Matrix World

Kenji Hiranabe
2020/9/29
updated 2023/3/23
 Matrix World : The Picture of All Matrices

I am happy to tell the history of Matrix World—the creation of Kenji Hiranabe in Japan. In April
2020 his friend Satomi Joba asked if I would send him a birthday message as a surprise. He was happy
(and very surprised). Kenji combines mathematics with art and with computing : three talents in one. I
was the one to be surprised when he sent Matrix World in its first form—without a name, without
many of the entries and ideas that you see now, but with the central idea of displaying the wonderful
variety of matrices.

Since that first form, Matrix World has steadily grown. It includes every property that would fit
and every factorization that would display that property. Interesting that the SVD is in the outer circle
and the identity matrix is at the center—it has all the good properties : the matrix I is diagonal,
positive definite symmetric, orthogonal, projection, normal, invertible, and square.

Lek-Heng Lim has pointed out the usefulness of matrices M that are symmetric and
orthogonal—kings and also queens. Their eigenvalues are 1 and −1. They have the form M = I − 2P
(P = symmetric projection matrix). There is a neat match between all those matrices M and all
subspaces of Rn . You may see something interesting (or something missing) in Matrix World. We
hope you will ! Thank you to Kenji.
Gilbert Strang
 Walk througn the "Matrix World"
There are many categories of matrices: Symmetric, Orthogonal, Singular, Invertible, Square, Diagonalizable, and more. This is a
map of the categories using a venn diagram. From the top to the bottom, I'm walking you through this diagram. All matrices (𝑚×𝑛) can
be factorized to 𝑨 = 𝑪𝑹 and also be factorized into 𝑨 = 𝑼𝜮𝑽𝐓 .
Square matrices (𝑛×𝑛) are either Invertible or Singular. Invertibility, indicated by the dotted line through the center, can be checked by
whether 𝑨 = 𝑳𝑼 has full pivots, det (𝑨) ≠ 𝟎 or all eigenvalues are nonzero.
An Invertible matrix is factorized into 𝑨 = 𝑸𝑹 using Orthogonal matrix 𝑸 (Gram-Schmidt). A Square matrix is either Diagonalizable
if it has 𝑛 independent eigenvectors.
A Diagonalizable matrix is factorized into 𝑨 = 𝑿𝚲𝑿"𝟏 with a Diagonal eigenvalue matrix 𝚲 and an Invertible eigenvector matrix 𝑿. A
Non-diagonalizable matrix has its Jordan normal form 𝑱 instead of the diagonal matrix 𝚲.
There is an important category Normal which includes Symmetric, Orthogonal, Anti-Symmetric matrices. A matrix is Diagonalizable
as 𝑸𝚲𝑸"𝟏 by an Orthogonal matrix 𝑸 when and only when it is Normal (𝑨𝐓 𝑨 = 𝑨𝑨𝐓 ).
Symmetric matrices (𝑺 = 𝑺𝐓 ) and Orthogonal matrices (𝑸"𝟏 = 𝑸𝐓 ) are important Normal matrices. Some matrices are both
Symmetric and Orthogonal. They include reflection matrices with eigenvalues 1 and -1.

All eigenvalues of an Orthogonal matrix have |𝜆| = 1. All eigenvalues of a Symmetric matrix are real. A Symmetric matrix is called
Positive Semidefinite if all eigenvalues 𝜆 ≥ 0, and Positive Definite if all eigenvalues 𝜆 > 0.
Projection matrices (𝑷𝟐 = 𝑷 = 𝑷𝐓 ) are Symmetric and Positive Semidefinite, with all the eigenvalues are 1 or 0.
The Identity is the only Invertible Projection matrix.

