Concept/Decision Theory/No. 0579
Local vs. Global Optima
Local vs. global optima is a distinction in mathematical optimization. A local optimum is best among nearby feasible choices; a global optimum is best among all feasible choices. Sewall Wright’s 1932 landscape diagram illustrated how small steps can lead to a lower peak.
- Evidence
- Well established
- Read
- 6 min
- Links
- 10 connections
01You've seen this when…
- in life
You adjust your departure time by five minutes each way to shorten your commute. Then you try a different train line and arrive fifteen minutes earlier.
- at work
Your team tests small changes to a ten-step signup flow. Completion barely improves. A prototype with two steps performs much better.
- out in the world
A transit agency keeps adjusting bus departure times. Riders still wait through long transfers because the route map stays fixed.
02The idea
After enough small improvements, you reach a point where every nearby change makes things worse. That point is a local optimum. A global optimum is the best choice across the entire set of feasible alternatives. Several choices can tie for either distinction.
Picture a landscape with several hills. Each upward step improves your position. Eventually you reach a summit. You can see higher ground across a valley, but reaching it requires a different route, possibly including some downhill steps. The same picture works upside down when the goal is to minimize cost: a search can settle into one valley while a deeper valley lies elsewhere.
The crucial word is nearby. In a continuous problem, nearby means a sufficiently small change in the variables. In a discrete problem, it usually means the alternatives reachable through a specified move, such as swapping two assignments or changing one setting.
That makes local optimality relative to the search method. A schedule can be unbeatable when the only permitted move swaps two shifts, yet improve when three shifts change together.
Global optimality also has boundaries: the objective, constraints and feasible set. The cheapest delivery plan under today’s staffing limits may differ from the cheapest plan available with another driver. Change the problem and its optimum can change.
03Why it matters
Small improvements are often cheap, measurable and easy to reverse. Teams can test them without rebuilding the whole operation. Marginal analysis helps make those incremental choices. Trouble starts when success with small changes becomes evidence that the entire design is best.
A mature product may need a different architecture. A carefully tuned schedule may need a different staffing model. A factory may have efficient movements inside a poor layout. The search has exhausted one neighborhood while leaving broader alternatives largely unexplored.
This is one reason to balance refinement with exploration. The explore-exploit trade-off asks how much effort to spend improving a known option and how much to spend discovering others. Searching from several starting points can reveal alternatives that repeated refinement of one starting point misses.
Wider search has a cost, though. It takes time, money and attention. A dependable local solution can be the sensible choice when the remaining gains are small or uncertain. Robustness vs. optimality adds another consideration: a plan that performs steadily across changing conditions can serve better than one tuned perfectly to a single forecast.
04A worked example
Consider an invented distributor choosing a depot and a departure schedule. It has two possible depots and two possible schedules. All four combinations meet customers’ delivery windows. For this simplified example, compare weekly operating costs and leave one-time switching costs aside:
- City depot, early departures. $12,000 per week.
- Outer depot, early departures. $13,000 per week.
- City depot, later departures. $12,500 per week.
- Outer depot, later departures. $10,000 per week.
The distributor currently uses the city depot with early departures. Its improvement process changes one decision at a time.
What it looks like The current setup has survived every available test. Moving the depot adds $1,000 a week. Changing the schedule adds $500. Both proposals get rejected.
What’s actually going on Changing both decisions together saves $2,000 a week. Depot and schedule interact: the outer depot works better with later departures. The current setup is a local minimum under the one-change-at-a-time rule. The $10,000 setup is the global minimum among these four modeled alternatives because every combination has been compared.
What would have helped Evaluating the paired change before ruling out either component. With only four combinations, the team can enumerate the whole feasible set. In a larger problem, it can test bundles of changes or start its search from several configurations. An actual decision would also include moving costs, disruption and uncertainty in the estimates. The toy example establishes the distinction; the fuller model determines whether switching pays.
05Where people trip up
- They leave the neighborhood undefined. A claim that a plan cannot improve needs a description of the changes tested. Adjusting prices by a few cents, adding a subscription and entering a new market cover very different neighborhoods. Write down what the search permits.
