From Statistics

to Data Mining
Master 1
Computational Colour and Spectral Imaging (COSI)
Cyber‐Physical Social System (CPS2)
Saint-Étienne, France

Fabrice MUHLENBACH
https://perso.univ-st-etienne.fr/muhlfabr/
e-mail: fabrice.muhlenbach@univ-st-etienne.fr 1
 Tutorial

• Linear Algebra → Matrix Multiplication

3 −6 2 4
➢ compute the matrix product 𝐴𝐵 with 𝐴 = and 𝐵 =
−2 4 1 2

➢ then compute the matrix product 𝐵𝐴

➢ what can you conclude?

From Statistics 2 DM –Tutorial F. Muhlenbach 2

 Tutorial

• Linear Algebra → Matrix Multiplication

➢ results:
0 0
𝐴𝐵 =
0 0
➢ but
−2 4
𝐵𝐴 =
−1 2

➢ we can observe that 𝐴𝐵 ≠ 𝐵𝐴

which is a result that is most often encountered

From Statistics 2 DM –Tutorial F. Muhlenbach 3

 Tutorial

• Linear Algebra → Matrix Determinant

➢ compute the matrix determinant of the (3 × 3) square matrix
3 2 −1
𝐶 = −1 2 3
1 1 1

From Statistics 2 DM –Tutorial F. Muhlenbach 4

 Tutorial

• Linear Algebra → Matrix Determinant

➢ Solution 1: we extend the matrix with the 2 first columns:
3 2 −1 3 2
𝐶 = −1 2 3 −1 2
1 1 1 1 1
➢ we compute the positive sum of the products in NW-SE diag. and the
negative sum of the products in SW-NE diag.
➢ the result gives:
3 × 2 × 1 + 2 × 3 × 1 + −1 × −1 × 1
− −1 × 2 × 1 − 3 × 3 × 1 − (2 × (−1) × 1)
=6+6+1+2−9+2
=8
From Statistics 2 DM –Tutorial F. Muhlenbach 5
 Tutorial

• Linear Algebra → Matrix Determinant

➢ Solution 2: we make the sum or subtraction (alternatively) of the
terms of the 1st row of the 3 × 3 -matrix 𝐶 with the determinants of
the (2 × 2)-square matrices extracted from 𝐶
3 2 −1
𝐶 = −1 2 3
1 1 1
2 3 −1 3 −1 2
➢ 3 × 𝑑𝑒𝑡 − 2 × 𝑑𝑒𝑡 + (−1) × 𝑑𝑒𝑡
1 1 1 1 1 1
➢ = 3× 2×1−3×1 −2× −1 × 1 − 3 × 1 − 1 × −1 × 1 − 2 × 1
➢ = 3 × −1 − 2 × −4 − 1 × −3
➢ = −3 + 8 + 3 = 8

From Statistics 2 DM –Tutorial F. Muhlenbach 6

 Tutorial

• Linear Algebra → Reduced row echelon form

1 −2 5 0 3
➢ the matrix 𝑅 = 0 1 3 0 0 is in row-echelon form
0 0 0 1 0
0 0 0 0 0
𝑟1 = 1 − 2 5 0 3
➢ The vectors ൞ 𝑟2 = 0 1 3 0 0 form a basis for the row space of 𝑅,
𝑟3 = 0 0 0 1 0
1 −1 0
and the vectors 𝑐1 = 0 , 𝑐2 = 1 , 𝑐4 = 0
0 0 1
0 0 0
form a basis for the column space of 𝑅

From Statistics 2 DM –Tutorial F. Muhlenbach 7

 Tutorial

• Linear Algebra → Reduced row echelon form

➢ exercise: reduce the matrix 𝐴 to row-echelon form
(first step for finding the bases for the row and column spaces)
 1 −3 4 − 2 5 4 
 2 −6 9 −1 8 2 
A= 
 2 −6 9 −1 9 7 
 
 −1 3 −4 2 −5 −4 
➢ solution: reduction 𝐴 to row-echelon form
1 −3 4 −2 5 4
0 0 1 3 −2 −6 
R=
0 0 0 0 1 5
 
0 0 0 0 0 0

From Statistics 2 DM –Tutorial F. Muhlenbach 8

 Tutorial

• Linear Algebra → Reduced row echelon form

➢ exercise: find a basis for the space spanned by the vectors
𝑣1 = (1, −2, 0, 0, 3), 𝑣2 = (2, −5, −3, −2, 6),
𝑣3 = (0, 5, 15, 10, 0), 𝑣4 = (2, 6, 18, 8, 6).
➢ solution: write down the vectors as row vectors first
 1 −2 0 0 3  1 −2 0 0 3
 2 −5 −3 −2 6  0 1
  3 2 0 
 0 5 15 10 0 0 0 1 1 0
   
