Canonical quotient-feature Fusion #
Quotient the semantic vector space by the linear span of every free input and named constant. The quotient map is then a canonical linear feature that annihilates all zero-cost data. Addition cannot create a new quotient direction, while each multiplication can create at most one.
Consequently, the dimension of the requested outputs modulo free data is a multiplication lower bound. No hand-designed feature coordinates are needed.
Linear span of all semantic values available without a multiplication or addition gate: free problem inputs and named scalar constants.
Equations
- Algebraic.Fusion.Arithmetic.Interaction.Linear.Quotient.freeSubmodule K constant problem = Submodule.span K (Set.range problem.inputs ∪ Set.range constant)
Instances For
Every free input belongs to the free-data submodule.
Every named constant belongs to the free-data submodule.
Canonical interaction certificate obtained from quotienting by all free semantic data.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Dimension of the span of requested outputs modulo free inputs and named constants.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Canonical quotient-output rank lower-bounds multiplication cost.
Linear independence modulo free data forces one multiplication per requested output.
Canonical quotient-output rank lower-bounds total nonconstant gate cost.
Canonical quotient-output rank lower-bounds raw circuit size.