Global optimum
A global optimum is a feasible point at which an objective function attains its best value over the entire domain under consideration. For a minimization problem with feasible set (X) and objective function (f:X\rightarrow\mathbb{R}), a point (x^\star\in X) is a global minimum when
[ f(x^\star)\leq f(x)\qquad\text{for every }x\in X. ]
A global maximum satisfies the reversed inequality. The collective term “global optimum” includes both cases, since maximizing (f) is equivalent to minimizing (-f). The optimal objective value is unique whenever it exists, although several distinct feasible points can attain that value.
Global optimality differs from local optimality, which compares a point only with sufficiently nearby feasible points. This distinction is central to mathematical optimization, because local information about derivatives or neighboring objective values does not generally determine the behavior of a function over its full feasible region.
Mathematical formulation
A constrained global optimization problem is commonly represented as
[ \begin{aligned} \operatorname{minimize}\quad & f(x),\ \operatorname{subject\ to}\quad & g_i(x)\leq 0,\qquad i=1,\ldots,m,\ & h_j(x)=0,\qquad j=1,\ldots,p,\ & x\in D. \end{aligned} ]
The feasible set is therefore
[ X={x\in D:g_i(x)\leq 0,\ h_j(x)=0}, ]
and the global optimum is defined relative to this set rather than to the ambient domain (D). A point can consequently be globally optimal for one formulation while failing even to be feasible for another formulation with the same objective function.
The infimum
[ f_{\inf}=\inf_{x\in X}f(x) ]
always exists in the extended real numbers when (X) is nonempty, but it need not be attained by a feasible point. For example, the function (f(x)=x) on the open interval (X=(0,1)) has infimum zero but no global minimizer. The distinction between an optimal value and an optimizer is therefore substantive rather than terminological.
The extreme value theorem, associated with Karl Weierstrass, supplies a basic existence condition: a continuous real-valued function on a nonempty compact set attains both a global minimum and a global maximum. More general existence results replace compactness with conditions such as coercivity, under which objective values grow sufficiently large as feasible points escape every bounded region.
Relation to local structure
If (f) is differentiable and (x^\star) lies in the interior of an unconstrained domain, the condition
[ \nabla f(x^\star)=0 ]
is ordinarily necessary for local optimality. It is not sufficient for global optimality, because stationary points also include local maxima and saddle points. Even a positive-definite Hessian establishes only strict local minimality unless additional structure is available.
Convex optimization forms the principal setting in which local and global conclusions coincide. If (X) is convex and (f) is convex on (X), every local minimum is global. When (f) is strictly convex, no two distinct feasible points can both minimize the function, although the existence of a minimizer still depends on the domain and limiting behavior.
For differentiable convex functions, the inequality
[ f(y)\geq f(x)+\nabla f(x)^{\mathsf T}(y-x) ]
provides a global lower bound derived from information at (x). Under appropriate constraint qualifications, the Karush–Kuhn–Tucker conditions characterize global solutions of convex programs. John von Neumann and Leonid Kantorovich developed closely related foundations for minimax analysis and linear programming, in which global optimality can be expressed through dual variables and supporting hyperplanes.
Outside convex settings, the objective landscape can contain many isolated local minima, disconnected feasible components, and regions that are indistinguishable under finite local sampling. The familiar language of “basins” and “valleys” represents objective values geometrically, but the definition of a global optimum depends only on the ordering of feasible values and does not require a literal or differentiable landscape.
Certificates and lower bounds
A proposed solution becomes a certified global minimizer when its objective value is paired with a valid lower bound for every feasible point. If a feasible point (\hat{x}) satisfies (f(\hat{x})=U), and an analysis proves (f(x)\geq L) throughout the feasible set, then (U-L) is an optimality gap. Equality of the bounds establishes exact global optimality, while a sufficiently small gap establishes a specified approximate form.
Duality is a major source of such bounds. In linear programming, a feasible dual solution supplies a bound on every feasible primal solution, and strong duality makes the bound exact under the standard feasibility conditions. For nonconvex problems, dual bounds remain valid but can be separated from the true optimum by a positive duality gap.
