lin-0224

18.7 The SID controller

Sheaf Infinitesimal Descent

  1. Estimate local predictive updates \(u_i\)

  2. Evaluate typed compatibility and effectivity witnesses

  3. Localize each witness to its incident cover simplices

  4. Solve for admitted tangent corrections \(a_i\)

  5. Retract \(x_i\leftarrow \operatorname {Ret}_{x_i}(u_i+a_i)\)

  6. Audit pair, triple-overlap, and global-realizability conditions

  7. Promote, damp, quarantine, collect data, or revise the cover

Method

Primitive

Structural consequence

Glue loss

scalar tradeoff

approximate pair agreement

Augmented Lagrangian

equality plus duals

solver-dependent feasibility

Projection

realized constraint

direct local feasibility

SID

descent sketch and repair lift

typed compatibility, effectivity, localization, and admission

Table 18.1 Similar linear algebra can implement different learning contracts.