𝑨𝐓 𝑨 for any matrix 𝑨 is Positive Semidefinite and is Positive Definite if and only if the columns of 𝑨 are independent.
If a matrix is Invertible, its inverse can be expressed as 𝑨"𝟏 = 𝑽𝜮𝑼𝐓 . Any matrix even if it is not Invertible nor even Square, there
exists a pseudoinverse 𝑨"𝟏 = 𝑽𝜮𝑼𝐓 .
Matrix World Matrix (𝑚×𝑛) Matrix Factorization
Appearing section
in 1.4 𝐴 = 𝐶𝑅 𝐴 = 𝑈Σ𝑉 > 7.1
𝑟𝑜𝑤 𝑟𝑎𝑛𝑘 = 𝑐𝑜𝑙𝑢𝑚𝑛 𝑟𝑎𝑛𝑘 𝑆𝑉𝐷: 𝑜𝑟𝑡ℎ𝑜𝑛𝑜𝑟𝑚𝑎𝑙 𝑏𝑎𝑠𝑖𝑠 𝑈, 𝑉 (in Linear Algebra for Everyone)
Linear Algebra
for Everyone 𝐴=
1 2 3 Square Matrix (𝑛×𝑛)
4 5 6 Invertible Singular
det(𝐴) ≠ 0, a𝑙𝑙 𝜆 ≠ 0 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝜆 = 0, det(𝐴) = 0
4.4 𝐴 = 𝑄𝑅 𝑇𝑟𝑖𝑎𝑛𝑔𝑢𝑙𝑎𝑟𝑖𝑧𝑒 𝐴 = 𝐿𝑈 2.3
Gram-Schmidt 𝑈 ℎ𝑎𝑠 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝑧𝑒𝑟𝑜 𝑟𝑜𝑤
2 1
𝐴=
0 2 Diagonalizable
1 1 6.2 𝐴 = 𝑋Λ𝑋 ?@ 𝐷𝑖𝑎𝑔𝑜𝑛𝑎𝑙𝑖𝑧𝑒 𝐴 = 𝑋𝐽𝑋 ?@ A7
𝐴=
0 2 𝐽 = 𝐽𝑜𝑟𝑑𝑎𝑛 𝑓𝑜𝑟𝑚
A5 Normal 𝐴=
0 1
𝐴#𝐴 = 𝐴𝐴# 2.4 Symmetric 0 1
𝐴=
1 1
𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙𝑖𝑧𝑎𝑏𝑙𝑒 𝑏𝑦 𝑜𝑟𝑡ℎ𝑜𝑔𝑜𝑛𝑎𝑙 𝑚𝑎𝑡𝑟𝑖𝑥 𝑆 = 𝑆 #, a𝑙𝑙 𝜆 𝑎𝑟𝑒 𝑟𝑒𝑎𝑙 −1 −1
Positive
𝐴 = 𝑄Λ𝑄> 6.3 𝑆 = 𝑄Λ𝑄> 6.3 0 1
Semidefinite 𝐽=
# 0 0
a𝑙𝑙 𝜆 ≥ 0, 𝑎𝑙𝑙 𝐴 𝐴 4.2 1 2
Orthogonal 4.4 2.4 Projection
𝑆=
2 0
𝑄 !" = 𝑄 # Permutation 𝑃 $ = 𝑃 = 𝑃 #, 𝜆 = 1 𝑜𝑟 0
a𝑙𝑙 |𝜆| = 1 𝑝𝑒𝑟𝑚𝑢𝑡𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑰
a𝑙𝑙 𝜆 𝑎𝑟𝑒 𝑟𝑜𝑜𝑡𝑠 𝑜𝑓 1
𝑰 O
0 1 $ 𝜆 0
𝑄= Diagonal Σ= 𝜎 0 Λ=
−1 0 0 𝜎$ 0 𝜆

