Linear Independence and Nonsingular Matrices

Q: If a square matrix is singular, is every column redundant?

No. Dependence guarantees that at least one column is a combination of the others, namely any column whose coefficient in the relation is non-zero. Columns with zero coefficients in every relation are not redundant. The number of columns that can be removed while preserving the column space is exactly $n - r$.

Q: What do the columns of a nonsingular matrix span?

All of $\mathbb{C}^n$. Since there are $n$ of them and they are independent, they form a basis, so every vector $\vec{b}$ is a combination of them — which is the geometric reason $A\vec{x} = \vec{b}$ is always consistent. This later joins the equivalence list as a further entry.

Q: Is checking independence cheaper than checking nonsingularity?

No, and that is the point: they are the same computation. Both are answered by one row reduction or one LU factorisation at $O(n^3)$ cost. What can be cheaper is exploiting structure — orthogonality, triangularity or diagonal dominance — which establishes any member of the list without the general test.

Q: How does the equivalence list grow?

Each new characterisation of nonsingularity proved later is appended. Subsequent rounds add invertibility, full rank, non-zero determinant, spanning of $\mathbb{C}^n$ by the columns, injectivity and surjectivity of the associated linear transformation, and the absence of zero as an eigenvalue. Every addition is a new route into the same equivalence class.

Q: What should I actually check on real data?

Not exact independence. Compute the condition number or the smallest singular value of the matrix and compare it with a threshold justified by the uncertainty in the entries. A matrix that is mathematically nonsingular but has condition number $10^{14}$ behaves like a singular one in binary64, and the columns are near-dependent for every practical purpose.

Overview

Two ideas developed for different purposes turn out to coincide. Nonsingularity was defined by the behaviour of a solution set: a square matrix is nonsingular when the homogeneous system it heads admits only the trivial solution. Linear independence was defined by the absence of redundancy: a set of vectors is independent when the only relation combining them into the zero vector is the trivial one.

Applied to the columns of a square matrix, these are the same condition. A relation of linear dependence on the columns of $A$ , with coefficients collected into a vector $\vec{x}$ , is exactly a solution of $A\vec{x} = \vec{0}$ . The trivial relation corresponds to $\vec{x} = \vec{0}$ . So the columns are independent precisely when that system has a unique solution — and for a square matrix, that is the definition of nonsingularity.

The proof is therefore a chain rather than a case analysis. Nonsingularity is equivalent to the homogeneous system having a unique solution by definition, and that in turn is equivalent to independence of the columns by the homogeneous-system test for independence. Two links, both already established, and the theorem follows without any new argument.

The value is interchangeability. Structural questions about redundancy among columns — whether one sensor channel duplicates another, whether one loading direction is reachable by combining others — can be answered by a nonsingularity test, and questions about the solvability of a system can be answered by inspecting the columns. This equivalence is the fifth entry on a list that continues to grow, and the practical technique is always to establish whichever member of the list is cheapest in the situation at hand.

Definition

Nonsingular Matrices Have Linearly Independent Columns

NSLIC

Let $A$ be a square matrix. Then $A$ is nonsingular if and only if the columns of $A$ form a linearly independent set. Equivalently, $A$ is singular if and only if its columns are linearly dependent.

The hypothesis that $A$ is square is essential for the statement as written, because nonsingularity is defined only for square matrices. Independence of the columns remains meaningful for rectangular matrices and is then called full column rank.

Nonsingular Matrix Equivalences, Round 2

NSME2

Let $A$ be a square matrix of size $n$ . The following statements are equivalent — each implies all the others.

$A$ is nonsingular.
$A$ row-reduces to the identity matrix $I_n$ .
The null space of $A$ contains only the zero vector, $\mathcal{N}(A) = \{\vec{0}\}$ .
The system $A\vec{x} = \vec{b}$ has a unique solution for every choice of $\vec{b}$ .
The columns of $A$ form a linearly independent set.

The list grows as the subject develops. Later rounds add invertibility, full rank, non-zero determinant, spanning of $\mathbb{C}^n$ by the columns, and the absence of zero eigenvalues.

Concepts

The chained proof

Some equivalences are proved by establishing each implication separately. This one is proved by chaining two existing equivalences, which is shorter and more informative.

