Optimal Routing for Federated Learning over Dynamic Satellite Networks: Tractable or Not?

The core question here is not whether federated learning works in space, but whether finding the optimal communication route through a constantly shifting satellite topology is even computationally solvable in reasonable time. That tractability question has direct consequences for whether in-orbit FL can ever run autonomously without pre-computed routing tables.

This connects loosely to the shortest-path generalization paper from mid-April, which used routing as a controlled test bed for ML reasoning and found that scaling path length broke model reliability. That work was about LLM problem-solving, not distributed training infrastructure, but both papers are circling the same underlying difficulty: graph-based routing problems get hard fast as scale increases. The satellite FL paper approaches this from the optimization side rather than the learning side, asking whether the problem class itself admits efficient solutions. Recent Modelwire coverage has otherwise focused on optimizer benchmarking and attention efficiency, neither of which connects meaningfully here.

If the authors or a follow-up group publish a polynomial-time approximation algorithm or a hardness proof (NP-hardness would be the significant result) within the next six months, that will determine whether practical in-orbit FL scheduling requires heuristics by necessity or whether exact methods remain on the table.