Column Space of a Nonsingular Matrix

Q: Why does this theorem require the matrix to be square?

Because it equates a statement about outputs with a statement about inputs, and only for a square matrix do the input and output spaces have the same dimension. A $2 \times 3$ matrix can have column space equal to all of $\mathbb{C}^2$ while its null space is a line, so it is surjective and degenerate at once. Squareness is what forces rank deficiency in one direction to imply rank deficiency in the other.

Q: Does a full column space guarantee the solution is unique?

For a square matrix, yes — because $\mathcal{C}(A) = \mathbb{C}^n$ implies nonsingularity, which implies a trivial null space and therefore a unique solution for every right-hand side. For a rectangular matrix it does not: a wide matrix with a full column space always has infinitely many solutions for every consistent target.

Q: How do I check whether a specific vector lies in the column space?

Row-reduce the augmented matrix $[\, A \mid \vec{b} \,]$ and look at the final column. If it contains a pivot, the system is inconsistent and $\vec{b} \notin \mathcal{C}(A)$. For repeated tests against the same matrix, it is more efficient to compute a basis for the left null space once and then check that $\vec{b}$ satisfies the corresponding homogeneous conditions.

Q: What exactly goes wrong when the matrix is singular?

Two things simultaneously. The reachable set shrinks to a proper subspace of dimension $r < n$, so a large family of targets is unattainable at any command level, and each attainable target is produced by an infinite family of commands differing by elements of the null space. In an engineering model these appear as lost authority and as unobservable command directions.

Q: Is a numerically near-singular matrix still covered by the theorem?

Mathematically yes: a matrix with a very small but non-zero smallest singular value is nonsingular, so the column space is genuinely all of $\mathbb{C}^n$. Practically the guarantee is hollow, because the commands required to reach targets near the deficient direction grow in proportion to the condition number and will exceed physical limits long before the theorem fails.

Q: Why is the converse direction of the proof harder than the forward direction?

The forward direction only has to spend information — nonsingularity already delivers a solution for every right-hand side, which is the definition of a full column space. The converse has to create information about the null space out of a statement about reachability, and it does so constructively, by solving for each standard unit vector and assembling a right inverse.

Q: Can I conclude the columns of a square matrix are a basis just from spanning?

Yes, for a square matrix. Spanning $\mathbb{C}^n$ means $\mathcal{C}(A) = \mathbb{C}^n$, which by this theorem makes $A$ nonsingular, which makes the columns linearly independent. A spanning set of $n$ vectors in $\mathbb{C}^n$ is automatically independent, and an independent set of $n$ vectors in $\mathbb{C}^n$ automatically spans.

Overview

The column space $\mathcal{C}(A)$ of an $m \times n$ matrix is the span of its columns, equivalently the set of all products $A\vec{x}$ . It is the reachable set of the matrix: the collection of every output the matrix is capable of producing. Because $A\vec{x} = \vec{b}$ is consistent precisely when $\vec{b}$ lies in that set, the column space is the exact boundary between the systems you can solve and the systems you cannot.

Restricting attention to square matrices sharpens the question. A square matrix takes $\mathbb{C}^n$ to $\mathbb{C}^n$ , so it is meaningful to ask whether the reachable set is all of the target space or only a proper subspace of it. The answer is governed by a single property already familiar from the study of homogeneous systems: nonsingularity. A square matrix is nonsingular if and only if its column space is the whole of $\mathbb{C}^n$ , with no exceptions and no side conditions.

This equivalence is more useful than it first appears because it links two very different kinds of evidence. Nonsingularity is established by row-reducing $A$ and observing that the result is the identity, or by observing that the null space contains only $\vec{0}$ — both statements about what the matrix annihilates. A full column space is a statement about what the matrix produces. The theorem says these two accounts never disagree for a square matrix, which is a genuinely strong claim: for rectangular matrices it fails in both directions.

Practically, the result is the guarantee that underwrites a large class of engineering models. If the matrix relating actuator commands to a system response is square and nonsingular, then every commanded response is achievable and achievable by exactly one command. If it is singular, there is a proper subspace of achievable responses, an infinite family of commands producing each one, and an entire complementary set of specifications that no command can meet. The distinction is not a numerical inconvenience — it is a design fault.

