Invertible Matrices

Can a matrix have a left inverse but no right inverse? This question sits at the heart of understanding square matrices. In infinite-dimensional settings the answer is yes. A shift operator can have a one-sided inverse without having a two-sided one. But in finite dimensions something remarkable happens. For square matrices, one-sidedness collapses. A single equation $BA = I$ forces $AB = I$ as well. This asymmetry between the finite and infinite worlds is not a technicality. It is a load-bearing fact used constantly, often without comment, throughout linear algebra. Invertible matrices matter because they are exactly the matrices representing bijective linear maps. These are the algebraic shadows of coordinate changes, symmetries, and reversible processes. Solving $Ax = b$ uniquely for every $b$ depends on this. So does changing basis, defining determinants and eigenvalues, and building the machinery of Rank of a Matrix. All of these lean on a precise understanding of what invertibility means and how to detect it. The set of invertible $n\times n$ matrices also carries its own rich structure. It forms a group, the general linear group $GL_n(F)$ , under matrix multiplication. This group is the algebraic home for symmetry itself. The big picture is this: invertibility of a square matrix $A$ is a single yes/no condition. It is equivalent to a long list of others. These include nonzero determinant, full rank, trivial kernel, columns forming a basis, and more. This article isolates the cleanest version of this equivalence. It proves the deep one-sided fact rigorously, using only Matrix Multiplication as Composition and known facts about linear maps on finite-dimensional spaces. It also shows exactly where the finite-dimensionality hypothesis is used, and what breaks without it. By the end you will have a complete proof of several key facts. For square matrices over a field, injectivity, surjectivity, and bijectivity of the associated linear map are all equivalent. A one-sided inverse is automatically two-sided. The invertible matrices form a group. We also preview how row reduction (Elementary Matrices and Row Operations) gives a computational test for invertibility.

Key takeaways

A square matrix $A$ is invertible if there exists $B$ with $AB = BA = I$ ; such $B$ is unique and written $A^{-1}$ .
In finite dimensions, a one-sided inverse of a square matrix is automatically a two-sided inverse — this fails in infinite dimensions.
Invertibility of $A \in F^{n\times n}$ is equivalent to: $\det A \ne 0$ , $\operatorname{rank}(A) = n$ , columns (or rows) form a basis, and the map $x \mapsto Ax$ is injective (equivalently surjective).
The invertible $n \times n$ matrices over a field $F$ form a group $GL_n(F)$ under multiplication, generally non-abelian for $n \ge 2$ .
$(AB)^{-1} = B^{-1}A^{-1}$ whenever $A$ and $B$ are invertible — order reverses, exactly as for composing invertible functions.
Key idea: invertibility of a square matrix is a rigid, all-or-nothing property fully witnessed by either side of the inverse equation, thanks to finite dimensionality.

Where this lives today

Invertible matrices are the algebraic backbone of change of basis: passing between coordinate systems in a vector space requires an invertible transition matrix, used throughout diagonalization and canonical form theory. In differential equations and dynamical systems, the invertibility of the flow map (or of a linear system's coefficient matrix) governs whether trajectories can be uniquely reversed in time. In cryptography, invertible matrices over finite fields (e.g., $GL_n(\mathbb{F}_2)$ ) underlie linear codes and classical cipher constructions such as the Hill cipher, where decryption literally requires computing $A^{-1}$ . In computer graphics and robotics, invertible matrices represent reversible geometric transformations (rotations, scalings, shears), and $GL_n(\mathbb{R})$ together with its subgroups (like $O(n)$ , $SL_n(\mathbb{R})$ ) parametrizes the symmetries used in rendering pipelines and kinematics. In statistics, invertibility of the design matrix's Gram matrix $X^TX$ is precisely the condition for a unique least-squares solution in linear regression. In quantum mechanics, invertible (indeed unitary) matrices represent reversible time evolution of finite-dimensional quantum states.

