Summary (LLM read my papers; human bias-correction applied)

My work sits at the interface of stochastic/nonconvex optimization and statistical inference. On one hand, I use modern optimization tools—especially stochastic and primal–dual methods—to compute statistical estimands that are naturally defined by optimization (e.g., minimax/robust objectives), and I derive sharp convergence rates that describe how fast these procedures reach reliable solutions. On the other hand, I use core statistical ideas—identifiability, stability, robustness, and careful modeling of noise/heterogeneity—to make difficult optimization problems easier: in particular, the statistical structure often rules out “bad” solutions and turns seemingly intractable nonconvex landscapes into ones where simple algorithms can provably find the right answer, enabling global (or near-global) convergence guarantees rather than only local stationarity. Finally, because many optimization-defined targets are nonsmooth or involve boundary/selection effects, the resulting estimators can exhibit nonstandard (non-normal) asymptotics; a key contribution of my work is to characterize these limits and provide principled uncertainty quantification so that the outputs of large-scale optimization pipelines come with interpretable, trustworthy statistical guarantees.