Definition

Column Space of a Matrix

CSM

Let $A$ be an $m \times n$ matrix with columns $A_1, A_2, \ldots, A_n$ . The column space of $A$ , written $\mathcal{C}(A)$ , is the span of the set of columns:
$\mathcal{C}(A) = \operatorname{span}(\{\, A_1, A_2, \ldots, A_n \,\})$
It is a subspace of $\mathbb{C}^m$ , and it consists of exactly those vectors expressible as $A\vec{x}$ for some $\vec{x} \in \mathbb{C}^n$ .

Some texts call this the range or image of $A$ . The three names denote the same subspace; this library uses column space when the object is a matrix and range when the object is a linear transformation.

Nonsingular Matrix

A square matrix $A$ of size $n$ is nonsingular if the homogeneous system $A\vec{x} = \vec{0}$ has only the trivial solution $\vec{x} = \vec{0}$ ; equivalently, if $\mathcal{N}(A) = \{\vec{0}\}$ . A square matrix that is not nonsingular is singular.

Nonsingularity is defined only for square matrices. Asking whether a $3 \times 5$ matrix is nonsingular is a category error, not a hard question.

Standard Unit Vector

SUV

The vector $\vec{e}_i \in \mathbb{C}^n$ whose $i$ th entry is $1$ and whose remaining entries are all $0$ . The set $\{\, \vec{e}_1, \vec{e}_2, \ldots, \vec{e}_n \,\}$ collects the columns of the identity matrix $I_n$ .

Concepts

Consistency is membership of the column space

The product $A\vec{x}$ is a linear combination of the columns of $A$ with the entries of $\vec{x}$ as coefficients. Therefore the equation $A\vec{x} = \vec{b}$ asks whether $\vec{b}$ can be written as such a combination, which is exactly the question of whether $\vec{b}$ lies in $\operatorname{span}(\{A_1, \ldots, A_n\}) = \mathcal{C}(A)$ . The system is consistent if and only if $\vec{b} \in \mathcal{C}(A)$ .

This restatement is the hinge of the whole page. It converts the analytic question "does this system have a solution?" into the geometric question "does this vector lie in this subspace?", and it makes the column space the single object that decides solvability for every possible right-hand side at once.

The equivalence and why it is not obvious

Theorem. Let $A$ be a square matrix of size $n$ . Then $A$ is nonsingular if and only if $\mathcal{C}(A) = \mathbb{C}^n$ .

The forward direction is short. If $A$ is nonsingular then $A\vec{x} = \vec{b}$ has a unique solution for every choice of $\vec{b}$ , so every $\vec{b} \in \mathbb{C}^n$ lies in $\mathcal{C}(A)$ ; combined with the containment $\mathcal{C}(A) \subseteq \mathbb{C}^n$ that holds by definition, the two sets are equal. The converse is the substantive half, because it must manufacture information about the null space from information about the column space. Note that the hypothesis square is doing real work: a $2 \times 3$ matrix can have column space equal to $\mathbb{C}^2$ while possessing a non-trivial null space, and a $3 \times 2$ matrix can have a trivial null space while its column space is a proper subspace of $\mathbb{C}^3$ .

The converse: building a right inverse

Suppose $\mathcal{C}(A) = \mathbb{C}^n$ . Each standard unit vector $\vec{e}_i$ then lies in the column space, so the system $A\vec{x} = \vec{e}_i$ is consistent; choose one solution $\vec{b}_i$ for each $i$ from $1$ to $n$ . Assemble those solutions as the columns of a matrix $B = [\, \vec{b}_1 \mid \vec{b}_2 \mid \cdots \mid \vec{b}_n \,]$ . Because matrix multiplication acts column by column, $A B$ has $i$ th column $A\vec{b}_i = \vec{e}_i$ , so $AB = I_n$ and $B$ is a right inverse for $A$ .

The argument closes by observing that $I_n$ is nonsingular and that a product of two square matrices is nonsingular only when both factors are. Since $AB = I_n$ is nonsingular, neither $A$ nor $B$ can be singular. The construction is worth noting in its own right: a surjective square matrix is handed an explicit right inverse, and that right inverse then turns out to be the two-sided inverse.

