# Pareto-Sparse Fractional Dynamics Layer

- ID: 2213
- Canonical URL: https://synthcore.org/idea/2213/pareto-sparse-fractional-dynamics-layer
- API JSON: https://synthcore.org/api/idea/2213.json
- API Markdown: https://synthcore.org/api/idea/2213.md
- Verification status: unverified
- Source: [arXiv:2608.12879](https://arxiv.org/abs/2608.12879)
- Category: regularization
- Solves: accuracy, generalization, sample-efficiency
- ML areas: world-model, regularization, training, optimizer
- Math tags: optimization, sparse-regression, fractional-analysis, numerical-analysis
- Ratings: usefulness 6/10; difficulty 6/10; novelty 6/10

## Idea description

Add a sparse, interpretable fractional-dynamics layer to a neural world model: candidate terms are evaluated through weak projections, while both their support and continuous derivative orders are selected by validation error versus model complexity. This avoids forcing the model to choose from a dense fixed dictionary containing many nearly collinear fractional orders.

## Mathematical statement

Let S be a subset of candidate weak features, where each feature is identified by a term type k and a continuous derivative order q_k in a branch-specific interval such as (0,1), the exact integer mode q_k=1, or (1,2). For fixed orders q, form the weak design matrix Phi(q), whose entry Phi_{jk}(q) is the integral of test function phi_j times feature G_k(u;q_k) over the space-time domain Q. Fit active coefficients c_S by ridge regression: the squared weak prediction error plus lambda times the squared coefficient norm. Evaluate every support/order pair by validation error and support size, retain nondominated solutions, and select the validation-error/complexity elbow. The branch distinction matters because q=1 is an exact integer operator rather than a value obtained by smoothly crossing between fractional definitions.

## Key formulas

- $$\hat c_S(q)=\arg\min_{c\in\mathbb{R}^{|S|}}\left\|y_w-\Phi_S(q)c\right\|_2^2+\lambda\|c\|_2^2$$
- $$\Phi_{jk}(q)=\int_Q\phi_j(t,x)\,G_k\!\left(u(t,x);q_k\right)\,dt\,dx$$
- $$\mathcal{P}=\left\{\big(E_{\mathrm{val}}(S,q),|S|\big): (S,q)\text{ is not dominated by any }(S',q')\right\}$$
- $$q_k\in(0,1)\ \text{or}\ q_k=1\ \text{or}\ q_k\in(1,2),\qquad \text{choose the validation-error/complexity elbow on }\mathcal{P}$$

## Implementation notes

(1) Integration point: use this as a residual or dynamics head on top of an encoder h_t for noisy trajectories. The encoder may be a GRU, Transformer, or MLP; the new head predicts a weak target y_w and represents the update as a sparse combination of fractional features. For a simpler MVP, use a neural field u_theta(t,x) and learn only the coefficients and orders jointly with theta. (2) Pseudocode: choose K candidate term types and J test functions; initialize a population of branch-labelled genomes g=(S,q), where S is a binary support mask and q contains one order per active feature. For each genome, compute the weak design matrix Phi_S(q), solve c=(Phi_S^T Phi_S+lambda I)^(-1)Phi_S^T y_w, evaluate training and validation error, and evolve genomes with differential-evolution mutation and crossover while never mutating q across incompatible branches. Collect nondominated (validation error, |S|) solutions and select the elbow. Freeze the selected support and orders, then fine-tune the encoder and coefficients by backpropagation. (3) Compute Phi by quadrature using adjoint fractional kernels; estimate c by the closed-form ridge solve; search only the orders q and support S. Normalize every feature column before regression and report unnormalized physical coefficients afterward. (4) First experiment: use noisy one-dimensional Burgers data with 5-20 trajectories, comparing a dense fixed-order library, L1 regression, and branch-aware Pareto search. Evaluate support precision and recall, recovered orders, rollout error, and parameter count. Success is recovery of the correct two or three terms at noise levels where dense-library regression selects spurious terms, together with lower long-horizon rollout error at equal encoder size.

## Disclaimer

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