Unverified 2026

Sparse Legendre Shape Hypernetwork

Usefulness6/10

Difficulty5/10

Novelty6/10

Architecture Sample efficiency Scalability Accuracy Mlp Graph nn World model Fine tuning Approximation theory Functional analysis Complex analysis

Source paper: Shape Holomorphy and Sparse Approximation of the Maxwell Electric Field Integral Operator arXiv:2609.00466 ⓘ · analyzed Sep 2, 2026

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 Not peer-reviewed.

Idea description

Use a sparse multivariate Legendre expansion as the geometry-to-network-weights map, rather than an unconstrained MLP that consumes all shape parameters. The hypernetwork predicts only coefficients for a selected set of polynomial multi-indices, allowing high-dimensional or countably parameterized shape uncertainty to be handled with a number of learned terms determined by coefficient decay rather than ambient dimension.

Formulas

$$W(\mathbf y)=\sum_{\nu\in\mathbb N_0^{(\mathbb N)}}c_\nu P_\nu(\mathbf y),\qquad P_\nu(\mathbf y)=\prod_{j\ge1}P_{\nu_j}(y_j),$$

$$\mathbf b\in\ell^p(\mathbb N),\quad 0<p<1\quad\Longrightarrow\quad (\|c_\nu\|)_{\nu}\in\ell^p,$$

$$\|c-c_N\|_{\ell^q}\le C_{p,q}N^{-(1/p-1/q)}\|c\|_{\ell^p},\qquad q>p,$$

$$\widehat W_N(\mathbf y)=\sum_{\nu\in\Lambda_N}\widehat c_\nu P_\nu(\mathbf y),\qquad \Lambda_N=\text{the }N\text{ indices with largest estimated }\|\widehat c_\nu\|.$$

Mathematical statement

The paper states that $\mathbf b=(b_j)\in\ell^p$, $0<p<1$, yields complex-parametric holomorphy of the pulled-back operator, current, and far field, and consequently $\ell^p$-summability of their Legendre coefficients. Let $P_\nu(\mathbf y)=\prod_jP_{\nu_j}(y_j)$ be tensor Legendre polynomials indexed by finitely supported multi-indices $\nu\in\mathbb N_0^{(\mathbb N)}$, and write a target operator or network parameter as $W(\mathbf y)=\sum_\nu c_\nu P_\nu(\mathbf y)$. The transferable consequence is $(c_\nu)\in\ell^p$. If coefficients are ordered by magnitude, the best $N$-term truncation obeys the Stechkin bound $\|c-c_N\|_{\ell^q}\le C_{p,q}N^{-(1/p-1/q)}\|c\|_{\ell^p}$ for any $q>p$. This supplies a principled sparse architecture and an empirical coefficient-selection rule.

Implementation notes

(1) Integrate this into a geometry-conditioned MLP or hypernetwork. The input is the first $K$ shape coordinates $y_j$; the output is either the prediction itself or a low-rank adapter $\Delta W(y)$ for a base neural operator. Do not learn a dense MLP over all monomials. Maintain a set $\Lambda$ of multi-indices and learn one coefficient tensor $C_\nu$ per index. (2) Pseudocode: initialize $\Lambda=\{0\}$; sample a shape batch $y_b$; evaluate normalized Legendre features $P_\nu(y_b)$; compute $\Delta W_b=\sum_{\nu\in\Lambda}C_\nu P_\nu(y_b)$; run the base network with $W_0+\Delta W_b$; backpropagate prediction loss plus $\lambda\sum_{\nu\in\Lambda}\|C_\nu\|_2^p$. Every $T$ steps, estimate validation sensitivity for candidate indices obtained by incrementing one active component, add the best candidates, and prune the smallest coefficients to keep $|\Lambda|=N$. This is the adapted $\widehat W_N$ formula. (3) The paper supplies the expected $\ell^p$ decay, not the value of $p$; estimate an effective $p$ by fitting the log-log ordered coefficient curve and compare it with validation error. Normalize Legendre polynomials and clip shape samples to valid deformations. (4) First experiment: use a 2D neural operator or MLP predicting solutions of a PDE on randomly perturbed domains with 32 shape modes. Compare a dense coordinate MLP, Fourier features, and sparse Legendre adapters at equal trainable parameter counts. Success is lower error at fixed parameter count, a validation curve close to $N^{-(1/p-1/2)}$, and no degradation when increasing available shape modes from 32 to 128.

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.