Fixed-policy linear evaluation
Sutton and colleagues’ ICML 2009 work studies . Ordinary off-policy TD with approximation can be unstable; residual-gradient approaches have distinct objective and sampling difficulties. GTD2/TDC use MSPBE and an auxiliary vector for linear-time updates.
Projected error
With and , appropriately weighted expectations give
Explicit off-policy ratios are omitted for readability here. This is not an uncorrected arbitrary-replay recipe. Coverage, invertibility, and expectation definitions belong to the assumptions.
Auxiliary and primary recursions
Track through
Common simplified GTD2 and TDC directions are respectively and . With accurate tracking, their expectations relate to the same MSPBE gradient; finite-step recursions differ.
The auxiliary recursion avoids per-step matrix inversion and a naive product of correlated single-sample estimates. Timescale and stability assumptions remain important.
Scope of the guarantees
The notes emphasize linear complexity and off-policy stability, but the guarantees concern fixed-policy linear evaluation. They are not universal global-convergence results for neural networks, changing target policies, or control. Runtime comparisons are configuration-specific.
Reading and a proposed check
Start with Bellman errors and projection geometry, then stochastic approximation. A Baird-style example could compare TD, residual gradients, and GTD2/TDC with explicit sampling and step sizes. No such experiment was rerun here.