CFR without Unbiasedness: Debit-Budgeted Public Feedback
The paper develops a deterministic target-transfer theorem for public feedback where coverage is partial. It proves that this setup allows for schedules to converge when balancing different regret matching strategies (RM and RM+). The core finding is a trace-level certificate that numerically bounds exploitability, linking delivered regret, discrepancy, and strategy path length to ensure predictable performance.