Ask the tutor: Why exactly does the finite-dimensionality of $F^n$ break down for the shift operator example, and where precisely in the Rank–Nullity proof does that assumption get used?

F

Intuition

Think of a square matrix $A$ as encoding a linear map $T_A : F^n \to F^n$ , $T_A(x) = Ax$ , in a fixed basis (see Matrix Multiplication as Composition). Composition of linear maps corresponds to matrix multiplication, and the identity map corresponds to $I$ . So " $A$ is invertible" means exactly " $T_A$ is an invertible function," i.e., $T_A$ has a two-sided inverse function. The content of this article is that for square matrices on a finite-dimensional space, checking only one direction of invertibility — say that $T_A$ is injective — is already enough to conclude the whole story: $T_A$ is bijective, and the inverse map is also linear, hence represented by an honest inverse matrix $A^{-1}$ . Picture $F^n$ as a rigid frame of $n$ independent directions (a basis). Applying $A$ moves this frame to the frame given by the columns of $A$ . If those images stay independent (never collapsing into a lower-dimensional subspace), the map cannot lose information, so it must also be onto: there's no "room" left over in an $n$ -dimensional target for the image of an $n$ -dimensional domain to be a proper subspace and still capture everything only if it fails to be injective. This finite counting argument — dimension in equals dimension out — is the engine behind the one-sided-inverse theorem.

Figure

Columns of

A

as images of the standard basis; invertibility means these images stay independent and span the plane

F

V

V

(2) \Rightarrow (4) \Rightarrow (5) \Rightarrow (3) \Rightarrow (1) \Rightarrow (2)

GL_n(F)

GL_n(F)

Worked examples

Example 1: Verifying invertibility via a genuine one-sided check. Let

A = \begin{pmatrix} 2 & 1 & 0 \\ 1 & 1 & 1 \\ 0 & 1 & 2 \end{pmatrix} \in \mathbb{R}^{3\times 3}.

We test invertibility by checking injectivity of $T_A$ directly — i.e., showing $Ax = 0 \Rightarrow x = 0$ — which by the theorem suffices, with no need to separately verify surjectivity or exhibit a two-sided inverse up front. Suppose $A(x_1,x_2,x_3)^T = 0$ :

2x_1 + x_2 = 0, \qquad x_1 + x_2 + x_3 = 0, \qquad x_2 + 2x_3 = 0.

From the first equation, $x_2 = -2x_1$ . Substitute into the third: $-2x_1 + 2x_3 = 0$ , so $x_3 = x_1$ . Substitute both into the second: $x_1 - 2x_1 + x_1 = 0$ , which gives $0 = 0$ — no new information, and the system is consistent for any $x_1$ . Set $x_1 = t$ : then $x_2 = -2t$ , $x_3 = t$ . Substitute back into the original equations to double check: first equation $2t + (-2t) = 0$ ✓; third equation $(-2t) + 2t = 0$ ✓; second: $t + (-2t) + t = 0$ ✓. So every choice of $t$ gives a solution $(t, -2t, t)$ . Taking $t = 1$ gives the nonzero vector $(1,-2,1)^T$ with $A(1,-2,1)^T = 0$ . Hence $\ker T_A \ne \{0\}$ , so $T_A$ is not injective. By the theorem, $A$ is not invertible — no right or left inverse can exist, without our needing to search for one. As a cross-check via determinant (previewing Rank of a Matrix and cofactor expansion):

\det A = 2(1\cdot 2 - 1\cdot 1) - 1(1\cdot 2 - 1\cdot 0) + 0 = 2(1) - 1(2) + 0 = 0,

confirming singularity independently.

Example 2: Exhibiting a one-sided inverse and getting the other side for free. Let

A = \begin{pmatrix} 1 & 2 \\ 0 & 3 \end{pmatrix}, \qquad B = \begin{pmatrix} 1 & -2/3 \\ 0 & 1/3 \end{pmatrix} \in \mathbb{Q}^{2\times 2}.

