CoCD: A Deterministic ZO Optimizer with Gradient Coherence
A new paper on arXiv proposes Coherent Coordinate Descent (CoCD), a deterministic zeroth-order optimization method that achieves O(1) query complexity per step by leveraging historical gradients through implicit landscape smoothing. The method is theoretically equivalent to Block Cyclic Coordinate Descent with warm starts, converting stale gradients into a computational asset. Error bounds reveal that large gradient coherence can improve convergence.
Key facts
- CoCD is a deterministic ZO optimizer
- It achieves O(1) query complexity per step
- CoCD is equivalent to BCCD with warm starts
- Historical gradients are used as a computational asset
- Error bounds show large coherence improves convergence
- The paper is on arXiv with ID 2605.14373
- It addresses sample inefficiency and high variance in ZO methods
- The method is budget-aware
Entities
Institutions
- arXiv