The singular case has definite structure

When a square $A$ is singular, the failure is not vague. Row-reduce $A$ and let $r$ be the number of pivot columns. Then $\dim(\mathcal{C}(A)) = r$ and $\dim(\mathcal{N}(A)) = n - r$ , with $r < n$ exactly when $A$ is singular. The reachable set is a proper subspace of dimension $r$ , so the set of unreachable targets is not a scattering of awkward vectors but everything outside an $r$ -dimensional subspace — overwhelmingly most of $\mathbb{C}^n$ .

At the same time, each target that is reachable is reachable in infinitely many ways, since any element of the null space can be added to a solution without changing the product. A singular square matrix therefore fails in two coupled ways simultaneously: too few outputs, too many inputs per output. This coupling is exactly what disappears once the matrix is square and nonsingular.

Position in the nonsingular matrix equivalences

This theorem is one clause of a growing list of mutually equivalent conditions on a square matrix $A$ of size $n$ : $A$ is nonsingular; $A$ row-reduces to $I_n$ ; $\mathcal{N}(A) = \{\vec{0}\}$ ; $A\vec{x} = \vec{b}$ has a unique solution for every $\vec{b}$ ; the columns of $A$ are linearly independent; $A$ is invertible; and $\mathcal{C}(A) = \mathbb{C}^n$ .

The engineering value of such a list is that it licenses free substitution between very cheap and very expensive tests. Establishing $\mathcal{C}(A) = \mathbb{C}^n$ directly would require examining every $\vec{b}$ ; instead one row-reduces $A$ once and reads the pivot count. Conversely, a modelling argument that shows some particular $\vec{b}$ is unreachable immediately proves the matrix singular, with no arithmetic on $A$ at all.

Columns, independence and spanning together

For a square matrix the two basis conditions on its columns coincide. The columns span $\mathbb{C}^n$ exactly when $\mathcal{C}(A) = \mathbb{C}^n$ , and they are linearly independent exactly when $\mathcal{N}(A) = \{\vec{0}\}$ . Because both conditions are equivalent to nonsingularity, a set of $n$ columns in $\mathbb{C}^n$ either has both properties or has neither.

That is a strong and often surprising fact: for $n$ vectors in an $n$ -dimensional space, checking independence alone is enough to conclude spanning, and checking spanning alone is enough to conclude independence. The saving in verification effort is substantial, and the same phenomenon reappears in general vector spaces once dimension is available.

Decision path: is every target reachable?

Confirm the matrix is squareNonsingularity is defined only for size

n \times n

. For a rectangular matrix, ask about rank and the column space separately.

Row-reduce $A$ and count pivot columnsLet

r

be the number of pivot columns of the reduced form. This is

\operatorname{rank}(A) = \dim(\mathcal{C}(A))

Is $r = n$ ?If yes,

A

reduces to

I_n

A

is nonsingular and

\mathcal{C}(A) = \mathbb{C}^n

: every target is reachable, uniquely.

If $r a m p; l t; n$ , describe the deficient column spaceThe reachable set is an

r

-dimensional subspace. Extract a basis from the pivot columns of the original matrix, or characterise it as a null space via the left null space.

Test a specific targetFor a given

\vec{b}

, row-reduce

[\, A \mid \vec{b} \,]

. A pivot in the final column proves

\vec{b} \notin \mathcal{C}(A)

and the specification is unattainable.

Equations

Column space as a span and as an image

EQ-CSNS-01

C (A) = span ({A_{1}, A_{2}, \dots, A_{n}}) = {A \vec{x} ∣ \vec{x} \in ℂ^{n}}

The two descriptions agree because $A\vec{x}$ is by definition the linear combination of the columns of $A$ weighted by the entries of $\vec{x}$ .

Consistency criterion

EQ-CSNS-02

A \vec{x} = \vec{b} is consistent \Leftrightarrow \vec{b} \in C (A)

Solvability for a particular right-hand side is exactly membership of the column space. This holds for matrices of any shape.

Column space of a nonsingular matrix

EQ-CSNS-03

A nonsingular \Leftrightarrow C (A) = ℂ^{n}

The central result of this page, for a square matrix $A$ of size $n$ . Both implications require the square hypothesis.

Right-inverse construction in the converse proof

EQ-CSNS-04

Right-inverse construction in the converse proof

Each $\vec{b}_i$ is any solution of $A\vec{x} = \vec{e}_i$ . Collecting them column-wise produces a right inverse directly from surjectivity.