By the definition of nonsingularity, $A$ is nonsingular exactly when the homogeneous system $A\vec{x} = \vec{0}$ has a unique solution. By the homogeneous-system criterion for linear independence, the columns of $A$ form a linearly independent set exactly when that same system has a unique solution. The two statements are equivalent to a common third statement, and therefore to each other. No new mathematics is introduced; the theorem records that two definitions from different chapters happen to be the same condition wearing different notation.

Why the two ideas coincide

The mechanism is the identification of a linear combination of the columns with a matrix-vector product. Writing $A = [\, \vec{A}_1 \mid \vec{A}_2 \mid \cdots \mid \vec{A}_n \,]$ , the product $A\vec{x}$ equals $x_1\vec{A}_1 + x_2\vec{A}_2 + \cdots + x_n\vec{A}_n$ . A vector in the null space is therefore precisely a list of coefficients producing a relation of linear dependence on the columns.

Read that correspondence in both directions. The null space of $A$ is the space of dependence relations among its columns. A trivial null space means no non-trivial relations, hence independence. A non-trivial null space means at least one relation, hence dependence, and the null vector itself names the redundancy: its non-zero entries mark which columns participate.

The square case removes one escape route

For a general $m \times n$ matrix, two counting results bound what is possible. If $n > m$ the columns are automatically dependent, since more vectors than the size of the space cannot be independent. If $n < m$ the columns cannot span $\mathbb{C}^m$ , though they may well be independent.

The square case $n = m$ is the only one in which independence and spanning can both hold, and where they hold, they hold together. That is why nonsingularity is a square-matrix concept and why the equivalence list is so long: in the balanced case, every reasonable notion of the matrix being “full strength” collapses into a single condition. It is also why an automatic dependence result is unavailable here — with $n = m$ the counting bound is silent and genuine computation is required.

Reading the null vector as a column relation

When a square matrix is singular, the reduction that established it also delivers a certificate. Each free variable, set to $1$ with the others set to $0$ , yields a non-zero null vector, and that vector's entries are the coefficients of an explicit relation among the columns.

This converts an abstract verdict into an actionable diagnosis. In a calibration matrix, the relation identifies which channels are linear combinations of others. In a stiffness matrix, it names the displacement pattern that costs no energy. In a design matrix, it names the collinear predictors. The verdict “singular” on its own tells an engineer only that something is wrong; the null vector tells them what.

The list as a proof strategy

An equivalence list is a substitution device. Any one statement may be assumed as a hypothesis and any other concluded, so the cost of a proof is the cost of the cheapest available route. Establishing independence of columns may be trivial in a construction where the columns are orthogonal by design; establishing a trivial null space may be trivial where a physical argument shows no zero-energy mode exists; establishing reduction to the identity may be trivial where the matrix is triangular with non-zero diagonal.

The same logic applies computationally. A single row reduction settles every statement on the list at once, so there is never a reason to run separate tests for independence, for invertibility and for unique solvability. Recognising that they are one question is what avoids paying three times for one answer.

Contrast: dependent columns of a singular matrix

The negated statement is equally useful. A singular square matrix has linearly dependent columns, so at least one column is a linear combination of the others and can be discarded without changing the column space. The column space of a singular $n \times n$ matrix therefore has dimension strictly less than $n$ , which means some vectors $\vec{b}$ lie outside it and the system $A\vec{x} = \vec{b}$ is inconsistent for those.

This closes the circle with the solution-set view. Singularity produces both failure modes at once: for $\vec{b}$ inside the column space, dependent columns mean infinitely many solutions; for $\vec{b}$ outside it, there are none. Which failure appears depends on $\vec{b}$ ; that a failure appears depends only on the columns.

Using the equivalence in practice

Confirm the matrix is squareNonsingularity requires

m = n

. For a rectangular matrix, ask instead whether the columns are independent, which is the property of full column rank.

Choose the cheapest entry pointOrthogonal columns give independence immediately; a triangular matrix with non-zero diagonal reduces to

I_n

on inspection; a known zero-energy mode establishes singularity without arithmetic.

Otherwise, row-reduce onceOne reduction of

A

settles every statement on the equivalence list. Compare the reduced form with

I_n

, or equivalently count the non-zero rows and compare

r

with

n

If singular, extract the relationSet the single free variable of a rank-deficient case to

1

and solve for the rest. The resulting null vector is the explicit relation of dependence among the columns.

Transfer the conclusionRead the verdict in whichever language the problem needs: redundant columns, non-trivial null space, non-unique solutions, or failure to reduce to the identity.

