Every number on this page is computed by the Volume II Chapter 3 notebook's own code; its random chords and starting plans use the lab's seed 26203. Your notes stay on this page until you leave it; copy the memo before you go.
Meridian's chief risk officer wants a ceiling on the value of any automation plan, computed without enumeration, and a price list the board can audit. You support her team, which must turn Worked Example 2.1's instance into a convex relaxation, solve it with a certificate, and explain what the prices say.
Use AXIOM to recognize convexity, solve the relaxation, audit its KKT conditions, bound the lumpy plans, and test the prices against re-solves.
Worked Example 3.1: both pledges dropped; the optimizer chooses the rounds and the payouts, keeping the $350 million floor and the $470 million end condition; ten years of the year-6 savings credited. By enumeration of all 64 schedules, three rounds in years 1 to 3 are best: $455 million paid out plus ten years of $122.5 million, $1,680 million. All amounts undiscounted.
In the feasible region explorer, enter the program in Volume I's form and in exponential form; check that the two agree at the 64 lumpy plans, and find the violating chord the explorer flags in the first.
Year-2 savings for fractional rounds (z1, z2). Volume I's form: the gap 14(1 − z1/2)(1 − z2/2), a product of decisions. Exponential form: the gap 14 × 2−(z1 + z2). Dashed: the chord between the two ends; concave savings stay on or above it.
Declare the relaxation in disciplined convex form, confirm that the convexity check passes, and solve it with the interior-point solver; watch the duality gap close in the duality demonstration.
The notebook writes out the barrier method of Section 3.6: each commitment hi(x) ≥ 0 gets a barrier log hi(x); for a weight t it maximizes t J(x) + Σ log hi(x), and the gap m/t (m commitments) is what the prices certify. The book's own solve used Clarabel, which returns after 15 iterations.
Not Meridian: a budget of 100 split between two channels with returns 3√x1 and 4√x2. Every price λ gives a ceiling q(λ) = 25/(4λ) + 100λ; the best price closes the gap.
Run the KKT verification module on the returned plan and rebuild the marginal values and prices of Figure 3.3(a).
Compare the ceiling with the 64 lumpy plans, compute the gap, and value the relaxation's plan rounded down and rounded up.
In the sensitivity analysis, move the credit from 0 to 60 years and add a capital cap; check the price bound of Theorem 3.5(iii) against re-solves, and find the kink at $360 million.
Shaded, the credits at which rounding the relaxation's rounds picks the wrong number.
Re-solves at caps 360 + s against the line 1,680 + price × s. A price bounds every re-solve when no re-solve lies above its line (Theorem 3.5(iii)).
A model that fails the convexity check in one form and passes in another (Table 3.2); a gap that closes to zero (Theorem 3.4); margins equal to the cost exactly where a round is partial (Theorem 3.5); and a ceiling that bounds every lumpy plan while rounding sometimes fails (Proposition 3.2).
The ceiling and the solver's certificate, the three-round plan's distance from it, the price list, the credits at which the relaxation proves the lumpy plan best, and a recommendation on whether further search is worth its cost.