Duality — the series
-
Two numbers that have to meet
Every linear program has a shadow — a second program built from the same numbers read the other way, whose minimum can never fall below the first's maximum. That much is a one-line calculation; the theorem is that the two numbers are always exactly equal.
-
What a constraint is worth
A linear program and its dual reach the same number. What the dual's variables are is a separate question, and the answer converts a solution into a rate for every constraint — piecewise constant, zero on the constraints that are not doing any work.
-
When one of the two numbers is missing
The duality theorem is usually quoted as an equality: a linear program and its dual reach the same number. That is one of four cases. A program can run away to infinity, or have no feasible point at all, and then its dual is forced into a matching failure. Every small program with coefficients from minus one to one has been classified, and the table has exactly four occupied cells out of nine.
-
The lines the optimum lies under
Change the resources a linear program is given and its best value changes too, tracing a graph. Every solution of the dual is a straight line lying above that graph, and the graph is exactly the lowest of those lines — a bent roof of finitely many planks. Require the answer to be in whole numbers and the roof stays where it was while the graph falls away beneath it in steps, and the space between is the part of the problem no price can see.
-
Prices at every corner
The duality theorem says a linear program's best value equals its dual's, and says nothing about how to find either. The simplex method finds both at once — it walks from corner to corner, and at each one asks the constraints that meet there for prices. A negative price names an edge that climbs; when none is negative, the prices are the proof.
-
The cube that takes every corner
The simplex method is fast on every program anybody meets in practice. In 1972 Victor Klee and George Minty squashed a cube so that the method, choosing the steepest edge each time, visits all of its corners — 2ⁿ − 1 moves in n variables, with the optimum one edge from the start.
-
A bound that may be off by a third
The shortest tour through a set of cities is hard to find, and a linear programme gives a lower bound for it in polynomial time: give every road a weight between nought and one, two at each city, at least two across every division of the map. On random cities the bound is almost always exact. On two triangles joined by three long paths it falls short by nearly a third, and whether a third is the worst it can ever do has been conjectured for decades and never proved.
-
Tours within half again of the best
Nobody can find the shortest tour through many cities quickly, but a tour at most half as long again as the best can be built in a few steps: the shortest tree, a cheapest pairing of the cities where the tree branches oddly, an Euler circuit, and shortcuts. Nicos Christofides found it in 1976, and for forty-five years nobody could guarantee better. A strip of cities shows the half is really lost, and Laurence Wolsey's reading of the same argument shows it bounds the linear programme too.
-
A relaxation that answers in halves
The cheapest set of points touching every edge of a graph is hard to find. Allow each point to be taken fractionally and the problem becomes a linear programme — and its best answers are never messier than halves. Every point gets 0, ½ or 1, the 0s and 1s can always be trusted, and the proof is a picture: double the graph into a bipartite one, where a matching decides everything exactly.
-
Two colonels and a linear programme
Two colonels divide their soldiers among three battlefields at once, and whoever sends more to a field wins it. Every division can be beaten by some other, so neither colonel has a safe plan — until they are allowed to choose at random. The best random choice is a linear programme, the opponent's best is its dual, and the theorem that the two meet is von Neumann's. With nine soldiers each, one optimal plan uses only seven divisions, and on every field it sends each number from nought to six exactly once.