Chapter 12
Optimization with inequality
constraints
Here we want to solve the following constrained maximization problem:
Maximize f(x)
subject togj(x) 0, j = 1, . . . , m, and x2X,
whereXis a non-empty subset ofRnandf,gj(·),j = 1, . . . , m, are functions
fromX toR. Here the constraint set is given by
C={x2X :gj(x) 0, j = 1, . . . , m}.
A point x⇤ 2 X is a constrained local maximum if there exists an open
ball B✏(x⇤) ⇢ Rn such that f(x⇤) f(x) for all x 2 B✏(x⇤)\C; x⇤ is a
constrained global maximum if it solves the problem above. A constrained minimization problem and its local and global minima can be defined anal-ogously.
12.1
Saddle point
Definition 12.1 (Saddle point). A pair (x⇤, ⇤)2X⇥Rn
+ is a saddle point if
(x, ⇤) (x⇤, ⇤) (x⇤, )
for all x2X and 2Rn
+, where
(x, ) = f(x) + g(x).
92CHAPTER 12. OPTIMIZATION WITH INEQUALITY CONSTRAINTS
12.1.1
Constrained global maximum and saddle points
The following result shows the equivalence of a constrained global maximum and a saddle point.
Theorem 12.1. If (x⇤, ⇤)2X⇥Rm
+ is a saddle point, then
• ⇤g(x⇤) = 0; • g(x⇤) 0; and
• x⇤ is a point of constrained global maximum.
Proof. By the second inequality in the definition of a saddle point, for all
2Rn
+
(x⇤, ) =f(x⇤) + g(x⇤) (x⇤, ⇤)
=f(x⇤) + ⇤g(x⇤) or, g(x⇤) ⇤g(x⇤)
or, ( ⇤)g(x⇤) 0.
Pick any j 2 {1, . . . , m}, let j = j⇤ + 1 and i = ⇤i for all i 6= j. This
gives gj(x⇤) 0. Repeating this procedure, we obtain gj(x⇤) 0 for all
j = 1, . . . , m. Consequently, ⇤·g(x⇤) 0. On the other hand, setting = 0
above, we get ⇤ ·g(x⇤)0. Therefore, ⇤·g(x⇤) = 0. Now, for all x2C, f(x)f(x) + ⇤g(x)
f(x⇤) + ⇤g(x⇤) =f(x⇤).
The first inequality holds since ⇤ 2 Rn
+ and gj(x) 0, j = 1, . . . , m. The second inequality holds by the first inequality in the definition of a saddle point. The equality holds because ⇤g(x⇤) = 0.
A converse of this result also holds as follows.
Theorem 12.2. Suppose x⇤ 2 X is a constrained global maximum where
X is convex,f and gjs, j = 1, . . . , m, are concave functions, and there exists
¯
x2X such thatgj(¯x)>0,j = 1, . . . , m(Slater’s condition). There exists a ⇤ 2Rn
+ such that
• ⇤g(x⇤) = 0, and
• (x⇤, ⇤) is a saddle point.
12.2
Kuhn-Tucker Conditions and Saddle Points
LetX ⇢Rn be open and f, gj,j = 1, . . . , m be C1 functions from X to R. A pair (x⇤, ⇤)2X⇥Rm
+ satisfies the Kuhn-Tucker conditions if
• @@xfi(x
⇤) +Pm j=1 ⇤j@g
j @xi(x
⇤) = 0, i= 1, . . . , n and
• g(x⇤) 0 and ⇤g(x⇤) = 0.
The condition ⇤g(x⇤) = 0 is known as the complementary slackness
condi-tion. It is equivalent to ⇤jgj(x⇤) = 0 for each j = 1, . . . , m. Hence it implies
that if gj(x⇤) > 0 then j⇤ = 0 but if gj(x⇤) = 0 then ⇤j can be either zero
or strictly positive.
The following result uses the fact that if X be an open convex set and f, gj, j = 1, . . . , m are C1 and concave functions from X to R, then any (x⇤, ⇤) 2 X ⇥ Rm
+ satisfying the Kuhn-Tucker conditions is maximizing (x, ⇤) =f(x) + ⇤g(x). This fact and the complementary slackness condi-tions imply that (x⇤, ⇤) must be a saddle point.
Theorem 12.3. LetX be an open convex set and let f,gj, j = 1, . . . , m be
C1 and concave functions fromX toR. If a pair (x⇤, ⇤)2X⇥Rm
+ satisfies the Kuhn-Tucker conditions then
• (x⇤, ⇤) is a saddle point and
• x⇤ is a point of constrained global maximum.
The converse is also easy to prove and left as an exercise.
Exercise 12.2.1. Let X be an open set and let f, gj, j = 1, . . . , m be C1 functions fromX toR. If a pair (x⇤, ⇤)2X⇥Rm
+ is a saddle point then it satisfies the Kuhn-Tucker conditions.
12.3
Necessary and su
ffi
cient conditions for
constrained local maximum
Now we show that if x⇤ 2 X is a constrained local maximum, then under suitable conditions, there exists ⇤ 2 Rm+ such that (x⇤, ⇤) satisfies the Kuhn-Tucker conditions.
