Formulating Applied Multivariable Optimization Models

From a Practical Goal to a Mathematical Model

Applied optimization begins before any calculation aimed at finding a best design. First, the situation must be expressed as a model: a collection of decision variables, an objective function, and constraints that describe which decisions are allowed. A model that omits a requirement or uses inconsistent units can produce a mathematically correct answer to the wrong problem.

The purpose of this tutorial is to formulate models, not to solve them. In particular, writing down a constraint does not establish that a model has a feasible solution or identify an optimum. Those are separate questions. Here the central task is to translate the stated goal and requirements carefully enough that the mathematical model represents them.

Decision Variables, Objectives, and Constraints

Decision variables represent quantities a planner or designer can choose. Each variable needs a clear meaning and a unit, such as hours, kilograms, hectares, or pallets per day. The unit is part of the interpretation: a variable representing daily flow is not interchangeable with one representing stored inventory.

The objective function assigns a value to each decision. It encodes what the model is intended to improve, such as maximizing estimated benefit or minimizing cost. An objective is not itself a requirement. For example, minimizing cost does not automatically impose a budget limit; a budget limit must be written as a constraint if the situation specifies one.

Constraints express required conditions and limits. An exact requirement is commonly represented by an equality. A maximum resource allowance is commonly represented by a “less than or equal to” inequality; a minimum performance requirement is commonly represented by a “greater than or equal to” inequality. Physical restrictions such as nonnegative quantities must also be included.

A common notation is to write equality constraints as \(g_i(\mathbf{x})=0\) and inequality constraints as \(h_j(\mathbf{x})\leq 0\), for specified functions \(g_i\) and \(h_j\). Constraints written with \(\geq\) are equally valid; they can be rearranged into the \(\leq 0\) form by multiplying by \(-1\). The form used should make the original requirement easy to interpret.

This notation is a template, not a recipe that determines the model by itself. The variables, objective, functions, and domains must come from the particular application. For a continuous calculus model, variables typically range over real values in specified intervals or regions. If the real decisions must be whole numbers, that integrality requirement is part of the model too.

A Reliable Formulation Process

Separating these steps helps prevent a frequent modeling error: putting a condition into the objective when it should be a constraint, or treating a preference as if it were a required condition. It also makes assumptions visible. For example, a model might assume a quoted cost per unit stays constant across the range of decisions; that is a modeling choice, not an algebraic consequence.

Worked Example: Choosing a Water Blend

Why the Concentration Constraint Has This Form

The blend condition in the example follows from the definition of concentration as mass divided by volume. If the source volumes are \(x\) and \(y\) cubic metres per day, their total volume is \(1000(x+y)\) litres per day. Their total contaminant mass is \(1000(2.0x+0.5y)\) milligrams per day. Therefore the blend concentration is:

Requiring this concentration to be no more than \(1.4\) gives \((2.0x+0.5y)/(x+y)\leq1.4\). Since the total volume is \(100\), the denominator is positive, so multiplication by \(x+y\) preserves the inequality direction. Equivalently, one can multiply by the positive conversion factor \(1000\) and write \(1000(2.0x+0.5y)\leq1.4\cdot1000(x+y)\). This derivation explains why the concentration requirement compares contaminant mass with the maximum permitted mass for the total volume. It also shows why the positivity of the total volume matters when removing a denominator.

Worked Example: Allocating Restoration Areas

Worked Example: Planning Deliveries to Two Clinics

Units, Domains, and Modeling Assumptions

A useful check is dimensional consistency: terms that are added must have compatible units, and the two sides of a physical equality or inequality must measure the same kind of quantity. In a cost constraint, for example, multiplying dollars per unit by units gives dollars. In a balance equation, both sides must represent the same flow or amount. Unit conversion factors should be written explicitly when needed, as in the water-blend example.

Domain restrictions are also part of the model. A volume, area, or delivery amount is normally nonnegative. A variable may also have an upper or lower bound from a capacity, design specification, or policy. If a decision counts indivisible items, it must be an integer. For instance, a model of whole vehicles cannot faithfully allow \(3.6\) vehicles unless the model deliberately uses a continuous approximation. Standard calculus methods for continuous variables apply to the continuous model, not automatically to an integer-restricted one.

Finally, distinguish what the scenario requires from what it merely suggests. An exact demand is an equality; a minimum target is a \(\geq\) inequality; an available-resource limit is a \(\leq\) inequality. A preference such as “use the cheaper source when possible” is not a substitute for defining an objective and the required constraints. State assumptions behind estimated objective formulas, too: an objective is only as representative as the relationship it encodes.

A Common Modeling Mistake

Suppose the two-source blend uses \(x=60\) and \(y=40\), with source concentrations \(2.0\) and \(0.5\) milligrams per litre. A tempting but incorrect calculation is to compare \(2.0x+0.5y=140\) directly with the limit \(1.4\). This compares a concentration-weighted volume with a concentration; the units do not match. It also leaves out the total volume used to define the blended concentration.

The correction is to divide total contaminant mass by total volume, or equivalently to compare mass with the allowed mass for the total volume. In this example, the blended concentration is \((2.0\cdot60+0.5\cdot40)/(60+40)=140/100=1.4\) milligrams per litre. The corresponding mass comparison is \(1000\cdot140=140{,}000\) milligrams per day versus \(1.4\cdot1000\cdot100=140{,}000\) milligrams per day. Both forms give the same valid check.