Positive 6.3
Definite
a𝑙𝑙 𝜆 > 0

𝑝𝑠𝑒𝑢𝑑𝑜𝑖𝑛𝑣𝑒𝑟𝑠𝑒 𝑓𝑜𝑟 𝑎𝑙𝑙 𝐴

𝐴?@ = 𝑉Σ?@𝑈 > 𝐴Q = 𝑉ΣQ𝑈 > 3.5, 7.4

Drawn by Kenji Hiranabe
with the help of Prof. Gilbert Strang
(v1.5, Mar.2th,2023)
Matrix World Matrix (𝑚×𝑛) Matrix Factorization
Appearing section
in 1.4 𝐴 = 𝐶𝑅 𝐴 = 𝑈Σ𝑉 > 7.1
𝑟𝑜𝑤 𝑟𝑎𝑛𝑘 = 𝑐𝑜𝑙𝑢𝑚𝑛 𝑟𝑎𝑛𝑘 𝑆𝑉𝐷: 𝑜𝑟𝑡ℎ𝑜𝑛𝑜𝑟𝑚𝑎𝑙 𝑏𝑎𝑠𝑖𝑠 𝑈, 𝑉 (in Introduction to Linear Algebra 6th edtion)
Introduction to
Linear Algebra 𝐴=
1 2 3 Square Matrix (𝑛×𝑛)
6th Edition 4 5 6 Invertible Singular
det(𝐴) ≠ 0, a𝑙𝑙 𝜆 ≠ 0 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝜆 = 0, det(𝐴) = 0
4.4 𝐴 = 𝑄𝑅 𝑇𝑟𝑖𝑎𝑛𝑔𝑢𝑙𝑎𝑟𝑖𝑧𝑒 𝐴 = 𝐿𝑈 2.3
Gram-Schmidt 𝑈 ℎ𝑎𝑠 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝑧𝑒𝑟𝑜 𝑟𝑜𝑤
2 1
𝐴=
0 2 Diagonalizable
1 1 6.2 𝐴 = 𝑋Λ𝑋 ?@ 𝐷𝑖𝑎𝑔𝑜𝑛𝑎𝑙𝑖𝑧𝑒 𝐴 = 𝑋𝐽𝑋 ?@ A5
𝐴=
0 2 𝐽 = 𝐽𝑜𝑟𝑑𝑎𝑛 𝑓𝑜𝑟𝑚
A2 Normal 𝐴=
0 1
𝐴#𝐴 = 𝐴𝐴# 2.4 Symmetric 0 1
𝐴=
1 1
𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙𝑖𝑧𝑎𝑏𝑙𝑒 𝑏𝑦 𝑜𝑟𝑡ℎ𝑜𝑔𝑜𝑛𝑎𝑙 𝑚𝑎𝑡𝑟𝑖𝑥 𝑆 = 𝑆 #, a𝑙𝑙 𝜆 𝑎𝑟𝑒 𝑟𝑒𝑎𝑙 −1 −1
Positive
𝐴 = 𝑄Λ𝑄> 6.3 𝑆 = 𝑄Λ𝑄> 6.3 0 1
Semidefinite 𝐽=
# 0 0
a𝑙𝑙 𝜆 ≥ 0, 𝑎𝑙𝑙 𝐴 𝐴 4.2 1 2
Orthogonal 4.4 2.4 Projection
𝑆=
2 0
𝑄 !" = 𝑄 # Permutation 𝑃 $ = 𝑃 = 𝑃 #, 𝜆 = 1 𝑜𝑟 0
a𝑙𝑙 |𝜆| = 1 𝑝𝑒𝑟𝑚𝑢𝑡𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑰
a𝑙𝑙 𝜆 𝑎𝑟𝑒 𝑟𝑜𝑜𝑡𝑠 𝑜𝑓 1
𝑰 O
0 1 $ 𝜆 0
𝑄= Diagonal Σ= 𝜎 0 Λ=
−1 0 0 𝜎$ 0 𝜆

