Problem Set 3

Due: May 29, 2026 at 11:59pm EST

Instructions for Completing the Problem Set:

• Please complete all the questions provided; however, note that some may not be graded. Unless
otherwise indicated, you must provide full proofs or justifications for your answers.
• You are welcome to reference any lemmas or theorems that were proven in the textbook or during
lectures. You may also reference lemmas or theorems that were explicitly stated but not proven, as
long as doing so does not trivialize the problem.

• Collaboration with your peers is allowed and encouraged! Working together can deepen your under-
standing. However, it is essential that you personally understand the solutions and are able to explain
them in your own words.
• All submissions must be written in your own words. This ensures you have internalized the
material and are strengthening your own reasoning and problem-solving abilities.

• Please refrain from using external tools such as AI. The purpose of these problem sets is not simply
to earn a grade, but to prepare you for the final exam and, more importantly, to help you truly
master the material. Think of each problem as an opportunity to practice, explore, and solidify your
understanding.
• All submissions should be uploaded through Gradescope by the specified deadline.

• Please note that as indicated in the syllabus, the submissions will be graded on both correctness and
presentation. Please make sure to:
– Write legibly.
– Write your proofs with logical flow and narrative.
– Justify all your steps. Incase you are not able to fully justify a step, identify what needs to be
done to complete the proof.

Remember: The goal here is learning, not just completing an assignment. Focus on understanding the
concepts deeply — the grades will naturally follow. We believe in your ability to grow and succeed!

1
Questions:
1. Prove or provide a counterexample for the following:
(a) Let γ : [a, b] → Rn be a parametric curve, then the unit tangent

γ ′ (t)
T (t) =
||γ ′ (t)||

and unit normal vector

T ′ (t)
N (t) =
||T ′ (t)||
is orthogonal for all t (provided they exist).
(b) Let γ : [a, b] → Rn be a parametric curve, if T ′ (t) ̸= 0 for all t, then the unit normal N (t) and the
acceleration vector γ ′′ (t) are parallel for all t.

2. (a) Let f : Rn → R be a differentiable function such f (tx) = tm f (x) for all t ∈ R, x ∈ Rn and some
fixed m ∈ N. Show that:
n
X ∂f
xi = mf (x)
i=1
∂xi

Hint: Consider the function g(t) = f (tx) and g ′ (1)

(b) Let f : Rn → R be a non-linear function such that f (tx) = tf (x) for all t ∈ R, x ∈ Rn . Show that
f is not differentiable at 0.

3. Prove or provide a counterexample for the following:

(a) Let f : Rn → Rm and g : Rm → Rk . Suppose all directional derivatives of f and g exist at a point
p ∈ Rn and f (p) ∈ Rm respectively (Dv f (p) and Du g(f (p)) exist for all v ∈ Rn and u ∈ Rm ). Then
all directional derivatives of g ◦ f exist at p.
(b) Let S ⊂ Rn be open and f : S → Rm . Suppose that all partial derivatives are bounded on S. Then
f is continuous on S.

4. Neural networks are used in machine learning to solve complicated problems. The basic idea is twofold.
First, design a mechanism to predict the correct output given some inputs. Second, update this mech-
anism using training data to improve the quality of predictions. A critical aspect of the second step is
backpropagation. Backpropagation is an application of the chain rule.

(Optional : Watch this video summarizing neural networks. You do not need any of the specific for-
mulas.)

One common application of neural networks is classification: you are given data where each data point
is assigned to one of m + 1 categories. If I give you a new data point without its category, can you
correctly predict which category it belongs to?
Here is an example of a neural network. Each node is called a neuron.

Page 2
 Input Hidden Hidden Output
layer layer 1 layer 2 layer
a2,1
a1,1 y1
X1 a2,2
a1,2 y2
X2 a2,3
a1,3 y3
a2,4

Each layer has links to the next layer, i.e. a neuron is a function of each neuron in the previous layer.
Here is a brief description of a general neural network with L hidden layers, n inputs, and m outputs.
• Input layer has n inputs X = (X1 , X2 , X3 , . . . , Xn ) ∈ Rn .
• There are L ∈ N+ hidden layers. The input layer is layer 0 and the output layer is the layer L + 1.
• Fix ℓ ∈ {1, . . . , L}. Hidden layer ℓ has N (ℓ) neurons aℓ = (aℓ,1 , aℓ,2 , . . . , aℓ,N (ℓ) ). Each component is
a real-valued C 1 function of the previous layer. That is, for every j ∈ {1, 2, . . . , N (ℓ)}, the quantity
aℓ,j is a real-valued C 1 function of aℓ−1 . If ℓ = 1 then a1,j is a real-valued C 1 function of the input
layer X.
• Output layer has m predicted outputs ŷ = (ŷ1 , ŷ2 , . . . , ŷm ). Each component is a real-valued C 1
function of hidden layer L. That is, for i ∈ {1, 2, . . . , m}, the quantity ŷi is a real-valued C 1 function
of aL .
Pm
• For a classification problem, we further have that ŷ1 , . . . , ŷm > 0 and i=1 ŷi ≤ 1. We interpret ŷi
as the probability that X is in class i.
Ignore the diagram’s neural network and assume that L = 2 and the parameters N (1), N (2), m, and n
are arbitrary.
(a) Briefly explain why ŷ can be considered a C 1 function of X. Specify the domain and codomain of
ŷ as a function of X.
(b) Training the network requires modifying the functions a1,i and a2,j . As we change these functions,
we need to know how these changes affect ŷ.
∂ ŷi
Use Leibniz notation to express in terms of the intermediate partial derivatives of consecutive
∂Xj
∂a2,l ∂a1,k
layers and .
∂a1,k ∂Xj
(c) A piece of training data is a fixed input x = (x1 , x2 , . . . , xn ) ∈ Rn with a fixed correct output class
i ∈ {1, . . . , m + 1}. We represent this class by the fixed vector y = ei ∈ Rm+1 . Once you’ve specified
the functions for a1 , a2 , and ŷ, your neural network predicts an output of ŷ(x) instead of the fixed
correct y. The cost function C : Rn → [0, ∞) for this piece of training data is
!  
Xm Xm
C(x) = − yk ln(ŷk ) − ym+1 ln 1 − yˆj 
k=1 j=1

with the convention that 0 ln(0) = 0. The value C(x) measures the quality of your prediction for
this piece of training data. Backpropagation quantifies how the cost function changes when you
modify neurons and links in your neural network. This step in machine learning algorithms helps
choose how to best modify the neural network.

Page 3
 ∂C(x) ∂ ŷ ∂a2,l
Use matrices to express in terms of ”intermediate” partial derivatives, that is, , ,
∂Xj ∂a2,l ∂a1,k
∂a1,k
and . No justification necessary.
∂Xj

Page 4