Dimension form of the criterion

EQ-CSNS-05

\dim (C (A)) = rank (A), C (A) = ℂ^{n} \Leftrightarrow rank (A) = n

A subspace of $\mathbb{C}^n$ equals $\mathbb{C}^n$ precisely when its dimension is $n$ , so full rank and full column space are the same condition.

Structure of the singular case

EQ-CSNS-06

A singular \Rightarrow \dim (C (A)) = r < n and \dim (N (A)) = n - r > 0

Deficient output and non-trivial input degeneracy arrive together, in exactly complementary amounts.

Column space of the worked example as a constraint

EQ-CSNS-07

C (A) = {\vec{y} \in ℂ^{3} ∣ y_{1} + y_{2} - y_{3} = 0}

A rank-two column space inside $\mathbb{C}^3$ is a plane through the origin, describable by a single linear condition drawn from the left null space.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Square matrix	The $n \times n$ matrix whose column space is under examination	n x n matrix over C
$n$	Size	Common row and column count of the square matrix	positive integer
$A_{i}$	Column of $A$	The $i$ th column of $A$ , regarded as a vector in $\mathbb{C}^n$	vector in C^n
$C (A)$	Column space	Span of the columns; the set of all reachable products $A\vec{x}$	subspace of C^n
$N (A)$	Null space	Set of vectors annihilated by $A$ ; trivial exactly when $A$ is nonsingular	subspace of C^n
$\vec{b}$	Target vector	Right-hand side of the system; the specification to be met	vector in C^n
${\vec{e}}_{i}$	Standard unit vector	Column $i$ of the identity matrix, used to probe the column space	vector in C^n
$I_{n}$	Identity matrix	Square matrix with ones on the diagonal and zeros elsewhere	n x n matrix
$r$	Rank	Number of pivot columns of the reduced form; the dimension of $\mathcal{C}(A)$	0 to n

Worked Numerical Example

Problem statement

A planar positioning stage is driven by three actuators. Each actuator contributes a fixed force direction, and the columns of the $3 \times 3$ matrix $A$ hold those directions in the stage frame. Determine whether every commanded force vector is achievable, and repeat the analysis after one actuator is re-aimed.

Assemble the actuator matrix
The three actuator direction vectors form the columns of $A$ . A command vector $\vec{x}$ of actuator efforts produces the resultant $A\vec{x}$ , so the achievable resultants are precisely $\mathcal{C}(A)$ .
$A = [\begin{matrix} 2 & 1 & 3 \\ 1 & - 1 & 0 \\ 3 & 0 & 3 \end{matrix}]$
Row-reduce and count pivots
Swap $R_1$ and $R_2$ to obtain a leading entry of $1$ , then apply $R_2 \to R_2 - 2R_1$ and $R_3 \to R_3 - 3R_1$ , scale the new second row by $\tfrac{1}{3}$ , and clear column $2$ above and below. The reduced row-echelon form has two pivot columns, so $r = 2$ and $A$ does not reduce to $I_3$ .
$Row-reduce and count pivots$
Diagnose the redundancy
Since $r = 2 < 3$ , the matrix is singular. The reduced form shows the dependence directly: the third column is the sum of the first two, which is confirmed against the original matrix by $A_1 + A_2 = (3, 0, 3)^{t} = A_3$ . The third actuator adds nothing the first two cannot already produce.
Describe the reachable set
The pivot columns $1$ and $2$ of the original matrix give a basis for $\mathcal{C}(A)$ , so the reachable set is the plane spanned by $(2,1,3)$ and $(1,-1,0)$ . Both satisfy $y_1 + y_2 - y_3 = 0$ , and that single condition describes the plane completely.
$C (A) = span ({[\begin{matrix} 2 \\ 1 \\ 3 \end{matrix}], [\begin{matrix} 1 \\ - 1 \\ 0 \end{matrix}]})$
Test an unreachable specification
Take the target $\vec{b} = \vec{e}_1 = (1,0,0)^{t}$ , a pure unit force along the first axis. Substituting into the plane condition gives $1 + 0 - 0 = 1 \ne 0$ , so $\vec{b} \notin \mathcal{C}(A)$ . Equivalently, row-reducing the augmented matrix leaves a pivot in the final column and the system is inconsistent.
$[\begin{matrix} 2 & 1 & 3 & 1 \\ 1 & - 1 & 0 & 0 \\ 3 & 0 & 3 & 0 \end{matrix}] ⟶ [\begin{matrix} 1 & 0 & 1 & 0 \\ 0 & 1 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}]$
Re-aim the third actuator
Change the third direction from $(3,0,3)$ to $(3,0,4)$ , giving a new matrix $B$ . Its determinant is $2(-4) - 1(4) + 3(3) = -3 \ne 0$ , and row reduction now yields $I_3$ , so $B$ is nonsingular.
$Re-aim the third actuator$
Conclude full reachability and solve for the former failure case
By the theorem, $\mathcal{C}(B) = \mathbb{C}^3$ , so every commanded force is achievable. The command that was previously impossible is now available and unique: solving $B\vec{x} = \vec{e}_1$ gives the first column of $B^{-1}$ . Direct substitution confirms it, since $2(\tfrac{4}{3}) + 1(\tfrac{4}{3}) + 3(-1) = 1$ and the remaining two components evaluate to zero.
$\vec{x} = [\begin{matrix} \frac{4}{3} \\ \frac{4}{3} \\ - 1 \end{matrix}], B \vec{x} = {\vec{e}}_{1}$

