RuleStatics
Certified thrust extremes of the course pattern
statics.tna.extremes
Our derivation from a construction or rule in the sources.
Statement
With the plan fixed, the least and the greatest thrust over all force distributions of the course pattern are a non-convex problem: the heights depend on the forces through the vertical equilibrium. In a square, oblong or regular bay every course is a hanging-chain arch between two rib nodes at the same height, and every rib a chain whose forces grow outwards. The heights then become linear in the reciprocals of the course and rib forces, and the only non-linearity left is x ↦ 1/x for each of them. On a box of values its graph is enclosed by tangents below and the secant above, which gives a linear relaxation; splitting the boxes (branch and bound) tightens it. Every linear programme returns a bound by duality, every network found is checked exactly, and the result is the best network together with a bound that no network of the pattern can beat. Where the relaxation is empty, no network exists: that certifies the thickness factor too.
Formula
Parameters
- relaxation
- tangents (≥ 5 per link) and secant of 1/x on each box; r decreasing outwards; thrust window from the incumbent
- bound_tightening
- optimisation-based bound tightening at the root, repeated until a round shrinks the boxes by less than 5 % (at most 10 rounds; until wave 6: 3 rounds, which left the greatest thrust of Ib at 32 courses 17 % open); interval propagation of the height rows, rib forces, outward growth and thrust window at every node
- incumbents
- each new incumbent is refined by the relaxation on a small box around it (trust region, grown while it pays), then repaired and polished on the exact model
- branching
- largest relative violation of φ·x = 1; split at the relaxation point or the geometric mean
- default_gap
- 0.001
- default_nodes
- 300 per extreme (the essay uses 1000)
- near_infeasible
- a relaxation point that is tight but a hair outside the sliver of admissible networks is walked towards the incumbent until it fits, then polished
- verified
- reduced heights equal the full vertical solve to 1e-8 m (property test); no random distribution beats the bounds (property test)
Modern note
The reduction holds where every course has equal end heights and all ribs are alike: square and regular polygonal bays by rotation, oblong rectangular bays by their two mirror axes (the courses of the long and the short compartments then carry thrusts in proportion to their spans, as p. 61 of the book says, and their resultant falls along the rib). The loads must be mirror-symmetric too, so the surface of each load cell averages both triangulations. It is checked at run time against the full solve; trapezoidal, irregular and sexpartite bays fall back to the search. The certificate is for the course pattern only; it says nothing about denser form diagrams.
Sources
„wird sich aber die Grösse der Schübe verhalten wie die Grösse der vorliegenden Rechteckseiten, was zur Folge hat, dass die Mittelkraft aus beiden Schüben in die Richtung der Rippe fällt“
Where implemented
@ungewitter/statics#certifiedThrustRange, reduceCourseNetwork