 2 6 18 8 6
0 0 0 0 0
➢ the nonzero row vectors in this matrix are
𝑤1 = (1, −2, 0, 0, 3), 𝑤2 = (0, 1, 3, 2, 0), 𝑤3 = (0, 0, 1, 1, 0)

From Statistics 2 DM –Tutorial F. Muhlenbach 9

 Tutorial

• Linear Algebra → Reduced row echelon form

➢ keeping in mind that 𝐴 and 𝑅 may have different column spaces, we
cannot find a basis for the column space of 𝐴 directly from the
column vectors of 𝑅
➢ however, if we can find a set of column vectors of 𝑅 that forms a
basis for the column space of 𝑅, then the corresponding column
vectors of 𝐴 will form a basis for the column space of 𝐴
➢ in the previous example, the basis vectors obtained for the column
space of 𝐴 consisted of column vectors of 𝐴, but the basis vectors
obtained for the row space of 𝐴 were not all vectors of 𝐴
→ transposition of the matrix

From Statistics 2 DM –Tutorial F. Muhlenbach 10

 Tutorial

• Linear Algebra → Reduced row echelon form

1 −2 0 0 3
➢ find a basis for the row space of  2 −5 −3 −2 6 
A=
 0 5 15 10 0
 
 2 6 18 8 6
consisting entirely of row vectors from 𝐴
➢ solution:  1 2 0 2 1 2 0 2 
 −2 −5 5 6  0 1 5 −10 
  
A =  0 −3 15 18
T
0 0 0 1 
   
 0 −2 10 8  0 0 0 0 
 3 6 0 6  0 0 0 0 
➢ the nonzero vectors in this matrix are
𝑤1 = (1, 2, 0, 2), 𝑤2 = (0, 1, 5, −10) and 𝑤3 = (0, 0, 0, 1)
From Statistics 2 DM –Tutorial F. Muhlenbach 11
 Tutorial

• Convex Optimization
➢ Exercise 1: we want to solve:
max 𝑥. 𝑦
ቊs. t. 𝑥 + 3𝑦 = 24

➢ Questions:
o what are the objective and constraint functions?
o solve this system by using the method of Lagrange multipliers

From Statistics 2 DM –Tutorial F. Muhlenbach 12

 Tutorial

• Convex Optimization
➢ Solution:
max 𝑥. 𝑦
ቊs. t. 𝑥 + 3𝑦 = 24

o objective function: max 𝑥. 𝑦

o constraint function: 𝑥 + 3𝑦 = 24 ⟺ 𝑥 + 3𝑦 − 24 = 0

o method of Lagrange multipliers:

→ ℒ 𝑥, 𝑦, 𝜆 = objective function to optimize − 𝜆 constraint
o Lagrange function to optimize:
ℒ 𝑥, 𝑦, 𝜆 = 𝑥. 𝑦 − 𝜆 𝑥 + 3𝑦 − 24 = 𝑥. 𝑦 − 𝜆𝑥 − 3𝜆𝑦 + 24𝜆
From Statistics 2 DM –Tutorial F. Muhlenbach 13
 Tutorial

• Convex Optimization
o ℒ 𝑥, 𝑦, 𝜆 = 𝑥. 𝑦 − 𝜆 𝑥 + 3𝑦 − 24 = 𝑥. 𝑦 − 𝜆𝑥 − 3𝜆𝑦 + 24𝜆
o optimizing a function → finding the critical points
→ finding when the derivative of the function is equal to zero
o we have 3 variables: 𝑥, 𝑦, 𝜆, therefore the function need to be derived
3 times
o there are 3 first order conditions:
𝜕ℒ
o =0⇔𝑦−𝜆 =0⇔𝜆 =𝑦 1
𝜕𝑥
𝜕ℒ 𝑥
o = 0 ⇔ 𝑥 − 3𝜆 = 0 ⇔ 𝜆 = 2
𝜕𝑦 3
𝜕ℒ
o = 0 ⇔ 𝑥 + 3𝑦 − 24 = 0 3
𝜕𝜆
From Statistics 2 DM –Tutorial F. Muhlenbach 14
 Tutorial

