Research Interests
My research focuses on the theoretical underpinnings of federated and decentralized learning, with a particular emphasis on developing and analyzing optimization algorithms. I am interested in understanding the convergence properties and generalization capabilities of these methods under realistic and challenging conditions, such as non-convex loss landscapes and data heterogeneity.
Publications
Online Learning with Non-convex Losses: New Condition to Achieve Small Dynamic Regret
IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Hyderabad, India, 2025 DOI
We study online learning with non-convex, time-varying loss functions. We show that Online Gradient Descent (OGD) can achieve dynamic regret comparable to the strongly convex case, under a novel condition defined within a ball of radius p around the initialization. This condition incorporates gradients, losses, and temporal variation to reflect the online nature of the problem. Under this assumption, OGD attains dynamic regret scaling sub-linearly with variation. Experimental results support our theory
Generalization of FedAvg Under Constrained Polyak-Łojasiewicz Type Conditions: A Single Hidden Layer Neural Network Analysis
Submitted
We study the FedAvg algorithm in Federated Learning, focusing on its optimization and generalization performance. Under new constrained conditions, we show that FedAvg converges linearly and applies to certain neural networks with sufficient width. We also prove that its generalization error improves optimally with more data and benefits further from having more clients.
Localized Growth Conditions for Decentralized FedAvg: Convergence to Global Optimal Points
Submitted
We prove the first linear convergence guarantee for Decentralized Federated Averaging (D-FedAvg) under a local PL-type growth condition, far weaker than standard global assumptions. This setting requires neither data homogeneity nor global minimizers—these properties emerge naturally as iterates stay in a bounded region. Our analysis introduces new drift-dynamics bounds to control consensus error, and experiments on diverse datasets confirm the theory
A PL-type Framework for Dynamic Regret in Non-Convex, Non-Smooth Online Composite Optimization
Submitted
Online learning plays a central role in addressing non-stationary data in signal processing and machine learning. In this work, we study non-convex online composite optimization with non-smooth regularization and analyze the Online Proximal Gradient (OPG) method. We introduce a new proximal Polyak-Lojasiewicz (PL)-type condition under which the dynamic regret scales as $\mathcal{O}(\texttt{D}_T)$, where $\texttt{D}_T$ is a measure of function variation. This result extends PL-type analysis to non-convex, non-smooth settings, marking a significant advance over prior work. Finally, we provide experiments that corroborate the theoretical guarantees.