Lectures 10 and 11: Constrained optimization

problems and their optimality conditions

Coralia Cartis, Mathematical Institute, University of Oxford

C6.2/B2: Continuous Optimization

Minor additions by Yuji Nakatsukasa

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 1/25
Problems and solutions

minimize f (x) subject to x ∈ Ω ⊆ Rn .

f : Ω → R is (sufficiently) smooth.
f objective; x variables.
Ω feasible set determined by finitely many (equality and/or
inequality) constraints.

x∗ global minimizer of f over Ω =⇒ f (x) ≥ f (x∗ ), ∀x ∈ Ω.

x∗ local minimizer of f over Ω =⇒
∃N (x∗ , δ) such that f (x) ≥ f (x∗ ), for all x ∈ Ω ∩ N (x∗ , δ).
• N (x∗ , δ) := {x ∈ Rn : )x − x∗ ) ≤ δ}.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 2/25
Example problem in one dimension

Example : min f (x) subject to a ≤ x ≤ b.

f(x)

x1 x2 x
a b
The feasible region Ω is the interval [a, b].
The point x1 is the global minimizer; x2 is a local
(non-global) minimizer; x = a is a constrained local minimizer.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 3/25
An example of a nonlinear constrained problem
2
√
min2 (x1 − 2) + (x2 − 0.5(3 − 5))2 subject to
x∈R

−x1 − x2 + 1 ≥ 0, x2 − x21 ≥ 0.
2.5

2 c
2

1.5
c1 Ω
1 contours of f

0.5 x∗
x2

−0.5

−1

−1.5

−2
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 3

∗
√ √ x1

x = 0.5(−1 + 5, 3 − 5); Ω feasible set.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 4/25
Optimality conditions for constrained problems

== algebraic characterizations of solutions −→ suitable for

computations.
provide a way to guarantee that a candidate point is optimal
(sufficient conditions)
indicate when a point is not optimal
(necessary conditions)

minimizex∈Rn f (x) subject to cE (x) = 0, cI (x) ≥ 0.

(CP)
f : Rn → R, cE : Rn → Rm and cI : Rn → Rp (suff.) smooth;
• cI (x) ≥ 0 ⇔ ci (x) ≥ 0, i ∈ I .
• Ω := {x : cE (x) = 0, cI (x) ≥ 0} feasible set of the problem.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 5/25
Optimality conditions for constrained problems

unconstrained problem −→ x̂ stationary point (∇f (x̂) = 0).

constrained problem −→ x̂ Karush-Kuhn-Tucker (KKT) point.
Definition: x̂ KKT point of (CP) if there exist ŷ ∈ Rm and
λ̂ ∈ Rp such that (x̂, ŷ, λ̂) satisfies
! !
∇f (x̂) = ŷj ∇cj (x̂) + λ̂i ∇ci (x̂),
j∈E i∈I

cE (x̂) = 0, cI (x̂) ≥ 0,
λ̂i ≥ 0, λ̂i ci (x̂) = 0, for all i ∈ I.

• Let A := E ∪ {i ∈ I : ci (x̂) = 0} index set of active constraints

at x̂; cj (x̂) > 0 inactive constraint at x̂ ⇒ λ̂j = 0. Then
i∈I∩A λ̂i ∇ci (x̂).
" "
i∈I λ̂i ∇ci (x̂) =
• J (x) = ∇ci (x)T i Jacobian matrix of constraints c. Thus
# $

j∈E ŷj ∇cj (x̂) = JE (x) ŷ and i∈I λ̂i ∇ci (x̂) = JI (x) λ̂.
T T
" "

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 6/25
Optimality conditions for constrained problems ...

x̂ KKT point −→ ŷ and λ̂ Lagrange multipliers of the equality

and inequality constraints, respectively.
ŷ and λ̂ −→ sensitivity analysis.

L : Rn × Rm × Rp → R Lagrangian function of (CP),

L(x, y, λ) := f (x) − y # cE (x) − λ# cI (x), x ∈ Rn .

Thus ∇x L(x, y, λ) = ∇f (x) − JE (x)# y − JI (x)# λ,

and x̂ KKT point of (CP) =⇒ ∇x L(x̂, ŷ, λ̂) = 0
(i. e., x̂ is a stationary point of L(·, ŷ, λ̂)).
• duality theory...

