The margin becomes a slope, and the slope becomes a condition. If another unit adds more benefit than cost, move toward it; if it adds less, turn back. Somewhere between the two, in the smooth country imagined by the model, the additional gain disappears. A question that once demanded a comparison of whole courses of action can now seem to ask only where a derivative vanishes. This is an extraordinary economy of thought. It is also a translation, and translations retain some features of their originals by surrendering others. What kind of truth is gained, and what kind lost, when a choice becomes a derivative set to zero?
The gain begins with a disciplined distinction between an amount and its rate of change. Total benefit can be large while the benefit of one further unit is small; average cost can fall while the next unit is already becoming more expensive. A derivative describes the local change in one quantity as another changes, with the specified remaining conditions held fixed. Its units matter. Dollars per additional loaf are not dollars per loaf averaged across all production, and an extra hour's output is not the value of the worker's whole day. The calculus makes these distinctions precise enough to combine. It can show how a constraint, an input price or a change in the objective alters the point at which further movement ceases to help.
Jevons's account of labor gives the abstraction a recognizable body. Production yields something wanted; continued exertion can become burdensome. His freely choosing worker extends labor while the utility of the additional product exceeds the additional burden. The comparison already contains rates within rates: how much another interval produces, and how much the resulting product is wanted. It also contains assumptions about freedom to vary the interval and the character of the work. Jevons acknowledges that hours cannot always be adjusted at will, that some work is itself attractive, and that a higher reward may induce either more effort or more leisure. The formal comparison sharpens the question. It does not make those alternatives disappear.
Take an invented undertaking whose benefit from q units is 12q − q² and whose variable cost is 2q. Its net gain is 10q − q². The slope is 10 − 2q, which reaches zero at five units. At four units the gain is twenty-four; at five it is twenty-five; at six it has fallen to twenty-four again. Because the gain is concave throughout this example, the summit is the best feasible point when any nonnegative quantity may be chosen. The equality of marginal benefit and marginal cost is doing useful work. It finds the place where an expansion that was improving the result ceases to improve it, and the curvature tells us that no higher summit has been missed.
The last qualification is essential. A zero derivative by itself says that the first-order change vanishes, not that the chosen point is best. The bottom of a bowl has a zero slope too. An objective with several hills may possess several stationary points, only one of which is highest. In a differentiable concave maximization problem over a convex feasible set, the appropriate first-order comparison can certify a global maximum. Without that structure, it may identify a candidate that must be compared with other candidates and boundaries. The theorem's conditions are part of its content. Omitting them does not make the result more general; it converts a precise statement into a misleading rule of thumb.
A boundary can stop the chooser while the slope remains positive. Give the invented undertaking a capacity of three units. Its best attainable gain is then twenty-one, and its derivative there is four. The business would gain from a little more capacity; it cannot obtain that gain by selecting an infeasible quantity. For a small relaxation of this particular cap, the value of the optimum initially rises at four per unit of capacity. That is the shadow value of the constraint in this example. It is a local statement: as capacity approaches five, the additional value falls to zero. The number four neither prices every enlargement nor demonstrates that capacity should be purchased at any cost.
Constrained optimization formalizes this distinction by putting the restrictions inside the comparison. Its multipliers attach a value to relaxing particular constraints under the conditions of the problem. At a differentiable interior optimum without binding restrictions, the ordinary derivative can vanish. At a constrained optimum the relevant balance includes the restrictions, while an unused resource may carry no scarcity value at all. The familiar conditions of stationarity, feasibility and complementary slackness make those relationships explicit. Their necessity requires the relevant regularity conditions; their sufficiency follows for the appropriate convex problem. A constrained optimum is not a failed attempt to reach an unconstrained summit. It is the best result among the alternatives that have actually been admitted.
Sometimes the next unit cannot be made arbitrarily small. If the undertaking must choose whole units, its fifth unit adds three to total benefit and costs two, so it improves the result by one. The sixth adds only one of benefit and still costs two, reducing the result by one. Five remains the answer, but the reason is a comparison of available finite changes. A derivative is an exact description of the smooth function, not an exact description of a decision set containing only integers. When purchases come in fleets, production comes in shifts or construction comes in whole bridges, the finite alternatives must be compared. Smooth approximations may illuminate those choices, provided the approximation does not quietly erase the choice that matters.
A fixed outlay exposes another limit of looking only at the slope. Suppose beginning the same invented undertaking requires a payment of thirty that can still be avoided by staying out. The best operating gain remains twenty-five, at five units, but entry now loses five compared with doing nothing. If the entry payment were twenty, entry would instead yield five. The first-order equation within the operating region is unchanged in both cases. It has answered how much to produce conditional on operating, while leaving the decision to operate unsettled. Once the thirty has been irrecoverably paid, it drops out of the comparison between future operating choices. The same item changes its place in the decision because the available alternatives have changed, not because its historical amount has been revised.
The mathematical power is larger than a collection of prohibitions. A model can make alternatives visible that unaided intuition obscures. It can show that the best available quantity rises when a constraint loosens, identify when an extra unit of one resource is useless without another, and reveal that an apparent economy saves money only by shifting a cost outside the objective. It can also force incompatible aspirations into the open. A person cannot spend the same hour twice simply because both uses are valuable. By requiring an explicit feasible set, optimization can expose a promise whose attractive parts cannot all be delivered together. The instrument earns its place when it clarifies the actual comparison.
What it cannot furnish on its own is the reason for choosing the thing to maximize. Profit, leisure, output, health and a distribution judged fair are different objectives, even when each admits a mathematical representation. Weights attached to them do not become morally compulsory when a solution is unique. Nor does a constraint become natural because it is entered as fixed: a budget may reflect a genuine resource limit, a decision made elsewhere or an institution that could itself be changed. The derivative reports the consequences of the specification. It does not absolve the author of the specification from answering why these benefits count, why these burdens are excluded and why this set of alternatives is treated as available.
Even a faithfully stated objective may leave out part of the experience through which a choice is made. The worker in a smooth model knows how another hour will feel; the actual worker may discover it only by working. A commitment can change later wants, a skill can be learned through an initially unrewarding effort, and a decision can disclose information that changes the next decision. Such features can sometimes be brought into a richer model. Their absence from a simple one is therefore neither proof that mathematics is incapable of describing them nor permission to behave as though they were absent from the world. The question is whether the simplification preserves what is decisive for this choice.
The derivative can locate the best point in a stated world. Someone must still answer for the world that was stated.
The truth gained is a conditional one with unusual reach: given this objective, these alternatives and these relations, the comparison follows. Its conditions can be inspected, its calculation repeated and its mistakes exposed. The truth lost arises when that conditional statement is mistaken for the whole event, and the chooser's uncertain purposes and circumstances disappear behind the elegance of the answer. The marginal method remains valuable precisely because its limits can be stated. The next plate asks how much of desire the theory needs to measure at all. Perhaps a ranking can do the work once assigned to an intensity. Even then, the mathematical order and the life ordered will remain distinguishable. The argument remains open.