Does global optimization always require random search?

No. Deterministic methods also exist, including approaches based on valid bounds and domain subdivision.

Is a global minimum always unique?

No. Several feasible points can share the same best value.

Can a local minimum also be global?

Yes. Global minima satisfy local minimality, and for convex objectives on convex feasible sets, every local minimum is global under the usual definition.

AI Glossary / Mathematics and optimization / Intermediate

Global optimization

Definition

Global optimization seeks the best objective value over the entire allowed set of choices. For minimization, a global minimizer x* satisfies f(x*) ≤ f(x) for every feasible x. This differs from a local minimizer, which is only required to be best among nearby feasible choices. The distinction matters when an objective has several valleys, because improving one candidate locally may leave a better region unexplored.

What it means

The word global refers to the stated feasible domain, not every imaginable real-world decision. Changing variable bounds or constraints changes the problem and can change its best solution. A global optimization method tries to explore beyond one local neighborhood, but a finite computation does not always provide proof that its best candidate is globally optimal. Some methods can produce bounds or certificates under specified assumptions. Others provide useful search heuristics whose results should be described as the best found within the search budget.

How it works

A problem specifies an objective, variables and feasible set. Methods may explore multiple starting points, maintain a population of candidates, use randomized moves, subdivide the domain or combine broad exploration with local refinement. Deterministic bounding methods can rule out regions when valid bounds show they cannot improve the incumbent solution. Stochastic methods instead use sampling rules to explore promising alternatives. The appropriate strategy depends on dimensionality, smoothness, constraints, evaluation cost and the level of assurance required. SciPy's global-optimization tools include several different strategies rather than one universally best solver.

Three practical examples

An engineering team can search a bounded design space where several distinct parameter combinations produce locally good performance.
An analyst can investigate hyperparameter settings using a defined validation objective while recognizing that better validation optimization can itself create selection bias.
A scheduling model can compare a current feasible plan with mathematical bounds to quantify how far it might be from the model's optimum, when an appropriate solver provides those bounds.

Think of it this way

Think of seeking the lowest valley in a mapped region. Walking downhill from one location may lead to a nearby valley, while finding the lowest one requires broader exploration or reliable information about the rest of the terrain. The analogy is limited: real optimization spaces can have thousands of dimensions, constraints can disconnect feasible regions, and evaluating one point may itself be expensive.

Worked example

Consider f(x) = x⁴ − (4/3)x³ − 4x² on the closed interval [−2, 3]. Its derivative is 4x³ − 4x² − 8x = 4x(x − 2)(x + 1), so interior stationary points are −1, 0 and 2. The second derivative is 12x² − 8x − 8. It is 12 at −1 and 24 at 2, identifying both as local minima, while its value at 0 is −8, identifying a local maximum. To determine the global minimum on the interval, also check the endpoints. The function values are f(−2) = 32/3, f(−1) = −5/3, f(0) = 0, f(2) = −32/3 and f(3) = 9. The smallest is −32/3 at x = 2. Because the function is continuous on a compact interval, it attains a minimum; because it is differentiable, any interior minimizer must occur among the stationary points. Comparing that complete candidate list therefore establishes global optimality here. A local method stopping at −1 would produce a valid local minimum but miss the lower value at 2. Checking several arbitrary grid points would not provide the same complete argument.

Illustrative case study

In a fictional design study, Cedar Components tunes a simulation with several locally favorable settings. Engineers run a broad search and locally refine promising candidates, while recording failed simulations and constraint violations. They label the selected design as the best feasible candidate found under the documented budget unless a valid certificate is available. A separate evaluation checks the design under perturbed operating conditions. The case illustrates responsible reporting, not a claim that one search run identifies the best real-world product.

Common misconceptions and mistakes

A successful solver status does not automatically certify global optimality. Trying many initial guesses provides evidence but is not generally a proof. A global optimizer can spend substantial effort on an objective that poorly represents the business goal. Constraints must be checked rather than treated as optional annotations. A continuous objective on a noncompact or open domain may have an infimum that is never attained. Finally, randomness does not make a method inherently global in a guaranteed sense, and deterministic behavior does not by itself establish a guarantee either.

Business use

State the feasible region and why its bounds are meaningful. Establish evaluation budgets, acceptable constraint violations and reporting requirements before searching. Compare feasible candidates using a consistent objective and account for noisy evaluations. Preserve seeds and settings where relevant, and repeat stochastic experiments when needed to understand variability. Ask whether the solver supplies a valid bound, an optimality gap or only a candidate. Validate the chosen result against conditions outside the optimization sample or simulation assumptions. Optimization quality and model validity are separate parts of a useful decision.

Check your understanding

Why must the endpoints be checked in the worked interval problem?

Answer: A constrained global minimum can occur at a boundary without a zero derivative. Interior stationary points alone do not cover all possible minimizers on the interval.

Related terms

Objective function
local minimum
feasible set
differential evolution
branch and bound
optimality gap

Learn next

restrict the worked domain to [−2, 0] and identify its new global minimum. Then compare an exhaustive stationary-point argument with a finite grid search. Explain what evidence each procedure supplies and why its conclusion depends on the domain and function assumptions.

Sources

Sources: SciPy, Optimization tutorial, Global optimization, https://docs.scipy.org/doc/scipy/tutorial/optimize.html ; SciPy, Optimization and root finding reference, https://docs.scipy.org/doc/scipy/reference/optimize.html . The polynomial, complete candidate comparison and fictional design study are original; no software benchmark or universal optimality guarantee is claimed.

Previous: Global minimum Next: Global serializability