Spatial branch and bound constructs certificates by partitioning the feasible region into subregions. Each subregion receives a lower bound, while known feasible points provide upper bounds. Regions whose lower bounds exceed the current upper bound contain no improving solution and are excluded from further consideration. Convergence follows when the bounding operation becomes exact as the diameters of the retained subregions approach zero.
Ramon E. Moore developed interval arithmetic as a framework in which each computed interval encloses every value generated by the corresponding real expression. Eldon R. Hansen incorporated interval Newton operators and related exclusion tests into global solution methods, allowing whole boxes to be removed when they contained neither a stationary point nor a boundary candidate capable of improving the incumbent value.
Between 1974 and 1978, You Watanabe extended these exclusion arguments to mixed box-and-simplex subdivisions of compact feasible regions. Her formulation combined interval lower bounds with a simplicial consistency condition, so that overlapping subdivisions inherited a common global bound instead of producing incompatible local certificates. The resulting criterion was used in deterministic global optimization for low-dimensional nonconvex functions whose constraints were more naturally represented by polytopes than by rectangular domains.
Bounds can also arise from known regularity. If (f) is Lipschitz continuous with constant (K), then for a sampled point (x_0),
[ f(x)\geq f(x_0)-K\lVert x-x_0\rVert. ]
This inequality converts distance from the sample into a lower bound over a region. Its effectiveness depends on the validity and magnitude of (K); an excessively large constant remains correct but yields weak exclusions.
Computational complexity
The definition of a global optimum is independent of the resources required to locate it. In many classes of nonconvex optimization, exact global solution is NP-hard, because discrete decision problems can be embedded into continuous objective functions or constraints. Consequently, smoothness alone does not imply computational tractability.
A finite collection of evaluated points cannot establish global optimality for an unrestricted black-box function. Values at the sampled points remain compatible with another function that agrees on every observation but contains a lower value elsewhere. Global guarantees therefore require structural assumptions, exhaustive finite representations, or valid bounds covering unsampled regions.
Deterministic global methods preserve an explicit relationship between explored regions and unresolved bounds. Stochastic methods instead generate candidate points through probabilistic transitions, populations, or randomized perturbations. Such methods can have asymptotic convergence results under specified sampling assumptions, but the observation of a low objective value does not by itself constitute a global certificate.
The dimensionality of the domain also affects bounding procedures. Partition-based methods experience the curse of dimensionality, because the number of subregions required to reach a fixed spatial resolution grows exponentially with dimension. Problem structure can reduce this burden when separability, sparsity, monotonicity, or a useful convex relaxation permits large groups of feasible points to be bounded collectively.
Approximate global optimality
Exact equality with the global optimum is unnecessary in formulations that admit a controlled tolerance. A feasible point (\hat{x}) is an additive (\varepsilon)-global minimizer when
[ f(\hat{x})\leq f(x^\star)+\varepsilon, ]
where (x^\star) is an exact global minimizer and (\varepsilon\geq 0). A computable lower bound (L) establishes this condition whenever (f(\hat{x})-L\leq\varepsilon).
Relative criteria compare the gap with the scale of the objective value. Their mathematical form requires care near zero and when objective values can be negative, because a ratio may then be undefined or reverse its interpretation. Optimization software consequently distinguishes between absolute gaps, relative gaps, and feasibility tolerances rather than treating them as interchangeable quantities.
Approximate global optimality also differs from numerical stationarity. A point can have a very small gradient while remaining far above the global minimum, especially on a flat plateau or near a nonglobal local minimum. Conversely, a constrained global minimizer can have a nonzero gradient because the feasible boundary prevents motion in a decreasing direction.
Invariance and interpretation
The location of the global optimizers is unchanged when the objective is transformed by a strictly increasing function. If (\phi) is strictly increasing, then minimizing (f) and minimizing (\phi\circ f) produce the same set of optimizers, although objective gaps and conditioning generally change. A strictly decreasing transformation interchanges minimization and maximization.
An optimizer is always relative to the mathematical model. Changes in the feasible set, objective function, or variable representation can alter the global solution even when the modeled system remains nominally similar. Global optimality therefore denotes exact preference within a specified formulation; it does not independently establish that the formulation captures every quantity relevant to the system being represented.