The finite interval, pure integrand, output-error measure and budget.
Pure nonnegative norm-like measure of one panel's absolute component errors, using the current whole integral. Infinity requests refinement; NaN and negative errors are invalid.
Pure finite integrand with one fixed positive vector dimension.
Optionalseparable?: {Optional exact decomposition error(value,e)=dot(weights(value),components(e)), with nonnegative fixed-dimension coordinates. The caller owns this mathematical premise; arbitrary measures retain the original full scan. Current weights are recomputed after every split and aggregate-addition roundoff is charged through error.
Fixed nonnegative panel-error coordinates independent of the whole integral.
Current nonnegative coordinate weights, recomputed once per integral.
Interval start, seconds.
Interval end after t0, seconds.
Nonnegative output-error budget in the measure's unit.
The Richardson-corrected integral vector.
Adaptive Simpson integration of a shooting seed's vector integrand.
A panel compares one Simpson rule with two half-panel rules. Their difference divided by 15 estimates the refined rule's error, and the same Richardson correction forms its value (Gander and Gautschi, Adaptive Quadrature--Revisited, BIT 40 (2000), 84--101, https://www.cs.purdue.edu/homes/wxg/selected_works/section_07/160.pdf). The caller measures each panel's componentwise absolute estimates in its output's unit. Those measured errors are summed, and the panel with the largest error is split until the global estimate meets the budget. A panel component whose error added to that component's magnitude (the coarse integral of its absolute value) leaves it unchanged in floating point is resolved and counts no error, Gander and Gautschi's termination test: no refinement can move the computed sum there. So a budget below the arithmetic's resolution ends once every panel is resolved, instead of splitting every panel toward machine width.