Result

The original three-actuator layout is singular: its reachable set is a two-dimensional plane inside a three-dimensional command space, so an entire family of specifications — including a pure force along the first axis — cannot be met at any effort level, while every reachable target admits a one-parameter family of commands. Re-aiming a single actuator makes the direction matrix nonsingular, which by the theorem promotes the reachable set to all of $\mathbb{C}^3$ and simultaneously makes each command unique. The change is a geometry decision, not a control-gain decision.

Applications & Industry Use

Robotics & motion control

Actuator authority in force allocation

For a square allocation matrix mapping actuator efforts to end-effector wrench, nonsingularity is exactly the condition that every commanded wrench is producible. A configuration where the matrix drops rank is a kinematic singularity: the reachable wrench set collapses to a subspace and the mechanism loses authority in the complementary directions, regardless of available power.

Power systems

Solvability of network injection problems

Linearised network models relate bus injections to state variables through a square admittance-derived matrix. A full column space guarantees that any admissible injection pattern corresponds to a state, so the linear solve is well posed; a deficient column space signals an ungrounded or islanded network in which certain injection patterns cannot be realised at all.

Chemical process engineering

Reachability of blend specifications

When a fixed set of feed streams is combined to hit target concentrations, the stream composition matrix is square when the number of streams equals the number of specifications. Nonsingularity means every specification within the model is attainable and the recipe is unique; a singular matrix means two streams are compositionally redundant and a whole family of targets is off limits.

Structural engineering

Statically determinate frames

For a pin-jointed frame with as many unknown member forces as equilibrium equations, the equilibrium matrix is square. Nonsingularity is the algebraic form of static determinacy: every load pattern is carried, by exactly one internal force distribution. Singularity indicates a mechanism, in which some load patterns cannot be equilibrated by any member forces.

Signal processing

Invertible transforms and perfect reconstruction

A square analysis matrix in a filter bank or transform coder must have full column space for perfect reconstruction, since the synthesis stage is a right inverse. The construction used in the proof of the theorem — solving for each standard unit vector in turn — is the abstract form of computing the synthesis basis one atom at a time.

Metrology & instrumentation

Sensor array observability

A square calibration matrix mapping physical quantities to sensor readings must be nonsingular for the array to distinguish every physical state. If it is singular, a whole subspace of physical states produces identical readings, and the corresponding directions in measurement space are never generated by any real input.

Design Considerations

Verify squareness before invoking the theorem

Both directions of the equivalence rely on the matrix being square. A wide matrix can be surjective while badly degenerate, and a tall matrix can be injective while missing most of its target space. For rectangular problems, use rank against the appropriate dimension: $\operatorname{rank}(A) = m$ for a full column space in $\mathbb{C}^m$ , $\operatorname{rank}(A) = n$ for a trivial null space.

Prefer one row reduction to many consistency tests

Testing $\vec{b} \in \mathcal{C}(A)$ for a specific target requires reducing $[\, A \mid \vec{b} \,]$ , which must be repeated for every new target. Reducing $A$ alone once and comparing the pivot count against $n$ settles the question for all targets simultaneously. Choose the second when the design question is about capability rather than a single load case.

