BB Method Fails Superlinear Convergence on Some Quadratic Problems

Dawei Li, Xiaotian Jiang, Mingyi Hong · July 23, 2026

The Barzilai-Borwein (BB) method is a popular optimization algorithm known for its practical speed, but its theoretical convergence behavior has been puzzling. A key open question was whether it achieves superlinear convergence (faster than linear) for almost every strictly convex quadratic problem and starting point. This paper answers that question with a definitive "no." For dimensions 4 and higher, the authors construct an open set (with positive measure) of quadratic problems and initial points where the BB method converges, but only linearly—not superlinearly. They prove that the gradient norm and error energy are bounded between two geometric sequences, ruling out superlinear speed-up. The proof relies on a computer-assisted analysis of a seven-cycle in the dynamics of the BB method in 4D. This matters because it shows that even for simple quadratic problems, the BB method can be slower than expected, which has implications for understanding its behavior in machine learning and large-scale optimization. Practically, it suggests that users should not rely on superlinear convergence guarantees for BB, and may need to consider alternative methods or preconditioning when faster convergence is critical.