Can several parameter settings be global minimizers?

Yes. They can attain the same minimum value.

Does a convex objective alone settle a constrained problem?

No. The feasible set also matters for the standard local-to-global guarantee.

Does failure to improve a solution prove it is globally minimal?

No. It may reflect the search method, its budget, or its neighborhood rather than a property of all feasible alternatives.

AI Glossary / Graphs and symbolic AI / Intermediate

Global minimum

Definition

A global minimum is the smallest value an objective function actually reaches over its entire feasible set. A point that reaches it is a global minimizer. In machine learning, the objective may be a training loss; in planning, it may be cost. The comparison covers every allowed solution, so a global-minimum claim is stronger than reporting the lowest value found during one search.

What it means

The function and feasible set together define the question. Minimizing training error, validation error, and a regularized objective are different problems. Even a proved minimum of one does not automatically minimize the others.
The minimum value is a number, while a minimizer is an input or parameter setting. There may be several minimizers. A problem may also have an infimum, meaning its greatest lower bound, without attaining that bound anywhere. This distinction matters when parameters can keep changing while the objective approaches a limit.

How it works

A global-minimum proof needs a feasible point and a reason no feasible point can do better. Exhaustive enumeration works for a sufficiently small finite set. Analytic inequalities, convexity, and valid lower bounds can provide other routes to certification.
For a convex function minimized over a convex feasible set, every local minimizer is global. That property does not mean every numerical iterate is optimal or that every optimization problem involving a convex-looking term is convex. Constraints and the complete objective must satisfy the relevant conditions. General nonconvex optimization may require more elaborate global methods or a carefully qualified best-found result.

Three practical examples

A least-squares teaching example derives the parameter that gives the smallest sum of squared residuals.
A small scheduling model enumerates all feasible assignments to certify its lowest modeled cost.
A solver report compares its best feasible cost with a valid lower bound to describe an unresolved optimality gap.

Think of it this way

Think of the lowest point in a defined landscape. Reaching the bottom of one valley does not establish that no deeper valley exists elsewhere. The analogy is limited because AI objectives may be discrete, high-dimensional, or statistically estimated, and the lowest modeled loss may not correspond to the best behavior outside the model.

Worked example

Suppose a constant predictor outputs w for two observed targets, 1 and 3. Its mean squared training loss is L(w) = ((w - 1) squared + (w - 3) squared) / 2. Expanding and completing the square gives L(w) = (w - 2) squared + 1.
Because the squared term is nonnegative, L(w) is at least 1 for every real w. At w = 2, it equals 1, proving the global minimum. The minimum is 1; the minimizer is 2. At w = 1, the loss is 2, so merely matching one training target does not minimize the average.
Now suppose a held-out target is 10. The squared error of the training minimizer on that target is (2 - 10) squared = 64. This does not invalidate the training optimum. It shows that the optimized objective and the later evaluation are different quantities. All data and calculations here are fictional teaching values.

Illustrative case study

In a fictional model-fitting class, the Beech team compares a solver's output with an exact analytic solution for a small quadratic. The solver stops at a nearby parameter because of its numerical tolerance. The team reports both the attained objective and the known minimum instead of rounding the distinction away.
For a larger nonconvex exercise, it uses the phrase “best result found” because no valid global certificate is available. The invented case illustrates precise reporting across two kinds of problems, rather than suggesting that a numerical method has failed whenever it returns an approximate solution.

Common misconceptions and mistakes

A zero gradient is not proof of a global minimum in a general problem. It can occur at a maximum or saddle point. A boundary minimum may also fail to have zero unconstrained gradient. Checking only a derivative equation can therefore miss the relevant answer.
Do not equate global training-loss minimization with good generalization. A flexible model can fit training examples while performing poorly elsewhere. Also distinguish a lower bound from an attained minimum: for L(w) = exp(-w) over real w, the infimum is zero, but no finite w achieves zero.

Business use

The concept helps teams evaluate claims that an AI or optimization system has found the “best” solution. Ask which cost is minimized, which decisions are allowed, and what evidence establishes globality. When proof is unavailable, a transparent limit is preferable to treating repeated unsuccessful searches as a certificate.
If a minimization solver has a feasible cost of 125 and a valid lower bound of 120, the true optimum lies between them, so the absolute gap is at most 5 in the objective's units. Such bounds help assess whether further search is worthwhile without claiming an exact optimum prematurely.

Check your understanding

In the constant-predictor example, which is the global minimum: 1 or 2?

Answer: The minimum loss value is 1. The parameter w = 2 is the global minimizer that attains it. Keeping the value separate from its location prevents ambiguity.

Related terms

Loss function
Convex optimization
Infimum (or greatest lower bound)
Generalization
Global minimizer
local minimum
feasible set
lower bound
infimum
optimality gap

Learn next

Learn next about regularization. Add the penalty w squared to the worked example's loss and complete the square again. Determine the new minimizer and explain why changing the objective can change the optimal parameter even when the two training targets stay the same.

Sources

Boyd and Vandenberghe, original Convex Optimization slides, optimal points and local/global distinction: https://web.stanford.edu/~boyd/cvxbook/bv_cvxslides.pdf Boyd and Vandenberghe, original book and teaching-resource page: https://web.stanford.edu/~boyd/cvxbook/

Previous: Global maximum Next: Global optimization