graph is genuinely constrained, though its constraint is misaligned with the localisation task.Source: §5.3 · p. 6, Table 5 · p. 7
5.4 A PRACTICAL TRANSFERABILITY TEST
- A Jaccard test using 15–20% of target-task data takes under an hour and separates all four evaluated pairs at J≥0.30, but the threshold is calibrated on few cases.The authors present it as a practical heuristic, not a proven guarantee.Source: Table 3 · p. 6, §5.4 · p. 7
6 DISCUSSION
Forward dynamics and inverse localisation optimize for different relationships: temporal co-evolution versus spatial constraints on an unseen source. The discussion recommends choosing structure according to its relevance to the inverse task and the available data.
Why forward and inverse objectives diverge
- Temporal coherence and spatial constraint are different objectives, and the forward-dynamics objective does not reward graphs for helping localise an unseen source.Green’s Function has the best non-learned transfer and out-of-distribution robustness, including a −0.7% acoustic OOD gap, because its structure is task-relevant by construction.Source: §6 · p. 7
Practical guidance
- Use physics-based structure when domain knowledge directly constrains the inverse task and data is limited; use flexible task-specific learning when more than 10 scenarios per sensor are available and accuracy is paramount.The suggested Jaccard threshold J ≥ 0.30 is calibrated on physics-based and NRI-learned structure and does not straightforwardly extend to other dynamics-derived estimators.Source: §6 · p. 7
Limitations
- Both evaluated domains are simulated, so real-world validation with sensor deployments remains the most direct test beyond simulation.The theorems cover attention-based localisers; extension to other masked architectures is argued qualitatively but not formally proven.Source: §6 · p. 7
7 CONCLUSION
Approximate structure can reduce sample complexity, but edge accuracy alone does not ensure that a graph learned from forward dynamics transfers to inverse localisation. The conclusion reports converging evidence for this gap and a Jaccard-based diagnostic calibrated on four domain/structure pairs.
7 CONCLUSION
- Approximate relational structure reduces sample complexity when its edge error satisfies Theorem 3’s condition, but that condition does not ensure task-objective alignment.The conclusion expects this gap to recur when forward and inverse objectives are not explicitly designed to share optimal structure.Source: §7 · pp. 7–8
- Across two physically distinct domains, four lines of evidence link failed transfer to a gap between edge accuracy and task alignment.The Jaccard-based test correctly predicts transfer for the four calibrated domain/structure pairs, but does not straightforwardly extend to two additional dynamics-derived estimators.Source: §7 · pp. 7–8
A.1 PROOF OF THEOREM 1
- The structured hypothesis class has effective dimension O(kn), since attention weights are supported on the graph’s edges.The proof uses bounded-loss uniform-convergence arguments and the non-positive middle term from the empirical-risk-minimiser decomposition.Source: §A.1 · pp. 9–10
A.2 PROOF OF THEOREM 2
Theorem 2’s proof gives a sample lower bound for identifying sparse graphs and an LASSO-based recovery guarantee under a restricted-eigenvalue assumption.
A.2 PROOF OF THEOREM 2
- For k-sparse graphs, the proof bounds the hypothesis count below by (n/k)^(2k), yielding a necessary sample count of Ω(k log n) under bounded signal-to-noise ratio.The lower bound follows from Fano’s inequality because each sample carries O(1) bits.Source: §A.2 · p. 10
- The proof models each node’s dynamics as a sparse linear regression with design matrix X_t and a k-sparse coefficient vector.The coefficient support corresponds to that node’s graph neighbours.Source: §A.2 · p. 10
- Under the restricted-eigenvalue assumption with constant κ, LASSO recovers each node’s coefficient support with probability at least 1 − δ/n.The stated sample requirement is m ≥ Ck log(n/δ)/κ².Source: §A.2 · p. 10
A.3 PROOF OF THEOREM 3
Theorem 3 bounds the dimension of a localiser using an estimated graph by kn + Δ, preserving an estimation-error benefit when Δ < n² − kn. In the linear-dynamics case described, exact graph recovery gives an end-to-end sample complexity of O(kn/ε²).
