Every number on this page is computed by the Volume II Chapter 4 notebook's own code; the book's 100 starts use seed 2026, the lab's fresh starts seed 26204. Your notes stay on this page until you leave it; copy the memo before you go.
Meridian's chief operating officer can automate plant lines in part, and her planners' optimizer returns different plans from different starts. The chief financial officer wants the inventory of plans, an audit of each, and a proof of the best one before the board meets. You support the planning team.
Use AXIOM to explore the landscape of partial rounds, run and audit local searches, bound them with Chapter 3's relaxation, and close the gap by branch-and-bound.
Worked Example 4.1: the board credits the year-6 savings for twenty years; a fraction z of a round costs $120z million and closes the fraction z/2 of the remaining efficiency gap. A plan is worth the cash it frees above $470 million plus twenty years of its year-6 savings: V(z) = 150 + Σ Sk − 120 Σ zk + 20 S6. All amounts undiscounted.
In the objective visualization, draw the value on the face where only years 4 and 5 vary; find the two corner peaks and the saddle, and read the curvature at each from the Hessian display.
The value is affine in each year's round when the others are fixed (Proposition 4.2(i)): part of a round is the blend of none and a whole round. Chapter 3's exponential form overstates part rounds (2−z ≤ 1 − z/2), so its relaxation is a ceiling on this model too.
In the multistart harness, run SLSQP and the interior-point solver from 100 seeded starts; compare their basin inventories and iteration counts.
Right: the lab's 100 starts (seed 26204), by their year-4 and year-5 rounds, coloured by where SLSQP ends. Iteration counts vary with SciPy's version (its SLSQP was rewritten in 1.16); where the runs end does not.
Run the KKT verification module on each local answer and on the saddle: residuals, the rank check for LICQ, and the reduced-Hessian test.
Load Lab 2.3's relaxation, solve it, and compute each local answer's distance from the ceiling.
With a $60 million capital cap, half a round, at ten years' credit, no whole round can run. Pricing the cap gives a ceiling; the best price leaves a gap that no price list closes.
In the relaxation, fix the year-5 round at none and at a whole round, re-solve, and continue until every node is settled or pruned; record the tree and the gap after each node.
Different answers from different starts, each passing the first-order audit (Theorem 4.2); a saddle that only the second-order test rejects (Theorem 4.1); local prices that change by basin (Theorem 4.3); and a ceiling that bounds what no run reached, with a tree that closes the gap (Theorem 4.4).
The inventory of local answers with their audits, the ceiling and each answer's distance from it, the branch-and-bound certificate, and a recommendation on partial automation.