Equations

Columns of a matrix

EQ-LINSM-01

A = [{\vec{A}}_{1} ∣ {\vec{A}}_{2} ∣ \dots ∣ {\vec{A}}_{n}], A \vec{x} = x_{1} {\vec{A}}_{1} + x_{2} {\vec{A}}_{2} + \dots + x_{n} {\vec{A}}_{n}

The identification on which the whole equivalence rests: a matrix-vector product is a linear combination of the columns weighted by the entries of the vector.

Nonsingularity and independent columns

EQ-LINSM-02

A nonsingular \Leftrightarrow {{\vec{A}}_{1}, {\vec{A}}_{2}, \dots, {\vec{A}}_{n}} independent

The theorem, for square $A$ only. Its negation is equally usable: $A$ is singular exactly when its columns are linearly dependent.

The chain of equivalences that proves it

EQ-LINSM-03

A nonsingular \Leftrightarrow A \vec{x} = \vec{0} unique \Leftrightarrow {{\vec{A}}_{1}, \dots, {\vec{A}}_{n}} independent

The first link is the definition of nonsingularity; the second is the homogeneous-system criterion for linear independence. Chaining two known equivalences replaces a two-part proof.

Null space as the space of column relations

EQ-LINSM-04

N (A) = {\vec{x} \in ℂ^{n} ∣ x_{1} {\vec{A}}_{1} + \dots + x_{n} {\vec{A}}_{n} = \vec{0}}

Every null vector is a relation of linear dependence on the columns, and every relation is a null vector. Nonsingularity is the statement that only the trivial relation exists.

Rank criterion for a square matrix

EQ-LINSM-05

A nonsingular \Leftrightarrow r = n \Leftrightarrow RREF (A) = I_{n}

For a square matrix of size $n$ , full rank, reduction to the identity, and independence of the columns are one condition. A single reduction decides all three.

A singular matrix and its column relation

EQ-LINSM-06

B = [\begin{matrix} 1 & 2 & 3 \\ 2 & 5 & 7 \\ 3 & 7 & 10 \end{matrix}], {\vec{B}}_{3} = {\vec{B}}_{1} + {\vec{B}}_{2}, B [\begin{matrix} - 1 \\ - 1 \\ 1 \end{matrix}] = \vec{0}

The third column is the sum of the first two, so the columns are dependent and $B$ is singular. The corresponding null vector states the same fact in solution-set language.

The equivalence list, round two

EQ-LINSM-07

A nonsingular \Leftrightarrow RREF (A) = I_{n} \Leftrightarrow N (A) = {\vec{0}} \Leftrightarrow \exists! \vec{x} \forall \vec{b} \Leftrightarrow {{\vec{A}}_{1}, \dots, {\vec{A}}_{n}} independent

Five equivalent statements about a square matrix. Establish whichever is cheapest in context and inherit the rest; a single row reduction establishes all five simultaneously.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Square matrix	The matrix under test, with equal row and column counts	n x n matrix
${\vec{A}}_{j}$	Column $j$	The $j$ -th column of $A$ , regarded as a vector in $\mathbb{C}^n$	vector in C^n
$n$	Size	Common number of rows and columns; also the number of columns whose independence is tested	positive integer
$\vec{x}$	Coefficient vector	Solution of the homogeneous system; its entries are the scalars of a relation of dependence on the columns	vector in C^n
$\vec{b}$	Constant vector	Right-hand side for which unique solvability is asserted	vector in C^n
$N (A)$	Null space	Space of all relations of linear dependence among the columns of $A$	subspace of C^n
$C (A)$	Column space	Span of the columns; equal to $\mathbb{C}^n$ exactly when $A$ is nonsingular	subspace of C^n
$r$	Rank	Number of non-zero rows of the reduced form; equals $n$ precisely when $A$ is nonsingular	0 to n
$I_{n}$	Identity matrix	The reduced row-echelon form reached by every nonsingular square matrix	n x n matrix

Worked Numerical Example

Problem statement

A three-channel instrument mixes three physical quantities. Two candidate designs give calibration matrices $A$ with rows $(1,2,3)$ , $(2,5,7)$ , $(3,7,11)$ and $B$ with rows $(1,2,3)$ , $(2,5,7)$ , $(3,7,10)$ — identical except for the last entry. Decide for each whether the columns are linearly independent, and translate the verdict into a statement about recovering the physical state from the readings.

