Explicit burn-in and step-size constants

Determine explicit values for the burn-in duration and the step-size cutoff required by the two-phase online SGD results for the two-unit bias-free ReLU regression model with weight decay.

Background

The main theorem establishes the existence of a source burn-in time and a sufficiently small step-size threshold under which the warm-SGD metastability result holds. These quantities are represented abstractly by constants such as the burn-in parameter and the step-size cutoff.

The paper explicitly leaves their numerical values unresolved; the simulations illustrate the mechanism but do not determine these proof-dependent constants.

References

Explicit values of the burn-in $L$ and the step-size cutoff $\eta_0$ are left open; the simulations test the mechanism, not these constants.

— Rare Gate Disagreements Can Limit Plasticity: When Gradient Flow Mispredicts Finite-Batch SGD  (2610.11475 - Zhao et al., 8 Oct 2026) in Section 6, Conclusion and Future Work, paragraph “Scope and extensions”