• Convex Optimization
o ℒ 𝑥, 𝑦, 𝜆 = 𝑥. 𝑦 − 𝜆 𝑥 + 3𝑦 − 24 = 𝑥. 𝑦 − 𝜆𝑥 − 3𝜆𝑦 + 24𝜆
𝑥
o 1 and 2 : 𝑦 = 4
3
𝑥
o 3 and 4 : 𝑥 + 3 × − 24 = 0
3
o ⇔ 𝑥 + 𝑥 − 24 = 0
o ⇔ 2𝑥 = 24
o ⇔ 𝑥 = 12
𝑥 12
o with 4 : 𝑦 = = =4
3 3 𝑥 = 12
o with 1 : 𝜆 = 𝑦 = 4 therefore ቐ 𝑦 = 4
𝜆=4
From Statistics 2 DM –Tutorial F. Muhlenbach 15
 Tutorial

• Convex Optimization
➢ Exercise 2: we want to solve:
min 2𝑥 + 2𝑦
ቊ
s. t. 𝑥. 𝑦 = 4
➢ Solution: → constraint: 𝑥. 𝑦 = 4 ⟺ 𝑥. 𝑦 − 4 = 0
o ℒ 𝑥, 𝑦, 𝜆 = 2𝑥 + 2𝑦 − 𝜆 𝑥. 𝑦 − 4 . FOC:
𝜕ℒ 2
o = 0 ⇔ 2 − 𝜆𝑦 = 0 ⇔ 2 = 𝜆𝑦 ⇔ 𝜆 = 1
𝜕𝑥 𝑦
𝜕ℒ 2
o = 0 ⇔ 2 − 𝜆𝑥 = 0 ⇔ 2 = 𝜆𝑥 ⇔ 𝜆 = 2
𝜕𝑦 𝑥
𝜕ℒ
o = 0 ⇔ 𝑥. 𝑦 − 4 = 0 3
𝜕𝜆

From Statistics 2 DM –Tutorial F. Muhlenbach 16

 Tutorial

• Convex Optimization
o ℒ 𝑥, 𝑦, 𝜆 = 2𝑥 + 2𝑦 − 𝜆 𝑥. 𝑦 − 4
2 2
o 1 and 2 : = ⇔𝑥=𝑦 4
𝑦 𝑥
o 3 and 4 : 𝑥 2 − 4 = 0 ⇔ 𝑥 2 = 4 ⟺ 𝑥 = 2 5
o 4 and 5 : 𝑦 = 2 and 𝜆 = 1
𝑥=2
o therefore ቐ𝑦 = 2
𝜆=1

From Statistics 2 DM –Tutorial F. Muhlenbach 17

 Tutorial

• Convex Optimization
➢ Exercise 3: we want to solve:
𝑓 𝑥, 𝑦 = max 𝑥 2 𝑦
൝
s. t. 𝑥 2 + 𝑦 2 = 1

o what is the graphical interpretation of 𝑥 2 + 𝑦 2 = 1?

o what is the graphical interpretation of 𝑥 2 𝑦?
o therefore, what is the graphical interpretation of
𝑓 𝑥, 𝑦 = max 𝑥 2 𝑦
൝ 2 2 ?
s. t. 𝑥 + 𝑦 = 1

From Statistics 2 DM –Tutorial F. Muhlenbach 18

 Tutorial

• Convex Optimization
➢ Solution: we want to solve:
𝑓 𝑥, 𝑦 = max 𝑥 2 𝑦
൝
s. t. 𝑥 2 + 𝑦 2 = 1
o graphical interpretation of 𝑥 2 + 𝑦 2 = 1:
𝑦 ❑ 𝑥2 + 𝑦2 = 1
❑ value for 𝑥: 𝑥𝐴
❑ ∀𝐴, 𝑥𝐴2 + 𝑦𝐴2 = 𝑥𝐴 , 𝑦𝐴 2 = 1
❑ value for 𝑦: 𝑦𝐴 ❑ 𝑥𝐴 , 𝑦𝐴 = 1
2 2 2
❑ 𝑥𝐴 + 𝑦𝐴 = 𝑥𝐴 , 𝑦𝐴 𝑦𝐴 𝐴 ⇔ 𝑂, 𝐴 = 1
❑ 𝑥𝐴 , 𝑦𝐴 = [𝑂, 𝐴] → circle with
𝑥 radius with the
𝑂 𝑥𝐴 size of 1
From Statistics 2 DM –Tutorial F. Muhlenbach 19
 Tutorial

• Convex Optimization
o graphical interpretation of 𝑥 2 + 𝑦 2 = 1: circle with radius =1
o 𝑥2 + 𝑦2 = 1 ⟺ 𝑦2 = 1 − 𝑥2
o ⟺ 𝑦 = 1 − 𝑥2 or 𝑦 = − 1 − 𝑥 2