Note that the counting bound is silent
Each matrix contributes $n = 3$ columns from $\mathbb{C}^3$ . Since $n = m$ , the automatic dependence result for $n > m$ does not apply and neither design can be dismissed by inspection. Both matrices are square, so the language of nonsingularity is available.
$A = [\begin{matrix} 1 & 2 & 3 \\ 2 & 5 & 7 \\ 3 & 7 & 11 \end{matrix}], B = [\begin{matrix} 1 & 2 & 3 \\ 2 & 5 & 7 \\ 3 & 7 & 10 \end{matrix}]$
Reduce $A$ : clear the first column
The pivot entry is already $1$ . Apply $-2R_1 + R_2$ , giving $(0,\ 5-4,\ 7-6) = (0, 1, 1)$ , and $-3R_1 + R_3$ , giving $(0,\ 7-6,\ 11-9) = (0, 1, 2)$ .
$[\begin{matrix} 1 & 2 & 3 \\ 0 & 1 & 1 \\ 0 & 1 & 2 \end{matrix}]$
Finish the reduction of $A$
Clear column 2 above and below with $-2R_2 + R_1$ and $-1R_2 + R_3$ , giving rows $(1, 0, 1)$ and $(0, 0, 1)$ . The third pivot is already $1$ ; clear column 3 with $-1R_3 + R_1$ and $-1R_3 + R_2$ . Every entry off the diagonal is now zero.
$RREF (A) = I_{3}$
Verdict for $A$
The reduced form is the identity, so $r = 3 = n$ and $A$ is nonsingular. By the equivalence, the columns of $A$ form a linearly independent set: no column is a combination of the other two, and the only relation of linear dependence is the trivial one.
$A nonsingular, N (A) = {\vec{0}}, {{\vec{A}}_{1}, \dots, {\vec{A}}_{n}} independent$
Reduce $B$ with the same operations
Applying $-2R_1 + R_2$ and $-3R_1 + R_3$ gives $(0, 1, 1)$ and $(0,\ 7-6,\ 10-9) = (0, 1, 1)$ . The two lower rows are now identical, so $-1R_2 + R_3$ produces a zero row. Clearing column 2 with $-2R_2 + R_1$ leaves the reduced form below.
$RREF (B) = [\begin{matrix} 1 & 0 & 1 \\ 0 & 1 & 1 \\ 0 & 0 & 0 \end{matrix}]$
Verdict for $B$ , and the explicit relation
Here $r = 2 < 3 = n$ , so $B$ is singular and its columns are linearly dependent. Column 3 is the non-pivot column, so $x_3$ is free; the rows read $x_1 + x_3 = 0$ and $x_2 + x_3 = 0$ . Setting $x_3 = 1$ gives the null vector $(-1, -1, 1)$ , which states that $\vec{B}_3 = \vec{B}_1 + \vec{B}_2$ . Direct addition confirms it: $(1,2,3) + (2,5,7) = (3,7,10)$ .
$- {\vec{B}}_{1} - {\vec{B}}_{2} + {\vec{B}}_{3} = \vec{0}, {\vec{B}}_{3} = {\vec{B}}_{1} + {\vec{B}}_{2}$
Translate into instrument terms
For design $A$ the three channels respond independently, so every physical state produces a distinct reading vector and the calibration solve has a unique answer for any measurement. For design $B$ the third channel responds exactly as the sum of the first two, contributing no information; equivalently, the physical states $\vec{x}$ and $\vec{x} + t(-1,-1,1)$ are indistinguishable for every $t$ .
$B (\vec{x} + t [\begin{matrix} - 1 \\ - 1 \\ 1 \end{matrix}]) = B \vec{x} \forall t \in ℂ$
Note how small the difference is
The two matrices differ in a single entry, $11$ against $10$ . Dependence is a knife-edge condition: it holds on a set of measure zero, so an instrument built to design $B$ would show independent-looking columns as soon as manufacturing tolerance perturbed the response. The engineering defect is not exact singularity but near-singularity, which must be assessed with a condition number rather than an exact test.
$\det (A) = 1, \det (B) = 0$

Result

Design $A$ is nonsingular, its columns are independent, and the calibration is well posed for every measurement. Design $B$ is singular, its third column is the sum of the first two, and an entire line of physical states maps to the same reading. Both conclusions came from one row reduction each; no separate independence test, invertibility test or solvability test was required, because for a square matrix these are one question.

