Bases and Nonsingular Matrices | KEVOS® Mathematics

Overview

A basis is the smallest set of vectors that still describes an entire vector space: linearly independent, so nothing in it is redundant, and spanning, so nothing outside it is needed. Constructing one from scratch is laborious — two separate properties must be checked, each of them a system of equations. The result on this page removes that labour entirely for the space $\mathbb{C}^m$ of complex column vectors of size $m$ : assemble any $m$ vectors as the columns of a square matrix, ask the single question "is this matrix nonsingular?", and the answer settles independence and spanning simultaneously.

The mechanism behind the equivalence is that both halves of the definition of a basis already have matrix translations. Linear independence of a set of columns is exactly the statement that the homogeneous system has only the trivial solution, which is the definition of nonsingularity. Spanning $\mathbb{C}^m$ is exactly the statement that the column space is the whole of $\mathbb{C}^m$ , which is a known property of nonsingular matrices. Putting the two translations side by side gives a criterion that is checked by one row reduction.

What this buys an engineer is a limitless supply of coordinate systems. The standard unit vectors are the obvious basis of $\mathbb{C}^m$ , but they are rarely the natural one for a physical problem: modal coordinates for a vibrating structure, principal axes for a stress tensor, colour primaries for a display, or a joint frame for a manipulator all correspond to some other nonsingular matrix. Every such matrix is licensed as a coordinate system by this theorem, and the coordinates of a vector in it are obtained by solving one linear system.

The second half of the page proves that those coordinates are unique. Spanning guarantees at least one representation of any vector as a linear combination of basis vectors; linear independence guarantees at most one. The combination — exactly one — is what makes a basis usable as an addressing scheme, and it is the foundation on which coordinate vectors, matrix representations of linear transformations and change of basis are all subsequently built.

Definition

Basis

A subset $B$ of a vector space $V$ is a basis of $V$ when it satisfies both of the following conditions.

$B$ is a linearly independent set: the only relation of linear dependence among its vectors is the trivial one, with every scalar zero.
$B$ spans $V$ : $\operatorname{span}(B) = V$ , so every vector of $V$ is some linear combination of vectors from $B$ .

Dropping either condition destroys the usefulness of the set. Without independence the description is redundant; without spanning it is incomplete.

A vector space generally has infinitely many bases. All of them have the same number of vectors, which is what makes dimension well defined.

Nonsingular Matrix

A square matrix $A$ of size $m$ is nonsingular when the homogeneous system $A\vec{x} = \vec{0}$ has only the trivial solution $\vec{x} = \vec{0}$ ; equivalently, when $\mathcal{N}(A) = \{\vec{0}\}$ . A square matrix that is not nonsingular is singular. Non-square matrices are neither.

Columns of a Nonsingular Matrix are a Basis

CNSMB

Let $A$ be a square matrix of size $m$ . Then the set of columns of $A$ is a basis of $\mathbb{C}^m$ if and only if $A$ is nonsingular. Both directions are needed in practice: the forward direction certifies nonsingularity from a known basis, the reverse manufactures bases from known nonsingular matrices.

Vector Representation Relative to a Basis

VRRB

Let $V$ be a vector space with basis $B = \{\vec{v}_1, \vec{v}_2, \ldots, \vec{v}_m\}$ and let $\vec{w}$ be any vector in $V$ . Then there exist scalars $a_1, a_2, \ldots, a_m$ such that $\vec{w} = a_1\vec{v}_1 + a_2\vec{v}_2 + \cdots + a_m\vec{v}_m$ , and those scalars are unique. The ordered tuple $(a_1, a_2, \ldots, a_m)$ is the coordinate description of $\vec{w}$ relative to $B$ .

Uniqueness fails immediately if $B$ is merely a spanning set. It is the independence half of the definition of a basis that does the work.

Concepts

Proving the equivalence in both directions

Suppose first that the columns of the square matrix $A$ form a basis of $\mathbb{C}^m$ . By definition a basis is linearly independent, and a square matrix whose columns are linearly independent is nonsingular — that statement is simply the definition of nonsingularity rewritten in terms of columns, since a relation of linear dependence among the columns is precisely a non-trivial member of $\mathcal{N}(A)$ .

