Search papers, labs, and topics across Lattice.
This paper investigates the behavior of stability-annealed smoothed-sign descent in the context of full-batch linear classification on separable data, revealing that the normalized iterates converge to the minimizer of a convex Burg-type barrier over a margin slice. By employing entropic mirror ascent on a concave dual objective, the authors establish a precise characterization of the static barrier geometry, including KKT conditions and endpoint limits. Experimental results corroborate the theoretical findings, demonstrating the accuracy of dual identities and the predicted dynamics in practical scenarios.
Stability-annealed smoothed-sign descent reveals a hidden pathway to optimal classification that challenges conventional gradient descent assumptions.
Adaptive gradient methods can favor max-margin separators that differ from gradient descent, yet a fixed positive numerical stability constant eventually changes the update geometry again. This paper studies the rate-controlled middle case for full-batch linear classification on separable data. For memoryless stability-annealed smoothed-sign descent with weighted exponential loss, we prove that the normalized iterates converge to the minimizer of a convex Burg-type barrier over a margin slice. The proof rewrites the dynamics exactly as entropic mirror ascent on a concave dual objective, controls the dual gap by a KL recursion, and yields an explicit S_t^{-1/2} normalized-iterate envelope. The static barrier geometry is fully characterized, including KKT conditions and both endpoint limits. Experiments validate the exact dual identities to floating-point error, illustrate the predicted path and rate diagram, and show an empirical fixed-epsilon crossover scaling in cumulative time. We further report robustness and boundary diagnostics for logistic tails, fixed-epsilon crossover, and adaptive-method variants, delineating the scope of the proved smoothed-sign theory.
