Errata

(Mathematical Introduction to Data Science by Sven A. Wegner)

May 1, 2025

n Page 8, Line −4:

n
P n
P n
P n
P
(xi − x)(yi − y) = x i yi − y xi − x yi + nxy
i=1 i=1 i=1 i=1

n Page 9, Line −3:

n
P n
P n
P
0= (axi + b − yi ) = a xi + nb − yi = anx + nb − ny,
i=1 i=1 i=1

n Page 13, Line 10: . . . constant random variable av.

2
n Page 13, Line −5: since x(n) 2 = var(x(n) ) + x(n) as the sum . . .
Page 15, Line −9: · · · + n22 xi xj E(Ei ) E Ej )
P
n
i<j
n Page 16, Line 10: . . . , i.e., f ∗ (x) = ha∗ , xi + b∗ . . .
n Page 14, Line −2: If the latter is the case, then sign(rxy ) = sign(hu, vi) = · · ·
n Page 17, Line13: . . . we calculate (with just for now φ(ã) = hã, X T Xãi):
n Page 23, Line9: In the picture it must be the z-axis.
n Page 23, Line−3: . . . and (w, b) = (w1 , . . . , wd , b) for . . .
 
n Page 24, Line16: P Yi (f ) = yi for all i = · · ·
Page 24, Line 18: L : f : Rd → (0, 1) | f logistic function → R

n

n Page 26, Line 13: h0 (t) = sig(t)+1

n Page 27, Line 18: . . . the data overlaps if for every w ∈ Rd+1 \{0} there exists . . .
n Page 27, Line 22: . . . , the rounded logisitic logistic regressor, . . .
n Page 27, Line −10: if sig(hw, > xi) > 1/2
n Page 37, Line 1:
1: function k-NN Classifier(D, k, x)
2: D0 ← D, A ← ∅
3: for j ← 1 to k do
4: z ∗ ← argminz∈D0 ρ(x, π1 (z))
5: A ← A ∪ {z ∗ }, D0 ← D0 \{z ∗ }
6: for y in Y do
7: N (y) ← #{a ∈ A | π2 (a) = y}
8: ` ← argmaxy∈Y N (y)
9: return `
Here, π2 (x, y) = y denotes the projection onto the second entry of (x, y) ∈ D and y ∗ is
the label of z ∗ .
n Page 39, Line 9: . . . and k < n.
n Page 39, Line 10: The calculation of the k-neareast neighbors of x can be implemented
such that at most (n · d · k)-many mulitplikations have to be carried out.
n Page 39, Line 14: In the Euklidean metric, it requires (d − 1)-many multiplications to
compute one distance if we omit the root, which we can do as it does not change the
argmin. This leads to
(d − 1) · (n + (n − 1) + · · · + (n − k + 1)) 6 C· d · k · n
multiplikationen with a suitable C ∈ N.
n Page 40, Line 3: . . . , we choose k-nearest neighbors x1 , . . . , xk of x and denote their
labels by y1 , . . . , yk .
 k
w(xi , x) · yi
P
i=1
n Page 40, Line 9: f : X → Y, f (x) = k
P
w(xi , x)
i=1

(x(i) (j)
1 − minj=1,...,n x1 )(b − a)
 