Treat near-singularity as a design signal, not a numerical nuisance

In floating-point work, the boundary between singular and nonsingular is not observable exactly. A matrix with a very small smallest singular value is nonsingular in principle, so the theorem applies and every target is technically reachable — but the commands required grow without bound as the target approaches an unreachable direction. Report a condition number alongside any reachability claim.

Adding columns fixes surjectivity but destroys uniqueness

If a square matrix is singular, appending an extra column can restore a full column space. The matrix is then wide, the null space is necessarily non-trivial, and every target is reachable in infinitely many ways. This is the standard over-actuated design: it buys reachability and robustness at the price of needing an allocation policy — typically a minimum-norm or weighted pseudoinverse — to pick a command.

Never use the determinant as the reachability test in floating point

The determinant is zero exactly when the matrix is singular, so mathematically it settles the question. Numerically it is a poor instrument: it scales as the $n$ th power of the entries, so it can underflow or overflow for well-conditioned matrices and can be far from zero for badly conditioned ones. Use a rank-revealing factorisation or the singular values.

Describe a deficient column space by its constraints

When the column space is a proper subspace, listing a spanning set answers "what can be reached" but not "why is this target refused". Expressing the same subspace as a solution set of homogeneous equations — one equation per basis vector of the left null space — gives an immediate, auditable acceptance test for any proposed specification.

Standards & Codes

Notation, interchange and numerical standards that govern how this material is written down, stored and computed in production systems.

Applicable standards, conventions and reference implementations
Reference	Title	Relevance to this topic
`ISO 80000-2`	Quantities and units — Part 2: Mathematics	Fixes the notation used here for matrices, the identity matrix $I_n$ , set membership and spans, so that statements such as $\mathcal{C}(A) = \mathbb{C}^n$ are read consistently across documents.
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Provides the practical instruments for this test: `xGETRF` for LU with partial pivoting, `xGECON` for a condition estimate, and `xGESVD` or `xGEQP3` when rank must be decided under rounding.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Defines the arithmetic in which the singular/nonsingular boundary becomes a tolerance decision, and the overflow behaviour that makes determinant-based reachability tests unreliable at moderate matrix sizes.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the matrices and set expressions on this page as structured markup rather than images, keeping them searchable and machine-readable.
`IEC 61131-3`	Programmable controllers — Part 3: Programming languages	Relevant where an allocation matrix is inverted inside a control program: the standard's data typing and execution-time constraints often force a precomputed, offline-verified inverse rather than an on-line rank test.

Material Selection

For a mathematical topic, "material" is the numeric representation: the scalar field, storage format and precision the computation is built from.

Representation and precision selection
Representation	Select when	Trade-off
Exact rational arithmetic	Small symbolic or integer models where the design question is whether the matrix is singular at all, not how close it is.	Gives an unambiguous answer with no tolerance, but says nothing about conditioning, which is usually the property that matters in service.
IEEE 754 binary64 with an LU factorisation	Default choice for engineering-sized square systems solved repeatedly against many right-hand sides.	Fast and reusable across targets, but a pivot that is merely small rather than zero will be accepted, so a condition estimate must accompany the result.
IEEE 754 binary64 with a singular value decomposition	Reachability decisions on measured or estimated matrices, where the entries carry uncertainty.	Gives a numerically defensible rank and a quantitative distance to singularity, at several times the cost of an LU factorisation.
IEEE 754 binary32	Embedded or accelerator implementations of a fixed, well-conditioned allocation matrix.	Halves memory and bandwidth, but roughly seven significant digits leaves little margin between a genuinely full column space and a numerically deficient one.
Fixed-point arithmetic with a precomputed inverse	Hard real-time controllers where the matrix is constant and verified offline.	Deterministic timing and no floating-point unit required, but the inverse must be scaled by hand and rank cannot be re-checked at run time.
Finite field arithmetic modulo a prime	Certifying nonsingularity of a large integer matrix cheaply, or coding-theory applications where the matrix is defined over a finite field.	Exact and fast with bounded operand size, but nonsingularity over a finite field neither implies nor is implied by nonsingularity over $\mathbb{C}$ for unlucky primes.

Manufacturing Notes

Implementation notes — how the result is actually produced by hand, by algorithm and by library, including cost and numerical behaviour.

