Unverified 2026

Hard-Rod Symmetry Invariant Module

Implementation & benchmark of arXiv:2607.18872 — Absence of hidden analytic conserved quantities in harmonically confined rods

Usefulness6/10

Difficulty5/10

Novelty6/10

Architecture Stability Generalization World model Rnn Ssm Regularization Dynamical systems Linear algebra Statistics

Source paper: Absence of hidden analytic conserved quantities in harmonically confined rods arXiv:2607.18872 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Construct a scalar feature or critic for oscillator-based neural dynamics that is invariant under the transformations imposed by free harmonic motion and elastic collisions. For finite-size rods, the module should represent only quantities compatible with common oscillator-phase rotations and momentum permutations, preventing a learned world model from inventing coordinate-dependent pseudo-conserved quantities that disappear after collisions.

Formulas

$$R_{\theta}:(x_i,p_i)\mapsto(x_i\cos\theta+p_i\sin\theta,\;p_i\cos\theta-x_i\sin\theta),\qquad c_{\phi}(R_{\theta}z)=c_{\phi}(z).$$

$$c_{\phi}(x,\Pi p)=c_{\phi}(x,p)\quad\forall\Pi\in S_N,\qquad E=\frac{1}{2}\sum_{i=1}^{N}(x_i^2+p_i^2).$$

$$E_{\mathrm{cm}}=\frac{1}{2}\left[\frac{(\sum_i p_i)^2}{N}+N\left(\frac{\sum_i x_i}{N}\right)^2\right],\qquad c_{\phi}(x,p)=g_{\phi}(E,E_{\mathrm{cm}}).$$

Mathematical statement

For N equal-mass rods with m = omega = 1, free motion rotates every phase-space pair by a common harmonic phase theta: x_i' = x_i cos(theta) + p_i sin(theta) and p_i' = p_i cos(theta) - x_i sin(theta). Elastic collisions exchange momenta, p' = Pi p, where Pi is a permutation matrix and x is unchanged at the collision instant. Constrain the network scalar c_phi(x,p) to satisfy c_phi(R_theta(x,p)) = c_phi(x,p) and c_phi(x,Pi p) = c_phi(x,p). The known invariants are total energy E = one half times the sum of x_i squared plus p_i squared, and center-of-mass energy E_cm = one half times the sum of the squared center-of-mass momentum per particle and N times the squared center-of-mass position. The paper's finite-rod result predicts that every analytic conserved scalar has the form c_phi(x,p) = g(E,E_cm), so there should be no independent conserved scalar direction beyond these two variables.

Implementation notes

Integration point: use this module as the scalar invariant head of a learned Hamiltonian or collision world model, an SSM state monitor, or a recurrent predictor. Let the model receive z = (x_1,...,x_N,p_1,...,p_N) and output a scalar c_phi(z). The simplest implementation is Monte Carlo group averaging over harmonic phases and momentum permutations. Pseudocode: values = []; for theta in sampled_phases: for perm in sampled_permutations: p_perm = p[perm]; x_rot = x*cos(theta) + p_perm*sin(theta); p_rot = p_perm*cos(theta) - x*sin(theta); values.append(mlp_phi(concat(x_rot,p_rot))); c = mean(values). Add a temporal conservation loss lambda_cons times (c(z_next)-c(z_current))**2 when training on simulated trajectories. Separately train a small probe g_psi(E,E_cm) to predict c from the two known energies, and report the held-out residual rather than using the probe as the main training objective. The paper supplies the symmetry transformations and the finite-rod completeness claim; empirical quantities are group-invariance errors, temporal conservation error, and probe residual. First experiment: simulate four trapped rods with unequal nonzero lengths and exact elastic collisions. Compare a vanilla MLP critic, phase-averaged critic, and phase-plus-permutation-averaged critic. Measure the maximum and RMS values of c(R_theta z)-c(z) and c(x,Pi p)-c(x,p), including collision states. The quantitative prediction is that sampled group-average symmetry error decreases approximately as O(K_theta^(-1/2) + K_Pi^(-1/2)) with K_theta phases and K_Pi permutations, while the best held-out predictor based only on E and E_cm should explain nearly all variation in a successfully learned conserved scalar. A persistent residual indicates either model misspecification, finite-step numerical errors, or a falsification of the claimed transfer.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.