n Page 41, Line 5: x̃(i) = a+ (j) (j)
, . . .
maxj=1,...,n x1 − minj=1,...,n x1
 x(i) − x(·) 
n Page 41, Line 7: x̃(i) = 1
σ1
1
,...
(1) (4)
n Page 42, Line 14: ρ(x , x ) = 3.681
n Page 46, Line 1: We discuss some of these methods in Exercise 3.10.
n Page 45, Line 28: We thus see that texts no. 1 and text no. 2 are significantly more
cosine similar than text no. 1 and text no. 3 or text no. 2 and text no. 3.
n Page 48, Line 26: The cosine distance, on the other hand, may appear here more natural,
as the scalar product increases if the frequency of the fixed word increases in the second
text.
n Page 52, Line 11: For finite subsets A, B ⊆ X we define . . .
n Page 53, Line 20:
1: function Linkage-based Clustering (X, ρ, D, δ)
2: k ← #D
3: for i ← 1 to k do
4: Ci ← {xi }
5: while mini6=j ρ(Ci , Cj ) 6 δ and k > 2 do
6: m←0
7: (i∗ , j ∗ ) ← argmini6=j ρ(Ci , Cj )
8: for ` ← 1 to k − 1 do
9: if ` = min(i∗ , j ∗ ) then
10: C` ← Ci∗ ∪ Cj ∗
11: if ` = max(i∗ , j ∗ ) then
12: m←1
13: C` ← C`+m
14: else
15: C` ← C`+m
16: k ← k−1
17: return C1 , . . . , Ck
n Page 54, Line −6: K : Ck → R
n Page 56, Line −16: The following pseudocode approximates a minimizer of the k-means
cost function.
n Page 56, Pseudocode:
1: function k-means (D, k, X, ρ)
2: µ1 , . . . , µk ← pairwise different points from X
3: for i ← 1 to k do
4: Ci ← {x ∈ D | i ∈ argminj=1,...,k ρ(x, µj )}
5: U ← True
6: while U = True do
7: U ← False
8: for i ← 1 to k do
9: µi ← µ(Ci )
10: for i ← 1 to k do
11: Ci0 ← {x ∈ D | i ∈ argminj=1,...,k ρ(x, µj )}
12: if Ci0 6= Ci
13: Ci ← Ci0
14: U ←True
15: return C1 , . . . , Ck
 In the lines 4 and 9 of the pseudocode we pick as a single i in the case that the armin is
not unique.
n Page 57, Line 7:
ρ(x, µ)2 ,
P P
µ(A) ∈ argmin respectively µ(A) ∈ argmin ρ(x, µ).
µ∈A x∈A µ∈X x∈A

n Page 57, Line −7: For j > 1 denote by (C1(j) , . . . , Ck(j) ) that clustering which the algorithm
produces in the j-th round. For j > 2 we have
k
ρ(x, µi )2
P P
K(C1(j) , . . . , Ck(j) ) = min
µ1 ,...,µk ∈X i=1 (j)
x∈C i

n Page 58, Line 3: . . . line 11 of Algorithm 4.9 . . .

n Page 58, Line 7: There, Picture 1b corresponds to the penultimate line in the estimate
and Picture 2a to the line above that.
n Page 58, Line 10: . . . as we have just moved point x3 from cluster C2 in Figure 1b to
cluster C1 in Figure 2a, . . .
n Page 59, Line −4: We assume that we start with the initial values µ1 = 2 and . . .
n Page 62, Line 11: A = (aij )i,j=1,...,n
n Page 62, Line 13: L = (`ij )i,j=1,...,n
n Page 63, Line −2: . . . In Example 5.7, λ2 6= 0 and there are no clusters (or, depending
on how one prefers to see it, one single cluster), in Example 5.8 . . .
n Page 67, Line 8: For the other direction let {v1 , . . . , vn } be a basis consisting of eigenvec-
tors corresponding to the λi and let U ⊆ Rn be a subspace with dim U = n − k + 1. By
construction U ∩span{v1 , . . . , vk } =
6 {0} and we can select 0 6= x = α1 v1 +· · ·+αk vk ∈ U .
Then it follows k k
λi αi2 λk αi2
P P
hx, M xi
= i=1k 6 ki=1 = λk ,
hx, xi P 2 P 2
αi αi
i=1 i=1

since the λi ’s are increasing. With this we get min06=x∈U hx,M xi

hx,xi 6λk which then leads to

hx, M xi
maxn min 6 λk .
U ⊆R x∈U
x6=0
hx, xi
dim U =n−k+1

n Page 67, Line −4: . . . and any corresponding eigenvector v2 of norm 1:

n Page 68, Line 14: Let G = (V, E) be a graph with deg(v) > 0 for all v ∈ V .
n Page 70, Line 10:
(xi − xj )2
X

{i,j}∈E
λ2 (L) = min n .
x6=0
x2i di
X
hDx,1i=0
i=1

n Page 70, Line 16:

LD1/2 1 = D−1/2 LD−1/2 D1/2 1 = D−1/2 L1 = D−1/2 01 = 0D1/2 1.
↑
Proposition
5.6

n Page 70, Line 18:

