Shuang Li

Perfectly rank-matched tensor factorizations still suffer from quartically flat optimization valleys whenever the underlying Fourier slice ranks are nonuniform.

Entropy-regularized RL, though non-convex in the traditional sense, exhibits a Bellman-induced Polyak--Łojasiewicz geometry that guarantees global convergence of Wasserstein Policy Gradient.

EOS-Bench reveals that the complexity of satellite scheduling can be systematically quantified, unlocking new insights into algorithm performance across thousands of scenarios.