- They require every step to pay immediately. Some better configurations require coordinated changes or a temporary decline during implementation. Test the complete package and its transition costs. A speculative destination still needs evidence before it justifies a costly crossing.
- They silently change the score. Optimizing one department’s costs can increase costs elsewhere. That is a system boundary problem. A local optimum within a properly defined problem and a department optimizing its own interests are distinct ideas. State whose outcomes the objective includes.
- They treat unavailable options as competitors. A plan requiring unavailable skills, legal permission or equipment lies outside the current feasible set. Constraint relaxation can help explore what changing those limits would unlock, provided the cost of changing them enters the comparison.
- They confuse stopping with proof. Finding no improvement in the trials performed gives limited evidence about untested alternatives. Satisficing means stopping at an acceptable solution; a local optimum requires the stronger claim that no allowed neighboring choice improves it. With several objectives, Pareto efficiency asks a different question: whether improving one outcome requires worsening another.
For a practical review, ask the team to name one broader class of alternatives its current improvement process excludes. Then choose a small test that can reveal whether that exclusion matters.
06When it isn’t a trap
Some problems provide a guarantee that local improvement is enough. In convex minimization, every local minimum is global. The objective has the relevant bowl-like shape, and the feasible set is convex: a straight line between any two allowed points stays within that set. These properties eliminate separate inferior valleys.
That guarantee belongs to the mathematical model. It cannot rescue an objective that leaves out customer harm or a constraint based on a mistaken assumption.
For messier decisions, proving global optimality may be impractical. The useful distinction is between a solution supported by the search performed and a claim covering every feasible alternative. Keeping that distinction explicit helps people decide how much further searching is justified.
07Roots
In 1932, geneticist Sewall Wright used a hill-covered landscape in a paper for the Sixth International Congress of Genetics. Different combinations of genes occupied different positions; height represented fitness. The drawing made a difficult question visible: how could a process of gradual improvement settle on one peak while higher peaks remained elsewhere?
The mathematical distinction predates Wright’s illustration. Calculus supplied methods for finding candidate maxima and minima near a point. Establishing the best value over an entire domain required further work, including examining boundaries and comparing other candidates. Wright gave that older distinction a memorable picture that traveled into discussions of search, design and strategy.
Computing made the problem operational. Algorithms could improve a solution repeatedly and still finish at an inferior local optimum. In 1983, Scott Kirkpatrick, C. Daniel Gelatt Jr. and Mario Vecchi described simulated annealing, drawing on the controlled cooling of materials. Its search sometimes accepts worse solutions, with that willingness decreasing over time. The method offered a way to explore beyond a local optimum, though a finite run does not generally certify the global answer.
08How solid is this?
The distinction and the guarantee for convex minimization are mathematical results. Applying the landscape metaphor to business or personal decisions requires specifying the objective, feasible alternatives and neighborhood; improvement alone provides no general proof of global optimality.
09Connections
- Often confused with Satisficing
- Countered by Explore-Exploit Trade-Off, Constraint Relaxation
- Can follow fromNonlinearity
- See also Emergence, Pareto Efficiency, Marginal Analysis, Robustness vs. Optimality, System Boundary, Theory of Constraints
10Origin and sources
A longstanding distinction in mathematical optimization, without a single credited inventor. Sewall Wright’s adaptive-landscape illustration (1932) gave multiple local peaks a widely used visual form.
- [1]Boyd, S., & Vandenberghe, L. (2004). Convex Optimization. Cambridge University Press.
- [2]Wright, S. (1932). The roles of mutation, inbreeding, crossbreeding and selection in evolution. Proceedings of the Sixth International Congress of Genetics, 1, 356–366.
- [3]Kirkpatrick, S., Gelatt, C. D., Jr., & Vecchi, M. P. (1983). Optimization by Simulated Annealing. Science, 220(4598), 671–680.
Suggest an edit· Updated 2026-10-02