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 7/25
An illustration of the KKT conditions
2
√
min2 (x1 − 2) + (x2 − 0.5(3 − 5))2 subject to
x∈R

−x1 − x2 + 1 ≥ 0, x2 − x21 ≥ 0. (∗)

2.5
c
√ √ 2 c
2
∗
x = 1
2
(−1+ 5, 3 − 5) : $ 1
Ω
1.5
• global solution of (∗), 1 ∗
∇ c1(x )
• KKT point of (∗). 0.5
∗
x

2
∗

x
∇ f(x )
∇f (x ) = (−5 + 5, 0)$ ,
∗ 0

√ −0.5
∇c1 (x ) = (1 − 5, 1)$ ,
∗
∇ c (x∗)
−1 2

∇c2 (x∗ ) = (−1, −1)$ . −1.5

−2
−2 −1 0 1 2 3
x
1
√
∇f (x ) = λ∗1 ∇c1 (x∗ )
∗
+ , with
λ∗2 ∇c2 (x∗ ) = = λ∗1 λ∗2 5 − 1 > 0.
: constraints are active at x∗ .
c1 (x∗ ) = c2 (x∗ ) = 0

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 8/25
An illustration of the KKT conditions ...
2
√
min2 (x1 − 2) + (x2 − 0.5(3 − 5))2 subject to
x∈R

−x1 − x2 + 1 ≥ 0, x2 − x21 ≥ 0. (∗)

x := (0, 0)$ 2
c
c1 Ω 2
is NOT a KKT point of (∗)! 1.5

c1 (x) = 0: active at x. 1
∇ c (0)
c2 (x) = 1: inactive at x.
1

0.5

2
x
=⇒ λ2 = 0 and
0
∇f (x) = λ1 ∇c1 (x),
−0.5 ∇ f(0)
with λ1 ≥ 0. ∇ c (0)
2
−1
⇓ −4 −3 −2 −1 0 1 2 3
x
1
√
Contradiction with ∇f (x) = (−4, 5 − 3)$ and
∇c1 (x) = (0, 1)$ .

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 9/25
Optimality conditions for constrained problems ...

In general, need constraints/feasible set of (CP) to satisfy

regularity assumption called constraint qualification in order
to derive optimality conditions.
Theorem 16 (First order necessary conditions) Under
suitable constraint qualifications,
x∗ local minimizer of (CP) =⇒ x∗ KKT point of (CP).

Proof of Theorem 16 (for equality constraints only): Let I = ∅.

Then the KKT conditions become: cE (x∗ ) = 0 (which is trivial
as x∗ feasible) and ∇f (x∗ ) = JE (x∗ )T y ∗ for some y ∗ ∈ Rm ,
where JE is the Jacobian matrix of the constraints cE .
Consider feasible perturbations/paths x(α) around x∗ , where
α (sufficiently small) scalar, x(α) ∈ C 1 (Rn ) and
x(0) = x∗ , x(α) = x∗ + αs + O(α2 ), s 4= 0 and c(x(α)) = 0(†) .
(†) requires constraint qualifications, namely, assuming the existence of s $= 0 with above properties.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 10/25
Optimality conditions for constrained problems ...

Proof of Theorem 16 (for equality constraints only): (continued)

For any i ∈ E , by Taylor’s theorem for ci (x(α)) around x∗ ,
0 = ci (x(α)) = ci (x∗ + αs + O(α2 ))
= ci (x∗ ) + ∇ci (x∗ )T (x∗ + αs − x∗ ) + O(α2 )
= α∇ci (x∗ )T s + O(α2 ),
where we used ci (x∗ ) = 0. Dividing both sides by α, we
deduce
0 = ∇ci (x∗ )T s + O(α),
for all α sufficiently small. Letting α → 0, we obtain
∇ci (x∗ )T s = 0 for all i ∈ E ,
and so JE (x∗ )s = 0. [In other words, any feasible direction s
(which is assumed to exist) satisfies JE (x∗ )s = 0.]

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 11/25
Optimality conditions for constrained problems ...

