Finding Hidden Spillover Networks with Gentle Randomized Nudges

When people, devices, or cells influence one another, treating one of them changes the outcomes of others. Statisticians call this interference, or spillover. Almost every method that estimates such effects assumes the spillover runs along a network we already know, such as a friendship graph or a distance-based radio map. That assumption is often wrong, and when it is, the estimates are badly biased. The closest existing remedy only works around the unknown network to estimate treatment effects. It never says who actually influences whom. Yet many experimenters are already in a position to recover that network directly. They run repeated randomized interventions, and these are almost never "hard" overrides that clamp a unit to a fixed value. They are soft interventions: changes to a unit's own behavior that leave its responsiveness to neighbors intact. Examples are a nudge that raises someone's baseline activity, a drug that dampens a cell's variability, or a transmit-power dither that scales a radio's signal.

This note proposes to treat repeated soft interventions as a measurement device for the hidden spillover graph. The method has two stages. A convex, closed-form stage reads the graph off how the outcomes' means and precision matrix (inverse covariance) change across randomized intervention regimes. A non-convex nonlinear model is then fitted starting from that estimate. The bet is that this ordering matters. It is the same lesson graph matching learned when convex relaxations proved to pick the right basin while only the non-convex objective holds the right answer. Here it is applied to cyclic spillover models, where mutual influence makes standard acyclicity tools useless and random-start fitting is likely to land in mirrored or diluted solutions.

Core Claim

Model a network of $n$ units observed over many rounds $t$. At equilibrium each unit's outcome depends on its neighbors' outcomes, its own soft intervention, and its own noise:

$$X_t = B X_t + u_t + \varepsilon_t ,$$

Here $X_t \in \mathbb{R}^n$ is the vector of outcomes, $B$ is the unknown spillover matrix (zero diagonal, $B_{ij}$ is the influence of unit $j$ on unit $i$, cycles allowed), $u_t$ is the experimenter's randomized mean shift, and $\varepsilon_t$ is independent noise with diagonal variance $D$. If the feedback is contractive so that $I-B$ is invertible, the outcome precision matrix decomposes as a sum of one rank-one term per unit:

$$\Theta = (I-B)^\top D^{-1}(I-B) = \sum_{j} d_j^{-1} r_j r_j^\top ,$$

where $r_j$ is the $j$-th row of $I-B$ written as a column and $d_j$ is unit $j$'s noise variance. A soft scale intervention multiplies unit $j$'s noise by $s$. Its effect on the precision is exactly $(s^{-2}-1)\,d_j^{-1} r_j r_j^\top$. That is a rank-one change whose leading eigenvector is unit $j$'s full set of incoming spillover weights, with sign and scale fixed by $r_{jj}=1$. This holds even when the graph has cycles.

Randomized mean shifts give a complementary moment. Each unit's own shift acts as an instrument for its outcome, which identifies the rows of $B$ by instrumental-variable regression. Both moments are linear in the data and can be combined in a sparse convex program. The claim has two parts. First, randomized soft interventions with known targets identify the interference graph in this linear cyclic model, with a convex estimator. Second, when spillover is nonlinear (saturation, thresholds, the logarithmic rate–interference relation in wireless), this estimate is the right starting point for a non-convex equilibrium model. Initialized there, the nonlinear fit recovers edges that random-start continuous optimization systematically misses.

The rank-one identity is the cyclic, known-target cousin of the result in Scalable Intervention Target Estimation in Linear Models. That paper shows precision differences between observational and interventional regimes are sparse and localized, and uses them to find unknown targets in linear DAGs. We supply that method's core object but invert its use. The targets are known because we randomize them, and we read edges, not targets, off the difference, in graphs that are cyclic by nature.

Research Program

The first strand is identification and estimation in the linear model. We will state conditions under which combined shift and scale designs identify $B$, and derive how the number of rounds $T$ needed scales with maximum in-degree. Density-dependent degradation is a known pain point for interference bounds, as Causal Inference under Networked Interference and Intervention Policy Enhancement reports for estimation and regret. We expect per-row sample cost to grow with in-degree rather than with $n$, because each scale intervention isolates one row. The convex estimator stacks the instrumental-variable moments with the eigen-structure of regime-wise precision differences under an $\ell_1$ penalty. We will also study design: which units to scale or shift per round, and by how much, given a budget.