Applications & Industry Use

Instrumentation & metrology

Certifying a sensor array design

A square mixing matrix must have independent columns for every physical state to produce a distinct reading. Checking nonsingularity at design time is the same test, and when it fails the null vector identifies the exact combination of quantities the array cannot resolve — usually pointing at two sensors placed too similarly.

Structural engineering

Detecting mechanisms from column dependence

A dependent set of columns in a restrained stiffness matrix means a displacement pattern exists that generates no restoring force. The relation among the columns is the mechanism itself, so the same computation that reports singularity also draws the picture of the missing restraint.

Statistics & econometrics

Identifiability of a fully specified model

When the number of predictors equals the number of independent observations, the design matrix is square and the model is identifiable exactly when its columns are independent. Failure is typically a dummy-variable trap or an accounting identity among the regressors, and the null vector names the offending combination directly.

Control systems

Controllability of a square system

The Kalman controllability matrix of an $n$ -state system with $n$ inputs is square, and the system is controllable exactly when its columns are independent. Dependence identifies a direction in state space that no admissible input can reach, which is a structural limitation of the actuator layout rather than of the controller.

Computer graphics

Validity of a coordinate frame

A local frame is stored as a square matrix whose columns are the basis directions. Independence is what makes the frame usable: it guarantees an invertible transformation between local and world coordinates. Degenerate frames arising from a collapsed normal or a zero-length tangent are exactly the singular case.

Chemical engineering

Independence of specification constraints

A blending specification with as many constraints as constituents gives a square matrix. Independent columns mean each constituent influences the specifications in a way the others cannot reproduce, so a unique recipe exists for any target. Dependence means the specification set is degenerate and admits a family of recipes.

Design Considerations

One reduction answers every question on the list

Independence of columns, triviality of the null space, unique solvability, full rank and reduction to the identity are one condition for a square matrix. Running separate tests wastes several times the necessary work and risks inconsistent verdicts when tolerances differ between routines. Perform one factorisation and report all the consequences from it.

Exact dependence is not the engineering question

Perturbation makes a dependent set independent, so an exact test on measured or computed data almost always returns a clean verdict for a design that is in practice unusable. The meaningful quantity is the distance to singularity — the smallest singular value — or the condition number, and the threshold must be justified against the data's own uncertainty.

Keep the null vector, not just the verdict

A report of singularity is an error message; the null vector is a diagnosis. It names the redundant channel, the unrestrained mechanism, the collinear predictor or the unreachable state direction, and it comes free from the reduction that produced the verdict. Discarding it forces a second investigation to recover information already computed.

Do not extend the statement to rectangular matrices

Nonsingularity is defined only for square matrices, so the equivalence as stated has no rectangular version. What generalises is independence of the columns, which for an $m \times n$ matrix is full column rank, $r = n$ . Note that this requires $n \le m$ ; with more columns than rows the columns are automatically dependent.

Design for independence rather than testing for it

Where the columns are chosen rather than measured — sensor placement, actuator layout, basis construction — enforcing orthogonality gives independence by construction and a condition number of one. Orthogonal non-zero columns are always independent, the inverse is the transpose or adjoint, and no test is needed at any point in the lifecycle.

Use the equivalence to pick the cheapest proof

In an argument, assume whichever member of the list the situation already supplies. A physical energy argument may establish a trivial null space directly; a construction may make the columns orthogonal and hence independent; a triangular structure may make reduction to the identity immediate. All are equally good starting points, and the list transports the conclusion to whichever form is needed.

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	Standardises the notation for matrix columns, the null space and the identity matrix used in every statement of the equivalence, and the typographic conventions distinguishing vectors from scalars.
`LAPACK reference`	Linear Algebra PACKage reference implementation	Provides the single computation that settles the whole equivalence list: `xGETRF` for the factorisation, `xGECON` for the condition estimate that quantifies near-singularity, and `xGESVD` where the distance to dependence is required.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Explains why an exact independence test is unusable on computed columns: rounding perturbs a dependent set into an independent one, so the verdict must be taken against a stated tolerance rather than against exact zero.
`ISO 5725-1`	Accuracy (trueness and precision) of measurement methods and results	Supplies the uncertainty framework needed to set that tolerance when the columns are measured quantities, such as sensor response vectors or identified mode shapes.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the equivalence statements, column relations and matrices on this page as structured markup so each is individually addressable by search and by assistive technology.

