Vector Representation

Q: Does the coordinate vector change if I reorder the basis?

Yes. The $i$th entry is by definition the weight on the $i$th basis vector, so permuting the basis permutes the entries. The vector itself is unaffected. This is why a basis for representation purposes must be treated as an ordered list, and why coordinate exchange formats always fix a component order.

Overview

An abstract vector space offers no arithmetic that a computer can execute. Polynomials, matrices, solution spaces of differential equations and spaces of piecewise functions all satisfy the same ten axioms, but none of them arrives with a list of numbers attached. Vector representation supplies that list. Choose an ordered basis $B = \{\vec{v}_1, \vec{v}_2, \dots, \vec{v}_n\}$ of a space $V$ ; every vector $\vec{w} \in V$ can be written as a linear combination of the basis vectors in exactly one way, and the column of coefficients is defined to be $\rho_B(\vec{w})$ .

Two claims have to be established before the definition is legitimate. The first is that the coefficients exist and are unique, which is exactly the defining property of a basis: spanning gives existence and linear independence gives uniqueness. Without both halves the recipe would either fail on some vectors or produce several answers for the same vector, and $\rho_B$ would not be a function at all. The second is that the resulting map respects the vector space operations, which follows from comparing the unique expansions of $\vec{x} + \vec{y}$ and $\alpha\vec{x}$ against those of $\vec{x}$ and $\vec{y}$ .

With those in place, $\rho_B$ turns out to be as well behaved as a map can be. Its kernel is trivial, because a vector whose coefficients all vanish is the zero vector, so it is injective. It is surjective, because any column of scalars can be used as weights to build a vector in $V$ that represents it. Injective and surjective together make $\rho_B$ an invertible linear transformation, and the inverse $\rho_B^{-1}$ reassembles a vector from its coefficients.

The engineering consequence is that any finite-dimensional vector space can be worked in as if it were a space of column vectors, provided the basis is recorded. Modal coordinates in structural dynamics, coefficient vectors in finite element analysis, transform coefficients in signal processing and orbital expansion coefficients in computational chemistry are all instances of $\rho_B$ applied to a domain-specific basis. The coefficients are meaningless without the basis, which is why every one of those fields treats the basis specification as part of the data.

Definition

Let $V$ be a vector space with ordered basis $B = \{\vec{v}_1, \vec{v}_2, \dots, \vec{v}_n\}$ . For $\vec{w} \in V$ , take the unique scalars $a_1, a_2, \dots, a_n$ with
$\vec{w} = a_1\vec{v}_1 + a_2\vec{v}_2 + \cdots + a_n\vec{v}_n$
and define the function $\rho_B \colon V \to \mathbb{C}^n$ entry by entry through $\left[ \rho_B(\vec{w}) \right]_i = a_i$ for $1 \le i \le n$ .

The subscript is not decoration. A different basis, or the same basis in a different order, defines a different function.

Coordinate Vector

The column $\rho_B(\vec{w}) \in \mathbb{C}^n$ is called the coordinate vector of $\vec{w}$ relative to $B$ . Its $i$ th entry is the weight applied to the $i$ th basis vector. Applying $\rho_B$ is referred to as coordinatising and applying $\rho_B^{-1}$ as un-coordinatising.

A coordinate vector is a description of a vector, not the vector itself. The vector $\vec{w}$ does not change when the basis does; only its description does.

Unique Representation Property

If $B$ is a basis of $V$ then every $\vec{w} \in V$ has exactly one expansion as a linear combination of the vectors of $B$ . Spanning supplies at least one expansion; linear independence forbids a second, since the difference of two expansions would be a non-trivial relation of linear dependence. This property is what makes the definition of $\rho_B$ well posed.

Concepts

Why the definition needs a basis and not merely a spanning set

The phrase find scalars in the definition looks permissive, and it would be fatal if the set $B$ were only a spanning set. Suppose $B$ spans but is dependent: then some vector has two different expansions, and the recipe returns two different columns for one input, so no function has been defined. Suppose instead that $B$ is independent but does not span: then some vector has no expansion, and the recipe is undefined on part of $V$ .

