TorchLean API

NN.MLTheory.CROWN.Proofs.GraphIBPBasicTheorems

GraphIBPBasicTheorems #

Basic theorems about the graph-level IBP engine:

Extract the underlying scalar value at index i from a vector tensor.

Instances For

    Componentwise validity of a 1D interval box: lo ≤ hi for every coordinate.

    Instances For

      If a box contains any point, then it is componentwise valid (lo ≤ hi).

      A valid box contains its lower endpoint lo.

      Validity is preserved by (definitional) casts of the vector dimension.

      Converting a valid Box into a FlatBox preserves validity.

      Converting a valid FlatBox into a Box preserves validity.

      Validity is preserved by interval addition on FlatBox (over ).

      Validity is preserved by interval subtraction on FlatBox (over ).

      Validity is preserved by ReLU-IBP on FlatBox (over ).

      Validity of the IBP.linear transfer rule (over ).

      theorem NN.MLTheory.CROWN.Graph.Theorems.graph_ibp_linear_valid_real (id : ) (ps : ParamStore ) (Xin : FlatBox ) (hXin : Xin.Valid) :
      match ibp_linear id ps Xin with | none => True | some Bout => Bout.Valid

      Validity of the graph-level linear IBP rule (over ), if the parameters are present.

      theorem NN.MLTheory.CROWN.Graph.Theorems.graph_ibp_matmul_valid_real (id : ) (ps : ParamStore ) (Xin : FlatBox ) (hXin : Xin.Valid) :
      match ibp_matmul id ps Xin with | none => True | some Bout => Bout.Valid

      Validity of the graph-level matmul IBP rule (over ), if the parameters are present.