Research
Network control and optimization
Models and algorithms that keep large networks stable, efficient and fair.
How much traffic a topology can carry, how to schedule and route under uncertainty, and how to share scarce resources among users who do not cooperate. The threads below range from directional wireless networks to game-theoretic resource sharing and market mechanisms, and to representation learning on graphs.
Sectorized wireless networks
Capacity and stability when nodes reshape the topology with directional sectors.
Sectorization lets a node activate several directional antennas at once, each covering a sector, which reshapes the network topology itself. We model a multi-hop mesh of sectorized nodes together with the interference the sectors create, and characterize its capacity region through an auxiliary graph that depends on how every node is sectorized.
Choosing the sectorization is then an optimization problem: we define the sectorization gain, which measures how far a given network flow can be scaled by sectorizing the nodes, and we give a distributed algorithm that finds a near-optimal sectorization. We also propose a class of sectorizations, even homogeneous sectorizations, that makes the auxiliary graph bipartite. Routing then becomes faster and the capacity region larger at the same time, and the more the network is sectorized, the stronger both effects. When sectors can be reconfigured over time, the network keeps this advantage while staying stable under unknown traffic.
Papers
Game theory and mechanism design
Satisfaction games and market mechanisms for sharing a network among users who do not cooperate.
Classical resource allocation assumes every user maximizes its own utility, and the outcome is a Nash equilibrium. A user rarely needs the maximum. It needs enough. Games in satisfaction form replace maximization with satisfaction constraints. We introduced efficient satisfaction equilibria for power control, designed distributed dynamics that provably reach them, extended the framework to fragile common-pool resources and to risk-aware players, and showed consistent gains over Nash outcomes.
The same lens applies to infrastructure sharing. Network slicing builds customized logical networks on shared infrastructure, where tenants with private valuations compete for the same capacity. We designed a market mechanism for slicing and characterized the competitive equilibria it induces and the decentralized dynamics that reach them.
Papers
Learning on graphs
What latent graph models can express, and generative models with identifiable communities.
Latent graph models embed nodes in a geometric space and explain edges through distances or inner products. This thread studies what such embeddings can express: spectral occupancy rather than rank as the right notion of capacity, compositional embeddings built on Aitchison geometry, and archetypal generative models whose communities are explainable and identifiable.
Papers