# Certified Radical Coordinate Layer

- ID: 3092
- Canonical URL: https://synthcore.org/idea/3092/certified-radical-coordinate-layer
- API JSON: https://synthcore.org/api/idea/3092.json
- API Markdown: https://synthcore.org/api/idea/3092.md
- Verification status: unverified
- Source: [arXiv:2609.02871](https://arxiv.org/abs/2609.02871)
- Category: architecture
- Solves: stability, generalization, accuracy
- ML areas: mlp, embedding, regularization
- Math tags: approximation-theory, algebra, linear-algebra
- Ratings: usefulness 5/10; difficulty 4/10; novelty 8/10

## Idea description

Replace Fourier or sinusoidal one-dimensional coordinate features with a trainable radical layer f(x)=sum_i c_i sqrt(P_i(x)), where every P_i is a strictly positive quadratic. A nonzero scalar output formed by such a layer has at most 2n distinct real zeros, providing an explicit bound on sign changes and suppressing uncontrolled ringing. Use the radical features as an input embedding for a conventional MLP or neural implicit field.

## Mathematical statement

The paper studies f(x)=sum_{i=1}^n c_i sqrt(P_i(x)), where c_i are real coefficients and P_i are positive quadratic polynomials on the real line. Its special-case theorem states that any non-identically-zero such function has at most 2n real zeros. For the general family f_j(x)=(P_j(x))^{alpha_j}, the extracted Wronskian argument defines W_k=det[f_j^{(r-1)}]_{r,j=1}^k. When the f_j have no zeros on the interval, the paper divides by the product of the f_i and multiplies column j by P_j^{k-1}, obtaining tilde W_k=(W_k/product_{i=1}^k f_i) product_{j=1}^k P_j^{k-1}. The resulting determinant has polynomial entries of degree at most (k-1)d, so its degree and the number of possible roots are bounded by k^2 d. The implementation uses the sharper 2n result for positive quadratics. Here n is the number of radical features, x is a scalar coordinate, c_i are learned amplitudes, and positivity of P_i is enforced parametrically.

## Key formulas

- $$f(x)=\\sum_{i=1}^{n}c_i\\sqrt{P_i(x)},\\qquad P_i(x)>0\\ \text{ for all }x\\in\\mathbb R,$$
- $$\#\{x\in\mathbb R:f(x)=0\}\le 2n\qquad\text{for }f\not\equiv0,$$
- $$W_k=\det\!\begin{pmatrix}f_1&f_2&\cdots&f_k\\ f_1'&f_2'&\cdots&f_k'\\ \vdots&\vdots&\ddots&\vdots\\ f_1^{(k-1)}&f_2^{(k-1)}&\cdots&f_k^{(k-1)}\end{pmatrix},\qquad \widetilde W_k=\frac{W_k}{\prod_{i=1}^{k}f_i}\prod_{j=1}^{k}P_j^{k-1},$$
- $$P_i(x)=u_i^2x^2+v_i^2x+(w_i^2+\varepsilon),\qquad \varepsilon>0.$$

## Implementation notes

(1) Integration point: implement a PyTorch or JAX module for scalar coordinates before an MLP or neural implicit field. Given x with shape B by 1, compute n positive quadratic features r_i(x)=sqrt(P_i(x)), concatenate them with normalized x, and feed the result to the MLP. This replaces Fourier features or a SIREN embedding rather than adding another loss to an existing sinusoidal network. Parameterize each quadratic as P_i(x)=u_i^2 x^2+v_i^2 x+(w_i^2+epsilon), with epsilon=10^{-4}, so the constant term is strictly positive. (2) Pseudocode: `x=normalize(x); for i in 1..n: P=u[i]^2*x^2+v[i]^2*x+(w[i]^2+eps); r[i]=sqrt(P); h=MLP(concat(x,r)); y=head(h); loss=task_loss(y,target); backprop(loss); optimizer.step()`. Initialize u, v, and w near zero with small random perturbations and use first-layer weight variance proportional to 1/n. (3) The mathematical result supplies the deterministic oscillation budget 2n; it does not require root counting during optimization. Empirically evaluate the learned scalar output on a dense ordered grid, count sign changes while ignoring values with absolute magnitude below 10^{-5}, and measure derivative variation with finite differences. Compare the observed count with the theoretical budget when the output is directly a radical sum; for an MLP after the embedding, report the count as an empirical diagnostic rather than claiming the theorem applies to the whole MLP. (4) First cheap experiment: fit noisy one-dimensional functions on [-1,1], including y=sin(2 pi x)+0.2 sin(30 pi x), using a parameter-matched ReLU MLP, Fourier features, SIREN, and the radical embedding with n equal to 4, 8, and 16. Measure train MSE, held-out MSE, extrapolation error, output sign changes, and sensitivity to label noise. A positive result is lower held-out or extrapolation error than the parameter-matched ReLU baseline, with substantially fewer spurious sign changes than Fourier features or SIREN and stable behavior across random seeds.

## Disclaimer

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