Proof of Theorem 16 (for equality constraints only): (continued)

Now expanding f , we deduce
f (x(α)) = f (x∗ ) + ∇f (x∗ )T (x∗ + αs − s∗ ) + O(α2 )
= f (x∗ ) + α∇f (x∗ )T s + O(α2 ).
Since x∗ is a local minimizer of f , we have f (x(α)) ≥ f (x∗ )
for all α sufficiently small. Thus α∇f (x∗ )T s + O(α2 ) ≥ 0 for all
α sufficiently small. Considering α > 0, we divide by α to
obtain ∇f (x∗ )T s + O(α) ≥ 0; now letting α → 0, we deduce
∇f (x∗ )T s ≥ 0. Similarly, considering α < 0, we obtain
∇f (x∗ )T s ≤ 0. Thus
∇f (x∗ )T s = 0 for all s such that JE (x∗ )s = 0. (1)
By rank-nullity theorem, (1) implies that ∇f (x∗ ) must belong
to the range space of JE (x∗ )T (ie, span of columns of
JE (x∗ )T ), and so ∇f (x∗ ) = JE (x∗ )T y ∗ for some y ∗ . The next
slide details this argument.
Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 12/25
An argument using linear algebra  
∇c1 (x∗ )T
∇c2 (x∗ )T
 
 
∇f (x∗ )T s = 0 for all s ∈ Rn s.t. JE (x∗ )s =0⇔  s = 0.
 
..

 . 

∇cm (x∗ )T
h i
Taking the QR factorization of JE (x∗ )T = ∇c1 (x∗ ) ∇c2 (x∗ ) · · · ∇cm (x∗ ) = QR:

JE (x∗ )T = Q R (Q ∈ Rn×m , n ≥ m)

" # " #
QT 0
JE (x∗ )s= 0 means QT s
=0⇔ T s= ⇔ s = Q⊥ d for some d ∈ Rn−m .
Q⊥ d
([Q Q⊥ ] is orthogonal; assuming JE (x∗ ) is full rank rank(JE (x∗ )) = m)
Now ∇f (x∗ )T s = 0 for all such s ⇔ ∇f (x∗ )T Q⊥ = 0 ⇔ ∇f (x∗ ) = Qd˜ for some
d˜ ∈ Rn−m ⇔ ∇f (x∗ ) = (QR)(R−1 d) ˜ = JE (x∗ )T y ∗ .
Optimality conditions for constrained problems ...

Proof of Theorem 16 (for equality constraints only): (continued)

By rank-nullity theorem, there exists y ∗ ∈ Rm and s∗ ∈ Rn such that
∇f (x∗ ) = JE (x∗ )T y ∗ + s∗ , (2)
where s∗ belongs to the null space of JE (x∗ ) (so JE (x∗ )s∗ = 0).
Taking the inner product of (2) with s∗ , we deduce
(s∗ )T ∇f (x∗ ) = (s∗ )T JE (x∗ )T y ∗ + (s∗ )T s∗ , or equivalently,
(s∗ )T ∇f (x∗ ) = (y ∗ )T JE (x∗ )s∗ + )s∗ )2 .
From (1) and JE (x∗ )s∗ = 0, we deduce (s∗ )T ∇f (x∗ ) = 0. Thus
)s∗ )2 = 0 and so s∗ = 0. Again from (2): ∇f (x∗ ) = JE (x∗ )T y ∗ . !

Let (CP) with equalities only (I = ∅). Then feasible descent

direction s at x ∈ Ω if ∇f (x)T s < 0 and JE (x)s = 0.
Let (CP). Then feasible descent direction s at x ∈ Ω if
∇f (x)T s < 0, JE (x)s = 0 and ∇ci (x)T s ≥ 0 for all i ∈ I ∩ A(x).

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 13/25
Constraint qualifications

Proof of Th 16: used (first-order) Taylor to linearize f and ci

along feasible paths/perturbations x(α) etc. Only correct if
linearized approximation covers the essential geometry of the
feasible set. CQs ensure this is the case.
Examples:
(CP) satisfies the Slater Constraint Qualification (SCQ) ⇐⇒
if ∃ x s.t. cE (x) = Ax − b = 0 and cI (x) > 0 (i.e., ci (x) > 0, i ∈ I ).
with convex domain
(CP) satisfies the Linear Independence Constraint
Qualification (LICQ) ⇐⇒ ∇ci (x), i ∈ A(x), are linearly
independent (at relevant x).
Both SCQ and LICQ fail for
Ω = {(x1 , x2 ) : c1 (x) = 1 − x21 − (x2 − 1)2 ≥ 0; c2 (x) = −x2 ≥ 0}.
TΩ (x) = {(0, 0)} and F (x) = {(s1 , 0) : s1 ∈ R}. Thus TΩ (x) 4= F (x).