Compute $BA$ directly:

BA = \begin{pmatrix} 1 & -2/3 \\ 0 & 1/3 \end{pmatrix}\begin{pmatrix} 1 & 2 \\ 0 & 3 \end{pmatrix} = \begin{pmatrix} 1\cdot 1 + (-2/3)\cdot 0 & 1\cdot 2 + (-2/3)\cdot 3 \\ 0\cdot 1 + (1/3)\cdot 0 & 0\cdot 2 + (1/3)\cdot 3 \end{pmatrix} = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}.

So $BA = I_2$ : $B$ is a left inverse of $A$ . By the theorem's one-sided-inverse claim, $A$ is automatically invertible and $AB = I_2$ as well, with $B = A^{-1}$ . We verify this directly rather than taking it on faith:

AB = \begin{pmatrix} 1 & 2 \\ 0 & 3 \end{pmatrix}\begin{pmatrix} 1 & -2/3 \\ 0 & 1/3 \end{pmatrix} = \begin{pmatrix} 1 + 0 & -2/3 + 2/3 \\ 0 + 0 & 0 + 1 \end{pmatrix} = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} = I_2.

Indeed $AB = I_2$ too, exactly as the theorem predicted — checking one side was already conclusive, and the arithmetic above is the concrete instance of the general argument in Step 4 of the proof ( $A = A(CD) = (AC)D = D$ type identity, here with $C = B$ ).

V = F^{\mathbb{N}}

Scope, edge cases & common traps

Situation	What happens	Why
$A$ square, $F^n$ finite-dim.	One-sided inverse $\Rightarrow$ two-sided	Rank–Nullity forces injective $\Leftrightarrow$ surjective
$A$ non-square ( $m \times n$ , $m \ne n$ )	One-sided inverses can exist without the other side	Domain/codomain dimensions differ; no contradiction from Rank–Nullity
Infinite-dimensional $V$	One-sided inverses need not be two-sided	Shift operators; see counterexample
$F$ not a field but a general ring	Theorem can fail	Rank–Nullity and basis arguments use field structure (division by nonzero scalars)
$\det A = 0$	$A$ singular, no inverse of either kind	Equivalent condition (6): columns dependent

Common trapThe classic error

Do not assume that checking $Ax = 0 \Rightarrow x=0$ "morally suggests" invertibility without invoking finite-dimensionality — the logical engine is Rank–Nullity, which requires $\dim V < \infty$ . Equally common: forgetting that $(AB)^{-1} = B^{-1}A^{-1}$ , not $A^{-1}B^{-1}$ — order reverses, just as with inverse functions, $(f\circ g)^{-1} = g^{-1}\circ f^{-1}$ . A converse trap: showing $AB$ invertible does not by itself give $A$ and $B$ individually invertible unless you also know $A, B$ are square (which they are here) and apply the theorem to each factor via $B(AB)^{-1} = A^{-1}$ -type manipulations — a step people often skip.

The mathematician behind it

Origins · 1 minArthur Cayley and how this was discovered

Arthur Cayley

1821–1895 · Victorian algebra, matrix theory

Arthur Cayley, a British mathematician and one of the most prolific algebraists of the nineteenth century, is credited with founding matrix theory as an independent algebraic subject. Working initially as a lawyer to support himself before securing an academic post at Cambridge, Cayley treated matrices as objects to be added and multiplied in their own right, rather than mere arrays of coefficients for a linear system. His 1858 memoir on the theory of matrices introduced matrix multiplication, the identity matrix, and — crucially for this article — the notion of a matrix inverse, alongside an early version of what is now called the Cayley–Hamilton theorem.

History

- 1858 — Cayley's "A Memoir on the Theory of Matrices" formalizes matrix algebra, multiplication, and inverses as objects of independent study.

- 1878 — Georg Frobenius develops rank and canonical forms, sharpening the equivalence between invertibility and non-vanishing determinant.

- 1888 — Giuseppe Peano gives the first fully axiomatic definition of a vector space, providing the setting in which "finite-dimensional" becomes precise.

