Counterexample Construction

Preparation

Evidence quality

Highlighted claims

• The paper constructs a finite-state counterexample where local-clock differential TD remains stable but global-clock differential TD is unstable. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
• The smallest explicit instance has 25 states. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
• The explicit counterexample uses m + 2 states with m greater than 22. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
• In the counterexample, A_eta is stable exactly for eta in two disjoint intervals. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
• The constructed P_pi is row-stochastic, irreducible, and aperiodic. — On the Divergence of Differential Temporal Difference Learning without Local Clocks
• The counterexample establishes existence and mechanism rather than frequency in typical benchmarks. — On the Divergence of Differential Temporal Difference Learning without Local Clocks