Is the dual function the dual problem?

No. The function assigns a value to multipliers; the dual problem optimizes that function under multiplier restrictions.

Can its value be negative infinity?

Yes, when the Lagrangian is unbounded below for those multipliers.

Does a dual optimum always recover a primal optimum?

No. Recovery and equality of optimal values require conditions beyond merely defining the function.

AI Glossary / Time series and signal processing / Advanced

Lagrange dual function

Definition

The Lagrange dual function is obtained by taking the infimum of a problem's Lagrangian over its original decision variables, while holding the multipliers fixed. For a minimization problem with appropriately signed inequality multipliers, its value supplies a lower bound on the optimal feasible objective value. Maximizing this bound defines the associated Lagrange dual problem.

What it means

A constrained optimization problem asks for the best decision that obeys its requirements. The dual function approaches the problem indirectly: assign weights to the constraints, minimize the resulting Lagrangian over the decision variables, and observe the bound that those weights produce.
This can help assess solution quality even when finding the best feasible decision is difficult. The dual function is a function of multipliers, not another name for the original objective or for the Lagrangian before minimization.

How it works

Write a minimization problem with objective f(x), inequalities gi(x) ≤ 0, and equalities hj(x) = 0. Form L(x, lambda, nu) = f(x) + sum lambda-i gi(x) + sum nu-j hj(x). Inequality multipliers are nonnegative for the usual lower-bound interpretation; equality multipliers are unrestricted.
Define q(lambda, nu) as the infimum of L over x in the objective and constraint functions' common domain, retaining any domain restrictions that have not been relaxed. The value can be negative infinity. Because it is a pointwise infimum of affine functions of the multipliers, q is concave, even when the original problem is not convex. Strong duality, meaning equal primal and dual optimal values, requires additional conditions.

Three practical examples

An optimization system compares the cost of a feasible schedule with a valid dual bound to assess how far the schedule could be from optimal.
A resource-allocation model uses multipliers to coordinate subproblems, with the dual function measuring the bound produced by the current multiplier choices.
A machine-learning researcher derives a dual formulation of a constrained training problem, checking domain and regularity conditions before claiming equivalence to the primal formulation.

Think of it this way

Think of finding a defensible floor beneath the cost of every allowed solution. Each multiplier choice can produce a different floor. Raising that floor improves the certificate, but it does not automatically supply a feasible solution sitting on top of it. A gap can remain between the best bound and the best attainable objective.

Worked example

Consider an original scalar problem: minimize (x − 3) squared subject to x ≤ 1, with x otherwise real. The best feasible point is x = 1, giving primal objective 4. Write the constraint as x − 1 ≤ 0 and use lambda ≥ 0.
The Lagrangian is L(x, lambda) = (x − 3) squared + lambda(x − 1). For fixed lambda, its derivative in x is 2(x − 3) + lambda, so the minimizing x is 3 − lambda/2. Substitution gives q(lambda) = 2lambda − lambda squared/4.
At lambda = 0, the bound is 0. At lambda = 2, the bound is 3. At lambda = 4, the bound is 4. Completing the square gives q(lambda) = 4 − (lambda − 4) squared/4, proving that its maximum is 4 at lambda = 4. This matches the feasible objective at x = 1, certifying optimality in this example.
At lambda = 2, the Lagrangian minimizer is x = 2, which violates the original constraint. Its dual value still provides a valid lower bound. This distinction shows why a dual-function evaluation must not be confused with a feasible primal solution.

Illustrative case study

Illustrative fictional case: Cedar Planning receives a feasible production schedule with objective cost 125 and a separately verified dual lower bound of 120. The team reports that the schedule's objective is at most five units above the unknown optimum, under the specified model.
It does not describe the lower bound as an executable schedule or assume the optimum is exactly 120. When model constraints change, both the feasible solution and bound are rechecked. This fictional case illustrates a mathematical certificate's scope rather than measured commercial savings.

Common misconceptions and mistakes

The dual function is concave, but that does not mean it is differentiable or finite everywhere. An infimum need not be attained, so replacing it casually with a minimum can hide an existence issue.
Weak duality does not require the primal problem to be convex, but strong duality is not automatic. Even convex problems need appropriate conditions for standard strong-duality guarantees. Also keep the inequality direction and multiplier sign consistent: reversing a constraint without adjusting the formulation can invalidate the claimed bound.

Business use

Dual functions can support optimization certificates, decomposition, and sensitivity analysis. They help distinguish a good-looking feasible answer from one with a quantified bound on its possible suboptimality.
Record which constraints are relaxed, the retained domain, multiplier restrictions, and how the infimum is evaluated. Approximate numerical minimization does not automatically give a certified lower bound on the infimum; its error must be controlled. Operational interpretation should also remain within the model's assumptions, rather than treating mathematical optimality as proof that the model captures every business requirement.

Check your understanding

What is the example's dual value at lambda = 2?

Answer: Three.

Why is that value not the original objective at a feasible optimum?

Answer: It comes from minimizing the Lagrangian, whose minimizer at that lambda violates x ≤ 1.

What certifies the optimum here?

Answer: A feasible objective value of four and a matching valid dual lower bound of four.

Related terms

Lagrangian
Dual problem
Duality gap
Convex optimization
Lagrange multiplier
primal problem
weak duality
strong duality
complementary slackness

Learn next

derive the same bound by completing the square before taking the infimum. Then study a problem with an equality constraint and distinguish unrestricted equality multipliers from nonnegative inequality multipliers.

Sources

Sources: Jung and colleagues, The Aalto Dictionary of Machine Learning, Lagrange dual function: https://aaltodictionaryofml.github.io/ADictML_Main.pdf . Boyd and Vandenberghe, Convex Optimization, chapter 5: https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf . Primary dictionary and textbook sections consulted. The scalar derivation and fictional certificate example are original and independently checked. Additional source checked during independent review: Boyd and Vandenberghe, official lecture slides, Duality sections 5-2, 5-3 and 5-21 onward: https://web.stanford.edu/~boyd/cvxbook/bv_cvxslides_original.pdf . The full textbook link timed out during this review.

Previous: Lagged features Next: Lagrangian