From Statistics 2 DM –Tutorial F. Muhlenbach 20

 Tutorial

• Convex Optimization
1
o graphical interpretation of 𝑥 2 𝑦: 𝑦 = 𝑧 ×
𝑥2

𝑧=3
𝑧=2
𝑧=1

𝑧=0

𝑧 = −1

𝑧 = −2

From Statistics 2 DM –Tutorial F. Muhlenbach 21

 Tutorial

• Convex Optimization
1
𝑓 𝑥, 𝑦 = max 𝑥 2 𝑦 ⇒ 𝑦 = 𝑧 ×
o graphical interpretation of ቐ 𝑥2
s. t. 𝑥 2 + 𝑦 2 = 1 ⇒ 𝑦 = ± 1 −𝑥2
o when 𝑓 𝑥, 𝑦 = 𝑧 ≤ 0.3, the
𝑧 = 1.0
constraint intersects with the
𝑧 = 0.4
𝑧 = 0.3 circle, some solutions for the 2
𝑧 = 0.2
𝑧 = 0.1 conditions exists
𝑧=0
o when 𝑓 𝑥, 𝑦 = 𝑧 ≥ 0.4, no
intersection → no solution
o maxima → tangency points
o 𝑧 ? graphically, 𝑧 ∈ 0.3; 0.4
From Statistics 2 DM –Tutorial F. Muhlenbach 22
 Tutorial

• Convex Optimization
➢ Solution: we want to solve:
𝑓 𝑥, 𝑦 = max 𝑥 2 𝑦
൝
s. t. 𝑥 2 + 𝑦 2 = 1
➢ constraint: 𝑥 2 + 𝑦 2 = 1 ⟺ 𝑥 2 + 𝑦 2 − 1 = 0
o ℒ 𝑥, 𝑦, 𝜆 = 𝑥 2 𝑦 − 𝜆 𝑥 2 + 𝑦 2 − 1 . FOC:
𝜕ℒ
o = 0 ⇔ 2𝑥. 𝑦 − 𝜆2𝑥 = 0 ⇔ 2𝑥. 𝑦 = 𝜆2𝑥 ⇔ 𝑥 ≠ 0, 𝜆 = 𝑦 1
𝜕𝑥
𝜕ℒ
o = 0 ⇔ 𝑥 2 − 𝜆2𝑦 = 0 ⇔ 𝑥 2 = 𝜆2𝑦 ⇔ 1 ⇒ 𝑥 2 = 2𝑦 2 2
𝜕𝑦
𝜕ℒ
o = 0 ⇔ 𝑥2 + 𝑦2 − 1 = 0 3
𝜕𝜆

From Statistics 2 DM –Tutorial F. Muhlenbach 23

 Tutorial

• Convex Optimization
o ℒ 𝑥, 𝑦, 𝜆 = 𝑥 2 𝑦 − 𝜆 𝑥 2 + 𝑦 2 − 1
o 3 and 2 : 2𝑦 2 + 𝑦 2 − 1 = 0 ⇔ 3𝑦 2 = 1
1 1
o ⇔ 𝑦2 = ⇒𝑦=± 4
3 3

1 2 2
o 2 and 4 : 𝑥 2 = 2𝑦 2 = 2 × ⇔ 𝑥 2 = ⇒ 𝑥 = ± 5
3 3 3

1
o 1 and 4 : 𝑦 = 𝜆 ⇒ 𝜆 = ±
3

2 1 2 1 2 1 2 1
o solutions: , , ,− , − , and − ,−
3 3 3 3 3 3 3 3

From Statistics 2 DM –Tutorial F. Muhlenbach 24

 Tutorial

• Convex Optimization
o ℒ 𝑥, 𝑦, 𝜆 = 𝑥 2 𝑦 − 𝜆 𝑥 2 + 𝑦 2 − 1
2 1 2 1 2 1 2 1
o solutions: , , ,− , − , and − ,−
3 3 3 3 3 3 3 3

o → the possible solutions are the four points where the contour lines
are tangent
o which one maximize the function 𝑓 𝑥, 𝑦 = max 𝑥 2 𝑦
o but 𝑦 cannot be negative because 𝑥 2 𝑦 will be negative
2 1 2 1 1 2
o 𝑓 , =𝑓 − , = → 2 solutions
3 3 3 3 3 3

o 𝑧 = 𝑥 2 𝑦 = 2Τ3 1Τ3 ≅ 0.3849 (reminder: 𝑧 ∈ 0.3; 0.4 )

From Statistics 2 DM –Tutorial F. Muhlenbach 25
 Tutorial

• Convex Optimization
o function to
maximize:
1
𝑧 = 2Τ3 1Τ3
1
→𝑦=𝑧× 2
𝑥

-1
o constraint:
0 1
𝑦 = ± 1 − 𝑥2
o solutions:
2 1 2 1
, , − ,
3 3 3 3

From Statistics 2 DM –Tutorial F. Muhlenbach 26