Fractional Ellipsoidal Sparse Attention
Source paper: Ellipsoidal Positivity Sets for Fractional Obstacle Problems with Quadratic Forcing arXiv:2609.00703 ⓘ · analyzed Sep 2, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace dense attention weights with a compactly supported anisotropic bump derived from the obstacle solution, using one learnable ellipsoid per attention head or feature group. Tokens outside the learned ellipsoid receive exactly zero weight, while tokens inside receive smoothly decaying weights according to a fractional exponent. The learned positive-definite matrix represents orientation, scale, and correlations between feature dimensions.
Formulas
Mathematical statement
Theorem 1.1 states that the decaying viscosity solution of the fractional obstacle problem min{u, (-Delta)^s u - (c - <Ax,x>)} = 0 is u(x) = K max{1 - <Bx,x>, 0}^{1+s}, where 0<s<1, c>0, A is symmetric positive definite, B is symmetric positive definite, and K>0. Here <Bx,x> means x^T B x. Its positivity set is the ellipsoid {x : x^T B x < 1}. The paper defines the fractional Laplacian by (-Delta)^s v(x) = c_(n,s) P.V. integral over R^n of (v(x)-v(z))/|x-z|^(n+2s) dz. For attention, use displacement d_ij = q_i - k_j, quadratic radius r_ij = d_ij^T B_h d_ij, and kernel weight w_ij = K_h (1-r_ij)_+^(1+s_h). Enforce positive definiteness with B_h = L_h L_h^T + epsilon I, where L_h is learned, epsilon>0, and s_h is fixed in (0,1) or parameterized as s_h = 0.05 + 0.90 sigmoid(a_h). Normalize with alpha_ij = w_ij divided by the sum of weights for query i plus delta.
Implementation notes
(1) Integration point: replace or augment the score-and-softmax operation in a multi-head self-attention layer. For head h, let q_i, k_j, and v_j be the query, key, and value vectors in R^{d_h}. Compute d_ij=q_i-k_j. Give each head a learned lower-triangular matrix L_h and define B_h=L_h L_h^T+epsilon I with epsilon=1e-4. Use a_h to define s_h=0.05+0.90 sigmoid(a_h). (2) Pseudocode: for each query i and candidate key j, compute r_ij=d_ij^T B_h d_ij; compute w_ij=K_h*max(1-r_ij,0)^(1+s_h); optionally multiply w_ij by exp(q_i^T k_j/sqrt(d_h)) if content matching is retained; normalize alpha_ij=w_ij/(sum_j w_ij+delta); output o_i=sum_j alpha_ij v_j. During an initial correctness experiment, enumerate all pairs. After learning stable ellipsoids, construct candidate pairs with an approximate range-search data structure and retain only r_ij<1. (3) Computed directly from the paper’s mathematics: compact support, the SPD quadratic metric, and the exponent 1+s_h. Estimated or tuned empirically: the eigenvalues of B_h, the active-neighbor count, K_h, and the initialization radius. Initialize B_h near an isotropic matrix so every query has neighbors, and add a mild penalty if the mean active count is below a minimum. (4) First experiment: train a 4-layer, 8-head transformer on CIFAR-10 image patches and on a synthetic task where each token depends on a randomly rotated local ellipse. Compare dense softmax attention, fixed square-window attention, and the ellipsoidal kernel. Report validation accuracy, loss at equal FLOPs, active neighbors per query, and wall-clock attention time. A successful result is at least two times fewer active pairs with no more than one percentage point accuracy loss, or better accuracy than fixed windows at the same active-neighbor budget. Ablate the learned matrix, the fractional exponent, and the content-score multiplier.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.