AE-777 (Lecture No.

10)
Optimization: Solution to Selected Exercises

Multi-Variable Optimization
Let L(u) ∈ R be a continuous function of u ∈
Rm, and let it possesses continuous partial deriva-
tives of up to second order with respect to the
elements of u = (u1, u2, . . . , um)T .

Necessary Conditions:
If there are no constraints on the possible val-
ues of u, then the following necessary condi-
tions must be satisfied by a set u = (u1, u2, . . . , um)T
which minimizes L(u):
∂L
Lu = =0 (1)
∂u
and
∂ 2L
Luu = ≥0 (2)
∂u2

A set u = (u1, u2, . . . , um)T which satisfies Eq.(1)

is called a stationary point of L(u).
Inequality (2) implies that Luu is a positive
semi-definite matrix, i.e., all the eigenvalues of
Luu are either greater than, or equal to zero.

Sufficient Condition:
A stationary point of L(u) is its minimum point,
if it satisfies the following sufficient condition
of minimization:
∂ 2L
Luu = 2
>0 (3)
∂u

Inequality (3) implies that Luu is a positive def-

inite matrix, i.e., all the eigenvalues of Luu are
greater than zero.

If at least one eigenvalue of Luu is zero, then

the sufficient condition of minimization is not
satisfied by a stationary point, and it is termed
a singular point.
Exercise 1
(This is Exercise 1 of Lecture 7.)
For the following function:

 − tan(u)
 (u ≤ 0)
L(u) =
(u ≥ 0)

 tan(u)

where u ∈ R, determine (if any) the

(a) stationary points.
(b) minimum points.

(a) Since Lu never vanishes anywhere in the

real space, L(u) does not have any stationary
points.

(b) L(u) achieves its minimum possible value,

L(u) = 0, at the point u = 0. Hence, u = 0 is
the only minimum point of L(u). However, this
minimum point satisfies neither the necessary,
nor the sufficient conditions of minimization,
because Lu is discontinuous at u = 0.

A plot of L(u) is shown in the following figure:

Exercise 2
(This is Exercise 2 of Lecture 7.)
For the following function:

L(u) = −3u3

where u ∈ R, find the stationary points (if any)

and determine if they are the minimum points.

Since L(u) has the derivatives,

Lu = −9u2 ; Luu = −18u

which are continuous throughout the real space,
we can test for the necessary and sufficient
conditions.

To find the stationary points, we solve the sta-

tionary condition, Eq.(1), as follows:

Lu = −9u2 = 0
whose only real root is u = 0. Hence the only
stationary point of L(u) is u∗ = 0.
To check whether the stationary point, u∗ = 0,
is also a minimum point, we apply the sufficient
condition, Eq.(3), for that point:

Luu(u∗) = −18u∗ = 0
Since the sufficient condition of minimization
is not satisfied by u∗ = 0, but the necessary
conditions are satisfied, it is a singular point of
L(u).
Exercise 3
(This is Exercise 3 of Lecture 7.)
For
L(u) = u3 − 3u2 + 3u − 4
where u ∈ R, find the stationary points (if any)
and determine if they are the minimum points.

Since L(u) has the derivatives,

Lu = 3u2 − 6u + 3 ; Luu = 6u − 6
which are continuous throughout the real space,
we can test for the necessary and sufficient
conditions.

To find the stationary points, we solve the sta-

tionary condition, Eq.(1), as follows:

Lu = 3u2 − 6u + 3 = 0
whose only real root is u = 1. Hence the only
stationary point of L(u) is u∗ = 1.
To check whether the stationary point, u∗ = 1,
is also a minimum point, we apply the sufficient
condition, Eq.(3), for that point:

Luu(u∗) = 6u∗ − 6 = 0
Since the sufficient condition of minimization
is not satisfied by u∗ = 1, but the necessary
conditions are satisfied, it is a singular point of
L(u).
Exercise 4
(This is Exercise 4 of Lecture 7.)
Consider the following function:

L(u) = u2 2
1 − 2u1 u2 + 4u2
where u = (u1, u2)T ∈ R2. Find the stationary
points (if any) and determine if they are the
minimum points.

To determine the stationary points, we take

the partial derivatives of L with respect to the
elements of u and apply Eq.(1) as follows:
∂L
Lu = = [2u1 − 2u2 , −2u1 + 8u2]
∂u
= [0, 0]
which is satisfied only at u1 = u2 = 0. Hence,
u∗ = (0, 0)T is the only stationary point of
L(u).