A.3 PROOF OF THEOREM 3
- An estimated graph with edge error Δ has at most kn + Δ edges, bounding the structured hypothesis class dimension by kn + Δ.The bound counts true edges and spurious edges, which are a subset of the symmetric difference.Source: §A.3 · p. 10
- Structure learning improves over the unstructured O(n²)-dimensional baseline precisely when Δ < n² − kn.The proof obtains the corresponding excess-risk bound by substituting the dimension bound into the Rademacher-complexity argument.Source: §A.3 · p. 10
- For linear dynamics where LASSO achieves Δ = 0 using O(k log n) dynamics samples, the end-to-end requirement is O(kn/ε²), below O(n²/ε²) for sparse graphs.The total combines dynamics-learning samples with the localisation requirement.Source: §A.3 · p. 10
B.1 LOCALISER ARCHITECTURES
This appendix specifies the localiser architectures and structure estimators, reports their out-of-distribution behavior, and documents the simulated domains and data protocol.
B.1 LOCALISER ARCHITECTURES
- All localisers map sensor readings and positions to source coordinates; the MLP concatenates inputs, attention variants aggregate sensor encodings, and NRI-Graph restricts attention using learned edges.Green’s Function uses sensor–query distance, while Learned Attention uses learned query, key, and value projections.Source: §B.1 · pp. 10–11
B.2 CORRELATION- AND MUTUAL-INFORMATION-BASED STRUCTURE ESTIMATORS
- The best Correlation-Graph setting has RMSE 18.61 ± 0.01, while the best Mutual-Information-Graph setting has RMSE 14.48 ± 0.00.Both estimators use the same NRI encoder inputs and downstream localiser architecture, isolating structure estimation as the varying factor.Source: §B.2 · p. 11
B.3 OUT-OF-DISTRIBUTION GENERALISATION
- Learned Attention generalises best in CFD, whereas Green’s Function generalises best acoustically; NRI-Graph’s small negative OOD gaps do not offset its poor absolute performance.The evaluation trains on selected operating conditions and tests on held-out conditions in both domains.Source: §B.3 · pp. 11–12
C.1 CFD DOMAIN
- The CFD domain contains 180 garage leak scenarios spanning leak positions, leak rates, and ventilation levels, simulated with buoyancy-driven hydrogen transport.The garage includes two longitudinal and two transverse lanes, structural columns, and three ventilation rates.Source: §C.1 · p. 12
C.2 ACOUSTIC DOMAIN
- The acoustic domain contains 180 source scenarios in a 2D room with obstacles, governed by an advection-diffusion-reaction equation.Scenarios vary source position, power, and absorption coefficient.Source: §C.2 · p. 12
C.3 TRAIN/VALIDATION/TEST PROTOCOL
- The pooled 180 scenarios use stratified 80/10/10 splits redrawn across 10 seeds; scaling experiments vary training data from 10 to 180 scenarios.For each scaling level, results average 10 random training-set draws.Source: §C.3 · p. 12
C.4 THRESHOLD AND CONNECTIVITY ABLATION
The ablation finds NRI-Graph performance largely insensitive to threshold choice, while denser top-k graphs worsen performance.
C.4 THRESHOLD AND CONNECTIVITY ABLATION
- Performance is flat across edge thresholds τ ∈ [0.2, 0.5] and degrades under denser top-k connectivity, indicating that which edges are retained, not how many, is the binding constraint.Source: §C.4 · p. 13, Table 8 · p. 13
- The F1-optimal threshold τ = 0.1 differs from the plateau by less than 0.1% relative.The ablation summary reports the same broad plateau across τ ∈[0.2, 0.5].Source: §C.4 · p. 13, Table 8 · p. 13
- The results indicate that which edges are retained, rather than graph density alone, is the binding constraint.This pattern is consistent with an approximation-error rather than estimation-error explanation.Source: §C.4 · p. 13