A phase-space calculation closely predicted how often a simulated chaotic double pendulum flipped on average, according to a new arXiv preprint. The strongest match came in the model’s central energy range, where the root-mean-square relative difference between the calculation and the full simulation ensemble was 0.42%. Across the reported range, from 2 < E < 6 in the study’s chosen units, the difference was 1.05%.
The result is about a long-run average, not a timetable for the next flip. The study asked whether a statistical theory could estimate the mean flip rate of a chaotic double pendulum; it did not predict the exact timing of an individual event.
Turning motion into a rate
The theory treats the pendulum at a fixed energy as a microcanonical ensemble — a statistical description of all allowed states at that energy. It then estimates the flip rate from the probability flux crossing a chosen dividing surface, rather than following one chaotic trajectory for an extremely long time.
The test system was an egalitarian double pendulum: both arms had length 1, both masses were 1, and gravity was set to 1 in the chosen units. The researchers compared the flux estimates with numerical simulations based on independently sampled initial conditions. At each energy, they used 1,000 starting states and propagated each trajectory for 10,000 time units.
Individual paths were integrated in MATLAB with a variable-step ode45 solver using Lagrange’s equations. The saddle-orbit calculations used a second-order Gauss–Legendre symplectic integrator.
A good average can hide uneven motion
When the analysis was restricted to trajectories selected as chaotic, the prediction still tracked the average rate, especially in the central energy range. The prediction-quality index was 3.11% for 2.8 ≤ E < 5, compared with 5.66% across 2 < E < 6. Near the low-energy boundary, the discrepancy rose to 10.81%.
That comparison needs care because the chaotic-only sample and the full microcanonical calculation describe different ensembles. Their averages were therefore not expected to be exactly the same, and the study’s subgroup-selection procedure and phase-space fractions were not fully reported.
The trajectories were not all equally active. At E = 5.3, a small subset had flip rates approaching three times the ensemble mean. The result also shows that rates varied sharply from trajectory to trajectory toward the upper part of the tested energy range.
The definition of a flip changes the answer
The dividing surface mattered. Across the considered energy range, theoretical rates based on the upright-arm criterion differed from those based on the exact saddle-orbit criterion by 13.46% in root-mean-square terms.
The exact saddle-orbit surface also separated successive counted events more cleanly. At E = 5, its residence-time distribution — the time trajectories remained in a counted state — showed a short-time gap across 30,000 detected flips. No comparable gap appeared with the upright-arm criterion.
A simplified analytic version of the saddle-orbit criterion closely tracked the exact calculation: the two theoretical predictions differed by 0.36% RMS over the tested energy range. The authors regard the approximate curve as a useful low-cost substitute within this model, while noting that the agreement was shown only for the tested energies and system.
Useful evidence, narrow test
The same flux approach also matched simulations of flips by the other arm. For arm 1 over 4 < E < 6, the full-range implementation-quality index was 1.12%, while the prediction-quality index was 6.96%; it was 11.16% in the lower part of that range and 4.61% in the higher part.
The evidence is limited to an equal-mass, equal-arm, undriven and nondissipative double pendulum. The reported arm-2 analysis covered 2 < E < 6 and the arm-1 analysis covered 4 < E < 6; energies above 6 and other parameter choices were not tested.
Regular islands and mixed regions of phase space were associated with high-rate tails, while the low-energy regime showed larger deviations. The simulations also used finite-duration trajectories and 1,000 initial conditions per energy, and the total number of energy values was not reported.
The preprint reports no independent physical experiment or external dataset validation, and it gives no formal confidence intervals, hypothesis tests or p-values. It therefore supports a statistical mean-rate calculation for this model, not a universal rule for double pendulums or a demonstration that regular trajectories follow the same ergodic approximation.
The open problem is to account quantitatively for regular parts of phase space and for the bias introduced by selecting chaotic trajectories. The authors also leave open why the approximate saddle curve remains accurate well above the saddle energy, and whether the framework extends to unequal masses or arm lengths, higher energies or other dynamical systems.
Paper data and sources
Original title: Flip rate prediction in the double pendulum
Authors: Peleg Haham, Barak Kol
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text