The second strand is the nonlinear model and its landscape. We replace $BX$ with a contractive residual map in the style of SCOUT: Cyclic Causal Discovery Under Soft Interventions with Unknown Targets. SCOUT fits cyclic equilibrium models with spectral normalization and spline-flow noise, optimizing relaxed adjacency and target matrices from random initialization. We keep its model class but fix the targets to the known randomized design, and replace random initialization with the convex estimate. The analogy we test comes from Graph Matching: Relax at Your Own Risk. That paper shows the convex relaxation's optimum is an uninformative interior point, yet using it to start the non-convex Frank–Wolfe solver lands in the correct basin.

We will measure basins directly. For many random starts, we record which stationary point is reached and how far it lies in edge space from the truth. We will characterize the failure modes we expect: transposed edges when spillover is near-symmetric, and edge mass smeared over two-hop paths. As an optional middle step, we will compare warm-starting against Gaussian-smoothing continuation, which Effiecient Graph Matching by Adjacency Matrix Based Model and Gaussian Smoothing Based Optimization derived in closed form for polynomial matching objectives. For our likelihood it would have to be estimated by sampling.

We deliberately avoid acyclicity penalties. Spillover graphs are cyclic. The DAG-specific pathologies that On the Convergence of Continuous Constrained Optimization for Structure Learning documents therefore do not arise: vanishing constraint gradients, and penalty weights pushed to $10^{14}$. What remains is the ordinary non-convexity of the equilibrium likelihood, which is the part this proposal targets.

The third strand is downstream use. The learned graph replaces the proxy network inside effect estimators and inside interference-aware optimizers. This tests whether recovering the graph buys anything beyond working around it.

Baselines

The naive reference point uses the observed proxy network as the interference graph. Concretely, this is the friendship graph in social data, or distance-thresholded links in wireless. All edge-recovery metrics are computed against the true latent graph, and downstream estimators run with this graph plugged in. Beating it should be easy whenever proxy and truth differ. It anchors how much the problem matters.

The strongest structure-learning competitor is SCOUT, run as a cyclic soft-intervention learner on round-by-unit outcome matrices. Unmodified, it would face unknown targets while we use known ones, which would be unfair. We therefore run two versions. One uses its native unknown-target mode. The other has the target matrix clamped to the true design, which differs from our method only in initialization. Both get the same data, the same nonlinear model capacity, and the same tuning budget, measured in hyperparameter trials and GPU-hours. On its synthetic graphs SCOUT reports near-perfect recovery (AUPRC ≈ 1.0, SCOUT). That setting is single-system SCMs, not spillover networks. For interference-graph recovery itself, there is no reported result found on Lacuna. Beating SCOUT means higher edge AUPRC and lower structural Hamming distance at matched $T$, with the margin concentrated in the regimes where we predict spurious basins.

The strongest simple linear competitor is per-row sparse instrumental-variable regression, using each unit's randomized shift as the instrument. It needs no scale interventions and no nonlinear model. It runs on the same rounds, with its sparsity level tuned on the same held-out likelihood. Beating it requires a gain that grows with nonlinearity strength.

The ablation that isolates the key ingredient is our nonlinear model started at random rather than from the convex estimate, with the same number of total optimizer steps. A second ablation drops the scale-intervention moments from the convex stage to measure what precision differences contribute.

For downstream effect estimation, the competitor is CauGramer from Causal Graph Transformer for Treatment Effect Estimation Under Unknown Interference. It learns cross-attention over expanded neighborhoods instead of trusting the proxy graph, and reports 28% to 80% lower $\epsilon_{AVG}$ and $\sqrt{\epsilon_{PEHE}}$ than prior methods on BlogCatalog and Flickr. We run it as published, and also with our learned graph replacing its anchoring adjacency. Beating it means lower $\sqrt{\epsilon_{PEHE}}$ at equal training budget.