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 14/25
Constraint qualifications...

Tangent cone to Ω at x: [See Chapter 12, Nocedal & Wright]

TΩ (x) = {s : limiting direction of feasible sequence} [‘geometry’ of Ω]
zk − x
s = lim k
where z k
∈ Ω , t k
> 0 , t k
→ 0 and z k
→ x as k → ∞.
k→∞ t
Set of linearized feasible directions: [‘algebra’ of Ω]
F (x) = {s : sT ∇ci (x) = 0, i ∈ E; sT ∇ci (x) ≥ 0, i ∈ I ∩ A(x)}
Want TΩ (x) = F (x) ←−[ensured if a CQ holds]

min(x1 ,x2 ) x1 + x2
s.t. x21 + x22 − 2 = 0.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 15/25
Optimality conditions for constrained problems ...

If the constraints of (CP) are linear in the variables, no constraint

qualification is required.

Theorem 17 (First order necessary conditions for linearly

constrained problems) Let (cE , cI )(x) := Ax − b in (CP). Then
x∗ local minimizer of (CP) =⇒ x∗ KKT point of (CP).

Let A = (AE , AI ) and b = (bE , bI ) corresponding to equality

and inequality constraints.
KKT conditions for linearly-constrained (CP): x∗ KKT point ⇔
there exists (y ∗ , λ∗ ) such that
∇f (x∗ ) = AT T ∗
E y + AI λ ,
∗

AE x∗ − bE = 0, AI x∗ − bI ≥ 0,
λ∗ ≥ 0, (λ∗ )T (AI x∗ − bI ) = 0.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 16/25
Optimality conditions for convex problems

(CP) is a convex programming problem if and only if

f (x) is a convex function, ci (x) is a concave function for all
i ∈ I and cE (x) = Ax − b.

• ci is a concave function ⇔ (−ci ) is a convex function.

• (CP) convex problem ⇒ Ω is a convex set.
• (CP) convex problem ⇒ any local minimizer of (CP) is global.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 17/25
Optimality conditions for convex problems

(CP) is a convex programming problem if and only if

f (x) is a convex function, ci (x) is a concave function for all
i ∈ I and cE (x) = Ax − b.

• ci is a concave function ⇔ (−ci ) is a convex function.

• (CP) convex problem ⇒ Ω is a convex set.
• (CP) convex problem ⇒ any local minimizer of (CP) is global.

First order necessary conditions are also sufficient for optimality

when (CP) is convex.
Theorem 18. (Sufficient optimality conditions for convex
problems: Let (CP) be a convex programming problem.
x̂ KKT point of (CP) =⇒ x̂ is a (global) minimizer of (CP). !

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 17/25
Optimality conditions for convex problems

Proof of Theorem 18.

f convex =⇒ f (x) ≥ f (x̂) + ∇f (x̂)$ (x − x̂), for all x ∈ Rn . (3)
"
(3)+[∇f (x̂) = A ŷ +
$
i∈Iλ̂i ∇ci (x̂)] =⇒
"
f (x) ≥ f (x̂) + (A ŷ) (x − x̂) + i∈I λ̂i (∇ci (x̂)$ (x − x̂)),
$ $

"
f (x) ≥ f (x̂) + ŷ $ A(x − x̂) + i∈I λ̂i (∇ci (x̂)$ (x − x̂)) (4).

Let x ∈ Ω arbitrary =⇒ Ax = b and c(x) ≥ 0.

Ax = b and Ax̂ = b =⇒ A(x − x̂) = 0. (5)

ci concave =⇒ ci (x) ≤ ci (x̂) + ∇ci (x̂)$ (x − x̂).