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 or integer arithmetic	Design-time certification of a small square matrix where the verdict must be unambiguous.	Definitive and tolerance-free, but silent about how close a nonsingular matrix is to singular — usually the more important question in practice.
IEEE 754 binary64 with LU and condition estimation	The general default for square matrices assembled from measured or computed quantities.	One factorisation settles the whole equivalence list at $O(n^3)$ cost, and the condition estimate adds only $O(n^2)$ ; but the verdict is threshold-dependent and the tolerance must be reported.
Singular value decomposition	The distance to dependence must be quantified for robustness, sensor placement or workspace design.	The smallest singular value is exactly the distance to the nearest singular matrix, at roughly an order of magnitude more work than a factorisation.
Orthogonal or unitary construction	The columns are being designed rather than measured — frames, modal bases, actuator directions.	Independence is guaranteed by construction with condition number one and the inverse available as the adjoint; but orthogonality may conflict with physical placement constraints.
Finite field arithmetic modulo a prime	Certifying nonsingularity of a large integer matrix cheaply, or working in coding and cryptographic settings.	A nonsingular result modulo a prime certifies independence over the rationals at low cost; a singular modular result is inconclusive, because an unlucky prime can create a spurious dependence.
Sparse factorisation with fill-reducing ordering	Large square systems from discretised physical models.	Makes the test tractable at a cost governed by fill-in rather than by $n^3$ , but a zero pivot may reflect the ordering as much as the matrix, so structural and numerical singularity must be distinguished.

Manufacturing Notes

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

One computation, five conclusions

A single LU factorisation with partial pivoting, at roughly $\tfrac{2}{3}n^3$ operations, decides every statement on the equivalence list. The pivots decide nonsingularity; their count gives the rank; a zero pivot triggers extraction of a null vector and hence a column relation. There is never a reason to compute an inverse, a determinant and a rank separately in order to answer what is one question.

Extracting the column relation

After reduction, each non-pivot column indexes a free variable. Setting one free variable to $1$ and the rest to $0$ , then back-substituting for the dependent variables, produces a null vector whose entries are the coefficients of an explicit relation among the columns. Verify it by forming the linear combination and confirming every entry is zero.

Structural shortcuts

A triangular matrix has independent columns exactly when no diagonal entry is zero, decidable in $O(n)$ . A matrix with orthogonal non-zero columns has independent columns by construction. A strictly diagonally dominant matrix is nonsingular and hence has independent columns. Each shortcut avoids the cubic test entirely and should be checked before it is incurred.

Library behaviour

Comparing numpy.linalg.matrix_rank(A) with A.shape[1] implements the independence test with an SVD-based tolerance. numpy.linalg.solve raises LinAlgError only on exactly singular input and will happily return meaningless values for a near-singular matrix, so numpy.linalg.cond should accompany it. SymPy answers exactly for symbolic matrices via Matrix.rank and Matrix.nullspace.

Hand technique

Scan for an obvious relation first — a zero column, a repeated column, or one column that is visibly the sum or multiple of others — since finding one settles the question immediately. Otherwise reduce, and stop the moment a zero row appears. For sizes two and three the determinant is a legitimate exact shortcut, though it should never be used as a numerical test.

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	Concluding a $4 \times 3$ matrix is nonsingular because its columns are independent.	Nonsingularity requires $m = n$ . For rectangular matrices the correct statement is full column rank, $r = n$ , which requires $n \le m$ .
Testing exact dependence on measured columns	high	Applying an exact rank test to floating-point sensor or simulation data, where rounding guarantees an independent verdict.	Use the smallest singular value or the condition number against a tolerance derived from the measurement uncertainty.
Running separate tests for the same property	medium	Computing a rank, a determinant and an inverse to check independence, invertibility and solvability as if they were distinct questions.	One factorisation settles all five statements. Separate routines also use different tolerances and can return contradictory verdicts on the same matrix.
Judging by determinant magnitude	high	Treating a small determinant as evidence of near-dependence among columns.	The determinant scales as the $n$ -th power of a uniform column scaling and carries no conditioning information. Use the condition number or the smallest singular value.
Checking row independence instead of column independence	medium	Reducing the transpose out of habit, or forming the matrix with the vectors as rows.	For a square matrix the verdict is the same, since row rank equals column rank, but the null space and any extracted relation then describe the rows. Fix the orientation before interpreting the certificate.
Assuming independence implies orthogonality	medium	Treating a nonsingular matrix as if its columns formed a well-conditioned frame.	Independent columns can be nearly parallel and the matrix arbitrarily ill conditioned. Measure the conditioning explicitly rather than inferring it from the nonsingularity verdict.
Discarding the null vector	medium	Reporting only that a matrix is singular and re-investigating the cause separately.	Capture the null vector from the same reduction. Its non-zero entries identify precisely which columns participate in the redundancy.
Expecting a counting shortcut in the square case	low	Transferring the automatic dependence result for $n > m$ to the case $n = m$ .	The counting bound is silent when $n = m$ ; this is exactly the case in which genuine computation is required and both outcomes are possible.

