Every time you pick the shortest queue, pack a suitcase to fit, or leave home at the last safe moment, you are optimizing — searching a space of choices for the one that scores best. Airlines route planes to burn the least fuel; a phone battery charges as fast as heat allows; a portfolio chases the most return for a fixed risk. Optimization is the branch of mathematics that makes "the best choice" precise and then finds it.

Three ingredients turn a vague wish into a solvable problem:

• Decision variables \mathbf{x} = (x_1, \dots, x_n) — the knobs you get to turn.
• an objective function f(\mathbf{x}) — the single number you want as small (or as large) as possible.
• constraints — the rules the choice must obey, carving out the feasible region of allowed \mathbf{x}.

Written out, the whole subject fits in one template — the standard form:

The g_i are inequality constraints (budgets, capacities, "at least this much"), the h_j are equality constraints (something must balance exactly). The set of \mathbf{x} satisfying every constraint is the feasible region \mathcal{F}; a point inside it is feasible, a point outside is infeasible. Optimization asks for the feasible point with the best objective value.

Minimize or maximize? Same problem

By convention we write everything as a minimization. That costs nothing, because maximizing f is exactly minimizing -f:

Flip the sign of the objective, solve the minimization, and flip the sign of the answer back. The location \mathbf{x}^\star of the optimum is identical either way — only the reported value changes sign. So a solver only ever needs to know how to go downhill.

Likewise, a "\ge" constraint becomes a "\le" one by negating it: g(\mathbf{x}) \ge 0 is the same rule as -g(\mathbf{x}) \le 0. Every optimization problem can therefore be squeezed into the one standard form above — which is why a single theory covers such a sprawling zoo of applications.

Local versus global — the central difficulty

A point \mathbf{x}^\star is a local minimum if it beats every feasible neighbour close by; it is a global minimum if it beats every feasible point everywhere. A landscape can have many local minima but only the deepest is global. Standing in a valley, you cannot tell by looking at your feet whether a deeper valley lies over the next ridge — and that, in one sentence, is why optimization is hard.

The picture below is an objective f(x) = \sin x + 0.15\,x restricted to a feasible interval. It dips into several local minima, each a little higher than the last because of the upward drift. Reveal the steps: the feasible interval, then the local minima, then the single global minimum among them. A greedy searcher that stopped at the first dip it found would still be wrong about the whole picture.

A worked set-up: the fence problem

A farmer has 100 m of fencing and wants to enclose the largest possible rectangular paddock against a straight river (so one side needs no fence). Turn the wish into standard form.

Variables. Let the two sides perpendicular to the river each be x and the side parallel be y.

Objective. Maximize the area A = xy — equivalently, minimize -xy.

Constraints. The fence used is 2x + y = 100 (an equality), and lengths are non-negative: x \ge 0, y \ge 0 (inequalities).

Solve. The equality lets us eliminate y = 100 - 2x, so A(x) = x(100 - 2x) = 100x - 2x^2. Then A'(x) = 100 - 4x = 0 \Rightarrow x = 25, giving y = 50 and area A = 1250 m². The feasible region for x is the interval [0, 50]; the optimum sits in its interior, and the boundary points x = 0 or x = 50 give zero area. This is the whole optimization workflow in miniature: name the variables, write the objective, list the constraints, then hunt the feasible region for the best point.