MATH 484 - REVIEW QUESTIONS 3

1. Definitions
(1) Convex set
(2) Convex function
(3) Convex hull
(4) Convex combination
(5) Segment between two vectors
(6) Jensen’s inequality
(7) AM-GM inequality

2. Computation
(1) Use Sylvester’s criterion to determine whether the following matrices are positive definite, negative
definite,
 or semidefinite.

1 0 0
(a) 0 1 0.
0 0 1
 
−1 0 0
(b)  0 −1 0 .
0 0 −1
 
1 1 1
(c) 1 2 2.
1 2 3 
1 0 1
(d) 0 −2 0.
1 0 2
(2) Describe the convex hull of the set {(0, 0), (1, 0), (1, 1), (2, 1)} as an intersection of closed affine
half-spaces. (No proof required).
(3) Describe the set {(x, y, z) ∈ R3 : |x| + |y| + |z| ≤ 1} as an intersection of closed affine half-spaces. (No
proof required).
2 2
(4) Use the Hessian to show that the function f (x, y) = ex +y is convex.
(5) Use Jensen’s inequality to minimize ex + ey + ez subject to x + y + z = 1.
(6) Use Jensen’s inequality to maximize xy subject to x + y = 1 and x, y > 0.
(7) Use Jensen’s inequality to maximize x4 y subject to x + y = 1 and x, y > 0.

3. Constructions
(1) Construct a function f : R2 → R with a critical point at (0, 0) that is
(a) A strict global maximizer
(b) A strict local minimizer
(c) A saddle point
(2) Construct a nonlinear function that is convex but not strictly convex.
(3) Construct a convex function f : R2 → R for which (0, 0) is a strict global minimizer, and ∇f (0, 0) is
not defined.

1
 4. Proofs
n
(1) Show that for every a ∈ R , the function x 7→ aT x is convex.
(2) Show that f (x) = |x| is convex using the definition of convexity.
(3) Prove Jensen’s inequality.

5. Solutions
5.1. Computation.
(1) (a) The principal minors are all 1 > 0, so the matrix is positive definite.
(b) The principal minors are ∆1 = −1 < 0, ∆2 = 1 > 0, and ∆3 = −1 < 0. Since the principal
minors alternate sign and ∆1 < 0, the matrix is negative definite.
(c) The principal minors are ∆1 = 1, ∆2 = 2 − 1 = 1, and ∆3 = 6 + 2 + 2 − 2 − 4 − 3 = 1. Since all
principal minors are positive, the matrix is positive definite.
(d) The principal minors are ∆1 = 1, ∆2 = −2, and ∆3 = −4 + 3 − (−2) = 1. Since ∆1 > 0 and
∆2 < 0, the principal minors do not follow the pattern for positive definiteness or negative
definiteness. Since ∆3 ≠ 0, the matrix has all nonzero eigenvalues; therefore, the matrix has at
least one positive and one negative eigenvalue and is indefinite.
(2) (This solution is easy to understand if you draw a picture.) We can take the four linear inequalities
y ≤ x, y ≤ 1, y ≥ 0, and y ≥ x − 1 and represent them as closed affine half-spaces; then, their
intersection gives us the convex hull of the set. In standard form as affine half spaces, the solution
sets of these inequalities are, respectively,
{(x, y) ∈ R2 : (−1, 1) · (x, y) ≤ 0},
{(x, y) ∈ R2 : (0, 1) · (x, y) ≤ 1},
{(x, y) ∈ R2 : (0, −1) · (x, y) ≤ 0},
{(x, y) ∈ R2 : (1, −1) · (x, y) ≤ 1}.
(3) For each of the eight vectors (k1 , k2 , k3 ) ∈ {−1, 1} × {−1, 1} × {−1, 1}, we can make an affine half
space
{(x, y, z) ∈ R3 : k1 x + k2 y + k3 z ≤ 1}.
Then, the intersection of all eight of these affine half-spaces gives the original set.
(4) The gradient of f is
" 2 2
#
2xex +y
∇f (x, y) = 2 2 .
2yex +y
2 2
When writing the Hessian, it is easier just to write f (x, y) instead of ex +y . We have
2f (x, y) + 4x2 f (x, y)
 
