CSympNet-ID: conformal-symplectic map learning for linearly damped Hamiltonian systems

What the math gives to ML

The paper provides a constructive way to learn dissipative one-step dynamics while enforcing the exact conformal-symplectic identity at every parameter value, rather than hoping a penalty term preserves it during optimization. Its transferable asset is the scaling-conjugacy factorization: a learned symplectic transport core is surrounded by analytically parameterized damping maps, making the global contraction factor positive, interpretable, and identifiable from irregular time steps. This suggests a structured transition module for physical world models, neural state-space models, and long-horizon sequence predictors where unconstrained networks accumulate qualitatively wrong contraction or expansion.