Conversely, suppose $A$ is nonsingular. Then the same translation gives linear independence of the columns immediately. For the spanning half, a nonsingular matrix has column space equal to all of $\mathbb{C}^m$ , because the system $A\vec{x} = \vec{b}$ is solvable for every right-hand side. Saying $\mathcal{C}(A) = \mathbb{C}^m$ is the same as saying the columns of $A$ span $\mathbb{C}^m$ . A linearly independent spanning set is a basis, and the argument closes.

A factory for bases of complex column space

The practical reading of the equivalence is generative rather than diagnostic. Any square matrix that row-reduces to the identity supplies $m$ vectors that are guaranteed to be a basis of $\mathbb{C}^m$ , with no independence check and no spanning check performed separately. This is how one produces bases that are adapted to a problem rather than adapted to the notation.

The standard unit vectors $\vec{e}_1, \vec{e}_2, \ldots, \vec{e}_m$ are then not a special case requiring their own proof: they are the columns of the identity matrix $I_m$ , which is nonsingular, so the theorem delivers them as a corollary. Every other basis of $\mathbb{C}^m$ is obtained from a different nonsingular matrix, and the matrix that converts between two such bases is itself nonsingular.

Where this sits in the equivalence chain

Nonsingularity is the hub of the first half of linear algebra. For a square matrix $A$ of size $n$ the following statements are all equivalent: $A$ is nonsingular; $A$ row-reduces to the identity matrix; $\mathcal{N}(A) = \{\vec{0}\}$ ; the system $A\vec{x} = \vec{b}$ has a unique solution for every choice of $\vec{b}$ ; the columns of $A$ are linearly independent; $A$ is invertible; $\mathcal{C}(A) = \mathbb{C}^n$ ; and — the addition made here — the columns of $A$ form a basis of $\mathbb{C}^n$ .

The value of such a list is that any one entry may be substituted for any other in an argument. A question about spanning becomes a question about a null space; a question about invertibility becomes a question about a pivot count. In engineering practice this is what lets a single numerical test, the rank or condition of a matrix, answer questions that were posed in quite different vocabularies.

Existence of a representation comes from spanning

The first half of the representation theorem is almost a restatement of the definition. If $B$ is a basis of $V$ , then $B$ spans $V$ , so any $\vec{w}$ in $V$ can be written as some linear combination of the vectors of $B$ . There is no content beyond the definition here, and no computation is implied — the theorem asserts existence, not a method for finding the scalars.

In $\mathbb{C}^m$ the method is nonetheless obvious. If $B$ consists of the columns of the nonsingular matrix $A$ , then finding the scalars means solving $A\vec{x} = \vec{w}$ , and nonsingularity guarantees that this system is consistent for every $\vec{w}$ . The abstract existence statement and the concrete linear solve are the same fact seen from two sides.

Uniqueness comes from linear independence

Suppose a vector $\vec{w}$ admits two representations, with scalars $a_1, \ldots, a_m$ and $b_1, \ldots, b_m$ . Subtracting one from the other gives $\vec{0} = (a_1 - b_1)\vec{v}_1 + (a_2 - b_2)\vec{v}_2 + \cdots + (a_m - b_m)\vec{v}_m$ , which is a relation of linear dependence on the basis vectors. Since a basis is linearly independent, every coefficient must vanish, so $a_i = b_i$ for each $i$ and the two representations were the same all along.

This pattern — assume two objects, subtract, obtain a relation of linear dependence, conclude the coefficients are zero — is the standard use of a linear independence hypothesis and recurs throughout the subject. The consequence for a basis is stark: a basis carries exactly enough raw material to build each vector once. Add a vector and representations proliferate; remove one and some vectors become unreachable.

From unique scalars to coordinate vectors

Because the scalars are unique, collecting them into a column produces a well-defined function from $V$ into $\mathbb{C}^m$ , written $\rho_B(\vec{w})$ . This map is the bridge between an abstract vector space and concrete arithmetic: polynomials, matrices and functions all acquire numeric coordinates once a basis is fixed, and every structural question about them becomes a question about column vectors.

