# Dyadic Harish-Chandra spectral layer

- ID: 2615
- Canonical URL: https://synthcore.org/idea/2615/dyadic-harish-chandra-spectral-layer
- API JSON: https://synthcore.org/api/idea/2615.json
- API Markdown: https://synthcore.org/api/idea/2615.md
- Verification status: unverified
- Source: [arXiv:2608.20313](https://arxiv.org/abs/2608.20313)
- Category: architecture
- Solves: accuracy, scalability, stability
- ML areas: graph-nn, cnn, embedding
- Math tags: representation-theory, harmonic-analysis, lie-groups, geometry
- Ratings: usefulness 6/10; difficulty 7/10; novelty 6/10

## Idea description

Construct a neural layer as a sum of equivariant spectral operators at dyadic frequency scales, with each scale represented by a smooth learnable multiplier instead of an unconstrained dense spectral table. Enforce derivative and off-diagonal decay constraints so high-frequency components cannot create arbitrarily large or spatially nonlocal responses. On a discretized homogeneous space, this gives a multiresolution equivariant alternative to a generic graph filter or convolution kernel.

## Mathematical statement

The complete-symbol construction uses $(T_a f)(x)=\int_{\mathbb{R}^n}a(x,\xi)\hat f(\xi)e^{ix\cdot\xi}\,d\xi$, where $a(x,\xi)$ is a smooth position-frequency symbol and $\hat f$ is the Fourier transform. The paper decomposes a symbol with a dyadic partition as $m=\sum_{j=0}^{\infty}m_j$, where $m_j$ is localized near frequency scale $2^j$. Lemma 5.2.10 gives $|(\partial^\alpha\ell_j)(H)|\le C_{\alpha,M}|H|^{-M}2^{j(a+r-M+|\alpha|)}$ for $H\ne0$. Here $\ell_j$ is the inverse-transform kernel at scale $j$, $\alpha$ is a multi-index, $M\ge0$ is a chosen decay order, $a$ is the dimension of the Abelian spectral variable $\mathfrak a$, $r$ is the rank of the symmetric space, and $C_{\alpha,M}$ is independent of $j$. The Abel transform is a support-preserving isomorphism $\mathcal A:C_c^\infty(K\backslash G/K)\to C_c^\infty(\mathfrak a)^W$, so radial equivariant kernels can be parameterized by Weyl-invariant functions. Adapt this with spectral bands $B_j(\lambda)$ and Weyl-invariant learned multipliers $s_j(\lambda)$: $y=\sum_j\mathcal F_{G/K}^{-1}[s_j(\lambda)B_j(\lambda)\mathcal F_{G/K}f(\lambda)]$.

## Key formulas

- $$(T_a f)(x)=\int_{\mathbb{R}^n}a(x,\xi)\,\hat f(\xi)\,e^{ix\cdot\xi}\,d\xi.$$
- $$m=\sum_{j=0}^{\infty}m_j.$$
- $$\left|\left(\partial^{\alpha}\ell_j\right)(H)\right|\leq C_{\alpha,M}\,|H|^{-M}\,2^{j(a+r-M+|\alpha|)},\qquad H\neq0.$$
- $$y=\sum_{j=0}^{J}\mathcal F_{G/K}^{-1}\!\left[s_j(\lambda)B_j(\lambda)\mathcal F_{G/K}f(\lambda)\right],\qquad s_j(w\lambda)=s_j(\lambda)\ \forall w\in W.$$

## Implementation notes

Integrate the layer between an input feature tensor on a discretized homogeneous space and a pointwise channel mixer. Use features $f\in\mathbb R^{N\times C}$, a precomputed approximation to the symmetric-space transform $\mathcal F_{G/K}$, spectral coordinates $\lambda_q$, and dyadic bands $B_j(\lambda_q)$. Begin with a compact quotient or finite homogeneous graph where graph-Laplacian eigenvectors approximate spherical eigenfunctions. For every scale $j=0,\ldots,J$, parameterize $s_j(\lambda_q)$ with a small MLP of invariant coordinates, such as Casimir eigenvalues or $\|\lambda_q\|$. Weyl invariance can be imposed by tying values across Weyl-related coordinates, or by using a radial multiplier $s_j(\lambda)=g_j(\|\lambda\|)$. Compute $z=\mathcal Ff$, then set $z_j(q)=s_j(\lambda_q)B_j(\lambda_q)z(q)$, sum over $j$, apply $y=\mathcal F^{-1}(\sum_j z_j)$, and finish with a channel projection, residual connection, and nonlinearity. Add a smoothness penalty $L_{\mathrm{sym}}=\sum_j\sum_q|\Delta_\lambda s_j(\lambda_q)|^2$, where $\Delta_\lambda$ is a finite-difference spectral derivative. Add a scale penalty $L_{\mathrm{scale}}=\sum_j2^{-2j(a+r-M)}\|s_j\|_2^2$ to discourage violation of the paper's dyadic bound. Use known $a$ and $r$ for the chosen space, and start with $M=2$ or $3$. Estimate transform quality by reconstructing random signals and estimate equivariance error using known group actions. The first experiment should use a small homogeneous graph or spherical rotation benchmark, comparing this layer with a standard graph convolution and an unconstrained spectral filter at equal parameter count. Record validation accuracy, equivariance error $\|L(T_gf)-T_g(Lf)\|/\|Lf\|$, peak memory, and runtime. Success means lower equivariance error with equal or better accuracy, and graceful accuracy retention when the number of dyadic bands is reduced.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.