Point Feature Descriptor via Directional Partition of Unity on Maps

What the math gives to ML

The paper turns directional feature aggregation into a smooth analysis operator: environment or feature signals are integrated against nonnegative softmax weights forming a partition of unity. Its transferable asset is not softmax itself, which is standard, but the functional-analytic treatment of the resulting weight family: the Gram matrix quantifies redundancy, the pseudoinverse reconstructs the orthogonal projection onto the represented directional subspace, and the induced seminorm measures how much directional information is retained. A neural implementation should therefore replace hard angular pooling with learnable smooth directional atoms and explicitly use Gram whitening or projected-energy preservation to prevent collapsed or redundant angular channels.