Types of Matrices: A Bestiary

Same Object, Different Structure

A matrix's shape and constraints determine its properties. A constraint is not merely a restriction; it is a type hint for which operations exist and which can be cheap.

Type	Constraint	Why it is useful
Square	same number of rows and columns	determinants and eigenvalues are defined, and a nonsingular square matrix has an inverse
Rectangular	$m \neq = n$	changes dimensionality, such as projecting 1,024 dimensions to 256
Diagonal	zero off the diagonal	scales each axis and needs only diagonal entries for many operations
Identity $I$	ones on the diagonal, zero elsewhere	multiplicative identity: $I A = A$
Zero	all entries are zero	additive identity for matrices of the same shape
Symmetric	$A = A^{T}$	common in covariance and energy models, with useful real eigendecomposition properties
Sparse	most entries are zero	stores and computes mainly nonzero entries to save memory and work

Why You Should Care

Identity matrices appear in regularization and linear-system formulas, diagonal matrices in scaling and SVD, and symmetric matrices in covariance, graph, and energy problems. Sparse storage is what lets enormous recommendation matrices and graphs fit in memory.

A rectangular matrix has no ordinary two-sided inverse, though least-squares solutions and pseudoinverses may apply. A square matrix can also be singular and noninvertible. Shape alone never guarantees an inverse.

Matrix types are mathematical type hints. Recognizing one reveals which operations are cheap and which assumptions fail. Do not densify a sparse matrix casually, and do not feed a nonsymmetric matrix to an algorithm that assumes symmetry.

Code

Bestiary, four types·python

import numpy as np

zero = np.zeros((3, 3))
identity = np.eye(3)
diagonal = np.diag([2, 5, 7])
print(diagonal)
# [[2 0 0]
#  [0 5 0]
#  [0 0 7]]

# Diagonal matrices scale: D @ v scales v[i] by D[i, i]
v = np.array([1, 1, 1])
print(diagonal @ v)        # [2 5 7]

# Symmetric — auto-symmetric construction
A = np.random.randn(3, 3)
S = (A + A.T) / 2          # any matrix + its transpose is symmetric
print(np.allclose(S, S.T)) # True

Same Object, Different Structure

Why You Should Care

Code

External links

Exercise

Progress

Comments 2