Theorem
5.13 hx, Lxi (∗) hD1/2 y, LD1/2 yi
λ2 (L) = min hx,xi = min
x6=0 y6=0 hD1/2 y,D1/2 yi
hx,D1/2 1i=0 hD1/2 y,D1/2 1i=0
2
P
hy,Lyi {i,j}∈E (yi −yj )
= min hy,Dyi = min Pn 2
y6=0 y6=0 i=1 yi di
hy,D1i=0 hDy,1i=0
n Page 71, Line 14: (ii) min(vol Sk , vol Skc ) = vol Skc and vol Skc − vol Sk+1
c = dk+1 hold
whenever r 6 k 6 n − 1.
n Page 71, Line 17:
n n
vol Skc − vol Sk+1
c
P P
= di − di = dk+1
i=k+1 i=k+2
n Page 72, Line 9:
n
P
hDx, 1i = di xi = . . .
i=1
n Page 72, Line −1 (and Page 73, Line 1):
vol S
· · · = vol S − vol S · vol S+vol S c
vol S
> vol S − vol S · 2 vol S ,

n Page 73, Line 9: . . . and our goal in the following will be to show λ2 > α2 /2, . . .
n Page 73, Line 12 (Equation (5.2)):
n
P
· · · and hDx, 1i = di xi = 0
i=1

n Page 73, Line 21:

 x1 −xr   x1 −xr   0

.. .. ..
.
 xr−1 −xr   x −x   . .
0

  r−1 r  
= −
  x −x  =: p − n.
0
 
 0   0
 xr+1.−xr   0
..   r
..
r+1 
.. . .
xn −xr x −x 0 r n

n Page 74, Line −3 (until top of page 75):

(xi − xj )2
P
{i,j}∈E
λ2 = Pn 2
i=1 xi di
2 2
P

{i,j}∈E (pi − pj ) + (ni − nj )
> Pn 2 2
↑ i=1 (pi + ni )di
(5.3)
(5.4) P
2 2
P
{i,j}∈E (pi − pj ) + {i,j}∈E (ni − nj )
= Pn 2
Pn 2
i=1 pi di + i=1 ni di
2 2
P P
{i,j}∈E (pi − pj ) {i,j}∈E (ni − nj )

> min P n 2 , P n 2
↑ i=1 pi di i=1 ni di
(5.5)
2 2
P P
{i,j}∈E (pi − pj ) {i,j}∈E (pi + pj )
 
= min Pn 2 · P 2
, . . .
i=1 pi di {i,j}∈E (pi + pj )

Z
 
=: min N , . . . .

n Page 75, Line 8:

n n
p2i di · (pi + pj )2 > p2i di · 2(p2i + p2j )
P P P P
N=
i=1 {i,j}∈E ↑ i=1 {i,j}∈E
(5.5)

n Page 76, Line 3:

 n−1 n j−1
2
p2k − p2k+1
P P P
··· = 1E (i, j)
↑
telescopic
i=1 j=i+1 k=i
sum
n Page 79, Line 13: Finally, we want to note that Theorem 5.21 together with Remark
5.17(i) provides an upper bound for the eigenvalue λ2 (L).
n Page 82, Line −5: Let a dataset D = {x1 , . . . , xn } ⊆ Rd be given, . . .
n Page 99, Line −3: Multiplication with V T from the left in
Corollary 5.22. Let G = (V, E) be a graph with deg(i) > 0 for all i ∈ V . Then for
the second smallest eigenvalue λ2 of the normalized Laplace matrix of G the estimate
λ2 6 2 holds. 
n Page 105, Zeile −1 and Page 106, Line 1:
 0.07 0.29 0.32 0.51 0.66 0.18 90.23   12.4 0.0 0.0 0.0 0.0   0.56 0.09 0.56 0.09 0.59 
0.13 90.02 90.01 90.79 0.59 90.02 90.06 0.0 9.5 0.0 0.0 0.0
 0.68 90.12 0.69 90.12 0.69 0.02
90.11 90.05 90.05 90.24 0.56 90.35   0.0 0.0 1.3 0.0 0.0 
  90.40 90.09 90.40 90.09 0.80 
A=  0.15
 0.59 0.65 90.25 90.33 90.09 0.11   0.0

0.0 0.0 0.0 0.0   0.41 0.09 0.40 0.09 90.80 
 0.41 90.07 90.03 0.10 90.02 90.78 90.43   0.0 0.0 0.0 0.0 0.0 