To test whether the stationary point is a min-

imum point, we take the second-order partial
derivatives of L with respect to the elements
of u as follows, and evaluate the matrix at the
stationary point, u∗ = (0, 0)T :
∂ 2L ∂
Luu = 2
= [2u1 − 2u2 , −2u1 + 8u2]
∂u
 ∂u 
2 −2
= 
 

−2 8

Finally, we find the eigenvalues of Luu at the

stationary point as follows to see whether Eq.(3)
applies:

det(sI − Luu) = (s − 1)(s − 4) − 1 = 0

or s1 = 0.6972 and s2 = 4.3028. Since both
the eigenvalues of Luu are positive at the sta-
tionary point, the sufficient condition of mini-
mization is satisfied, and the stationary point
is a minimum point.
Exercise 5
(This is Exercise 5 of Lecture 7.)
Consider the following function:

L(u) = −u2
1 + 2u u
1 2 + 3u2
2
where u = (u1, u2)T ∈ R2. Find the stationary
points (if any) and determine if they are the
minimum points.

Determination of the stationary points:

∂L
Lu = = [−2u1 + 2u2 , 2u1 + 6u2]
∂u
= [0, 0]
which is satisfied only at u1 = u2 = 0. Hence,
u∗ = (0, 0)T is the only stationary point of
L(u).

Evaluation of Luu at the stationary point, u∗ =

(0, 0)T :
∂ 2L ∂
Luu = 2
= [−2u1 + 2u2 , 2u1 + 6u2]
∂u
 ∂u 
−2 2
= 
 

2 6

The eigenvalues of Luu at the stationary point:

det(sI − Luu) = (s + 1)(s − 3) − 1 = 0

or s1 = −1.2361 and s2 = 3.2361. Since one
eigenvalue of Luu is positive while the other
is negative at the stationary point, neither the
necessary, nor the sufficient conditions of mini-
mization are satisfied, and the stationary point,
u∗ = (0, 0)T , is not a minimum point.

(Such a stationary point where some eigenval-

ues of Luu are negative, and the others are
positive is called a saddle point.)
Optimization with Equality Constraints
Consider the minimization of L(x, u) ∈ R with
respect to the control variables, u ∈ Rm, sub-
ject to a set of static equality constraints ex-
pressed in the following vector form:

f (x, u) = 0 (4)
where x ∈ Rn are the state variables, and f :
Rn × Rm → Rn is a smooth mapping function,
i.e., it possesses continuous partial derivatives
with respect to both x and u. It is further as-
sumed that L(x, u) possesses continuous par-
tial derivatives with respect to both x and u.

The constraint equation, Eq.(4), should be

such that it uniquely yields x, given u.

The constrained minimization problem is posed

in terms of the following function called the
Hamiltonian, which combines the equality con-
straints with the objective function to be min-
imized:
H(x, u, λ) = L(x, u) + λT f (x, u) (5)
and λ ∈ Rn are the Lagrange multipliers. Then
the problem of minimizing the Lagrangian, L(x, u),
with respect to u, subject to the equality con-
straint, f (x, u) = 0, is equivalent to the prob-
lem of minimizing the Hamiltonian, H, with
respect to u.

The stationary condition in the presence of a

small but arbitrary variation in u is now ex-
pressed as follows:
dH = Hxdx + Hudu = 0 (6)
where
∂H
Hx = ∈ R1×n (7)
∂x
and
∂H
Hu = ∈ R1×m (8)
∂u
Please pay careful attention to the derivatives
in Eqs. (7) and (8). Recall from Lecture 7
that the derivative of a column vector of size
p with respect to a column vector of size ℓ is
a matrix of the dimensions p × ℓ. H is a scalar
(i.e., a vector of size unity), hence both Hx
and Hu must be row vectors.

For determining the Lagrange multipliers, λ,

the following condition is imposed at the sta-
tionary point:
Hx = Lx + λT fx = 0 (9)
This implies that
λT = −Lxfx−1 (10)
The substitution of Hx = 0 in Eq.(6) gives the
following stationary condition:
Hu = Lu + λT fu = 0 (11)
Equations (9) and (12), along with the con-
straint equation, f (x, u) = 0, are the necessary
conditions of constrained minimization.
These equations constitute the 2n + m scalar
equations to be solved for x ∈ Rn, λ ∈ Rn, and
u ∈ Rm, to determine the stationary points.

Exercise 6
For the minimization of
 
2 0
1 2 1 T
L(x, u) = x + u  u

2 2 0 1
with respect to u ∈ R2, where x ∈ R, subject
to the constraint

f (x, u) = x + (2 , −1)u + 2 = 0
find the stationary points, (x∗, u∗), (if any) and
determine if they are the minimum points.

