Local vs Global Optimization

A local optimum is something you can improve that looks better when viewed narrowly but does nothing for the bigger picture. Goldratt’s example: speeding up a workstation that feeds a slower one. It is faster in some sense, but the factory still cannot produce more, because output is set by the slowest step. A global optimum advances the whole system’s goal — its throughput. CF inherits Goldratt’s lesson that most improvements are local-only: as he put it, optimization away from the constraint is wasted. The reliable way to optimize globally is to find what limits the goal and improve that, which is focusing on the constraint.

Strictly, a change can be both local and global; “local optimum” is a label for things judged not globally important. The point is to separate what only helps locally from what also moves the goal.

CF extends this from factories to thinking. It argues most of the factors people weigh in a decision are local optima sitting on excess capacity — already good enough with wide margin — so refining them changes no real outcome. This grounds CF’s critique of Bayesian-style weighted scoring: such methods raise a score whenever any local factor “improves,” even though throughput did not rise, so the higher score is an error. Better to give minor factors slack, leave them alone, and spend attention on the few factors that actually determine success.