The map depends on the basis, and on its ordering. Two engineers who choose different bases for the same space will assign different coordinate vectors to the same physical object, related by multiplication by a nonsingular change-of-basis matrix. Recording which basis is in force is therefore part of recording the data, in the same way that recording units is part of recording a measurement.

Decision path: does a set of $m$ vectors form a basis of $\mathbb{C}^m$ ?

Count the vectorsIf the set does not contain exactly

m

vectors of size

m

, it cannot be a basis of

\mathbb{C}^m

: fewer cannot span, more cannot be independent.

Assemble the square matrix $A$ Place the vectors as the columns of

A

, in a fixed order. The order will become the order of the coordinates.

Row-reduce $A$ Apply Gauss-Jordan elimination. The question is whether the reduced row-echelon form is the identity matrix

I_m

Is $\operatorname{RREF}(A) = I_m$ ?If yes,

A

is nonsingular and the columns are a basis. If no, some column is a linear combination of the others and the set is neither independent nor spanning.

Obtain coordinates on demandFor a target vector

\vec{w}

, solve

A\vec{x} = \vec{w}

. The unique solution is the coordinate vector

\rho_B(\vec{w})

Check conditioning before trusting the coordinatesA basis can be technically valid yet numerically poor. Estimate the condition number of

A

before using its coordinates in floating point.

Equations

Candidate basis assembled as a square matrix

EQ-BNSM-01

A = [{\vec{v}}_{1} ∣ {\vec{v}}_{2} ∣ \dots ∣ {\vec{v}}_{m}] \in M_{m m}

The $m$ candidate vectors of $\mathbb{C}^m$ become the columns of a single square matrix. Every subsequent test is a test on $A$ .

The basis criterion

EQ-BNSM-02

{{\vec{v}}_{1}, {\vec{v}}_{2}, \dots, {\vec{v}}_{m}} is a basis of ℂ^{m} \Leftrightarrow A is nonsingular

The central equivalence. Two properties — independence and spanning — collapse into one matrix property.

Nonsingularity as a row-reduction test

EQ-BNSM-03

Nonsingularity as a row-reduction test

The chain of equivalent formulations actually used in computation. The first is the practical test; the last two are the independence and spanning halves of the basis definition.

Unique representation relative to a basis

EQ-BNSM-04

\vec{w} = a_{1} {\vec{v}}_{1} + a_{2} {\vec{v}}_{2} + \dots + a_{m} {\vec{v}}_{m}, a_{1}, a_{2}, \dots, a_{m} unique

Existence follows from spanning; uniqueness follows from linear independence. Neither half is available from a merely spanning set.

Coordinates as the solution of a linear system

EQ-BNSM-05

A \vec{x} = \vec{w} \Rightarrow ρ_{B} (\vec{w}) = \vec{x} = A^{- 1} \vec{w}

In $\mathbb{C}^m$ the abstract coordinate map is a linear solve. Write $A^{-1}\vec{w}$ to reason about it; compute it by factorisation and substitution, not by forming the inverse.

Uniqueness argument in one line

EQ-BNSM-06

\vec{0} = \sum_{i}^{=} (a_{i} - b_{i}) {\vec{v}}_{i} \Rightarrow a_{i} - b_{i} = 0 (1 \leq i \leq m)

Two representations of the same vector are subtracted to give a relation of linear dependence; independence forces every coefficient to zero.

Standard unit vectors as the columns of the identity

EQ-BNSM-07

I_{m} = [{\vec{e}}_{1} ∣ {\vec{e}}_{2} ∣ \dots ∣ {\vec{e}}_{m}], ρ_{{{\vec{e}}_{i}}} (\vec{w}) = \vec{w}

