{ "corollary_3_4": { "dual": "sum_k ||M_k||_nuclear", "muon_norm": "max_k ||W_k||_spectral" }, "corollary_3_5": { "hybrid_norm": "max{||W||_msp,||u||_infinity}" }, "definition_5_1": { "approximate_sd": "N(t)=integral nu diverges; norm(theta)/N bounded; liminf alignment r(t)>=1" }, "theorem_3_1": { "learning_rate": "integral_0^infinity eta(t) dt=infinity", "soft_margin": "varphi^{-1}(log(1/L(theta)))/||theta||^L" }, "theorem_3_2": { "conclusion": "normalized-direction limit points are directions of KKT points for the same norm" }, "theorem_3_6": { "adam": "epsilon=0, c1>=c2 (beta1<=beta2), decaying nonsummable eta", "margin_norm": "l_infinity" } }