Theorem 12.4. Let X ⇢ Rn be open and f, gj be C1 functions from X to R. Suppose x⇤ 2 X is a constrained local maximum of f subject to the k inequality constraints gj(x) b
94CHAPTER 12. OPTIMIZATION WITH INEQUALITY CONSTRAINTS
generality, suppose that the first k0 of these constraints are binding and the remainingk k0 constraints are not binding. Further suppose that the Jacobian matrix corresponding to thek0 binding constraints has full rankk0. Form the Lagrangian
L(x, ) =f(x)
k
X
j=1
j(gj(x) bj).
Then there exist multipliers ⇤
1, . . . , ⇤ksatisfying the Kuhn-Tucker conditions,
i.e.,
• @L
@xi(x
⇤, ⇤) = 0, i= 1, . . . , n;
• ⇤
j(gj(x⇤, ⇤) bj) = 0, j = 1, . . . , k;
• ⇤
j 0,j = 1, . . . , k;
• gj(x⇤) b
j 0, j = 1, . . . , k.
Proof. Since the gjs are continuous functions, there exists an open ball B
around x⇤ such that for all x 2 B, gj(x) < b
j for j = k0 + 1, . . . , k. Note that x⇤ maximizes f on B for the constraint set {x 2 B : gj(x) = b
j, j =
1, . . . , k0}(If there was anotherx⇤⇤ 2Bfor whichfhad a higher value on this constraint set, then this would give a higher value for the original constrained maximization problem in B, contradicting that x⇤ is a local maximum).
Since this is a local maximum subject to equality constraints, and the non-degenerate constraint qualification corresponding to this problem is satisfied (the Jacobian matrix corresponding to these constraints has full rank), there existµ⇤1, . . . , µ⇤k0 such that
• @Lˆ
@xi(x
⇤, µ⇤) = 0, i= 1, . . . , n;
• gj⇤(x⇤) =b
j, j = 1, . . . , k0,
where ˆLx, µ=f(x) Pk0
j=1µj(gj(x) bj)).
Now consider the Lagrangian for the original problem, L(x, ) =f(x) Pk
j=1 j(gj(x) bj). Set ⇤j = µ⇤j for j = 1, . . . , k0 and ⇤j = 0 for j =
k0+ 1, . . . , k. Then observe that (x⇤, ⇤) satisfies all Kuhn-Tucker conditions
except one, ⇤j 0, j = 1, . . . , k0. We now prove that this condition is also satisfied.
Consider the system of k0 equations gj(x) = b
j, j = 1, . . . , k0 in k0 +n
small, we can find x such that g1(x) = b1 t and gj(x) = b
j, j = 2, . . . , k0.
This implies that there exists aC1 curvex(t) fort2[0,✏) such thatx(0) =x⇤
and g1(x(t)) = b1 t and gj(x(t)) =b
j, j = 2, . . . , k0. Let v=x0(0). By the
chain rule, Dg1(x⇤)v= 1 and Dgj(x⇤)v= 0, j = 2, . . . , k0. Since x(t) lies
in the constraint set for all t 2 [0,✏) and x⇤ maximizes f in the constraint
set, f must be nonincreasing along x(t). Therefore,
d
dtf(x(t)) t=0 =Df(x
⇤)v0.
LetDxL(x⇤) be the derivative ofL(x, ) evaluated at x⇤. Sincex⇤ is a local
maximum of f subject to the k equality constraints,
0 =DxL(x⇤)v
=Df(x⇤)v
k
X
j=1
jDgj(x⇤)v
=Df(x⇤)v 1Dg1(x⇤)v =Df(x⇤)v+ 1.
Since Df(x⇤)v 0,
1 0. Similarly, we can show that j 0 for j =
2, . . . , k0.
An analogous statement can be proven for Kuhn-Tucker conditions be-ing necessary for constrained local minimum. It must be noted that the inequality constraints here are of the form gj(x) b
j.
Exercise 12.3.1. State and prove the version of Theorem 12.4 when some of the constraints are equality constraints.
The Kuhn-Tucker conditions are also sufficient for a local maximum pro-vided the relevant second order condition is satisfied: supposeg representsk inequality constraints andh representsm equality constraints and let and µ be the corresponding multipliers. Suppose (x⇤, ⇤, µ⇤) satisfy the
Kuhn-Tucker conditions (with mixed constraints). Further, the Hessian matrix of the Lagrangian with respect toxat (x⇤, ⇤, µ⇤) is negative definite at the
96CHAPTER 12. OPTIMIZATION WITH INEQUALITY CONSTRAINTS
An analogous statement can be proven for Kuhn-Tucker conditions being sufficient for constrained local minimum provided the relevant second order condition is satisfied. This second order condition is that the Hessian matrix of the Lagrangian with respect to x at (x⇤, ⇤, µ⇤) is positive definite at the linear constraint set {v 6= 0 : Dgk0(x⇤)v = 0, Dh(x⇤)v = 0} where gk0 represent the binding inequality constraints. It must be noted that the inequality constraints here are of the form gj(x) b
j. This holds when the