The identity matrix is nonsingular, so the standard unit vectors are a basis. In that basis a vector is its own coordinate vector, which is why the choice is invisible until another basis is introduced.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Candidate matrix	Square matrix of size $m$ whose columns are the vectors under test	m x m complex matrix
$m$	Size	Common length of the vectors and the number of them required for a basis of $\mathbb{C}^m$	positive integer
${\vec{v}}_{i}$	Basis vector	The $i$ -th column of $A$ ; a member of the candidate basis	element of C^m
$\vec{w}$	Target vector	An arbitrary vector to be expressed in terms of the basis	element of C^m or of V
$a_{i}$	Coordinate	The unique scalar multiplying $\vec{v}_i$ in the representation of $\vec{w}$	complex scalar
$B$	Basis	The ordered set $\{\vec{v}_1, \ldots, \vec{v}_m\}$ , treated as a coordinate system	subset of V
$ρ_{B} (\vec{w})$	Coordinate vector	Column of the unique coordinates of $\vec{w}$ relative to $B$	element of C^m
$N (A)$	Null space	Set of vectors annihilated by $A$ ; trivial exactly when $A$ is nonsingular	subspace of C^m
$C (A)$	Column space	Span of the columns of $A$ ; equals $\mathbb{C}^m$ exactly when $A$ is nonsingular	subspace of C^m
$I_{m}$	Identity matrix	Square matrix with ones on the diagonal; its columns are the standard unit vectors	m x m matrix

Worked Numerical Example

Problem statement

A three-node thermal network has a conductance matrix whose columns describe how a unit temperature at each node distributes heat flux. Confirm that these three flux patterns form a basis of $\mathbb{C}^3$ , then find the unique combination that reproduces a measured flux vector.

Assemble the candidate matrix
The three flux patterns are $\vec{v}_1 = (2, 1, 0)$ , $\vec{v}_2 = (1, 3, 1)$ and $\vec{v}_3 = (0, 1, 2)$ , written as columns. The tridiagonal pattern is typical of a chain of nodes in which only neighbours are coupled.
$A = [\begin{matrix} 2 & 1 & 0 \\ 1 & 3 & 1 \\ 0 & 1 & 2 \end{matrix}]$
Row-reduce to test nonsingularity
Swap rows 1 and 2 so that the leading entry is $1$ , then clear column 1 with $R_2 \to R_2 - 2R_1$ . Continue left to right, clearing each pivot column above and below. The reduction terminates at the identity.
$Row-reduce to test nonsingularity$
Conclude that the columns are a basis
Because $A$ row-reduces to $I_3$ , the matrix is nonsingular. The basis criterion then gives the conclusion in one step: $\{\vec{v}_1, \vec{v}_2, \vec{v}_3\}$ is a basis of $\mathbb{C}^3$ . No separate spanning argument and no separate independence argument are required.
State the target vector
A measurement campaign reports the composite flux vector $\vec{w} = (5, 2, 3)$ . The engineering question is which mixture of the three node patterns produces it, and whether that mixture is the only one.
$\vec{w} = [\begin{matrix} 5 \\ 2 \\ 3 \end{matrix}]$
Solve for the coordinates
Row-reduce the augmented matrix $[\, A \mid \vec{w} \,]$ . Since the coefficient part reduces to $I_3$ , the final column of the reduced array holds the solution directly.
$[\begin{matrix} 2 & 1 & 0 & 5 \\ 1 & 3 & 1 & 2 \\ 0 & 1 & 2 & 3 \end{matrix}] ⟶ [\begin{matrix} 1 & 0 & 0 & 3 \\ 0 & 1 & 0 & - 1 \\ 0 & 0 & 1 & 2 \end{matrix}]$
Verify the linear combination
The coordinates are $a_1 = 3$ , $a_2 = -1$ , $a_3 = 2$ . Substituting back into the original vectors confirms the arithmetic component by component: $3(2) - 1(1) + 2(0) = 5$ , $3(1) - 1(3) + 2(1) = 2$ , $3(0) - 1(1) + 2(2) = 3$ .
$3 [\begin{matrix} 2 \\ 1 \\ 0 \end{matrix}] - 1 [\begin{matrix} 1 \\ 3 \\ 1 \end{matrix}] + 2 [\begin{matrix} 0 \\ 1 \\ 2 \end{matrix}] = [\begin{matrix} 5 \\ 2 \\ 3 \end{matrix}]$
Record the coordinate vector
Collecting the three scalars in the order fixed by the columns of $A$ gives the coordinate vector of $\vec{w}$ relative to $B = \{\vec{v}_1, \vec{v}_2, \vec{v}_3\}$ . Reordering the basis would permute this vector, which is why the ordering must be recorded alongside it.
$ρ_{B} (\vec{w}) = [\begin{matrix} 3 \\ - 1 \\ 2 \end{matrix}]$
Argue uniqueness rather than assume it
The representation is unique because the reduction produced no free columns: every one of the three variable columns is a pivot column, so the system has exactly one solution. Equivalently, a second representation would subtract to give a non-trivial relation of linear dependence among the columns, contradicting nonsingularity.