4xyf (x, y)
Hf (x, y) = .
4xyf (x, y) 2f (x, y) + 4y 2 f (x, y)
We can try to show that Hf (x, y) ≻ 0 using Sylvester’s criterion. We have ∆1 = (2 + 4x2 )f (x, y),
which is positive, since both factors are positive. When computing ∆2 , we will end up with a
f (x, y)2 factor on the outside, so we can just do the determinant computation
without worrying
about the f (x, y) factors. This
gives us f (x, y)2 (2 + 4x2 )(2 + 4y 2 ) − (4xy)2 . After simplification,
∆2 = f (x, y)2 4 + 16(x2 + y 2 ) . Hence, ∆2 > 0, as both factors are positive. Therefore, by Sylvester’s
criterion, Hf (x, y) ≻ 0, so f is a convex function.
(5) We write f (x) = ex . We know that
 
1 1 1 x+y+z 1
3( f (x) + f (y) + f (z)) ≥ 3f = 3f ( ) = constant.
3 3 3 3 3
By Jensen’s inequality, this constant lower bound is achieved when x = y = z. (In fact, since f is
strictly convex, this is the only way to achieve the constant lower bound.) Therefore, the solution is
(x, y, z) = ( 13 , 13 , 13 ).
2
 (6) Rather than maximizing xy, we will minimize − log(xy) = − log x − log y, as the solution to both
problems will be the same. We write f (x) = − log x. Since f ′′ (x) = x12 > 0, f is a convex function,
so we can use it with Jensen’s inequality. We have
 
1 1 x+y 1
2( f (x) + f (y)) = 2f = 2f ( ) = constant.
2 2 2 2
To achieve this lower bound, we let x = y. Therefore, the solution is (x, y) = ( 21 , 12 ).
(7) (This problem is hard, and it is much easier to do using the AM-GM inequality.) We will make several
adjustments to the objective function in order to arrive at something( that is easier to optimize.
Each time we change the objective function, we will not change the solution to the optimization
problem. Rather than maximizing x4 y, we will minimize − log(x4 y) = −4 log x − log y. The global
minimizer of this objective function does not change even if we change it by a constant; therefore,
we can minimize −4 log x − log y − 5 log 5 + 4 log 4 instead. By logarithm rules, we can rewrite the
objective function as −4 log( 45 x) − log(5y).
Now, we write f (x) = −5 log x, which is a convex function. Then, we see that our objective
function satisifes
   
4 5 1 4 5 1
f x + f (5y) ≥ f · x + · 5y = f (x + y) = f (1) = constant.
5 4 5 5 4 5
To achieve this lower bound, we should let the values 54 x = 5y. Therefore, we let x = 4y, so that
x = 45 and y = 15 .

5.2. Constructions.
2 2
(1) For these questions,
  we can use functions
 of the form f (x, y) = ax + by . If f is of this form,
2ax 2a 0
Hf (x, y) = , and Hf (x, y) = . By looking at the gradient, we can see that all such
2by 0 2b
functions f have a unique critical point at (0, 0).
(a) We would like to make Hf (0, 0) ≺ 0. Then, −Hf (0, 0) ≻ 0, so −f is convex, implying that the
strict local minimizer (0, 0) of −f is also a global minimizer of −f . This in turn implies that
(0, 0) is a strict local maximizer of f . In order to do this, we can choose a = b = −1, so that
Hf (0, 0) has negative eigenvalues −2 and −2.
(b) We would like to make Hf (0, 0) ≻ 0. Then, Hf (0, 0) ≻ 0, so f is convex, implying that the
strict local minimizer (0, 0) of f is also a global minimizer of f . In order to do this, we can
choose a = b = 1, so that Hf (0, 0) has positive eigenvalues 2 and 2.
(c) We would like to make Hf (0, 0) indefinite. To do this, we can let a = 1 and b = −1, so that
Hf (0, 0) has eigenvalues 2 > 0 and −2 < 0, making Hf (0, 0) indefinite, and making (0, 0) a
saddle point.
(2) One such example is f (x) = max{0, |x| − 1}. (A simpler example is g(x) = |x|, but f is more
interesting.) f is not strictly convex because we can choose x = 1 and y = −1, in which case
f (x) = f (y) = 0. Then, f ( 12 x + 12 y) = f (0) = 0 = 12 f (x) + 12 f (y). Since the convexity equality is
not strict even when we choose x ̸= y and t = 12 , f is not a strictly convex function. However, f
is a convex function. This can be proven using the definition, but it is easier to use the fact that
f (x) = max{x − 1, 1 − x, 0}. Since f is the maximum of three linear (and therefore convex) functions,
f is convex.
(3) For this, we can use the function f (x, y) = |x| + |y|. It is easy to see that (0, 0) is a strict global
minimizer, since f (0, 0) = 0, and f (x, y) > 0 for all (x, y) ̸= (0, 0). To show that f (x, y) is convex, we
need to show that for each pair (x1 , y1 ), (x2 , y2 ) ∈ R2 and t ∈ [0, 1], we have f (tx1 + (1 − t)x2 , ty1 +
(1 − t)y2 ) ≤ tf (x1 , y1 ) + (1 − t)f (x2 , y2 ). By the triangle inequality, we have
f (tx1 + (1 − t)x2 , ty1 + (1 − t)y2 ) = |tx1 + (1 − t)x2 | + |ty1 + (1 − t)y2 |