Cost of the two tests

Deciding nonsingularity by LU factorisation with partial pivoting costs about $\tfrac{2}{3}n^3$ operations and yields a reusable factorisation. Deciding it by full reduction to reduced row-echelon form costs roughly $n^3$ . Testing a single target for membership of the column space by reducing an augmented matrix costs the same order but must be repeated per target, so for more than a handful of right-hand sides the factorisation dominates on cost as well as on information.

Constructing the right inverse in practice

The proof constructs $B$ by solving $A\vec{x} = \vec{e}_i$ for each $i$ . That is precisely how a matrix inverse is computed in practice: factor $A$ once, then perform $n$ triangular solves against the columns of $I_n$ , for a total of about $2n^3$ operations. Libraries expose this as xGETRI after xGETRF; the theorem is the statement that surjectivity alone is enough to make the procedure succeed.

Library behaviour and what it actually reports

numpy.linalg.matrix_rank compares singular values against a tolerance scaled by the largest singular value and the matrix dimensions, so a full column space is reported as

\operatorname{rank}(A) = n

rather than as a boolean. numpy.linalg.inv raises only on exact singularity detected during factorisation, so it will silently return a meaningless inverse for a severely ill-conditioned matrix. Check numpy.linalg.cond before trusting the result.

Verifying a claimed inverse

Do not verify by re-deriving the inverse. Form the residual $AB - I_n$ and compare its norm against machine epsilon times the norm of $A$ times the norm of $B$ . A residual much larger than that bound indicates either a coding error or a matrix so ill-conditioned that the computed inverse carries no useful digits, and the two cases are distinguished by the condition estimate.

Exploiting a known structure

Many engineering allocation matrices are structured — symmetric positive definite, banded, or orthogonal. Each structure carries its own cheap certificate of nonsingularity: a completed Cholesky factorisation for positive definiteness, a banded LU for banded matrices, and $A^{*}A = I_n$ for unitary matrices. Testing the structural property is usually cheaper and more numerically robust than a general rank determination.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Applying the equivalence to a rectangular matrix	high	Treating "full column space" and "nonsingular" as interchangeable without checking that the matrix is square.	Confirm $m = n$ first. For rectangular matrices, state the result in rank form and specify which dimension the rank is compared against.
Concluding surjectivity from a non-empty column space	high	Observing that some targets are reachable and inferring that all are; the column space is never empty, since it always contains $\vec{0}$ .	Compare the dimension of the column space to $n$ , not to zero. Only $\operatorname{rank}(A) = n$ establishes $\mathcal{C}(A) = \mathbb{C}^n$ .
Using original columns of the reduced matrix as a basis	medium	Taking the pivot columns from the reduced row-echelon form instead of the corresponding columns of the original matrix.	Row operations change the column space. Use the reduced form only to identify which column indices are pivots, then extract those columns from the original $A$ .
Determinant used as a numerical singularity test	high	Computing $\det(A)$ in floating point and comparing against zero or a fixed threshold.	Use a condition estimate or the smallest singular value, both of which are scale-aware, instead of a determinant whose magnitude scales as the $n$ th power of the entries.
Assuming a singular matrix makes every system inconsistent	medium	Reading singularity as "unsolvable" rather than as "not solvable for every target".	A singular square matrix still solves $A\vec{x} = \vec{b}$ for every $\vec{b}$ inside $\mathcal{C}(A)$ — with infinitely many solutions. Always test the specific target.
Mistaking a right inverse for a proof in the rectangular case	medium	Carrying the proof's construction over to a wide matrix, where $AB = I_m$ can hold with $B$ not a left inverse.	The step that closes the proof is that $A$ and $B$ are both square, so a one-sided inverse forces a two-sided one. Without squareness that step is invalid.
Adding actuators or columns without an allocation policy	medium	Restoring a full column space by widening the matrix, then finding the command undefined because the null space is non-trivial.	Specify the selection rule — minimum-norm, weighted pseudoinverse, or a constrained optimisation — at the same time as the extra column is added.
Ignoring the coupling between deficient output and degenerate input	low	Diagnosing only the unreachable targets and overlooking the equally large family of commands producing identical outputs.	Compute both $\dim(\mathcal{C}(A)) = r$ and $\dim(\mathcal{N}(A)) = n - r$ , and report the null space basis as the set of commands with no effect.

FAQs

Why does this theorem require the matrix to be square?

