MSML604 Homework 4

Peeyush Dyavarashetty
March 2025

1 Question 1
1.1 Question 1.a
The domain x ∈ Rn is convex, since any line segment between x1 , x2 ∈ Rn i.e
θx1 + (1 − θ)x2 stays in R⋉ , where θ ∈ [0, 1]. So,

• g1i (x) = A(i) x − b(i) is convex because it is an Affine function.

• g2i (x) = ||A(i) x − b(i) ||p is convex because norm of convex function (affine
function here) is convex if p ≥ 1.
• f (x) = maxi∈[m] ||A(i) x − b(i) ||p = maxi∈[m] g2i (x) where g2i (x) is convex
∀ i ∈ [m], implying that f (x) is convex since from the condition pointwise
maximum and supremum.
Therefore, f (x) is convex.

1.2 Question 1.b

We know that sum of eigenvalues of a square matrix is its trace. Therefore,
n
X
f (X) = λi (X) = trace(X) (1)
i=1

Given, f : Sn → R, where Sn is a set of n × n symmetric matrices and to know

whether the f is a convex function

• We should prove that Sn is a convex set.

• ∀x, y ∈ domf, f (θx + (1 − θ)y) ≤ θf (x) + (1 − θ)f (y) ∀θ ∈ [0, 1]
Checking whether Sn is a convex set: Line between 2 matrices A, B ∈ Sn should
also be in Sn . Here, we know that A = AT and B = B T and for θ ∈ [0, 1],

(θA + (1 − θ)B)T = θAT + (1 − θ)B T = θA + (1 − θ)B

1
This proves that Sn is a convex set. Now, to prove for convex function, taking
A, B ∈ Sn and θ ∈ [0, 1]

f (A) = trace(A)
f (B) = trace(B)
f (θA + (1 − θ)B) = trace(θA + (1 − θ)B)
= θtrace(A) + (1 − θ)trace(B)
= θf (A) + (1 − θ)f (B)

Here, f (θx + (1 − θ)y) = θf (x) + (1 − θ)f (y) ∀θ ∈ [0, 1], proving that f is a
convex function.

1.3 Question 1.c

Here,

f (X) = λmin (X) = −λmax (−X) (2)

Since, X is symmetric, −X is symmetric. We know that λmax (M ) is a convex

function for M ∈ Sn and the negation of convex function is concave. Therefore,
here, f is a concave function.
Reason for λmax (M ) to be a convex function:

λmax (M ) = sup||v||=1 v T M v

Here, we can observe that λmax (M ) is linear w.r.t the matrix M . Therefore, it
is an affine function, which is convex.

2 Question 2
2.1 Question 2.a
Pn
Given total budget B = 1, which is i=1 Ii = 1, We need to

max E[RT ]
subject to V ar(RT ) ≤ σT2
Xn
Ii = 1
i=1

Since it is a maximization function, we can convert to

min − E[RT ]
subject to V ar(RT ) ≤ σT2
Xn
Ii = 1
i=1

2
The problem is that we do not know whether any P
of these minimization problems
n
or constraints are convex. Now, we know RT = i=1 Ri · Ii
n
X
E[RT ] = E[ Ri · Ii ]
i=1
n
X n
X
= E[Ri ]Ii = ri Ii
i=1 i=1

Here, E[RT ] is a convex function since it is an affine function.

n
X
V ar(RT ) = V ar( Ri · Ii )
i=1
n n
 2
 X  
X
=E Ri · Ii − E Ri · Ii
i=1 i=1
n
 X 2 
=E (Ri − ri ) · Ii
i=1
n
X n
X 
=E (Ri − ri )Ii (Rj − rj )Ij
i=1 j=1
n X
X n 
=E (Ri − ri )Ii (Rj − rj )Ij
i=1 j=1
X n
n X  
= E (Ri − ri )Ii (Rj − rj )Ij
i=1 j=1
Xn X n
= Σi,j Ii Ij = I T ΣI
i=1 j=1

 
I1
 I2 
where Σ is a covariance matrix and I =  . .
 
 .. 
In
Since the covariance matrix ∇2 V ar(RT ) = Σ is symmetric and positive semi-
definite, implying that the second derivative is positive. So V ar(RT ) − σT2 is
convex.
Since investment cannot be negative Ii ≥ 0 ∀i ∈ [n]. Therefore, the convex opti-
 T
mization for this problem, where I = I1 I2 . . . In , and Σ is covariance matrix

3
is
n
X
minI1 ,I2 ,...,In − ri Ii
i=1
subject to IΣI − σT2 ≤ 0
T

Xn
Ii − 1 = 0
i=1
− Ii ≤ 0

2.2 Question 2.b

min V ar(RT )
subject to E[RT ] ≥ Rmin
Xn
Ii − 1 = 0
i=1
− Ii ≤ 0
Now, we know that
min V ar(RT ) = IΣI T
n
X
subject to E[RT ] = ri Ii ≥ Rmin
i=1
n
X
Ii − 1 = 0
i=1
− Ii ≤ 0
where both equations are convex. Using these equations, we can convert the
following convex optimization problem to the following
minI IΣI T
n
X
min
subject to R − ri Ii ≤ 0
i=1
n
X
Ii − 1 = 0
i=1
− Ii ≤ 0

3 Question 3
Given sample information s1 , s2 , . . . sm ∈ Rn , we need to find non-negative
weights w = (w1 , w2 , . . . , wm ) ∈ Rm+ such that sum of the weights wi is 1.

4
Therefore, the conditions are
wi ≥ 0 ∀ i ∈ [m] =⇒ −wi ≤ 0 ∀i ∈ [m] (3)
Xm
wi = 1 (4)
i=1

Here, if wi > 1, then the equation 4 will not be satisfied. Therefore,

wi ≤ 1 ∀i ∈ [m] (5)
Pm
Therefore, the convex optimization problem, where s̄(w) = i=1 wi · si is
m
X
minw1 ,w2 ,...,wm ||si − s̄(w)||2
i=1
subject to − wi ≤ 0 ∀i ∈ [m]
wi − 1 ≤ 0 ∀i ∈ [m]
Xm
wi − 1 = 0
i=1

To prove this convex optimization problem is convex:

• The minimization function can be written as
Xm Xm Xm Xm m
X
si − wj · sj = wj · si − wj · sj
i=1 j=1 2 i=1 j=1 j=1 2
m
X m
X
= wj · (si − sj )
i=1 j=1 2

Xm Xm
= wj · (si − sj )
i=1 j=1 2

– Here, for a particular i, since fij (w1 , w2 , . . . , wm ) = si − sj is a con-

stant as they are given and is not a function of weights, they are
convex.
Pm Pm
– Now, for a particular i, j=1 wj ·(si −sj ) = j=1 wj ·fij (w1 , w2 , . . . , wm )
is convex because of non-negative weighted sum of convex functions
wj ∀j ∈ [m] as we know that wj ≥ 0. Since si ∈ Rm , si is convex.
– Norm of a convex function is a convex Pm function. Therefore, for a
particular i, fi (w1 , w2 , . . . , wm ) = || j=1 wj · (si − sj )|| is a convex
function.
Pm
– Sum of convex functions of i, which is i=1 fi (w1 , w2 , . . . , wm ) are
convex since non-negative weighted sum of convex functions.
• Inequality constraint functions are convex as they are affine.

• Equality function is affine, which is 1T w − 1, where 1 = 1 1

 T
...1