A basis is precisely the condition that excludes both defects. This is worth internalising because it is the same pair of failure modes seen elsewhere in the subject: dependence causes multiply-defined answers, failure to span causes undefined ones, and it is exactly the pair that injectivity and surjectivity exclude for an inverse map. The unique representation property is the vector-space statement of that principle.

The coordinate map is linear

Let $\vec{x}$ and $\vec{y}$ have expansions with coefficients $a_i$ and $b_i$ . Adding the two expansions and collecting terms on each basis vector produces an expansion of $\vec{x} + \vec{y}$ with coefficients $a_i + b_i$ . Uniqueness then says those are the coefficients of $\vec{x} + \vec{y}$ , so $\rho_B(\vec{x} + \vec{y}) = \rho_B(\vec{x}) + \rho_B(\vec{y})$ . The same argument with a scalar multiple gives $\rho_B(\alpha\vec{x}) = \alpha\rho_B(\vec{x})$ .

Note where the work happens. The collecting step uses commutativity, associativity and distributivity in $V$ ; the conclusion uses uniqueness. Linearity of $\rho_B$ is therefore not an assumption about coordinates but a consequence of the axioms of $V$ combined with the defining property of a basis.

Injectivity and surjectivity

For injectivity, compute the kernel. If $\rho_B(\vec{u}) = \vec{0}$ then every coefficient in the expansion of $\vec{u}$ is zero, so $\vec{u}$ is a sum of zero multiples of the basis vectors, which is the zero vector of $V$ . The kernel is trivial and $\rho_B$ is injective.

For surjectivity, take any column $\vec{v} \in \mathbb{C}^n$ and use its entries as weights: define $\vec{u} = [\vec{v}]_1\vec{v}_1 + \cdots + [\vec{v}]_n\vec{v}_n$ . By construction the unique expansion of $\vec{u}$ has exactly those coefficients, so $\rho_B(\vec{u}) = \vec{v}$ . Every column is hit, and the range is all of $\mathbb{C}^n$ .

Coordinatising is an invertible linear transformation

A linear transformation that is both injective and surjective is invertible, so $\rho_B$ has an inverse and that inverse is itself a linear transformation. The inverse has an explicit and cheap formula: $\rho_B^{-1}$ applied to a column simply forms the linear combination of the basis vectors weighted by its entries. Un-coordinatising is therefore always easier than coordinatising, which in general requires solving a system.

The asymmetry is worth remembering in implementation. Building a vector from coefficients costs one pass over the basis; extracting coefficients from a vector costs a solve, unless the basis has been chosen to make it cheap. Orthonormal bases exist precisely to remove that cost, replacing the solve by a sequence of inner products.

The representation depends on the ordering, not just the set

A basis for the purposes of representation is an ordered list. Permuting the list permutes the entries of every coordinate vector, so two engineers using the same set of basis vectors in different orders will exchange numerically incompatible columns while both being correct. This is not a subtlety to be tidied away; it is why coordinate conventions are published in standards and why file formats for modal, spectral and geometric data always fix a component order.

The same point applies to scaling. Replacing a basis vector by a non-zero multiple of itself is still a basis, but every coordinate vector changes accordingly. A basis specification is therefore a list of vectors in a fixed order with fixed normalisation, and anything less leaves the coefficients ambiguous.

Computing the representation in a concrete space

When $V$ is $\mathbb{C}^n$ itself, coordinatising relative to a basis $B$ means solving $M\vec{a} = \vec{w}$ , where $M$ is the matrix whose columns are the basis vectors. Since $B$ is a basis, $M$ is nonsingular and the solution is unique, giving $\rho_B(\vec{w}) = M^{-1}\vec{w}$ as a statement of uniqueness. In practice one factorises $M$ once and solves for each vector to be coordinatised.

For an abstract space the same idea applies after one preliminary step: express both the basis vectors and the target in terms of a convenient reference basis, such as the monomials for polynomials or the single-entry matrices for $M_{mn}$ . Coordinatising relative to an awkward basis then becomes a numerical system in the reference coordinates, which is the standard route to any representation in practice.

