Trace-Free Hodge Feature Mixer
Implementation & benchmark of arXiv:2608.04237 — The trace-free Beurling--Ahlfors transform and the Bourgain--Brezis problem for Hodge systems
Source paper: The trace-free Beurling--Ahlfors transform and the Bourgain--Brezis problem for Hodge systems arXiv:2608.04237 ⓘ · analyzed Aug 31, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Build a parameter-free spectral channel mixer whose channels are arranged as components of an l-form and whose multiplier is the trace-free Beurling--Ahlfors transform. At every nonzero spatial frequency it mixes the exact and coexact channel subspaces with opposite signs, preventing a uniform channel-direction bias and preserving a structured cancellation property. Insert it as a residual branch before a convolution, MLP, or attention block, with one learned scalar gate controlling its strength.
Formulas
Mathematical statement
The paper defines the trace-free Beurling--Ahlfors transform on l-forms in R^n as S=((n-l)/n)P-(l/n)P^{\perp}, where P is the Hodge projection onto the exact l-form subspace, P^{\perp}=I-P is the complementary coexact projection, n is the spatial dimension, and l is the form degree with 1<=l<=n-1. In Fourier coordinates, for nonzero frequency \xi, use P(\xi)=(\xi\wedge\iota_{\xi})/|\xi|^2 and P^{\perp}(\xi)=I-P(\xi), where \xi\wedge is exterior multiplication and \iota_{\xi} is contraction by \xi. The coefficients are trace-free because dim(range P)=C(n-1,l-1), dim(range P^{\perp})=C(n-1,l), and ((n-l)/n)C(n-1,l-1)-(l/n)C(n-1,l)=0. The construction is an order-zero Fourier multiplier, so it does not change spatial frequency magnitudes; it only applies opposite signed gains to complementary geometric channel subspaces. The paper further states that its scalar multipliers have oddness under suitable orthogonal reflections, motivating an optional antisymmetric-kernel regularizer.
Implementation notes
1. Integration point: use a 2D or 3D convolutional feature map X with C(n,l) channel groups interpreted as l-form coefficients; for the cheapest MVP choose n=2,l=1, so every spatial location has two channels and S(\xi) is a 2x2 matrix. Insert the mixer as a residual branch after normalization and before the next convolution or MLP. 2. Pseudocode: compute the FFT X_hat=FFT(X); for every nonzero frequency xi, construct the wedge/contraction matrices W=xi wedge and C=contraction by xi on the chosen l-form basis, set P=W@C/(||xi||^2+epsilon), Pperp=I-P, S=((n-l)/n)*P-(l/n)*Pperp, then Y=X+alpha*IFFT(S@X_hat). Set the zero-frequency multiplier to zero or to the identity residual path. 3. The projector formula and trace-free coefficients come directly from the mathematics; epsilon, alpha, normalization placement, and channel grouping are engineering choices. To test the reflection-cancellation claim, optionally add L_ref=lambda_ref*mean ||K-R(K)||^2 for learned spatial kernels K and reflection R, or compare the fixed S mixer against a random trace-zero mixer. 4. First experiment: train a small CIFAR-10 ResNet with 32/64 channels, comparing the baseline, a learned 1x1 mixer with matched parameters, and this fixed Hodge mixer at equal FLOPs. Measure early loss descent, gradient norm variance, final accuracy, and robustness to input translations/reflections. A positive result is lower gradient variance or faster loss reduction at equal accuracy, with ablations over alpha and the zero-frequency treatment.
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.