One extra hour of overtime might cost a familiar hourly rate. The fifth hour could require a supervisor, a special shift allowance or a higher wage. Customer delays may carry a modest cost at first and a severe consequence after a deadline is missed.
Real business costs have thresholds, switches and curves. Why, then, are so many powerful optimization models built from straight lines?
Because linear relationships are a useful mathematical language, and because some apparently nonlinear business behavior can be expressed exactly with additional variables and constraints.
Linear does not mean simplistic
A linear objective adds contributions that vary proportionally with the decisions. If each kilometer driven costs a certain amount, total distance cost is a linear sum. If producing one additional unit consumes a fixed quantity of material, total material use is linear too.
Linear relationships are easy to inspect: the coefficients tell us what each decision contributes. They are also supported by decades of work on linear and mixed-integer optimization. But an operation rarely contains only smooth, proportional costs.
Suppose overtime is free of premium charges for the first two hours, then costs $30 per hour beyond that. The premium part can be written as 30 max(0, h − 2), where h is total overtime in hours. The expression has a bend at two hours. It can be represented in a suitable linear model by introducing a new nonnegative variable p and requiring p ≥ h − 2. When the objective minimizes 30p, with p ≥ 0, the optimal value makes p exactly the excess hours.
The model still captures the threshold. It has simply expressed the calculation in a form the solver can use efficiently.
When the switch really matters
Some relationships are discontinuous. Opening a warehouse may incur a fixed charge, regardless of whether it handles ten orders or ten thousand. Starting a machine may trigger setup time. A truck route may need an additional driver as soon as a legal limit is exceeded.
Mixed-integer models can introduce a yes-or-no variable for such decisions. The variable acts like a switch: pay the fixed cost if the facility opens, activate the setup if the machine changes product, or restrict the schedule if a shift is selected.
These formulations are not always cheap computationally. Sometimes they require careful bounds, extra variables or a more suitable solver family. A poorly chosen approximation can make a model more complicated without making its decisions more useful.
Approximate the world, not the business purpose
A highly detailed cost curve may not be necessary when the underlying data are uncertain. A piecewise-linear approximation can be more transparent and easier to maintain, provided it preserves the trade-offs near the choices the business will actually make.
For example, a transport team may estimate carbon emissions from distance, vehicle class and load rather than simulate every engine event. That approximation can still support meaningful route comparisons. But if two routes differ mainly in steep gradients, congestion or electric charging conditions, a more detailed representation may become important.
The appropriate level of mathematics depends on the decision, not on a desire to make the model elegant.
There is a productive discipline in asking which parts of reality must be represented exactly, which can be approximated and which can be left outside the model. Better modeling often means finding the simplest representation that preserves the consequences people care about.