Result

The three node flux patterns constitute a valid coordinate system for $\mathbb{C}^3$ , and the measured flux decomposes as $3\vec{v}_1 - \vec{v}_2 + 2\vec{v}_3$ and in no other way. The negative coordinate is physically meaningful: node 2's pattern is subtracted, indicating that the measured state is not a simple superposition of positive node excitations. Because the decomposition is unique, these three numbers may be stored, transmitted or differenced as a faithful proxy for the full flux vector.

Applications & Industry Use

Colour science & display engineering

Primaries as a basis of tristimulus space

A display's three primaries define a matrix whose columns are their tristimulus values. That matrix is nonsingular for any usable display, so the primaries form a basis of the three-dimensional colour space and every colour inside the gamut has a unique set of drive values. A near-singular primary matrix — primaries chosen too close together — is exactly the condition that makes the drive values numerically unstable and the gamut vanishingly thin.

Robotics & kinematics

Frame definition and singularity avoidance

A robot frame is specified by three basis vectors assembled into a rotation or affine matrix. Nonsingularity is the condition that the frame is genuinely three-dimensional rather than degenerate, and the unique-representation property is what allows a task-space point to be converted into frame coordinates without ambiguity. Manipulator singularities are precisely configurations where the relevant matrix loses that property.

Structural dynamics

Modal coordinates

The mode shapes of a discretised structure are assembled as the columns of a modal matrix. When that matrix is nonsingular the modes form a basis of the displacement space, so any deflected shape has a unique modal decomposition. Truncating the modal series is then a controlled approximation with a known error, rather than a guess, because the omitted coordinates are well defined.

Communications & signal processing

Signal constellations and transform bases

A block transform is a nonsingular matrix whose columns form a basis of the signal space, so any block of samples has a unique transform-domain representation and the transform is exactly invertible. Compression and equalisation schemes rely on this: coefficients may be quantised or shaped in the transform domain and the original block reconstructed to within quantisation error alone.

Geospatial & surveying

Datum and coordinate frame transformation

Converting between geodetic reference frames uses a nonsingular transformation matrix whose columns are the images of the source axes. Because the matrix is nonsingular, the conversion is invertible and no position is lost or duplicated. Ill-conditioned transformations, arising from nearly parallel axes in a poorly defined local frame, produce coordinates that are formally unique but practically unreliable.

Process & chemical engineering

Component bases for blend specification

When a product is blended from a set of feedstocks, the composition vectors of those feedstocks form the columns of a matrix. If the matrix is nonsingular, every achievable target composition corresponds to exactly one recipe, which makes the blend controller well posed. Singularity signals that two feedstocks are compositionally redundant and the recipe is not determined by the target alone.

Design Considerations

Nonsingularity is a yes-or-no property; conditioning is a matter of degree

A matrix is either nonsingular or not, but that binary answer is silent about usefulness. A basis whose vectors are nearly parallel is still a basis, yet coordinates computed in it amplify measurement error by roughly the condition number of the matrix. Before adopting a basis for numerical work, estimate $\kappa(A)$ and treat a value above about $10^{8}$ in binary64 as a warning that the coordinates carry fewer significant digits than the data.

Prefer an orthonormal basis when one is available

If the basis vectors can be chosen mutually orthogonal and of unit length, the coordinate computation degenerates from a linear solve into a sequence of inner products, and the condition number becomes exactly one. The Gram-Schmidt procedure converts any basis into an orthonormal basis of the same span. The extra cost is paid once; the stability benefit is paid on every subsequent coordinate computation.

Fix and document the ordering

A basis as a set has no order, but a coordinate vector does. Two systems that agree on the basis and disagree on its ordering will exchange coordinate vectors that are permutations of one another, and the mismatch is silent because both are dimensionally valid. Record the ordering in the interface specification, not in a comment.

