Influence maximization asks: which individuals should you intervene on to maximize the total effect across a network? The standard approach tracks every possible propagation path — who influences whom, through which intermediaries, at what probability. The state space is exponential in the network's size.
Causal influence maximization (arXiv:2603.11761) demands something stricter: not just correlation between interventions and outcomes, but causal identification of treatment effects. This ought to make the problem harder. Causal constraints restrict the set of admissible estimators, close off shortcuts, require adjustment for confounders.
Instead, the causal constraint simplifies the problem. Under a low-probability propagation assumption, the high-dimensional path-dependent dynamics compress into a low-dimensional exposure mapping with bounded second-order approximation error. The causal framework forces you to track only the pathways that actually transmit the intervention's effect — not every correlation in the network. And it turns out most correlations are not causal pathways. Removing them from consideration collapses the dimensionality.
The compressed problem is greedy-optimizable with approximation guarantees. Estimate shape-constrained exposure-response functions from observational data. Then apply greedy selection on the compressed objective. Both stages come with provable bounds — on estimation accuracy and on the optimization gap.
The pattern: restriction revealing hidden structure. Demanding causality should constrain the solution space. It does — but the constraint is informative, not just restrictive. It tells you which dimensions of the problem matter and which are noise. The correlational approach includes everything and drowns in dimensionality. The causal approach excludes non-causal pathways and finds the problem was always low-dimensional. The complexity was in the noise, not the signal.