Young-Projected Tensor Channels

Integrate this into the first tensor-valued hidden block of an O(n)-aware geometric network, tensor MLP, graph neural network, or geometric transformer. For every rank-two feature X with shape batch by nodes by n by n by channels, compute the symmetric projection X_s=(X+transpose(X))/2 and alternating projection X_a=(X-transpose(X))/2 along the two geometric indices. Apply independent channel-mixing maps f_s and f_a, then reconstruct Y=f_s(X_s)+f_a(X_a). Do not feed the raw tensor to the learned map in the main condition; this tests whether representation separation itself provides the benefit. For rank-three features, implement antisymmetrization by summing signed permutations and setting components with repeated indices to zero, following the wedge-product rule. A compact variant retains only symmetric and alternating channels and uses a learned scalar gate g, with Y=g f_s(X_s)+(1-g) f_a(X_a). The mathematical quantities are computed exactly; no eigenvalues or representation multiplicities need to be estimated for the rank-two test. Validate equivariance by drawing random orthogonal matrices Q and checking that F(Q X Q-transpose) is close to Q F(X) Q-transpose. Start on a small 2-D or 3-D elasticity dataset with a tensor MLP or message-passing baseline matched for parameter count. Measure validation error, equivariance error, parameter count, and activation memory. Success is lower error at equal parameters or at least 30 percent fewer tensor parameters at unchanged error, with no increase in equivariance residual.