Method

$$$$MPNNLayer(X ; A) == sigma(A * X * W), where sigma(is a non -

linearity)

Method

$$$$ISAB(X) == MultiheadAttention(X, MultiheadAttention(I, X))

Method

$$$$p_theta(PAP**top) == p_theta(A), (for all permutation matrices)

P

Method

$$$$MPNNLayer(X ; A) == sigma(A * X * W), where sigma(is a non -

linearity)

Method

$$$$ISAB(X) == MultiheadAttention(X, MultiheadAttention(I, X))

Method

$$$$p_theta(PAP**top) == p_theta(A), (for all permutation matrices)

P

Method

$$$$MPNNLayer(X ; A) == sigma(A * X * W), where sigma(is a non -

linearity)

Method

$$$$ISAB(X) == MultiheadAttention(X, MultiheadAttention(I, X))

Method

$$$$p_theta(PAP**top) == p_theta(A), (for all permutation matrices)

P

Method

$$$$MPNNLayer(X ; A) == sigma(A * X * W), where sigma(is a non -

linearity)

Method

$$$$ISAB(X) == MultiheadAttention(X, MultiheadAttention(I, X))

Method

$$$$p_theta(PAP**top) == p_theta(A), (for all permutation matrices)

P