# Normalized Scheduling-Degree Truncation

- ID: 1541
- Canonical URL: https://synthcore.org/idea/1541/normalized-scheduling-degree-truncation
- API JSON: https://synthcore.org/api/idea/1541.json
- API Markdown: https://synthcore.org/api/idea/1541.md
- Verification status: unverified
- Source: [arXiv:2607.28490](https://arxiv.org/abs/2607.28490)
- Category: architecture
- Solves: scalability, accuracy, stability
- ML areas: world-model, ssm, rnn, regularization, inference-speedup
- Math tags: control-theory, dynamical-systems, linear-algebra, approximation-theory, stability
- Ratings: usefulness 6/10; difficulty 5/10; novelty 5/10

## Idea description

Use normalized scheduling variables and explicitly cap the degree of their products in a neural LPV or mixture-of-dynamics model. Instead of allowing every multiplicative interaction between scheduling coordinates and past or future features, retain only monomials below a chosen degree threshold. This produces a controllable approximation knob between a purely linear model and a full lifted predictor, while avoiding unstable extrapolation caused by poorly scaled high-degree features.

## Mathematical statement

Let p_k be the scheduling vector in R^{n_p}, and let tilde p_k be its normalized version, with componentwise mean mu_i and scale s_i. Construct a lifted feature vector containing monomials phi_alpha(tilde p)=product_i tilde p_i^{alpha_i}, where alpha is a multi-index and |alpha| is its total degree. The paper removes rows of the past-data lift Z_P and future-input lift U_F containing more than user-defined numbers h_Z and h_U of scheduling multiplications. The neural analogue uses A(tilde p)=A_0+sum_{|alpha|<=h} phi_alpha(tilde p)A_alpha and z_{k+1}=A(tilde p_k)z_k+B(tilde p_k)u_k. The number of monomials through degree h is binomial(n_p+h,h), giving a direct complexity prediction. Increasing h should reduce approximation error until omitted scheduling interactions are negligible, after which the combinatorial feature cost grows without meaningful accuracy gains.

## Key formulas

- $$\tilde p_{k,i}=\frac{p_{k,i}-\mu_i}{s_i},\qquad \phi_\alpha(\tilde p_k)=\prod_{i=1}^{n_p}\tilde p_{k,i}^{\alpha_i},\qquad |\alpha|=\sum_{i=1}^{n_p}\alpha_i.$$
- $$\widehat z_{k+1}=A_0z_k+\sum_{\alpha:\,|\alpha|\le h}A_\alpha\phi_\alpha(\tilde p_k)z_k+B(\tilde p_k)u_k.$$
- $$N_{\mathrm{monomials}}(n_p,h)=\binom{n_p+h}{h}.$$
- $$\mathcal{E}(h)=\frac{1}{N}\sum_{k=1}^{N}\|z_{k+1}-\widehat z_{k+1}^{(h)}\|_2^2.$$

## Implementation notes

1. Exact integration point: insert the truncation in a scheduling-conditioned MLP, state-space model, or latent world model immediately before the dynamics matrices are generated. Normalize scheduling channels using training-set statistics, enumerate monomials up to degree h, and use them as inputs to a linear hypernetwork or factorized bilinear dynamics layer. 2. Pseudocode: estimate mu and s on the training split; compute tilde p=(p-mu)/s; generate every multi-index alpha with |alpha|<=h; form phi_alpha; calculate A(tilde p)=A_0+sum_alpha phi_alpha A_alpha; calculate B(tilde p) similarly or with a lower degree; update z_{k+1}=A(tilde p_k)z_k+B(tilde p_k)u_k; and train with one-step and multi-step latent losses. Add spectral normalization or rescale A(tilde p) so its estimated spectral radius is below a target rho_star<1 when bounded stable rollouts are desired. 3. Computed from the paper: scheduling normalization and removal of rows or features with excessive scheduling products. Estimated empirically: degree h, coefficient regularization, and spectral-radius target. 4. First cheap experiment: use a two-dimensional Duffing or cart-pole system with one or two time-varying parameters, comparing h=0,1,2,3,4 against a full multiplicative MLP and a GRU. Measure validation one-step error, 50-step rollout error, feature count, and runtime. The prediction is that the first degree at which residual error stops falling identifies a truncation point; feature count must follow binomial(n_p+h,h). For the spectrally constrained version, the empirical latent Jacobian spectral radius should remain below rho_star and stable-trajectory rollout norms should decay approximately as rho_star^t.

## Disclaimer

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