Positive 6.3
Definite
a𝑙𝑙 𝜆 > 0

𝑝𝑠𝑒𝑢𝑑𝑜𝑖𝑛𝑣𝑒𝑟𝑠𝑒 𝑓𝑜𝑟 𝑎𝑙𝑙 𝐴

𝐴?@ = 𝑉Σ?@𝑈 > 𝐴Q = 𝑉ΣQ𝑈 > 4.5, 7.3
Drawn by Kenji Hiranabe
with the help of Prof. Gilbert Strang
(v1.5, Mar.2th,2023)
Matrix World Matrix (𝑚×𝑛) Matrix Factorization
𝐴 = 𝐶𝑅 𝐴 = 𝑈Σ𝑉 >
𝑟𝑜𝑤 𝑟𝑎𝑛𝑘 = 𝑐𝑜𝑙𝑢𝑚𝑛 𝑟𝑎𝑛𝑘 𝑆𝑉𝐷: 𝑜𝑟𝑡ℎ𝑜𝑛𝑜𝑟𝑚𝑎𝑙 𝑏𝑎𝑠𝑖𝑠 𝑈, 𝑉

𝐴=
1 2 3 Square Matrix (𝑛×𝑛)
4 5 6 Invertible Singular
det(𝐴) ≠ 0, a𝑙𝑙 𝜆 ≠ 0 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝜆 = 0, det(𝐴) = 0
𝐴 = 𝑄𝑅 𝑇𝑟𝑖𝑎𝑛𝑔𝑢𝑙𝑎𝑟𝑖𝑧𝑒 𝐴 = 𝐿𝑈
Gram-Schmidt 𝑈 ℎ𝑎𝑠 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝑧𝑒𝑟𝑜 𝑟𝑜𝑤
2 1
𝐴=
0 2 Diagonalizable
1 1 𝐴 = 𝑋Λ𝑋 ?@ 𝐷𝑖𝑎𝑔𝑜𝑛𝑎𝑙𝑖𝑧𝑒 𝐴 = 𝑋𝐽𝑋 ?@
𝐴=
0 2 𝐽 = 𝐽𝑜𝑟𝑑𝑎𝑛 𝑓𝑜𝑟𝑚
Normal 𝐴=
0 1
𝐴#𝐴 = 𝐴𝐴# Symmetric 0 1 1 1
𝐴=
𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙𝑖𝑧𝑎𝑏𝑙𝑒 𝑏𝑦 𝑜𝑟𝑡ℎ𝑜𝑔𝑜𝑛𝑎𝑙 𝑚𝑎𝑡𝑟𝑖𝑥 𝑆 = 𝑆 #, a𝑙𝑙 𝜆 𝑎𝑟𝑒 𝑟𝑒𝑎𝑙 −1 −1
Positive
𝐴 = 𝑄Λ𝑄> 𝑆 = 𝑄Λ𝑄> 0 1
Semidefinite 𝐽=
# 0 0
a𝑙𝑙 𝜆 ≥ 0, 𝑎𝑙𝑙 𝐴 𝐴 1 2
𝑆=
Orthogonal Projection 2 0
𝑄 !" = 𝑄 # Permutation 𝑃 $ = 𝑃 = 𝑃 #, 𝜆 = 1 𝑜𝑟 0
a𝑙𝑙 |𝜆| = 1 𝑝𝑒𝑟𝑚𝑢𝑡𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑰
a𝑙𝑙 𝜆 𝑎𝑟𝑒 𝑟𝑜𝑜𝑡𝑠 𝑜𝑓 1
𝑰 O
0 1 $ 𝜆 0
𝑄= Diagonal Σ= 𝜎 0 Λ=
−1 0 0 𝜎$ 0 𝜆