Do not compute the inverse to obtain coordinates

Writing $\rho_B(\vec{w}) = A^{-1}\vec{w}$ is correct as algebra and wasteful as computation. Forming $A^{-1}$ costs roughly three times a factorisation, consumes an extra dense array and is less accurate than solving directly. Factor once with $L U$ and apply the factors to each right-hand side; the algebraic notation and the executed algorithm need not coincide.

Reuse the factorisation across many vectors

The expensive step is establishing nonsingularity and factoring $A$ , at about $\tfrac{2}{3}m^3$ operations. Converting each additional vector to coordinates then costs only $2m^2$ . Systems that convert thousands of vectors into a fixed basis should hold the factorisation in memory rather than recomputing it, which changes the cost model from cubic per vector to quadratic.

A basis of the wrong size is not a near miss

Fewer than $m$ vectors cannot span $\mathbb{C}^m$ , and more than $m$ cannot be linearly independent in it. Neither situation is repairable by a better choice of vectors. When a candidate set has the wrong cardinality the matrix is not square, the nonsingularity test does not apply, and the correct tools are rank and dimension rather than this criterion.

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, column vectors and the identity matrix, and the convention that a basis is written as an ordered set when coordinates are intended.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Governs the arithmetic in which the nonsingularity test is executed, and therefore the reason that a computed pivot of magnitude $10^{-16}$ cannot be distinguished from an exact zero without a stated tolerance.
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Supplies the routines actually used: `xGETRF` for the factorisation that certifies nonsingularity, `xGETRS` for the coordinate solve, and `xGECON` for the condition estimate that decides whether the coordinates are trustworthy.
`IEC 61966-2-1`	Multimedia systems and equipment — Colour measurement: sRGB	Specifies a concrete nonsingular 3×3 primary matrix and its inverse, a working example of a basis of a three-dimensional space chosen for physical rather than notational convenience.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the matrices and coordinate vectors on this page as semantic markup, so basis notation remains machine-readable and accessible rather than being rendered as images.

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 integer or symbolic bases where the pivot pattern, and hence the basis property itself, must be certified without doubt.	Gives an unambiguous answer to the nonsingularity question, but intermediate fractions can grow rapidly and there is no notion of conditioning to warn about a poorly chosen basis.
IEEE 754 binary64 (double precision)	Default choice for engineering coordinate systems built from measured or designed quantities.	Roughly sixteen significant decimal digits, enough to tolerate condition numbers up to about $10^{8}$ while retaining half the digits; the nonsingularity test becomes a tolerance decision.
IEEE 754 binary32 (single precision)	Graphics pipelines and embedded frame transformations where throughput and memory bandwidth dominate and the basis is known to be well conditioned by construction.	About seven significant digits; adequate for orthonormal rotation frames, marginal for general bases, and unsafe for any matrix with a condition number above roughly $10^{3}$ .
Complex scalars	Bases for signal or modal spaces where phase information is intrinsic, such as complex mode shapes of a damped structure.	Doubles storage and roughly quadruples the arithmetic per operation, but avoids the artificial pairing of real and imaginary parts that a real formulation would impose.
Orthonormal basis stored as a unitary matrix	Repeated coordinate conversion is required and the basis may be chosen freely.	Coordinates come from inner products instead of a solve, the condition number is exactly one and the inverse is the adjoint; the cost is an up-front orthogonalisation and loss of any physical meaning attached to the original vectors.
Factored storage ( $L U$ with pivots) instead of the matrix	The basis is fixed at configuration time and thousands of vectors will be converted at run time.	Amortises the cubic cost across all conversions and preserves accuracy, at the price of holding a factorisation and permutation alongside, and of invalidating the cache whenever the basis changes.
Fixed-point arithmetic	Embedded controllers converting sensor readings into a fixed frame without a floating-point unit.	Deterministic timing and low gate count, but the scaling of each basis vector must be analysed by hand and any near-singular basis will overflow or lose all precision silently.

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 basis test

Deciding nonsingularity of an $m \times m$ matrix costs about $\tfrac{2}{3}m^3$ floating-point operations by $L U$ factorisation with partial pivoting, which is the same work as solving one system. Full Gauss-Jordan reduction to the identity costs roughly $m^3$ , about 50% more, and is preferred only when the reduced form itself is wanted. There is no cheaper general test: any method that certifies a basis must in effect eliminate.