- 1920s — the abstract group-theoretic viewpoint (Emmy Noether's structural algebra) recasts

GL_n(F)

as a fundamental example of a group acting on a vector space.

- 1940s–1950s — numerical linear algebra (Turing, von Neumann) turns invertibility testing and matrix inversion into central problems of scientific computing.

Does a square matrix need both

AB=I

and

BA=I

checked to confirm invertibility?

No — this is the central point of this article. For a square matrix $A$ over a field acting on a finite-dimensional space, if you find any $B$ with $AB = I$ or any $B$ with $BA = I$ , that single equation already forces $A$ to be invertible with $A^{-1} = B$ , and the other equation follows automatically. The proof runs through Rank–Nullity: one-sided inverse gives injectivity or surjectivity of $T_A$ , and on a finite-dimensional space these two properties are equivalent for maps from a space to itself, so bijectivity — and hence full invertibility — comes for free.

Why does this one-sided shortcut fail for non-square matrices?

Because Rank–Nullity in that setting reads $\dim(\ker T) + \dim(\operatorname{range} T) = n$ where $n$ is the dimension of the domain, while the codomain may have a different dimension $m$ . If $m \ne n$ , injectivity and surjectivity are no longer forced to coincide — an $m\times n$ matrix with $m > n$ can be injective without being surjective (its image is a proper subspace of the larger codomain), and one with $m < n$ can be surjective without injective. One-sided inverses for rectangular matrices exist in different senses (left vs. right pseudo-inverses) but are genuinely different objects.

GL_n(F)

abelian?

Only when $n = 1$ (or in degenerate cases like $F$ having very few elements combined with small $n$ , but generically no). For $n \ge 2$ , matrix multiplication is non-commutative: for instance swapping rows via one elementary matrix then scaling via another gives a different result depending on order. So $GL_n(F)$ is a genuinely non-abelian group for $n\ge2$ , which is part of why it plays such a rich role as a source of examples in group theory and representation theory.

How does the determinant relate to this theorem?

The determinant gives a single-number test: $A$ is invertible if and only if $\det A \ne 0$ . This is consistent with, and can be derived alongside, the equivalences in the theorem — e.g., columns of $A$ form a basis if and only if they are linearly independent if and only if $\det A \ne 0$ . The determinant is often the fastest hand-computation route to check invertibility for small matrices, while the injective/surjective formulation is the conceptual reason the determinant test works at all.

Does the theorem hold over rings that are not fields, like

\mathbb{Z}

Not in general, and the argument used here relies on field-specific facts, particularly the basis-extension and Rank–Nullity arguments valid for vector spaces over a field. Over $\mathbb{Z}$ , for instance, invertibility as a matrix over $\mathbb{Z}$ additionally requires $\det A = \pm 1$ (not merely nonzero), because you need $1/\det A$ to lie back in $\mathbb{Z}$ . The general theory of modules over rings replaces "basis" with more delicate notions (free modules, torsion), and one-sided invertibility results become subtler.

A

and

B

are both invertible

n\times n

matrices, is

A+B

necessarily invertible?

No. Invertibility is not preserved under addition. For example, with $A = I_2$ and $B = -I_2$ , both are invertible, but $A + B = 0$ , the zero matrix, which is singular (it sends every vector to $0$ , so it is not injective for $n \ge 1$ ). Invertibility is a multiplicative property, tied to the group structure of $GL_n(F)$ , and there is no reason to expect it to survive under the unrelated operation of addition.

What is the fastest way in practice to test if a matrix is invertible?

Row reduce $A$ to row echelon form using elementary row operations (see Elementary Matrices and Row Operations); $A$ is invertible exactly when this process produces $n$ nonzero pivot rows, equivalently when the reduced row echelon form is $I_n$ . This is computationally the standard method, and it simultaneously computes $A^{-1}$ if you augment with the identity matrix and row-reduce the pair $[A \mid I]$ to $[I \mid A^{-1}]$ .