Positive
Definite
a𝑙𝑙 𝜆 > 0

𝑝𝑠𝑒𝑢𝑑𝑜𝑖𝑛𝑣𝑒𝑟𝑠𝑒 𝑓𝑜𝑟 𝑎𝑙𝑙 𝐴

𝐴?@ = 𝑉Σ?@𝑈 > 𝐴Q = 𝑉ΣQ𝑈 >

Drawn by Kenji Hiranabe
with the help of Prof. Gilbert Strang
(v1.5, Mar.2th,2023)
Matrix World ⾏列 (𝑚×𝑛)
⾏列の積分解
in 1.4 𝐴 = 𝐶𝑅 𝐴 = 𝑈Σ𝑉 > 7.1 解説の節番号
教養の線形代数 行 𝑟𝑎𝑛𝑘 = 列 𝑟𝑎𝑛𝑘 𝑆𝑉𝐷: 単位直交基底 𝑈, 𝑉 (教養の線形代数)

1 2 3 正⽅⾏列 (𝑛×𝑛)
𝐴=
4 5 6 可逆（正則） ⾮可逆（特異）
det 𝐴 ≠ 0, ∀𝜆 ≠ 0 det(𝐴) = 0, ∃𝜆 = 0
4.4 𝐴 = 𝑄𝑅 三角化 𝐴 = 𝐿𝑈 2.3
グラム・シュミット法 𝑈 はゼロ行を持つ

𝐴=
2 1 対⾓化可能
0 2 6.2 𝐴 = 𝑋Λ𝑋 ?@ 対⾓化 𝐴 = 𝑋𝐽𝑋 ?@ A7
1 1 0 1
𝐴= 𝐴= 𝐽 = ジョルダン標準形
0 2 A5 正規 0 1
𝐴#𝐴 = 𝐴𝐴# 対称 2.4 𝐴=
1 1
直交⾏列で対⾓化可能 −1 −1
𝑆= 𝑆%, ∀𝜆∈ℝ
𝐴 = 𝑄Λ𝑄> 6.3 半正定値
𝑆 = 𝑄Λ𝑄> 6.3 𝐽=
0 1
∀𝜆 ≥ 0, ∀ 𝐴# 𝐴 0 0
1 2
直交 4.4 4.2 射影 𝑆=
2 0

𝑄 !" = 𝑄 # 置換 2.4 𝑃 $ = 𝑃 = 𝑃 #, 𝜆 = 1 𝑜𝑟 0

∀ |𝜆| = 1 I の並び替え
∀ 𝜆 は1の根
𝑰 O
0 1 $ 𝜆 0
𝑄= 対⾓ Σ= 𝜎 0 Λ=
−1 0 0 𝜎$ 0 𝜆

正定値 6.3
∀𝜆 > 0

すべての𝐴に対する擬似逆行列

Drawn by Kenji Hiranabe 𝐴 ?@

= 𝑉Σ ?@ >
𝑈 𝐴Q = 𝑉ΣQ𝑈 > 3.5, 7.4
with the help of Prof. Gilbert Strang
(v1.5, Mar.2th,2023)
Matrix World ⾏列 (𝑚×𝑛)
⾏列の積分解
𝐴 = 𝐶𝑅 𝐴 = 𝑈Σ𝑉 >
行 𝑟𝑎𝑛𝑘 = 列 𝑟𝑎𝑛𝑘 𝑆𝑉𝐷: 単位直交基底 𝑈, 𝑉

1 2 3 正⽅⾏列 (𝑛×𝑛)
𝐴=
4 5 6 可逆（正則） ⾮可逆（特異）
det 𝐴 ≠ 0, ∀𝜆 ≠ 0 det(𝐴) = 0, ∃𝜆 = 0
𝐴 = 𝑄𝑅 三角化 𝐴 = 𝐿𝑈
グラム・シュミット法 𝑈 はゼロ行を持つ