Performing the check by hand

Work strictly left to right, one pivot column at a time, scaling each pivot row to a leading one before clearing the column above and below. The test terminates early and negatively the moment a column has no available non-zero entry at or below the current pivot row — at that point the matrix is singular and there is no need to complete the reduction. For $3 \times 3$ and smaller, evaluating the determinant is faster, but the determinant scales badly and gives no basis for the columns when it is non-zero.

Library behaviour

NumPy exposes numpy.linalg.solve, which factors and substitutes in one call and raises on an exactly singular factor, and numpy.linalg.cond for the conditioning question that matters more in practice. SciPy's scipy.linalg.lu_factor and lu_solve separate the two phases so that a fixed basis can be factored once. SymPy's Matrix.rref and Matrix.inv answer the exact question over the rationals. Avoid numpy.linalg.inv followed by a matrix-vector product; it is slower and less accurate than the direct solve.

Verifying a claimed coordinate vector

Substitute the coordinates back into the original basis vectors and compare against the target componentwise, rather than checking the reduced system. An arithmetic slip during elimination produces a reduced array that is internally consistent but no longer equivalent to the input, so only a residual computed against the original data detects it. In floating point, report the residual norm relative to the norm of the target instead of testing for exact equality.

Detecting a near-singular basis before it does damage

Compute a condition estimate at the time the basis is established, not at the time coordinates are first used. A reciprocal condition estimate below the machine epsilon means the factorisation has effectively no correct digits; a value near $10^{-8}$ in binary64 means about half the digits are lost. Recording the estimate alongside the basis lets downstream code decide whether to fall back to an orthogonalised basis or refuse the conversion.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Checking only linear independence for a set of the wrong size	high	A set of fewer than $m$ independent vectors is declared a basis of $\mathbb{C}^m$ because the independence test passed.	Confirm the cardinality is exactly $m$ before applying the criterion; independence alone certifies a basis only for the span of the set, not for $\mathbb{C}^m$ .
Applying the criterion to a non-square matrix	high	Nonsingularity is defined only for square matrices, but the word is used loosely for any full-rank matrix.	For rectangular matrices use rank and dimension: the columns are independent when the rank equals the column count, and they span when the rank equals the row count.
Exact-zero pivot test in floating point	high	Nonsingularity decided by comparing a computed pivot against zero, so rounding promotes a structurally dependent column to independent.	Use a condition estimate or a rank-revealing factorisation with an explicit tolerance; treat the binary nonsingularity question as a continuous conditioning question.
Assuming uniqueness for a spanning set that is not a basis	high	A redundant set of more than $m$ vectors is used as a coordinate system, so representations exist but are not unique.	Reduce the spanning set to a basis first; the presence of any free column in the reduction is the signal that coordinates will not be unique.
Silent basis reordering across an interface	medium	Two modules agree on the basis vectors but store them in different orders, so coordinate vectors are permuted without any dimensional error.	Treat the basis as an ordered tuple in the data contract and validate the ordering, for example by transmitting a checksum of the basis matrix alongside the coordinates.
Adopting a technically valid but ill-conditioned basis	medium	Basis vectors chosen from physically similar configurations are nearly parallel, so the matrix is nonsingular but badly conditioned.	Estimate the condition number when the basis is chosen and prefer an orthogonalised or better-separated set; document the achievable accuracy of the coordinates.
Forming the explicit inverse for repeated conversions	medium	$A^{-1}$ computed once and multiplied against every vector, on the assumption that this is the efficient route.	Store an $L U$ factorisation instead. It is cheaper to compute, cheaper to apply for a single right-hand side and numerically superior.
Confusing the coordinate vector with the vector itself	medium	In the standard basis the two coincide, so code written against that special case breaks when a non-standard basis is introduced.	Type or name the two quantities distinctly and require the basis identifier to accompany any coordinate vector that leaves the module.
Assuming a nonsingular matrix over the rationals stays nonsingular modulo a prime	low	A modular check is used to certify a basis, and the chosen prime happens to divide the determinant.	Certify with several primes, or verify the determinant is non-zero modulo the prime before trusting the modular result.

FAQs

Why is it enough to check nonsingularity rather than checking independence and spanning separately?