For the wireless setting, the reference is UWMMSE from Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation. It reports average sum-rate 83.21 versus classical WMMSE's 82.94 on Rayleigh ad-hoc topologies given full channel knowledge. Here the target is not to beat it but to approach it while using only measured outcomes under power dithering.

Why not just…?

Why not just use the observed network? Cai et al. show this misspecification is costly: decoupling from the proxy graph cut effect-estimation error by 28% to 80% on BlogCatalog and Flickr (CauGramer). Our semi-synthetic experiments vary the overlap between proxy and true graph from identical to disjoint. That shows where the proxy is good enough. We expect it to win only near full overlap, and we will report that regime honestly.

Why not just estimate effects without recovering the graph, as CauGramer does? For effect estimation alone this may well be sufficient, and it is a real threat on that metric. But CauGramer still needs an initial adjacency to bound its receptive field. It also cannot answer the questions experimenters care about: whom to target, or which links to cut. The decisive test feeds our learned graph into CauGramer's anchor. If error does not drop relative to its default anchor, then graph recovery adds nothing for estimation, and the contribution narrows to structure discovery and design.

Why not just use the linear instrumental-variable estimate and stop? This is the most serious alternative. In a truly linear world it is consistent, convex, and cheap. The non-convex stage earns its place only if nonlinear spillover biases the linear estimate enough to matter. We sweep a single nonlinearity knob, the saturation strength in the spillover function, from zero upward. If the linear estimator matches the full method across the realistic range, we will say so. The contribution then reduces to the identification result and the design study, which is still useful but smaller.

Why not hard interventions, clamping a unit to a fixed value? In network experiments you can rarely override a person's or radio's behavior completely. Doing so also removes that unit's incoming edges from the data, which destroys exactly the row the scale intervention reveals. Equivalence-class results show soft interventions lose no identifiability relative to hard ones (Characterizing and Learning Equivalence Classes of Causal DAGs under Interventions). Targeted Reduction of Causal Models argues that overriding variables breaks feedback equilibria. We include a hard-knockout design arm at equal budget to quantify the difference directly.

Proposed Experiments

The synthetic suite uses Erdős–Rényi and scale-free spillover graphs with cycles, following the generators used across Stable Differentiable Causal Discovery and SCOUT. It varies $n$ from 20 to 1,000, mean in-degree, nonlinearity strength, noise family (Gaussian, Gumbel, exponential), and rounds $T$. Metrics are edge AUPRC, structural Hamming distance, sign accuracy, and the fraction of random restarts reaching the truth's basin.

The semi-synthetic suite follows the CauGramer protocol on BlogCatalog and Flickr, and the peer-effect setup on the Wave1 Add Health network from Jiang and Sun. In both, the true interference graph is generated to differ from the observed one. Soft interventions are applied over repeated rounds, and both graph recovery and downstream $\epsilon_{AVG}$ and $\sqrt{\epsilon_{PEHE}}$ are scored.

The wireless suite uses Rayleigh-fading ad-hoc topologies as in UWMMSE. Each transmitter's small power dithers act as scale interventions, and only per-link rates are observed. We learn the effective interference graph, hand it to UWMMSE, and compare sum-rate to the full-channel-knowledge figure of 83.21.

Three checks are decisive. First, the warm-started nonlinear model must beat both its random-start ablation and clamped-target SCOUT on edge AUPRC in the high-nonlinearity regime. Basin analysis must show those competitors failing for the predicted reasons, transposed or smeared edges, rather than from undertraining. Second, the gap over sparse instrumental-variable regression must grow monotonically with nonlinearity. Flat curves would falsify the case for the non-convex stage. Third, the rank-one precision prediction must hold empirically in the linear regime. That means the eigenvector of each regime's precision difference must match the true incoming row, with error shrinking roughly as $1/\sqrt{T}$, the usual rate for sample covariance estimates. If it fails even there, the identification argument is wrong and the program stops.