Because it equates a statement about outputs with a statement about inputs, and only for a square matrix do the input and output spaces have the same dimension. A $2 \times 3$ matrix can have column space equal to all of $\mathbb{C}^2$ while its null space is a line, so it is surjective and degenerate at once. Squareness is what forces rank deficiency in one direction to imply rank deficiency in the other.

Does a full column space guarantee the solution is unique?

For a square matrix, yes — because $\mathcal{C}(A) = \mathbb{C}^n$ implies nonsingularity, which implies a trivial null space and therefore a unique solution for every right-hand side. For a rectangular matrix it does not: a wide matrix with a full column space always has infinitely many solutions for every consistent target.

How do I check whether a specific vector lies in the column space?

Row-reduce the augmented matrix $[\, A \mid \vec{b} \,]$ and look at the final column. If it contains a pivot, the system is inconsistent and $\vec{b} \notin \mathcal{C}(A)$ . For repeated tests against the same matrix, it is more efficient to compute a basis for the left null space once and then check that $\vec{b}$ satisfies the corresponding homogeneous conditions.

What exactly goes wrong when the matrix is singular?

Two things simultaneously. The reachable set shrinks to a proper subspace of dimension $r < n$ , so a large family of targets is unattainable at any command level, and each attainable target is produced by an infinite family of commands differing by elements of the null space. In an engineering model these appear as lost authority and as unobservable command directions.

Is a numerically near-singular matrix still covered by the theorem?

Mathematically yes: a matrix with a very small but non-zero smallest singular value is nonsingular, so the column space is genuinely all of $\mathbb{C}^n$ . Practically the guarantee is hollow, because the commands required to reach targets near the deficient direction grow in proportion to the condition number and will exceed physical limits long before the theorem fails.

Why is the converse direction of the proof harder than the forward direction?

The forward direction only has to spend information — nonsingularity already delivers a solution for every right-hand side, which is the definition of a full column space. The converse has to create information about the null space out of a statement about reachability, and it does so constructively, by solving for each standard unit vector and assembling a right inverse.

Can I conclude the columns of a square matrix are a basis just from spanning?

Yes, for a square matrix. Spanning $\mathbb{C}^n$ means $\mathcal{C}(A) = \mathbb{C}^n$ , which by this theorem makes $A$ nonsingular, which makes the columns linearly independent. A spanning set of $n$ vectors in $\mathbb{C}^n$ is automatically independent, and an independent set of $n$ vectors in $\mathbb{C}^n$ automatically spans.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section CRS, Subsection CSNSM, Theorem CSNSM. Licensed under the GNU Free Documentation License v1.2.
Strang, G. Linear Algebra and Its Applications, 4th edition. Cengage Learning, 2006.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013.
Anderson, E. et al. LAPACK Users' Guide, 3rd edition. Society for Industrial and Applied Mathematics, 1999.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
Higham, N. J. Accuracy and Stability of Numerical Algorithms, 2nd edition. Society for Industrial and Applied Mathematics, 2002.

AI Suggested Questions

Construct a $3 \times 3$ singular matrix whose column space is a specified plane through the origin, and give the left null space vector that certifies it.
Show me a wide matrix that is surjective but has a non-trivial null space, and a tall matrix that is injective but not surjective, to demonstrate why squareness is essential here.
How does the condition number of a nonsingular allocation matrix bound the actuator effort required to reach a unit target in the worst direction?
Compare the operation counts of deciding nonsingularity by LU factorisation, by full row reduction and by singular value decomposition for a $500 \times 500$ matrix.
Explain how the right-inverse construction in the proof relates to computing a matrix inverse by $n$ triangular solves against the columns of the identity.
If a square matrix has rank $n-1$ , what is the geometry of the set of right-hand sides for which the system remains consistent, and how would I test membership efficiently?

Related Calculators

Column Space Calculator

Return a basis for $\mathcal{C}(A)$ from the original columns, together with the homogeneous conditions that characterise it.

Nonsingularity & Rank Checker

Decide whether a square matrix is nonsingular, report the rank, the condition estimate and the distance to singularity.

Linear System Solver

Solve $A\vec{x} = \vec{b}$ , test whether $\vec{b}$ lies in the column space and report the full solution set.

Matrix Inverse Calculator

Compute $A^{-1}$ by reduction of $[\, A \mid I_n \,]$ with an explicit residual check on $AA^{-1} - I_n$ .