Centrally Extended Fourier Mode Mixer
Implementation & benchmark of arXiv:2607.17168 — On a super-Virasoro group, a semigroup of annuli, and Gauss--Berezin integral operators
Source paper: On a super-Virasoro group, a semigroup of annuli, and Gauss--Berezin integral operators arXiv:2607.17168 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Construct a neural mixing layer on Fourier or positional modes using a small set of exponentiated Virasoro generators instead of a dense mode-to-mode matrix. The generator coefficients are shared across all inputs, while the Lie bracket fixes how different mode shifts interact; an optional central channel captures the special coupling between modes whose indices sum to zero.
Formulas
Mathematical statement
The paper defines the Virasoro algebra with basis elements \(L_n\) for \(n\in\mathbb{Z}\) and central element \(\zeta\), satisfying \([L_n,L_m]=(n-m)L_{n+m}+\frac{1}{12}(n^3-n)\delta_{n+m,0}\zeta\) and \([L_n,\zeta]=0\). Here \(n,m\) are integer Fourier-mode indices, \(\delta_{n+m,0}\) is one only for opposite modes, and \(\zeta\) commutes with every generator. Use the truncated adjoint action on feature coordinates \(x_k\): \((G_nx)_{n+k}=(n-k)x_k\), together with a central-coordinate update \((G_nx)_\zeta=\frac{1}{12}(n^3-n)x_{-n}\). \(G_n\) is the finite-dimensional matrix representation of \(\operatorname{ad}(L_n)\), \(x_k\) is the feature at mode \(k\), and \(x_\zeta\) is a carried scalar channel. The neural transformation is \(x'=\exp(A)x\), where \(A=\sum_{n\in S}\theta_nG_n\), \(S\) is a small retained-mode set, and \(\theta_n\) are learned coefficients. The exact infinite-dimensional commutator is approximated after truncating modes to \(|k|\leq K\), so boundary leakage must be controlled by padding or projection.
Implementation notes
Integrate this as a replacement for the dense token-mixing or spectral-mixing matrix in a residual block. Transform an input sequence h of shape B by T by d with a real FFT along the sequence axis, retain modes k from -K through K, append one scalar central channel per feature group, and apply the same mode operator independently to each group. Choose S equal to {-s,...,s}, build sparse G_n using the rule (G_n x) at index n+k equals (n-k)x_k, and add the central row using the opposite-mode formula. In pseudocode: x = fft(h); x = pad_modes(x,K); A = 0; for n in S: A += theta[n] * G[n]; y = solve(I - 0.5*A, (I + 0.5*A) @ x); y = ifft(y); out = h + alpha*y. Use the Cayley transform for a cheap stable approximation to the exponential, or use a matrix exponential for a small MVP. The coefficients theta_n are learned; K, s, alpha, and padding are hyperparameters. Measure boundary leakage as the norm of components sent outside the retained mode range divided by the output norm. Start on synthetic periodic sequence regression and Long Range Arena ListOps with a 2-layer residual model. Compare against a dense Fourier mixer and a standard Transformer token mixer at equal parameter count. Success means lower error on sequence lengths longer than training, fewer parameters at matched error, and no exploding residual norms over 10,000 steps.
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.