𝐴=
2 1 対⾓化可能
0 2 𝐴 = 𝑋Λ𝑋 ?@ 対⾓化 𝐴 = 𝑋𝐽𝑋 ?@
1 1 0 1
𝐴= 𝐴= 𝐽 = ジョルダン標準形
0 2 正規 0 1
𝐴#𝐴 = 𝐴𝐴# 対称 𝐴=
1 1
直交⾏列で対⾓化可能 −1 −1
𝑆= 𝑆%, ∀𝜆∈ℝ
𝐴 = 𝑄Λ𝑄> 半正定値 𝑆 = 𝑄Λ𝑄> 𝐽=
0 1
∀𝜆 ≥ 0, ∀ 𝐴# 𝐴 0 0
1 2
直交 射影 𝑆=
2 0
置換 𝑃 $ = 𝑃 = 𝑃 #, 𝜆 = 1 𝑜𝑟 0
𝑄 !" = 𝑄 #
∀ |𝜆| = 1 I の並び替え
∀ 𝜆 は1の根
𝑰 O
0 1 $ 𝜆 0
𝑄= 対⾓ Σ= 𝜎 0 Λ=
−1 0 0 𝜎$ 0 𝜆

正定値
∀𝜆 > 0

すべての𝐴に対する擬似逆行列

Drawn by Kenji Hiranabe 𝐴 ?@

= 𝑉Σ ?@ >
𝑈 𝐴Q = 𝑉ΣQ𝑈 >
with the help of Prof. Gilbert Strang
(v1.5, Mar.2th,2023)
 Linear Algebra 矩阵 (𝑚×𝑛) 矩阵分解
for Everyone
中的 1.4 𝐴 = 𝐶𝑅 𝐴 = 𝑈Σ𝑉 > 7.1 对应章节
行秩 = 列秩 𝑆𝑉𝐷: 单位正交基底 𝑈, 𝑉 (在 Linear Algebra for Everyone 中)

矩阵世界 𝐴=
1 2 3 方阵 (𝑛×𝑛)
4 5 6 可逆 奇异
det(𝐴) ≠ 0, a𝑙𝑙 𝜆 ≠ 0 ∃𝜆 = 0, det(𝐴) = 0
4.4 𝐴 = 𝑄𝑅 三角化 𝐴 = 𝐿𝑈 2.3
格拉姆-施密特 𝑈 有一个零行
2 1
𝐴=
0 2 可对角化
1 1 6.2 𝐴 = 𝑋Λ𝑋 ?@ 对角化 𝐴 = 𝑋𝐽𝑋 ?@ A7
𝐴=
0 2 𝐽 = 约旦型
A5 正规 𝐴=
0 1
𝐴#𝐴 = 𝐴𝐴# 2.4 对称 0 1
𝐴=
1 1
可通过正交矩阵对角化 𝑆 = 𝑆 #, 所有𝜆都是实数 −1 −1
𝐴 = 𝑄Λ𝑄> 6.3 半正定 𝑆 = 𝑄Λ𝑄> 6.3 0 1
𝐽=
# 0 0
a𝑙𝑙 𝜆 ≥ 0, 𝑎𝑙𝑙 𝐴 𝐴 4.2 1 2
正交 4.4 𝑆=
2.4 投影 2 0
𝑄 !" = 𝑄 # 𝑃 $ = 𝑃 = 𝑃 #, 𝜆 = 1 𝑜𝑟 0
置换
a𝑙𝑙 |𝜆| = 1
𝑰的置换 𝑰 O
a𝑙𝑙 𝜆 𝑎𝑟𝑒 𝑟𝑜𝑜𝑡𝑠 𝑜𝑓 1 $ 𝜆 0
0 1
𝑄= 对角 Σ= 𝜎 0 Λ=
−1 0 0 𝜎$ 0 𝜆

正定 6.3
a𝑙𝑙 𝜆 > 0

所有𝐴的伪逆

𝐴?@ = 𝑉Σ?@𝑈 > 𝐴Q = 𝑉ΣQ𝑈 > 3.5, 7.4

Drawn by Kenji Hiranabe
with the help of Prof. Gilbert Strang
(v1.4.7, Oct.25,2022)
Translator: Kefang Liu