Mathematical Optimization

Must an Optimum Satisfy KKT? Intuition and Proofs for LICQ, MFCQ, and Slater

A constraint can block all feasible motion while its gradient says nothing. Ramp limits and short multiplier proofs separate KKT necessity, convex sufficiency, and optimizer differentiability.

An optimum need not satisfy KKT. For a continuously differentiable problem, a local minimum admits KKT multipliers under a suitable constraint qualification (CQ). Convexity answers a different question: whether an existing KKT point is globally optimal.

The distinction becomes concrete through a ramp limit, a degenerate constraint, and short proofs for LICQ, MFCQ, and Slater.

1. A ramp limit: why the gradient need not vanish

Suppose increasing equipment output improves the objective, but this step’s hard ramp limit allows only outputs between 0.4 and 0.6. Minimizing the negative output selects the upper bound:

min.4≤a≤.6(−a)a*=.6,λ=1,−1+λ=0

The objective gradient is −1 even at the optimum. Increasing output would still improve the objective, but the active upper bound blocks that direction. Its gradient is +1. A multiplier of 1 balances the objective gradient in the stationarity equation. The lower bound has slack, so its multiplier is zero.

KKT expresses this balance with the gradients of the active constraints. It uses first-order information. That is also its limitation: a constraint may block actual motion without its gradient expressing the obstruction.

2. A global optimum that fails KKT

minxx,subject tox2≤0

The only feasible point is x* = 0. It is the unique global minimum, and both the objective and constraint function are convex. Yet stationarity cannot hold:

L(x,λ)=x+λx21+2λx=0⇒1+2λ·0=0

At zero, every finite multiplier leaves the left side equal to 1. The actual constraint prevents any motion, whereas its first-order approximation admits every direction:

g(0)+g′(0)d=0≤0

Here the objective decreases by a first-order amount when moving left; infeasibility appears through the squared displacement. The derivative of the constraint at zero misses that second-order obstruction.

Write the same feasible set as the equality x = 0 instead. Its gradient is 1, and a multiplier of −1 gives stationarity:

L(x,ν)=x+νx,1+ν=0⇒ν=−1

Thus the existence of KKT multipliers can depend on the constraint representation, even when the objective and feasible set are unchanged.

3. What “the optimum satisfies KKT” means

Consider real continuous variables and functions that are C1 in a neighborhood of the candidate:

minx∈ℝnf(x)gi(x)≤0, i = 1, …, mhj(x)=0, j = 1, …, p

The statement means that there exist nonnegative inequality multipliers and unrestricted equality multipliers satisfying all four conditions. Stationarity is

∇f(x*)+∑iλi∇gi(x*)+∑jνj∇hj(x*)=0

The remaining conditions are primal feasibility, dual feasibility, and complementary slackness:

gi(x*)≤0,hj(x*)=0λ≥0,λigi(x*)=0for every i

An inequality is active when it holds with equality. Define

I(x*)={i:gi(x*)=0}

Complementary slackness sets the multiplier of every inactive inequality to zero. It does not require an active inequality to have a positive multiplier.

4. Linear independence: at least one nonzero coefficient

A family of vectors is linearly independent if the only combination yielding the zero vector uses zero for every coefficient:

∑kckvk=0⇒every coefficient is zero

Dependence means there is a coefficient vector with at least one nonzero component whose combination is zero. It does not require every coefficient to be nonzero. For example,

v1=(1,0),v2=(0,1),v3=(2,0)2v1+0v2−v3=0

The coefficients (2, 0, −1) are a nonzero vector. The zero middle coefficient does not rescue independence. With vectors arranged as columns, independence is equivalent to full column rank.

5. The shared proof: exclude an abnormal Fritz John multiplier

For a smooth problem, a local minimum satisfies the Fritz John necessary conditions without a CQ. Evaluate the gradients below at x*, restrict the inequality sum to the active set, and set inactive multipliers to zero:

α∇f+∑i∈Iλi∇gi+∑jνj∇hj=0α≥0,λ≥0,(α,λ,ν)≠0

Feasibility and complementary slackness also hold. The difference from KKT is the extra objective multiplier α. If it is positive, divide all multipliers by it:

∇f+∑i∈Iλiα∇gi+∑jνjα∇hj=0

This is KKT. The difficulty is the abnormal case α = 0: the objective disappears and constraint gradients alone satisfy a nontrivial cancellation relation. LICQ and MFCQ rule out that case.

