ISyE 6669 HW 3 Solutions

Fall 2021

1. Consider the following linear optimization problem

min x+y
s.t. x + y = 1,
x ≤ 0, y ≤ 0.

Does this problem have an optimal solution? Is this problem feasible?

Explain your answer.
Solution: This problem does not have an optimal solution and is infea-
sible. We have that
x ≤ 0, y ≤ 0
which implies that
x + y ≤ 0.
But we also have the constraint:

x + y = 1.

It is not possible for

x+y ≤0
and
x + y = 1.
Thus the problem is infeasible.
2. Consider the following optimization problem

min (x · sin(x))2
s.t. x ∈ R.

(a) Find all the global minimum solutions. Explain how you find them.
Hint: there may be multiple ones.
(b) Is there any local minimum solution that is not a global minimum
solution?
(c) Is the objective function f (x) = (x · sin(x))2 a convex function on R?

1
Solution:

(a) The global minimum solutions are x = 0 and x = πn, where n ∈ Z.

The optimal objective value is 0. We find the global minima by
noting that
(x · sin(x))2 ≥ 0.
Thus the global minima are the x values such that

(x · sin(x))2 = 0.

Thus, the problem becoming a root-finding one. Clearly, x = 0 makes

(x · sin(x))2 = 0. Also, x = πn, where n ∈ Z makes sin(x) = 0 and
hence (x · sin(x))2 = 0
(b) There is no local minimum solution that is not a global minimum
solution.
(c) The objective function f (x) = (x · sin(x))2 is not a convex function
on R. We show that

f (αx + (1 − α)y) > αf (x) + (1 − α)f (y)

for a specific x, y ∈ R and α ∈ [0, 1]. Let x = 0, y = π and α = 12 .

We have that
π
f (αx + (1 − α)y) = f (0 + )
2
π π
= ( · sin( ))2
2 2
π 2
=( )
2
π2
=
4
Meanwhile, we have:
1 1
αf (x) + (1 − α)y = f (0) + f (π)
2 2
1
= 0 + (π · sin(π))2
2
=0

Note that
π2
> 0,
4
so f (x) is not a convex function on R

2
3. Consider the following optimization problem
1
min
x
s.t. x ≥ 0.

Does this problem have an optimal solution? Why?

Solution: This problem does not have an optimal solution. Note that for
x ≥ 0, x1 is monotonically decreasing and we have that
1
lim = 0.
x→∞ x
Thus, note that for any x1 ≥ 0, we are always able to find an x2 > x1 ≥ 0
so that
1 1
< .
x2 x1
Hence, no optimal solution exists.
4. Consider the following problem

min x + f (x)
s.t. x ∈ R,

where the function f (x) is defined as

⎧
⎪0,
⎪ −1 < x < 1
⎪
⎨1, x=1
f (x) = .
⎪2,
⎪
⎪ x = −1
+∞, x > 1 or x < −1
⎩

(a) Is the objective function a convex function defined on R? Explain

your answer by checking the criterion of convexity.
(b) Find an optimal solution, or explain why there is no optimal solution.
Solution:

(a) We have an objective function g : R → R ∪ {+∞} given by:

⎧
⎪x,
⎪ −1 < x < 1
⎪
⎨2, x=1
g(x) = .
⎪
⎪
⎪ 1, x = −1
+∞, x > 1 or x < −1
⎩

Method 1: We will show that function g(x) is convex using the defi-
nition of a convex function, i.e. we will show that

3
 g(λa + (1 − λ)b) ≤ λg(x) + (1 − λ)g(x)
for any a, b ∈ R and any λ ∈ (0, 1).
– If a, b ∈ (−1, 1) then λa + (1 − λ)b ∈ (−1, 1) and:
g(λa + (1 − λ)b) = λa + (1 − λ)b = λg(a) + (1 − λ)g(b)
We can also notice that in this case objective function is equal
to x + 0 which is linear function, and thus convex.
– If a = −1 and b ∈ (−1, 1) then λ(−1) + (1 − λ)b ∈ (−1, 1) and:
g(−λ+(1−λ)b) = −λ+(1−λ)b ≤ λ+(1−λ)b = λg(−1)+(1−λ)g(b)
– If a ∈ (−1, 1) and b = 1 then λa + (1 − λ)(1) ∈ (−1, 1) and
g(λa+(1−λ)(1)) = λa+1−λ ≤ λa+2(1−λ) = λg(a)+(1−λ)g(1)
– If either of the points a or b belongs to (−∞, −1) ∪ (1, +∞) then
value of the function g(x) at that point will be +∞. In that case
we have λg(a) + (1 − λ)g(b) = +∞ and:
g(λa + (1 − λ)b) ≤ +∞ = λg(a) + (1 − λ)g(b)
Method 2: Lets look at function h : [−1, 1] → R given by:

⎧
⎪x, −1 < x < 1
⎨
h(x) = 2, x = 1 .
⎪
1, x = −1
⎩

Its plot is given by:

Figure 2: Plot of function h

4
 and we can see it is a convex function on its domain. If function
h(x) is convex on its domain dom(h) = [−1, 1] then its extension
g : R → R ∪ {−∞} given by:

(
h(x), x ∈ [−1, 1]
g(x) = .
+∞, x∈/ [−1, 1]

is also convex (same argument as in the last bullet point from Method
1).
(b) Problem does not have a solution.
Value x∗ ∈
/ (−1, 1) can’t be the solution, because for example

⎧
⎨1,
⎪ x = −1
∗
g(0) = 0 < g(x ) = 2, x=1 .
⎪
+∞, x ∈
/ [−1, 1]
⎩

Value x∗ ∈ (−1, 1) can’t be solution because we can always find

x∗∗ ∈ (−1, 1) such that x∗∗ < x∗ and thus:

g(x∗∗ ) = x∗∗ < x∗ = g(x∗ ).

– if −1 < x < 1 then the optimal value is −1 +  where  is a very

small positive number
– if x = 1 then the optimal value is 2
– if x = −1 then the optimal value is 1
– if x > 1 or x < −1 then optimal value is +∞ which doesn’t
exist

5. For each of the statements below, state whether it is true or false. Justify
your answer.
(a) Consider the optimization problem

min f (x) s.t. g(x) ≥ 0.

Suppose the current optimal objective value is v. Now, if I change the

right-hand-side of the constraint from 0 to 1 and resolve the problem,
the new optimal objective value will be less than or equal to v.
(b) Consider the following optimization problem:

min f (x)4 s.t. x ∈ X

where f (x) is a nonconvex function and X is a non-empty set. Sup-

pose at a feasible solution x∗ ∈ X, f (x∗ ) = 0, then x∗ must be a
global optimal solution.

5
(c) Consider the following optimization problem
(P ) max f (x)
s.t. gi (x) ≥ bi , ∀i ∈ I.
Suppose the optimal objective value of (P) is vP . Then, the La-
grangian dual of (P) is given by
min{L(λ) : λ ≥ 0},
(D) (1)
P
where L(λ) = maxx {f (x) + i∈I λi (gi (x) − bi )}. Furthermore, sup-
pose the optimal objective value of (D) is vD , then vP ≤ vD .
Solution:
(a) False.
Restricting the solution space cannot result in a better solution. Note
that any solution of the new problem is also a feasible solution of the
original problem because
g(x) ≥ 1 =⇒ g(x) ≥ 0.
Let x̂ be the optimal solution of the new problem, and v̂ = f (x̂) be
the optimal value. Since x̂ is also feasible to the original problem,
v ≤ f (x̂) (* v is optimal value of original) (2)
∴ v ≤ v̂ (3)
(b) True.
Note that f (x)4 ≥ 0, ∀x. Since f (x∗ ) = 0, we have f (x)4 ≥ f (x∗ ), ∀x.
Thus, by the definition of global optimum, x∗ is a global optimal so-
lution.

(c) True.
Let S = {x : gi (x) ≥ bi , ∀i ∈ I}. Let x∗ ∈ S be the optimal solution
of (P). Thus, vP = f (x∗ ). Also, let λ∗ ≥ 0 be the optimal solution of
(D) and so vD = L(λ∗ ). Now, for any λ ≥ 0,
X
L(λ) = max{f (x) + λi (gi (x) − bi )} (4)
x
i∈I
X
≥ max{f (x) + λi (gi (x) − bi )} (5)
x∈S
i∈I
≥ max{f (x)} (6)
x∈S
= vP (7)
where Eq.(5) follows because the feasible space is being restricted
(refer to Q5 (a) if unclear) and Eq.(6) follows because λ ≥ 0 and
x ∈ S imply that λi (gi (x) − bi ) ≥ 0 for all i ∈ I. Since this holds for
all λ ≥ 0, it also holds for λ∗ . Therefore, vD = L(λ∗ ) ≥ vP .