Privacy-Budget Adjoint Mirror Descent optimizes DP noise schedules
Most DP training spreads its privacy budget uniformly across iterations, and this work proves that allocation, not just total budget, can be optimized with provable stability guarantees. Practitioners training differentially private models — from healthcare to finance teams subject to data-protection rules — can use the square-root allocation law or the PB-AMD method to squeeze more utility from the same certified privacy cost. The reduction to projected gradient descent comparisons makes the approach directly implementable.