Self-attention recast as a many-body Dyson equation: fractal nodal surfaces survive where mean-field pooling erases them
Synopsis
A self-attention layer is rewritten as a real-space Dyson iteration whose softmax kernel acts as a two-body self-energy, while DeepSets symmetric sum-pooling corresponds to a Hartree mean field. Fitting a fractal backflow determinant, DeepSets blurs the nodal surface into a smooth envelope, whereas self-attention recovers the nested fractal pockets. At a close probe separation the exact off-diagonal velocity Jacobian norm is 123, self-attention gives 125, and DeepSets saturates near 8.
Interpretation
Symmetric sum-pooling and self-attention are placed inside one two-body Dyson equation: the former is a Hartree mean field with uniform 1/N weights, the latter a state-dependent nonlocal two-body self-energy. DeepSets-style pooling had reached sub-meV accuracy on interatomic potentials yet struggled with correlated wavefunctions and excited states, and that gap lacked a unified physical explanation. On a fractal backflow determinant, DeepSets collapses the nodal surface into a smooth mean-field envelope, while a PsiFormer-inspired self-attention layer restores both the nested fractal pockets and the dipolar FTLE rosettes.
Self-attention normalizes each particle's query-key inner products through a softmax, producing a state-dependent two-body kernel that weights the value vectors it passes. DeepSets averages every particle's value vector globally, so all particles share one background potential and the off-diagonal Jacobian decays as 1/N. In an N=6, two-dimensional, open-boundary thermal Slater determinant, at probe-ring separation rij=0.05 the exact off-diagonal velocity Jacobian norm is 123, the self-attention fit is 125, and DeepSets saturates near 8 for rij<0.8.
Perspective
The demonstrated setting is an N=6, two-dimensional, open-boundary thermal Slater determinant targeting a strongly fractal backflow determinant, with both models sharing identical encoders and decoders and the attention variant using H=4 heads with dk=32. The text extends the argument to two-index tight-binding and Kohn–Sham Hamiltonian matrix elements and to multi-state Hamiltonians for excited-state crossings as a next direction, not as a validated result.
The fractal backflow wavefunction is itself a candidate form proposed from scaling arguments, with no underlying Hamiltonian constructed. The quantitative comparison is confined to a single N=6 open-boundary setting, and self-attention gives 0.021 for a distant spectator where the exact value is 0.070. The O(N²) cost of dense self-attention is interpreted as the irreducible price of unscreened pairwise interactions.
