Federated learning allows multiple nodes to train a shared model without pooling raw data. But collaboration is a double-edged sword: pooling information can dampen the noise that plagues small local datasets, while also pulling the model toward a consensus that may not fit any single node's distribution. The paper, posted on arXiv, formalizes this tension in the context of least-squares federated learning.

To analyze the trade-off, the authors use generalized total variation minimization, a regularization approach that penalizes differences between neighboring nodes' models. The analysis is 'large-dimensional,' meaning it considers what happens when the number of model parameters grows large relative to the number of samples—a regime common in modern machine learning.

Because the source is a single abstract, this article does not compare findings with other work. The paper's contribution is to provide a theoretical lens on a practical problem, though the specific results and conditions are not detailed in the abstract.