A multiplier is a local rate, under hypotheses
An optimization model often contains a prescribed resource level . Once we have solved the problem at one value, must we solve it again from the beginning when changes slightly? The multiplier can predict the first-order change in the optimal value, but only after we say which optimum is being followed and why it moves smoothly. We will prove this for a regular, strict local minimum. A later example shows why a global value need not inherit the same derivative.
Let and be twice continuously differentiable near , and consider
The right-hand side is the parameter. Use the Lagrangian . At suppose is stationary with multiplier , so . Write and . In particular, includes the curvature of the constraints, not only that of .
Theorem 2.3.1
(A smooth branch of strict local minima)
Assume has full row rank and for every nonzero satisfying . Then, for near , there is a unique stationary pair near , depending continuously differentiably on . Each is a strict local minimum. The branch value satisfies
This is a statement about the selected local branch; it does not assert that this branch is globally optimal.
Proof
(Implicit functions from the bordered system)
The stationarity and feasibility equations , have, with respect to at the base point, the derivative
If , then and . Taking the inner product of the second equation with gives . Positive definiteness on forces . Then , and full row rank forces . Thus the square matrix is nonsingular, and the implicit function theorem gives the unique stationary branch near . Full row rank of and positive definiteness of on persist in a sufficiently small neighborhood by continuity, so the equality-constrained second-order sufficient condition makes each branch point a strict local minimum.
Differentiating gives . The chain rule and stationarity now yield
Proof X-ray
Where each hypothesis enters
Full row rank is used twice: once to remove from the kernel of , and again, through , to turn stationarity into the value derivative. Second-order positivity is needed only on , and the one non-obvious move is the inner product with : because , the multiplier term vanishes and only the tangent curvature remains. The value formula then closes without any optimization argument, since along the branch. Nothing in the proof compares the branch with distant feasible points, so global optimality is neither used nor obtained; and if loses rank, as in the last worked example below, the first step already fails.
For a small change , the result reads . With our sign convention the predicted marginal cost is minus the multiplier.
Common mistake
(Reading a multiplier’s sign without the convention)
Saying “a positive multiplier means the value increases with ” is meaningless until the constraint is written down. With the value derivative is ; with the multiplier changes sign and the derivative is . Check the convention before interpreting a sign.
Computing the response
Differentiate stationarity and feasibility in a direction . At the base point, the derivatives and solve
The objective and constraint functions here do not depend directly on except through the displayed right-hand side. For a parameter occurring elsewhere, differentiating the corresponding equations produces additional terms. Since , the local value has Hessian .
When is positive definite on the whole space, the bordered system can be solved by elimination.
Proposition 2.3.2
(Response matrices when )
Under the hypotheses of the theorem, assume in addition on . Then is positive definite and
Proof
(Eliminating the primal response)
For , full row rank gives , so ; hence is positive definite and invertible. The first block row of the response system gives . Substituting into gives , so and . Since is arbitrary, these are the stated derivatives, and gives the last formula.
This formula is convenient, but is stronger than the theorem requires: the bordered system can be nonsingular even when is singular or indefinite off the tangent space.
Common mistake
(Using the Hessian of the objective instead of the Lagrangian)
The curvature condition is on , not on . With a linear objective and a curved constraint, as in the circle example below, and the second-order test would wrongly appear to fail; all the curvature comes from the constraint term.
Worked examples
Worked example 2.3.3
(Splitting a required total)
Minimize subject to . Stationarity gives and , hence , and . Therefore and . Here and , so and . The Hessian is positive definite, so this feasible stationary point is the unique global minimum for each .
Worked example 2.3.4
(Curvature supplied by the constraint)
Minimize subject to , with . Stationarity gives and , so , and feasibility gives . The two stationary points are , with .
At we have and , so the theorem applies. The branch value is , and
At , and , so and the proposition predicts ; directly, gives the same number. The other stationary point has and , so the tangent curvature is negative and it is the maximizer, not a branch of minima.
The next two items show what the theorem does not cover.
Counterexample mode
Smooth local branches can exchange global optimality
Minimize subject to . At , both and are strict local minima, with value zero; their equality gradients have full rank and their tangent curvatures are positive. Each therefore has a smooth local branch. Yet the global value is not differentiable at zero. Indeed , while the feasible points and show . If had a derivative at zero, its right difference quotients would force that derivative to be at most , whereas its left difference quotients would force it to be at least . The contradiction reflects competing local branches; neither local multiplier alone differentiates the global value at the switch.
A second failure is loss of regularity in a more general parametrization.
Worked example 2.3.5
(When the constraint gradient vanishes)
Minimize subject to . The feasible set is , so the global value is , which is not differentiable at . At the only feasible point is , where : full row rank fails. Worse, stationarity has no solution at , so no multiplier exists at all. A multiplier formula must not be applied merely because the value is finite.
Practice with solutions
Quick check
(Minimize subject to . Find the solution, multiplier, value, and value derivative.)
Use the convention .
Solution · Answer
Stationarity gives , , and feasibility gives , so . Thus and , with . The strictly convex objective gives a unique global minimum.
Quick check
(Minimize subject to . Verify the hypotheses of the theorem at , although the objective Hessian is indefinite, and compute the branch value.)
Test the curvature only on .
Solution · Answer
Here and . On the quadratic form is for . The feasible problem has and is minimized at , so , with and . The negative curvature does not contradict the theorem, which requires positivity only along tangent directions; the proposition with does not apply, since .
Quick check
(Minimize subject to . Write the constraint as , find its multiplier, and compare it with the derivative of the value.)
Form .
Solution · Answer
Stationarity for gives , so . The value is , whose derivative is . With the original convention , the multiplier would be and the derivative .
Quick check
(For the problem in the first worked example, use the derivative at to estimate the value change when increases by . Compare it with the exact change.)
Recall .
Solution · Answer
, so the first-order estimate is . Exactly, . The discrepancy is the quadratic term, not a failure of the first-order formula.