=⇒ ∇ci (x̂)$ (x − x̂) ≥ ci (x) − ci (x̂).
=⇒ λ̂i (∇ci (x̂)$ (x − x̂)) ≥ λ̂i (ci (x) − ci (x̂)) = λ̂i ci (x)≥ 0,
since λ̂ ≥ 0, λ̂i ci (x) = 0 and c(x) ≥ 0.
Thus, from (4) and (5), f (x) ≥ f (x̂). !

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 18/25
Example: Optimality conditions for QP problems

A Quadratic Programming (QP) problem has the form

minimizex∈Rn c# x + 12 x# Hx s. t. Ax = b, Ãx ≥ b̃. (QP)
H symm. pos. semidefinite =⇒ (QP) convex problem.
The KKT conditions for (QP):
x̂ KKT point of (QP) ⇐⇒ ∃ (ŷ, λ̂) ∈ Rm × Rp such that

H x̂ + c = A# ŷ + Ã# λ̂,
Ax̂ = b, Ãx̂ ≥ b̃,
λ̂ ≥ 0, λ̂# (Ãx̂ − b̃) = 0.

“An example of a nonlinear constrained problem” is convex;

removing the constraint x2 − x21 ≥ 0 makes it a convex (QP).

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 20/25
Example: Duality theory for QP problems

For simplicity, let A := 0 and H 8 0 in (QP): primal problem:

minimizex∈Rn c$ x + 21 x$ Hx s. t. Ãx ≥ b̃. (QP)
The KKT conditions for (QP):

H x̂ + c = Ã$ λ̂,
Ãx̂ ≥ b̃,
λ̂ ≥ 0, λ̂$ (Ãx̂ − b̃) = 0.

Dual problem:
maximize(x,λ) − 21 xT Hx + b̃T λ s.t. − Hx + Ã$ λ = c and λ ≥ 0.
Optimal value of primal pb=optimal value of dual pb (provided
they exist).

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 21/25
Optimality conditions for nonconvex problems

• When (CP) is not convex, the KKT conditions are not in

general sufficient for optimality
−→ need positive definite Hessian of the Lagrangian function
along “feasible” directions.

• More on second-order optimality conditions later on.

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 19/25
Second-order optimality conditions

• When (CP) is not convex, the KKT conditions are not in

general sufficient for optimality.
Assume some CQ holds. Then at a given point x∗ : the set of
feasible directions for (CP) at x∗ :
% &
F (x∗ ) = s : JE (x∗ )s = 0, sT ∇ci (x∗ ) ≥ 0, i ∈ A(x∗ ) ∩ I .

If x∗ is a KKT point, then for any s ∈ F (x∗ ),

T T ∗ T
∗ ∗
λi sT ∇ci (x∗ )
"
s ∇f (x ) = s JE (x ) y + i∈A(x∗ )∩I

= (JE (x∗ )s)T y ∗ + i∈A(x∗ )∩I λi sT ∇ci (x∗ )

"

T ∗
"
= i∈A(x )∩I
∗ λ i s ∇ci (x ) ≥ 0. (6)

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 22/25
Second-order optimality conditions...

If x∗ is a KKT point, then for any s ∈ F (x∗ ), either

sT ∇f (x∗ ) > 0
−→ so f can only increase and stay feasible along s
or sT ∇f (x∗ ) = 0
−→ cannot decide from 1st order info if f increases or not
along such s.
From (6), we see that the directions of interest are:
JE (x∗ )s = 0 and sT ∇ci (x∗ ) = 0, ∀i ∈ A(x∗ ) ∩ I with λi > 0.
i > 0},
F (λ∗ ) = {s ∈ F (x∗ ) : sT ∇ci (x∗ ) = 0, ∀i ∈ A(x∗ ) ∩ I with λ∗
where λ∗ is a Lagrange multiplier of the inequality constraints.
Then note that sT ∇f (x∗ ) = 0 for all s ∈ F (λ∗ ).

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 23/25
Second-order optimality conditions ...

Theorem 19 (Second-order necessary conditions)

Let some CQ hold for (CP). Let x∗ be a local minimizer of
(CP), and (y ∗ , λ∗ ) Lagrange multipliers of the KKT conditions
at x∗ . Then
sT ∇2xx L(x∗ , y ∗ , λ∗ )s ≥ 0 for all s ∈ F (λ∗ ),

where L(x, y, λ) = f (x) − y T cE (x) − λT cI (x) is the

Lagrangian function and so
"m "p
∇2xx L(x, y, λ) = ∇2 f (x) − j=1 yj ∇2 c j (x) − i=1 λi ci (x)].

