# Born-Structured Bilinear Neural Operator

- ID: 2505
- Canonical URL: https://synthcore.org/idea/2505/born-structured-bilinear-neural-operator
- API JSON: https://synthcore.org/api/idea/2505.json
- API Markdown: https://synthcore.org/api/idea/2505.md
- Verification status: unverified
- Source: [arXiv:2608.18262](https://arxiv.org/abs/2608.18262)
- Category: architecture
- Solves: accuracy, memory, stability
- ML areas: mlp, training, regularization
- Math tags: functional-analysis, operator-theory, approximation-theory, linear-algebra
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Build each nonlinear correction in an inverse neural operator from explicit bilinear products of learned operator features, following the inverse Born expansion instead of using an unconstrained pointwise MLP. Use a square activation to implement multiplication exactly, and truncate the interaction order so the model has a controllable polynomial structure.

## Mathematical statement

The paper represents an operator-expansion term as $$\sum_{m,n=1}^{N_{3}}\widetilde{C}_{mn}\langle\eta_{1}e^{\mathbf{i}k\theta\cdot\bullet},\Psi_{n}\rangle_{X}\langle\eta_{2}e^{-\mathbf{i}k\hat{x}\cdot\bullet},\Psi_{m}\rangle_{X}.$$ Here $X$ is the spatial or function domain, $\eta_{1}$ and $\eta_{2}$ are input functions or feature fields, $k$ is a frequency, $    heta$ and $\hat{x}$ are directions, $\Psi_{n}$ and $\Psi_{m}$ are basis functions, $\langle\cdot,\cdot\rangle_{X}$ is the inner product on $X$, $N_{3}$ is the truncation dimension, and $\widetilde{C}_{mn}$ are expansion coefficients. The paper states that this sum can be represented by a fixed-small-depth network applied to the four quantities $\eta_{2}(y),e^{\mathbf{i}k    heta\cdot y},\eta_{1}(z),e^{-\mathbf{i}k\hat{x}\cdot z}$, using a square activation $\sigma_{2}$. The key algebraic identity is exact multiplication: $$ts=\frac{(t+s)^{2}-(t-s)^{2}}{4}.$$ Thus a two-branch square-activation module can compute products without approximating multiplication through a wide generic MLP. In the adapted neural-operator block, $a_{n}=\langle\eta_{1},\Psi_{n}\rangle_{X}$ and $b_{m}=\langle\eta_{2},\Psi_{m}\rangle_{X}$ are projected feature coefficients, and the learned bilinear correction is $q=\sum_{m,n}\widetilde{C}_{mn}b_{m}a_{n}$; complex inputs can be handled by splitting real and imaginary parts.

## Key formulas

- $$\sum_{m,n=1}^{N_{3}}\widetilde{C}_{mn}\langle\eta_{1}e^{\mathbf{i}k\theta\cdot\bullet},\Psi_{n}\rangle_{X}\langle\eta_{2}e^{-\mathbf{i}k\hat{x}\cdot\bullet},\Psi_{m}\rangle_{X}$$
- $$ts=\frac{(t+s)^{2}-(t-s)^{2}}{4}$$
- $$q(y)=\sum_{m=1}^{N_{3}}\sum_{n=1}^{N_{3}}\widetilde{C}_{mn}\,b_{m}(y)a_{n}(y),\qquad a_{n}(y)=\langle\eta_{1}(y,\cdot),\Psi_{n}\rangle_{X},\quad b_{m}(y)=\langle\eta_{2}(y,\cdot),\Psi_{m}\rangle_{X}$$
- $$u_{r+1}=u_{r}+\mathcal{L}_{r}(u_{r})+\mathcal{B}_{r}(u_{r},u_{r}),\qquad \mathcal{B}_{r}(u,v)=\Phi_{r}\!\left[\operatorname{vec}\!\left(\sum_{m,n}\widetilde{C}^{(r)}_{mn}b_{m}(u)a_{n}(v)\right)\right]$$

## Implementation notes

1. Integration point: take a small Fourier neural operator or convolutional operator used for an inverse problem, and replace its pointwise channel-mixing MLP after each spectral convolution with a Born block. The block receives a feature field $u\in\mathbb{R}^{B\times C\times H\times W}$ and computes $N_{3}$ learned basis projections $a_n(u)$ and $b_m(u)$ using two $1\times1$ convolutions or low-rank Fourier projections. 2. Pseudocode: `a = proj_a(u)` and `b = proj_b(u)` with shapes `[B,N3,H,W]`; form pairwise sums and differences across the two feature-index axes; compute `prod = ((a[:,None] + b[None,:])**2 - (a[:,None] - b[None,:])**2) / 4`; contract `q = einsum('mn,bmnhw->bchw', C, prod)`; apply an output projection and residual update `u_next = u + linear(u) + q`. For a quadratic self-interaction, use two independently projected copies of $u$; for a cross-interaction, use the current iterate and the original measurement embedding. 3. The paper-derived components are the bilinear coefficient contraction and exact square-activation multiplication. Estimate the basis projections, coefficients $C$, residual linear operators, and truncation dimension $N_{3}$ by gradient descent. Use real-valued channels for real data, or represent each complex feature as concatenated real and imaginary channels and implement complex multiplication with four real bilinear products. 4. First experiment: train on a $64\times64$ synthetic 2-D inverse-scattering or Poisson inverse-coefficient dataset with additive Gaussian noise, comparing a standard FNO of equal width and parameter count against one or four Born blocks. Measure reconstruction MSE, error versus noise level, parameter count, and peak activation memory. The expected signal is equal-or-better reconstruction at lower parameter count, smoother degradation as noise increases, and faster loss descent at equal FLOPs. Ablate the bilinear block, the square identity, and the truncation dimension $N_{3}$.

## Disclaimer

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