FAQs

Why are nonsingularity and column independence the same thing?

Because a matrix-vector product $A\vec{x}$ is a linear combination of the columns of $A$ weighted by the entries of $\vec{x}$ . A relation of linear dependence on the columns is therefore exactly a solution of $A\vec{x} = \vec{0}$ . Independence says only the trivial relation exists; nonsingularity says only the trivial solution exists. They are one condition in two notations.

Does the theorem apply to rectangular matrices?

Not as stated, since nonsingularity is defined only for square matrices. Independence of the columns generalises perfectly well and is called full column rank, $r = n$ . For $n > m$ the columns are automatically dependent, and for $n < m$ they may be independent but cannot span the whole space.

If a square matrix is singular, is every column redundant?

No. Dependence guarantees that at least one column is a combination of the others, namely any column whose coefficient in the relation is non-zero. Columns with zero coefficients in every relation are not redundant. The number of columns that can be removed while preserving the column space is exactly $n - r$ .

What do the columns of a nonsingular matrix span?

All of $\mathbb{C}^n$ . Since there are $n$ of them and they are independent, they form a basis, so every vector $\vec{b}$ is a combination of them — which is the geometric reason $A\vec{x} = \vec{b}$ is always consistent. This later joins the equivalence list as a further entry.

Is checking independence cheaper than checking nonsingularity?

No, and that is the point: they are the same computation. Both are answered by one row reduction or one LU factorisation at $O(n^3)$ cost. What can be cheaper is exploiting structure — orthogonality, triangularity or diagonal dominance — which establishes any member of the list without the general test.

How does the equivalence list grow?

Each new characterisation of nonsingularity proved later is appended. Subsequent rounds add invertibility, full rank, non-zero determinant, spanning of $\mathbb{C}^n$ by the columns, injectivity and surjectivity of the associated linear transformation, and the absence of zero as an eigenvalue. Every addition is a new route into the same equivalence class.

What should I actually check on real data?

Not exact independence. Compute the condition number or the smallest singular value of the matrix and compare it with a threshold justified by the uncertainty in the entries. A matrix that is mathematically nonsingular but has condition number $10^{14}$ behaves like a singular one in binary64, and the columns are near-dependent for every practical purpose.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section LI, Theorem NSLIC and Theorem NSME2. Licensed under the GNU Free Documentation License v1.2.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013.
Higham, N. J. Accuracy and Stability of Numerical Algorithms, 2nd edition. Society for Industrial and Applied Mathematics, 2002.
Anderson, E. et al. LAPACK Users' Guide, 3rd edition. Society for Industrial and Applied Mathematics, 1999.
Strang, G. Linear Algebra and Its Applications, 4th edition. Thomson Brooks/Cole, 2006.

AI Suggested Questions

Give a 4x4 matrix whose columns are independent but whose condition number exceeds $10^{12}$ , and explain what a binary64 solve does with it.
Show why the counting bound for automatic dependence is silent exactly in the square case, and what that implies for the equivalence list.
For a singular 5x5 matrix with rank 3, how many columns can be removed without changing the column space, and which ones?
Trace how the nonsingular matrix equivalence list grows through invertibility, determinant and eigenvalue characterisations.
Compare establishing independence via orthogonality by construction against testing it numerically after the fact for a sensor array design.
How does full column rank generalise this theorem to rectangular matrices, and what fails when there are more columns than rows?

Related Calculators

Nonsingular Matrix Checker

Row-reduce a square matrix, compare with $I_n$ , and report both the nonsingularity verdict and the independence of the columns.

Linear Independence Checker

Test any set of column vectors by the rank criterion and return an explicit relation of dependence when one exists.

Condition Number Estimator

Quantify how close independent columns are to dependent using the condition number and smallest singular value.