Abstract
This paper develops a mathematical framework for the data-driven discovery of quasi-invariants in Hamiltonian–dissipative dynamical systems and studies the relation between empirical drift minimization and the emergence of approximate invariant manifolds. The central premise is that, when exact analytic integrals of motion are unavailable or inaccessible, one may instead search over a class of smooth observables and identify functions whose drift under the perturbed dynamics is asymptotically small. In this setting, we formulate a rigorous program that connects the output of a learning algorithm to genuine dynamical structure: if an observable is selected by minimizing a suitable empirical drift functional and its measured drift is of order , then the observable behaves as a quasi-invariant of order , and its level sets approximate invariant manifolds over timescales that grow polynomially or exponentially with respect to the perturbation regime, depending on the ambient stability theory. The objective is therefore not merely predictive accuracy, but the recovery of dynamically meaningful quantities. The significance of this perspective lies in its synthesis of classical perturbation theory, geometric mechanics, and computational learning. Traditional Hamiltonian methods, including normal form theory, KAM theory, Nekhoroshev estimates, and adiabatic invariance, establish the persistence of approximate constants of motion and long-lived geometric structures under small perturbations. However, in realistic systems with dissipation, stochastic forcing, nonlinear couplings, or high dimensionality, explicit analytic construction of such invariants is often prohibitively difficult. A data-driven method based on drift minimization offers a complementary route: it searches directly in function space for observables that are nearly conserved along trajectories, thereby revealing hidden phase-space organization even when closed-form formulas are unavailable. In this sense, learned observables are elevated from empirical surrogates to candidates for mathematically interpretable dynamical invariants. The framework is especially well motivated by applications in celestial mechanics near the triangular Lagrange points and . In Trojan dust clouds, as well as in spacecraft trajectory design near libration regions, the underlying dynamics are shaped by a Hamiltonian backbone together with weak nonconservative effects such as Poynting–Robertson drag, solar radiation pressure, thermal recoil, collisions, control inputs, and stochastic perturbations. These effects destroy exact conservation laws while preserving long-lived coherent structure. We argue that observables learned from simulation or observational data can detect this residual structure by identifying near-invariant level sets that organize slow drift, transport channels, clustering regions, and metastable motion. The conjectural principle articulated here is that empirical drift minimization is not just a numerical heuristic; under suitable regularity, approximation, and sampling assumptions, it should converge toward a true quasi-invariant structure associated with the perturbed flow. This paper is intended as the first stage of a larger theorem-building program. The present version provides the conceptual and mathematical architecture for such a result by clarifying the relevant function spaces, the drift operator, the empirical approximation procedure, and the geometric meaning of low-drift observables. It also identifies the principal analytic challenges of the full proof, including approximation error, sampling error, stability of learned level sets, and the transfer from empirical smallness of drift to long-time control along trajectories. The broader claim is that data-driven discovery, when properly integrated with perturbative and geometric analysis, can become a legitimate method for uncovering invariant structure in complex dynamical systems. In this way, this paper situates machine learning not as a replacement for analytic mechanics, but as a mathematically disciplined partner in the study of long-time evolution, metastability, and phase-space geometry.