Because both properties have the same matrix translation. Linear independence of the columns is precisely the statement that the null space is trivial, which is the definition of nonsingularity; spanning $\mathbb{C}^m$ is precisely the statement that the column space is all of $\mathbb{C}^m$ , which is a theorem about nonsingular matrices. One row reduction therefore settles both.

Does this criterion work for a set of vectors that is not the right size?

No, and the failure is not marginal. A set of fewer than $m$ vectors cannot span $\mathbb{C}^m$ and a set of more than $m$ cannot be linearly independent in it, so only a set of exactly $m$ vectors can be a basis. If the count is wrong the assembled matrix is not square and nonsingularity is undefined.

Are the standard unit vectors a special basis?

They are special only in being convenient. They are the columns of the identity matrix, which is nonsingular, so they qualify as a basis by exactly the same criterion as any other. Their distinguishing feature is that the coordinate vector of any vector relative to them is the vector itself, which is why the choice of basis is invisible until a second basis appears.

What actually fails if the set is only a spanning set and not a basis?

Existence of a representation survives, but uniqueness is lost. A redundant spanning set admits a non-trivial relation of linear dependence, and adding any multiple of that relation to a valid representation produces another valid representation. Coordinates then cease to be well defined, and every construction that depends on them — coordinate vectors, matrix representations, change of basis — collapses.

How do I compute coordinates relative to a basis in practice?

Assemble the basis vectors as the columns of $A$ and solve $A\vec{x} = \vec{w}$ by $L U$ factorisation with partial pivoting. If many vectors will be converted, factor once and reuse the factors, which reduces the marginal cost per vector from cubic to quadratic. Only form $A^{-1}$ if the inverse itself is the deliverable.

Can a basis be numerically bad even though it is mathematically valid?

Yes, and this is the usual practical failure. Nonsingularity is a binary property, but the accuracy of coordinates computed in a basis degrades in proportion to the condition number of the basis matrix. Vectors that are nearly parallel give a nonsingular matrix with a huge condition number, so the coordinates are exact in principle and meaningless in floating point.

Why does the uniqueness proof subtract two representations?

Subtracting turns the assumption of two representations into a relation of linear dependence on the basis vectors, with coefficients equal to the differences of the two coordinate sets. Linear independence forces every such coefficient to be zero, so the two coordinate sets agree. This manoeuvre — obtain a relation of linear dependence, then conclude the scalars vanish — is the standard way any independence hypothesis is used.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section B, subsections BNSM and VR. Licensed under the GNU Free Documentation License v1.2.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
IEEE 754-2019, IEEE Standard for Floating-Point Arithmetic. Institute of Electrical and Electronics Engineers.
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.
IEC 61966-2-1:1999, Multimedia systems and equipment — Colour measurement and management — Part 2-1: Default RGB colour space (sRGB).

AI Suggested Questions

Construct a 4x4 integer matrix that is nonsingular but has a condition number above 10^6, and show how badly the coordinates of a unit vector degrade in binary64.
Explain why the columns of a nonsingular matrix span the whole of C^m without appealing to the existence of an inverse.
Given two bases of C^3, derive the change-of-basis matrix between them and verify it is nonsingular.
Show what goes wrong with the uniqueness proof if the set is spanning but linearly dependent, using an explicit three-vector example in C^2.
Compare the cost of computing coordinates for 10000 vectors by LU reuse against forming the inverse once, for a 200x200 basis matrix.
How does the Gram-Schmidt procedure convert an ill-conditioned basis of C^m into an orthonormal one, and what happens to the condition number?

Related Calculators

Basis Checker

Test whether a set of $m$ vectors forms a basis of $\mathbb{C}^m$ by assembling the square matrix and reporting its reduced row-echelon form.

Coordinate Vector Calculator

Compute the unique coordinates $\rho_B(\vec{w})$ of a target vector relative to a supplied basis, with a residual check against the original data.

Matrix Condition Number Calculator

Estimate the conditioning of a basis matrix and report how many significant digits of the coordinates survive.

Change-of-Basis Calculator

Build the nonsingular matrix that converts coordinates from one basis of $\mathbb{C}^m$ into another and apply it to a list of vectors.