≤ t|x1 | + (1 − t)|x2 | + t|y1 | + (1 − t)|y2 | = t(|x1 | + |y1 |) + (1 − t)(|x2 | + |y2 |) = tf (x1 , y1 ) + (1 − t)f (x2 , y2 ).
Therefore, f is convex.
3
 Now, the function ϕ(t) obtained from f restricted to the line {(t, 0) : t ∈ R} is equal to ϕ(t) = |t|.
Hence, ϕ′ (0) is undefined. However, if ∇f (0, 0) existed, then we would have ϕ′ (0) = ∇f (0, 0) · (1, 0).
Therefore, ∇f (0, 0) does not exist.
5.3. Proofs.
(1) Let f (x) = aT x for each x ∈ Rn . We observe that for each t ∈ [0, 1] and pair x, y ∈ Rn ,
f (tx+(1−t)y) = aT (tx+(1−t)y) = taT x+(1−t)aT y = tf (x)+(1−t)f (y). Hence, f (tx+(1−t)y) ≤
tf (x) + (1 − t)f (y), and in fact the ≤ is an equality.
(2) Let f (x) = |x|. We want to show that for each pair x, y ∈ R and t ∈ [0, 1], we have f (tx + (1 − t)y) ≤
tf (x) + (1 − t)f (y). By the triangle inequality,
f (tx + (1 − t)y) = |tx + (1 − t)y| ≤ t|x| + (1 − t)|y| = tf (x) + (1 − t)f (y),
which is exactly what we needed to show.
(3) Let f be a convex function with a domain C. We need to show that if k ≥ 1, x1 , . . . , xk ∈ C, and
λ1 , . . . , λk are nonnegative values summing to 1, then f (λ1 x1 + · · · + λk xk ) ≤ λ1 f (x1 ) + · · · + λk f (xk ).
We induct on k.
For k = 2, Jensen’s inequality is equivalent to the definition of convexity. Now, suppose that k ≥ 3
and that the inequality holds for k − 1. We consider two cases. If λk = 1, then λ1 = · · · = λk−1 = 0,
so we just need to show that f (xk ) ≤ f (xk ), which is clearly true. Otherwise, we do something
sneaky. We define
λ1 λk−1
y= x1 + · · · + xk−1 .
1 − λk 1 − λk
Since λk < 1, this is legal. We note that by the induction hypothesis,
 
λ1 λk−1 λ1 λk−1
f (y) = f x1 + · · · + xk−1 ≤ f (x1 ) + · · · + f (xk−1 ).
1 − λk 1 − λk 1 − λk 1 − λk
(We can do this, because the coefficients are all nonnegative and add up to 1.) Also, by convexity, we
have
f ((1 − λk )y + λk xk ) ≤ (1 − λk )f (y) + λk f (xk ) = λ1 x1 + · · · + λk−1 xk−1 + λk xk .
Since (1 − λk )y = λ1 x1 + · · · + λk−1 xk−1 , this gives us
f (λ1 x1 + · · · + λk xk ) ≤ λ1 f (x1 ) + · · · + λk f (xk ),
which is what we needed to show.