The following Hamiltonian function:

H = L(x, u) + λT f (x, u)
is to be minimized with respect to u = (u1, u2)T ,
where λ ∈ R is the Lagrange multiplier.
Finding the partial derivatives of H w.r.t. x and
u and then applying the necessary conditions,
we have:
Hx = x + λ = 0
or λ = −x, and
 
2 0
Hu = uT   + λ(2, −1) = (0, 0)
 
0 1
which yields

uT = −λ(1, −1) = x(1, −1)

Substituting the equality constraint,

x = −(2 , −1)u − 2
we get

uT = −[(2, −1)u + 2](1, −1)

or

3u1 − u2 = −2
u1 − u2 = −1
solving which, we have the following stationary
point:
 
 −1/2 
 
u∗ =

 1/2 


x∗ = −(2 , −1)u∗ − 2 = −1/2

λ = −x∗ = 1/2
To see whether the stationary point satisfies
the Legendre-Clebsch condition, we find the
second-order partial derivatives of H w.r.t. x
and u:
Hxx = 1
  
0
Hux =   ; Hxu = (0, 0)
 
0
 T  
2 0 2 0
Huu =   =
   

0 1 0 1
The partial derivatives of the equality constraint
are derived as follows:

fx = 1 ; fu = (2 , −1)
Now we apply the Legendre-Clebsch condition
at the stationary point as follows:

(Luu)f =0 = Huu − Huxfx−1fu − fuT (fxT )−1Hxu

+ fuT (fxT )−1Hxxfx−1fu
   
2 0 2
=  +  (2 , −1)
   
0 1 −1
 
6 −2
= 
 

−2 2
√
whose eigenvalues are s1,2 = 4 ± 2 2, which
are both real and positive. Since the Legendre-
Clebsch condition is satisfied at the stationary
point, it is also the minimum point.
Exercise 7
Find the minimum points (if any) of the fol-
lowing Lagrangian function,
1 3 1 3 1
L(x, u) = u1 − u2 + u2 − u1 u2 − u2 u
1 2
3 3 2 1 2
1 2
+ (x1 + x2
2 ) − x1 u1 − x2 u2
2
with respect to u = (u1, u2)T ∈ R2, where x =
(x1, x2)T ∈ R2, subject to the constraint
 
1 1
f (x, u) = x −  u = 0
 
0 1
Finding the partial derivatives of H w.r.t. x and
u and then applying the necessary conditions,
we have:

Hx = (x1 − u1, x2 − u2) + λT = 0

or
   
 λ1 
   x1 − u1 
 
λ= =−

 λ   x −u 
 
2 2 2
Hu = (u21 + u1 − u22 − 2u1 u2 − x1 , −u22 − 2u1 u2 − u21 − x2 )
 
1 1
− λT
0 1
= (u21 + u1 − u22 − 2u1 u2 − x1 − λ1 , −u22 − 2u1 u2 − u21 − x2 − λ1 − λ2 )
= (0, 0)

which, after substituting for λ, yields

u2 2
1 = u2 + 2u1 u2
u2
1 + u2 + 2u u = x − u − u
2 1 2 1 1 2

Substituting the equality constraint gives the

following:

x1 = u1 + u2 ; x2 = u2
from which it follows that u∗1 = u∗2 = 0 is the
only stationary point,
   
 0 
   0 
 
u∗ = ; x∗ =

 0 
 
 0 


and λT = (0, 0).

To see whether the stationary point satisfies
the Legendre-Clebsch condition, we find the
second-order partial derivatives of H w.r.t. x
and u:
 
1 0
∂ T
Hxx = Hx =  =I
 
∂x 0 1
 
−1 0
∂ T
Hux = H =  = −I
 
∂x u 0 −1
 
−1 0
∂ T
Hxu = Hx =   = −I
 
∂u 0 −1
 
∂ T 2u1 + 1 − 2u2 −2u2 − 2u1
Huu = Hu =  
∂u −2u2 − 2u1 −2u2 − 2u1
u1 =0,u2 =0
 
1 0
=  
0 0
The partial derivatives of the equality constraint
are derived as follows:
 
1 1
fx = I ; fu = − 
 

0 1
and the concerned matrix products in the Legendre-
Clebsch condition:
 
1 1
Huxfx−1fu = 
 

0 1
 
1 0
fuT (fxT )−1Hxu = 
 

1 1
  
1 0 1 1
fuT (fxT )−1Hxxfx−1fu = 
  
 
1 1 0 1
 
1 1
= 
 

1 2
Now we apply the Legendre-Clebsch condition
at the stationary point as follows:

(Luu)f =0 = Huu − Huxfx−1fu − fuT (fxT )−1Hxu

+ fuT (fxT )−1Hxxfx−1fu
   
1 0 1 1
=  −
   

0 0 0 1
   
1 0 1 1
−  +
   

1 1 1 2
 
0 0
= 
 

0 0
whose eigenvalues are both zero, s1,2 = 0.
Since the Legendre-Clebsch condition is not
satisfied at the stationary point, the station-
ary point is a singular point of the function
subject to the given constraint.