Why does Fritz John appear? With no equalities, a direction that strictly decreases the objective and every active constraint would contradict local minimality:

∇f⊤d<0,∇gi⊤d<0for i ∈ I

For a sufficiently small positive step, the active inequalities become strict, inactive inequalities keep their slack by continuity, and the objective decreases. A theorem of alternatives converts the absence of such a direction into a nonzero nonnegative combination of these gradients equal to zero. That combination is Fritz John.

For nonlinear equalities, satisfying their linearized equations alone does not make a straight step feasible. If their gradients are independent, the implicit function theorem supplies feasible curves with the required tangent. If the equality gradients are dependent, their dependence already supplies an abnormal Fritz John relation. The unrestricted equality multipliers reflect these equality directions.

LICQ: independence prevents cancellation

LICQ requires the following entire family to be linearly independent:

{∇gi:i∈I}∪{∇hj:j = 1, …, p}

Suppose the Fritz John objective multiplier were zero. Then

∑i∈Iλi∇gi+∑jνj∇hj=0

Independence forces every coefficient to be zero. Together with α = 0, that contradicts the nonzero-multiplier condition. Hence α is positive and normalization gives KKT.

LICQ also makes the KKT multipliers unique at this point: subtracting two stationarity equations would give another zero combination of the same independent gradients. Neither argument requires a convex objective or convex constraints.

MFCQ: project the cancellation onto an inward direction

MFCQ requires independent equality gradients and a direction d such that

∇hj⊤d=0for every j∇gi⊤d<0for every i ∈ I

The direction preserves the equalities to first order and moves strictly inward for every active inequality. It need not improve the objective.

Again suppose α = 0 in Fritz John. Take the inner product with d. The equality terms vanish, leaving

∑i∈Iλi∇gi⊤d=0

Every inequality multiplier is nonnegative and every corresponding directional derivative is strictly negative. A positive multiplier would make the sum strictly negative. Therefore all inequality multipliers are zero. Independence of the equality gradients then forces all equality multipliers to zero. This contradicts nontriviality, so normalization gives KKT.

MFCQ permits some dependence among inequality gradients. For example,

minx(−x),x≤0,2x≤0d=−1,1d<0,2d<0−1+λ1+2λ2=0

At the optimum zero, the active gradients 1 and 2 are dependent, so LICQ fails. The direction −1 still moves inward for both, so MFCQ holds. The KKT multipliers can be (1, 0) or (0, 1/2). Dependence need not destroy KKT; it can merely destroy multiplier uniqueness. LICQ implies MFCQ, but the reverse implication fails.

Slater: convexity produces the inward direction

Assume the inequality functions are differentiable and convex, the equalities are affine, and a strict feasible point exists:

gi(x¯)<0for every iAx¯=b

For this MFCQ-based proof, remove redundant equality rows first. Feasibility ensures their right sides are consistent, so this does not change the feasible set. The remaining equality matrix has full row rank. Later, setting removed equality multipliers to zero lifts the obtained KKT relation back to the original representation.

From a candidate local minimum, choose

d=x¯−x*,Ad=0

The equality relation follows because both endpoints satisfy the same affine equations. A differentiable convex function lies above its tangent plane. For every active inequality this gives

∇gi(x*)⊤d≤gi(x¯)−gi(x*)gi(x*)=0⇒∇gi(x*)⊤d≤gi(x¯)<0

The direction to the strict feasible point therefore decreases every active constraint already at first order. Apply the MFCQ argument to obtain KKT.

This necessity proof uses convexity of the constraint functions, not of the objective. A nonconvex smooth objective can also have KKT multipliers at a local minimum under this constraint structure. Convexity of the objective is needed for the sufficiency result below.

Strict feasibility without that structure is insufficient:

minx(−x),x3≤0x*=0,x¯=−1,−1+λ·0≠0

The feasible set is the half-line x ≤ 0, the optimum is zero, and −1 is a strict feasible point. But the constraint derivative at zero vanishes, making stationarity impossible. The cubic constraint function is not convex on the real line. Even a convex feasible set does not repair this nonlinear representation.

