Federated optimization usually assumes clients share a common model, but many problems have a more complex structure. In federated minimax—or Stackelberg—settings such as distributionally robust optimization and adversarial training, there is an inner maximization variable that is derived from each client's own local data. This variable is not a single global quantity; it is specific to the client.
The challenge is that these inner variables can differ not only in value but also in meaning and dimension from one client to the next. That breaks the typical assumption of a shared parameter space, making standard federated minimax methods difficult to apply directly.
The preprint on arXiv introduces a method called Personalized Projected Ascent to handle this heterogeneity. The name suggests each client performs an ascent step that is projected onto its own local structure, rather than forcing all clients into one common formulation. Because the abstract is truncated, the article does not provide convergence guarantees or experimental details, so those remain open questions from this source alone.