Work

The Quadratic Coefficient Moment Governs a Leading Error Channel in Palindromic Compositions with an Explicit Velocity-Dependent Kick

Year: 2026
Type: preprint
Abstract: Palindromic composition of a symmetric second-order velocity-Verlet kernel is a standard technique for constructing high-order geometric integrators for separable Hamiltonian systems. When a velocity-dependent perturbation is inserted explicitly into the kick evaluations, however, the resulting elementary map is no longer self-adjoint and generically acquires a local defect of order O(τ^2), where τ is the sub-step length. We show, by a Taylor expansion followed by an explicit propagation of the stage defects through the exact flow, that the leading velocity-dependent contribution to the local error of a composition is H^2 Q d(y_0) + O(H^3), Q = \sum_{i=1}^s w_i^2, where H is the macro-step, s is the number of stages, {w_i} are the composition weights summing to one, y_0 = (q_0, p_0) denotes the initial state, and d(y_0) is the leading local defect vector evaluated at the initial state. For fixed final time T, the corresponding leading global contribution is HQ η_T(y_0) + O(H^2), where η_T is the transported defect coefficient and y_0 is the initial state. Thus Q, rather than the nominal order of the separable-kernel composition, governs a leading error channel introduced by this explicit velocity-dependent treatment. The mechanism is examined in a 28-body Solar System model containing Newtonian gravity, solar and terrestrial J_2 terms, and a Schwarzschild-type first-order post-Newtonian correction. In a one-year same-model test, a 3-minute equal-step calculation is used as a fine reference and four 30-minute compositions are compared. After accounting for the nonzero leading defect of the reference, the normalised Mercury errors of the constrained 63-stage, constrained 127-stage, and formally 10th-order 35-stage schemes agree within 0.5%. The equal-step 63-stage scheme, although it minimises Q, does not follow the same normalised relation at this time scale, indicating that uncancelled conservative and mixed next-order terms are no longer negligible. Centennial comparisons with JPL DE440 provide an independent model-consistency indication, with the equal-step scheme reducing the Mercury discrepancy from 5.64 × 10^-5 to 4.26 × 10^-6 AU. Controlled fixed-15-stage families spanning ranges of Q and S_3 show that the Mercury error scales with Q (R^2 > 0.998 for each Cartesian component) and that the error vectors remain collinear to within 0.99996. An exact linear benchmark confirms the reversal of accuracy ranking between conservative and velocity-dependent regimes, with error ratios consistent with the Q ratios.
Author: Jiaqi Shen
Language: English
Open Access status: green
OpenAlex ID: W7236259366
Cites: 18
Cited by:
Related to: 10