6. Necessity and sufficiency are different directions

Assumptions Valid conclusion
Smooth functions, local minimum, suitable CQ KKT multipliers exist; convexity is unnecessary.
Smooth convex objective and inequalities, affine equalities, a KKT pair The primal point is a global minimum; Slater is unnecessary for this direction.
Smooth convex problem with Slater, and an attained optimum Global optimality is equivalent to the existence of KKT multipliers.
Nonconvex problem, KKT alone The point may be a local minimum, maximum, or saddle.

For example, zero is a KKT point of the following problem, with both bound multipliers zero, but is a maximum:

min−1≤x≤1(−x2),f′(0)=0

Convex sufficiency has a short proof. Fix a KKT multiplier pair. The Lagrangian is convex in x, so its stationary point minimizes it globally. For any feasible x,

f(x)≥L(x,λ*,ν*)L(x,λ*,ν*)≥L(x*,λ*,ν*)=f(x*)

Feasibility and nonnegative multipliers give the first inequality. Convexity and stationarity give the second. Complementary slackness gives the equality at the candidate. This proves global optimality without invoking Slater.

Failure of LICQ, MFCQ, or the strict-feasibility version of Slater removes those sufficient guarantees of multiplier existence. It does not by itself prove that KKT fails.

7. Affine constraints: why LP and convex QP are a useful special case

With only affine constraints, any attained local minimum of a differentiable objective admits KKT multipliers:

Ax=b,Gx≤h∇f(x*)+A⊤ν+G⊤λ=0

To see the reason, linearized feasible directions satisfy the actual constraints for sufficiently small positive steps. There can be no feasible direction with a negative objective directional derivative at a local minimum. Farkas’ lemma then represents the negative objective gradient as a nonnegative combination of active inequality normals plus an unrestricted combination of equality normals. This is stationarity; assigning zero multipliers to inactive inequalities gives complementarity.

This polyhedral argument needs neither LICQ nor a strictly feasible point. For example,

minxx,x≤0,−x≤0x*=0,λ1=0,λ2=11+λ1−λ2=0

Both constraints are active and their gradients are dependent. No point satisfies both inequalities strictly. Nevertheless, the displayed multipliers satisfy KKT.

For a QP with a symmetric positive semidefinite Hessian and affine constraints, KKT is therefore necessary and sufficient for an attained optimum. A positive definite Hessian also makes the primal optimum unique whenever the feasible set is nonempty. Redundant constraints can still make the multipliers nonunique.

8. KKT does not guarantee that the optimizer is differentiable

Consider a parameterized ramp QP:

a*(η)=arg min.4≤a≤.6(12a2−ηa)a*(η)=min{.6,max{.4,η}}

For every η, the optimum is unique and admits KKT multipliers. Near η = 0.6, the action derivative is 1 on the left and 0 on the right. The ordinary derivative at the boundary does not exist. At that point, the upper bound is active but λ*upper = 0, so strict complementarity fails.

For implicit differentiation, collect the primal and dual variables in y* and write the equality part of KKT as a residual:

F(y*(η),η)=0Fy∂y*∂η=−Fη

Using the inverse formula requires a continuously differentiable residual and a nonsingular Fy at the solution. With sufficiently smooth problem data, LICQ, a second-order sufficient condition, and strict complementarity form a familiar sufficient set for a smooth local primal–dual solution branch. They are sufficient conditions, not an assertion that each is necessary for every differentiable primal solution map.

There are three separate checks: whether KKT multipliers exist, whether KKT certifies the intended optimum, and whether the chosen solution branch can be differentiated.

References

  • Boyd, S., and Vandenberghe, L. (2004). Convex Optimization, Chapter 5, especially §§5.2.3 and 5.5.3. Authors’ official book page.
  • Farina, G. (2025). Lagrange multipliers and KKT conditions, MIT 6.7220/15.084, Lecture 7. Affine normal cones, LICQ, and Slater. Official lecture PDF.
  • Sun, Y. (2024). Notes on constraint qualifications, STATS 606. Fritz John and MFCQ. Official course note.
  • Amos, B., and Kolter, J. Z. (2017). OptNet: Differentiable Optimization as a Layer in Neural Networks. ICML, PMLR 70, 136–145. Paper and metadata.