lin-0189
15.2 The tangent Bellman residual
For an admitted state direction \(v\in T_sX\),
\[ \delta _T(\theta ;s,v) = D_s\delta _B(\theta ;s)[v]. \]
This asks whether the two Bellman routes respond coherently to a local change of state representation. It is not the parameter gradient of the scalar TD loss.
Definition
15.1
GIRL architectural interface
A GIRL realization exposes:
the base Bellman residual \(\delta _B\);
the tangent residual \(\delta _T\) on declared probes;
a cover of the transition, reward, bootstrap, and representation branches;
a decision-relevant quotient; and
an admission controller for cross-regime use.
For isotropic random probes with \(\mathbb E[vv^\top ]=I\), writing \(J_s=D_s\delta _B\) gives
\[ \mathbb E_v\lVert J_sv\rVert ^2 = \operatorname {tr}(J_s^\top J_s) = \lVert J_s\rVert _F^2. \]
JVPs therefore estimate the squared tangent residual without materializing a full Jacobian.