Combining conditional message weights #
A normalized conditional kernel extends a weight on an earlier tuple. The kernels may depend on earlier coordinates, and positivity is needed only on realized inputs.
Relabelling a finite message space preserves a weight certificate.
Equations
Instances For
A message on a subsingleton output space has zero cost.
Equations
- Algebraic.Cutwidth.Aggregate.Geometry.Entropy.WeightBound.subsingleton key = { weight := fun (x : Y) => 1, nonneg := ⋯, mass := ⋯, positive := ⋯, log_bound := ⋯ }
Instances For
A parent-independent probability weight is a conditional kernel.
Equations
- bound.toConditional parent = { weight := fun (x : Z) => bound.weight, nonneg := ⋯, mass := ⋯, positive := ⋯, log_bound := ⋯ }
Instances For
A kernel depending on selected coordinates can be read from a larger parent key.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Multiplying an earlier weight by a normalized conditional kernel adds their costs.
Equations
Instances For
A deterministic readout cannot increase the certified message cost.
Equations
Instances For
A constant message has a zero-cost point-mass certificate.
Equations
- Algebraic.Cutwidth.Aggregate.Geometry.Entropy.WeightBound.const value = (Algebraic.Cutwidth.Aggregate.Geometry.Entropy.WeightBound.subsingleton fun (x : X) => ()).map fun (x : Unit) => value
Instances For
An arbitrary Boolean message has the uniform one-bit certificate.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Product weights add costs even when the two actual messages are dependent.
Equations
- first.prod second = (second.pairConditional (first.toConditional right)).mono ⋯
Instances For
Coordinatewise probability weights add their costs without an independence premise.
Equations
- One or more equations did not get rendered due to their size.