Theorem 20 (Second-order sufficient conditions)

Assume that x∗ is a feasible point of (CP) and (y ∗ , λ∗ ) are
such that the KKT conditions are satisfied by (x∗ , y ∗ , λ∗ ). If
sT ∇2xx L(x∗ , y ∗ , λ∗ )s > 0 for all s ∈ F (λ∗ ), s 4= 0,
then x∗ is a local minimizer of (CP). [See proofs in Nocedal & Wright]

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 24/25
Necessary and sufficient conditions for x∗ : summary

minimise f (x)
x
subject to cE (x) = 0, cI (x) ≥ 0.

Lagrangian: L(x, y, λ) = f (x) − yj cj (x) − λi ci (x)

P P
j∈E i∈I
(
cE (x) = 0,
x∈R x ∈ Rn x ∈ Rn , cE (x) = 0 x ∈ Rn ,
cI (x) ≥ 0
( ( (
f ′ (x∗ ) = 0 ∇f (x∗ ) = 0 ∇x L(x∗ ) = 0
nec. ∀s s.t. ∇cTj s = 0
f ′′ (x∗ ) ≥ 0 ∇2 f (x∗ ) ⪰ 0 sT (∇2xx L(x∗ ))s ≥ 0
( ( ( (
f ′ (x∗ ) = 0 ∇f (x∗ ) = 0 ∇x L(x∗ ) = 0 ∇cTj s = 0
suff. ∀s =
̸ 0 s.t.
f ′′ (x∗ ) > 0 ∇2 f (x∗ ) ≻ 0 sT (∇2xx L(x∗ ))s > 0 λi (∇ci )T s ≥ 0
Illustration: we need sT (∇2xx L(x∗ , y ∗ , λ∗ ))s ≥ 0 not sT (∇2xx f (x∗ ))s ≥ 0
Consider

minimise x1 + x2
x 1


1 − (x2 + x2 ) = 0, or
1 2 0.5

subject to √ √
1 − ((x + 2)2 + ((x + 2)2 ) = 0
1 2 0

-0.5

In both cases, ∇2 f (x∗ ) = 0 for

x∗ = − √12 11 , but x∗ is local (global)
  -1

√1 > 0 for blue, while

-1.5

minimum λ = 2
maximum λ = − √12 < 0 for red. Thus
-2

" # -2.5

−2 -2.5 -2 -1.5 -1 -0.5 0 0.5 1

∇2xx L = −λ = 2λI, so
−2
∇2xx L ⪰ 0 and ∇2xx L
 0.
Example: Trust-region subproblem
For the TRS,
1
minimise f (x) = g T x + xT Hx
x 2
2 2
subject to ∆ − ∥x∥ ≥ 0.

The Lagrangian is L(x, λ) = g T x + 12 xT Hx − λ(∆2 − ∥x∥2 ).

▶ Note: For fixed λ ≥ 0, minx L(x, λ) ≤ f (x∗ ) (V useful: gives lower bound!)
▶ The KKT conditions (roughly, 1st order necessary) are λ̃ ≥ 0,
∇x L = g + Hx + λ̃x = 0 ⇔ (H + λ̃I)x = −g (with λ̃ = 2λ 1 ),
Complementarity λ̃(∆ − ∥x∥) = 0, Feasibility ∥x∥ ≤ ∆.
▶ According to 2nd order conditions, at soln we require sT ∇2xx Ls ≥ 0 for all
s ∈ F (λ̃). Satisfied if (H + λ̃I) ⪰ 0; such KKT multiplier λ̃ is unique

1
Corrected 16/5/23. This was a ’minor’ typo in that the KKT conditions as stated were correct,
but ∇x ∥x∥2 = 2x so λ ̸= λ̃.
Some simple approaches for solving (CP)

Equality-constrained problems: direct elimination (a simple

approach that may help/work sometimes; cannot be
automated in general)
Method of Lagrange multipliers: using the KKT and second
order conditions to find minimizers (again, cannot be
automated in general)
[see Pb Sheet 4]

Lectures 10 and 11: Constrained optimization problems and their optimality conditions – p. 25/25