# Prescribed-Order Equilibrium Vector Field

- ID: 2019
- Canonical URL: https://synthcore.org/idea/2019/prescribed-order-equilibrium-vector-field
- API JSON: https://synthcore.org/api/idea/2019.json
- API Markdown: https://synthcore.org/api/idea/2019.md
- Verification status: unverified
- Source: [arXiv:2608.07351](https://arxiv.org/abs/2608.07351)
- Category: dynamics
- Solves: stability, accuracy, generalization
- ML areas: training-dynamics, world-model, rnn
- Math tags: differential-geometry, algebra, dynamical-systems
- Ratings: usefulness 5/10; difficulty 5/10; novelty 7/10

## Idea description

Constrain a neural vector field to vanish to order at least k at a designated anchor state c. The network predicts smooth coefficient functions, while a fixed degree-k monomial gate supplies the required vanishing behavior. This exactly enforces the equilibrium and suppresses all local drift terms below order k, potentially improving stability and extrapolation near known rest states.

## Mathematical statement

Let V = R^n with coordinates x_1,...,x_n, and let I_0 be the ideal of smooth functions vanishing at the origin. For integer k >= 1, the paper studies F_k = I_0^k X_V, where X_V is the module of smooth vector fields. For a multi-index alpha = (alpha_1,...,alpha_n), define |alpha| = sum_i alpha_i and x^alpha = product_i x_i^(alpha_i). The iterated Hadamard lemma states that I_0^k is generated by all monomials x^alpha with |alpha| = k. Therefore every vector field in F_k can be written as f(x) = sum_{|alpha|=k} x^alpha g_alpha(x), where each g_alpha(x) is an R^n-valued smooth function. Because every monomial has total degree k, f(x) = O(||x||^k) near the origin, and all derivatives of total order less than k vanish there. For an anchor c, substitute z = x - c. The neural adaptation is f_theta(x) = sum_{|alpha|=k} z^alpha g_alpha,theta(z). The Lie-bracket structure is also preserved: products of order-k coefficients and their derivatives produce terms of order at least 2k-1, which is at least k, so the constrained vector fields remain closed under brackets.

## Key formulas

- $$\mathcal{F}_{k}=\mathcal{I}_{0}^{k}\mathfrak{X}_{V}$$
- $$|\alpha|=\alpha_{1}+\cdots+\alpha_{n},\qquad x^{\alpha}=x_{1}^{\alpha_{1}}\cdots x_{n}^{\alpha_{n}}$$
- $$\mathcal{I}_{0}^{k}=\left\langle x^{\alpha}:|\alpha|=k\right\rangle_{C^{\infty}_{V}},\qquad \mathcal{F}_{k}=\left\langle x^{\alpha}\frac{\partial}{\partial x_j}:|\alpha|=k,\ 1\le j\le n\right\rangle_{C^{\infty}_{V}}$$
- $$f_{\theta}(x)=\sum_{|\alpha|=k}(x-c)^{\alpha}g_{\alpha,\theta}(x-c),\qquad f_{\theta}(x)=O(\|x-c\|^{k})\ \text{as }x\to c$$

## Implementation notes

(1) Integrate the construction into the drift function of a Neural ODE, latent world model, or residual recurrent block. Choose a known equilibrium or anchor state c and compute z = x - c. Enumerate A_k = {alpha in N^n : |alpha| = k}. The coefficient network outputs one n-dimensional vector g_alpha(z) for each alpha. The final vector field is the sum of z^alpha times these coefficient vectors. For k >= 1 this guarantees f(c) = 0; for k >= 2 it also guarantees vanishing Jacobian at c, and generally all derivatives of total order below k vanish. (2) Pseudocode: `z = x - c`; `coeff = MLP(z)`; split `coeff` into vectors `g[a]`; set `f = zeros_like(x)`; for each `a` in `A_k`, compute `monomial = product(z[i] ** a[i])` and update `f += monomial * g[a]`; return `f`. Use `x_next = x + dt * f(x)` in a residual block or pass f to an ODE solver. (3) The mathematical part is the fixed monomial basis and the exact vanishing-order guarantee. Estimate the empirical local order by sampling unit vectors u and fitting the slope of log(||f(c + r u)||) versus log(r) over decreasing radii r; the slope should approach k. For large n or k, use a factorized polynomial feature map or a sampled subset of multi-indices to control the O(n choose k) basis size. (4) First experiment: train a small recurrent latent dynamics model or Neural ODE on damped-oscillator trajectories with the known rest state as c. Compare an unconstrained MLP against k = 1, 2, and 3 versions at matched hidden width and approximately matched FLOPs. Measure one-step prediction error, long-horizon rollout error, drift from c under zero-input integration, Jacobian norm near c, and gradient stability. A successful result is lower long-horizon error or much smaller equilibrium drift at equal training loss; a failure is a substantial parameter and compute increase without improved local or rollout behavior.

## Disclaimer

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