Foundation
The Karush–Kuhn–Tucker (KKT) conditions are a set of first-order necessary conditions that any local optimum of a constrained optimization problem must satisfy. They generalize the classical Lagrange multiplier method to handle inequality constraints, and their dual variables carry an economic interpretation as shadow prices — the marginal value of relaxing each constraint. In power-system operations, the KKT multipliers of OPF bus-balance constraints are the locational marginal prices (LMPs) at the heart of every modern electricity market.
Consider a general constrained optimization problem. The objective f, the inequality constraints gi, and the equality constraints hj are continuously differentiable functions. The feasible set is the intersection of all constraint regions. A point x* is a local minimum if f(x*) ≤ f(x) for all feasible x in a neighborhood of x*. We want necessary conditions that x* must satisfy — conditions that turn an infinite-dimensional search problem into a finite system of equations.
Introduce dual variables: λi ≥ 0 for each inequality and μj ∈ ℜ for each equality. The Lagrangian function combines the objective and constraints into a single expression. Each λi prices the inequality gi ≤ 0: if the constraint is binding (gi = 0 at the optimum), λi tells us how much the optimal cost would change per unit relaxation of that constraint. Each μj prices the equality hj = 0 similarly. The Lagrangian is the engine that converts constraint information into gradient information.
At any local optimum x* (under standard regularity conditions called constraint qualifications), there exist multipliers λ* ≥ 0 and μ* such that the following four conditions hold simultaneously.
The gradient of the Lagrangian with respect to x vanishes. Geometrically: the gradient of f at the optimum lies in the span of the constraint gradients. There is no direction in which we can move to decrease f without violating constraints.
The constraints themselves must hold at the optimum. The optimum is a feasible point.
Inequality multipliers are non-negative. This reflects the economic interpretation: relaxing a binding constraint (allowing more) can only improve the optimal cost, never worsen it.
For each inequality, either the multiplier is zero or the constraint is active. If a constraint is non-binding (gi < 0 at the optimum, room to spare), its multiplier is zero — relaxing it does not help. If a constraint is binding (gi = 0), the multiplier can be positive, capturing the marginal benefit of further relaxation. Either way, the product vanishes.
The KKT conditions are necessary for any local optimum, but they are also sufficient when the problem is convex: f and the gi are convex, the hj are affine. Under convexity, any KKT point is automatically a global optimum. This is the key reason linear programs, quadratic programs, and second-order cone programs admit polynomial-time algorithms — solving the KKT system is equivalent to finding the global optimum.
For non-convex problems like AC OPF, KKT remains necessary but not sufficient: multiple KKT points may exist, some corresponding to local minima, others to saddle points or local maxima. Interior-point solvers like Ipopt converge to a single KKT point, with no guarantee it is the global optimum — a key motivation for convex relaxations such as SOCWR and SOCBF.
The Lagrangian leads naturally to a dual optimization problem. Define the dual function as the infimum of the Lagrangian over x. For any feasible primal x and any λ ≥ 0, μ, the Lagrangian is a lower bound on f(x), and hence q(λ, μ) is a lower bound on the primal optimal. The dual problem maximizes this bound.
Weak duality always holds: dual optimal ≤ primal optimal. Strong duality (equality) holds for convex problems under regularity conditions (e.g., Slater’s condition: strict feasibility). When strong duality holds, solving the dual is equivalent to solving the primal, and the optimal dual variables are exactly the KKT multipliers.
For DC OPF and other convex OPF formulations, strong duality holds and the KKT multipliers of the bus-balance constraints have a direct economic interpretation: the LMP at bus i is the marginal cost of delivering one additional unit of power to that specific bus, in equilibrium with the dispatch and network constraints. LMPs vary spatially when transmission constraints bind — congested networks exhibit price differences across the grid that signal where new generation or transmission is most valuable.
For non-convex AC OPF, the duality gap is generally nonzero and LMPs computed from KKT multipliers are only approximate. Convex relaxations (SOCWR, SOCBF) yield exact LMPs when the relaxation is tight, and approximate but consistent LMPs otherwise. Wholesale electricity markets typically clear on DC OPF for tractability, then settle on AC for accuracy — accepting the LMP approximation as a practical compromise.
The KKT conditions are necessary under constraint qualifications — technical conditions ensuring that the constraint geometry is well-behaved near the optimum. The most common are:
In practice, well-posed OPF formulations satisfy these conditions. Pathological cases (redundant constraints, degenerate networks) may fail LICQ and cause numerical solvers to stall or produce non-unique multipliers — sources of LMP instability in real-time markets.