0.51 0.48 90.51 90.48 90.00

0.07 0.73 90.67 0.00 90.00 0.00 0.00 0.0 0.0 0.0 0.0 0.0 0.48 90.51 90.48 0.51 90.00
0.55 90.09 90.04 0.17 0.17 90.11 0.78 0.0 0.0 0.0 0.0 0.0
| {z } | {z } | {z }
U Σ VT
 0.07 0.29 0.32 
0.13 90.02 90.01
 0.68 90.05  12.4
90.11
   0.56 0.09 0.56 0.09 0.59 
=  0.15
 0.59 0.65 
 9.5 90.12 0.69 90.12 0.69 0.02 .
 0.41 90.07 90.03  1.3 90.40 90.09 90.40 90.09 0.80
0.07 0.73 90.67
0.55 90.09 90.04

n Page 106, Zeile 6:

 0.15 1.97 0.15 1.97 0.56   0.07 0.29 
0.92 0.01 0.92 0.01 0.94 0.13 90.02
 4.84 0.03 4.84 0.03 4.95   0.68 90.11  h ih i
 12.4 0.56 0.09 0.56 0.09 0.59
Ǎ = =
   0.15 0.59 
 0.36 4.03 0.36 4.03 1.20   9.5 90.12 0.69 90.12 0.69 0.02
 2.92 90.00 2.92 90.00 2.98   0.41 90.07 
90.34 4.86 90.34 4.86 0.65 0.07 0.73
3.92 0.02 3.92 0.02 4.00 0.55 90.09

n Page 107, Line 15:

 0.07 0.29 90.23 
..
.

0.13 90.02 90.06 " # 0.56 0.09 0.56 0.09 0.59

 0 
 0.68 12.4 ..
90.11 90.35  9.5 90.12 0.69 90.12 0.69 0.02
= [0 1 0 0 ] 0.15
 0.59 0.11 
  . ..  . 
..
.

 0.41 90.07 90.43 

1.3 .. . 0
1
.
..

0.07 0.73 0.00 0.48 90.51 90.48 0.51 0.00 0

0.55 90.09 0.78
..
.

n Page 107, Line −2:

u2 = 0.29 · Abbie − 0.02 · Bailey + · · · −0.09 · Gladys,

n Page 109, Line 10:

Abbie −→
Bailey −→  0.07 0.29

Alien Casablanca Star Wars Titanic The Matrix
Catherine −
−→

0.13 90.02
−→

→
−→
−

Darlene −
→

→ 0.68 90.11 h ih i
  12.4 0.56 0.09 0.56 0.09 0.59
Ǎ =  0.15 0.59 
9.5 90.12 0.69 90.12 0.69 0.02
0.41 90.07
 
Elena −→
 
0.07 0.73
Fatima −→
Gladys −→
0.55 90.09

n Page 109, Line -3: . . . and V̌ = {v1 , v2 }.

n Page 229, Line 13: Alternatively, with sigmoid activation, . . .
P(A ∩ B)
n Page 282, Line 6: (ii) For A, B ∈ Σ with P(B) 6= 0, P(A|B) := ...
P(B)
n Page 283, Line 19:
kx−µk2
1
ρ(x) = 1
(2πσ 2 )d/2
e− 2σ 2 respectively ρ(x) = · 1B (x),
λd (B)

n Page 286, Line −5: For A = A1 × · · · × Ad ⊆ Rd with Ai ∈ Bd we calculate

n Page 288, Line 2:
Z
1 (s − t)2
t2

(ρ1 ∗ ρ2 )(s) = √ exp − 2a
exp − 2b dt
2π ab
ZR
1 b(s2 − 2st + t2 ) + at2

= √ exp − 2ab
dt
2π ab
ZR
1 t2 (b + a) − 2stb + bs2

= √ exp − 2ab
dt
2π ab
ZR
1 t2 (b + a)/c − 2stb/c + bs2 /c

= √ exp − dt
2π ab R
2ab/c
Z
(t − (bs)/c)2 − (sb/c)2 + s2 (b/c)

=√ p1 exp − dt
2πc 2π(ab/c) R
2ab/c
Z
(sb/c)2 − s2 (b/c) (t − (bs)/c)2

= √ 1 exp(+ )p 1 exp − dt
2πc 2ab/c 2π(ab/c) R
2ab/c
(sb/c)2 c2 − s2 (b/c)c2
= √1 exp(+ )
2πc 2abc
s2 (b2 − bc)
= √1 exp(+ 2abc
)
2πc
2
= √1 s
exp(− 2c ),
2πc