Procedure: computing $\rho_B(\vec{w})$

Fix the ordered basisRecord

B = \{\vec{v}_1, \dots, \vec{v}_n\}

as a list, with the ordering and normalisation stated explicitly.

Confirm $B$ is a basisCheck independence and spanning, or check that the count matches

\dim(V)

and one of the two holds. A dependent set makes the coefficients ambiguous.

Write the defining equationSet

\vec{w} = a_1\vec{v}_1 + \cdots + a_n\vec{v}_n

with the

a_i

unknown.

Reduce to a numerical systemExpress everything in a convenient reference basis and equate components, producing

n

equations in the

n

unknown scalars.

SolveThe coefficient matrix is nonsingular because

B

is a basis, so the solution exists and is unique.

Verify by un-coordinatisingRebuild

a_1\vec{v}_1 + \cdots + a_n\vec{v}_n

and confirm it equals

\vec{w}

. This one check catches almost every arithmetic error.

Equations

Unique expansion relative to a basis

EQ-VR-01

\vec{w} = a_{1} {\vec{v}}_{1} + a_{2} {\vec{v}}_{2} + \dots + a_{n} {\vec{v}}_{n}

Existence of the scalars comes from spanning, uniqueness from linear independence. Both are needed for the coordinate map to be a function.

Definition of the coordinate map

EQ-VR-02

ρ_{B} : V \to ℂ^{n}, [ρ_{B} (\vec{w})]_{i} = a_{i}, 1 \leq i \leq n

The $i$ th entry of the coordinate vector is the weight applied to the $i$ th basis vector, which is why the ordering of $B$ is part of the definition.

Linearity of the coordinate map

EQ-VR-03

ρ_{B} (\vec{x} + \vec{y}) = ρ_{B} (\vec{x}) + ρ_{B} (\vec{y}), ρ_{B} (α \vec{x}) = α ρ_{B} (\vec{x})

Proved by comparing unique expansions. It is what allows any computation built from the two operations to be carried out on coordinates instead.

Trivial kernel

EQ-VR-04

K (ρ_{B}) = {\vec{0}} \Rightarrow ρ_{B} injective

A vector all of whose coefficients vanish is the zero vector, so no two distinct vectors share a coordinate column.

Un-coordinatising

EQ-VR-05

Un-coordinatising

The inverse map is an explicit linear combination, so rebuilding a vector from its coordinates is always cheaper than extracting them.

Coordinatising inside complex n-space

EQ-VR-06

Coordinatising inside complex n-space

When $V = \mathbb{C}^n$ , the coordinate vector is the solution of a nonsingular linear system whose coefficient matrix holds the basis vectors as columns.

Coordinatisation is an isomorphism

EQ-VR-07

ρ_{B} invertible \Rightarrow V ≅ ℂ^{n}, n = \dim (V)

Injective and surjective together make $\rho_B$ an invertible linear transformation, which identifies $V$ structurally with a space of columns.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$V$	Vector space	The abstract space in which the vectors of interest live	vector space of dimension n
$B$	Ordered basis	The list of basis vectors, fixed in order and normalisation, that defines the representation	ordered set of n vectors
${\vec{v}}_{i}$	Basis vector	The $i$ th member of the ordered basis	vector in V
$\vec{w}$	Vector being represented	The element of $V$ whose coefficients are sought	vector in V
$a_{i}$	Coordinate	The unique weight applied to $\vec{v}_i$ in the expansion of $\vec{w}$	complex number
$ρ_{B}$	Vector representation map	The invertible linear transformation sending a vector to its coordinate column	linear map V to C^n
$Un-coordinatisation$	Un-coordinatisation	Rebuilds a vector from a column of coefficients by forming a linear combination	linear map C^n to V
$M$	Basis matrix	Matrix whose columns are the basis vectors, used when $V$ is a space of columns	nonsingular n x n matrix
$n$	Dimension	Number of basis vectors, and the height of every coordinate column	positive integer

Worked Numerical Example

Problem statement

A profile fitting routine stores quadratic shapes internally as coefficients relative to the non-standard basis $B = \{1 + x, \ x + x^2, \ 1 + x^2\}$ of $P_2$ . Verify that $B$ is a basis, coordinatise two profiles, confirm that the coordinate map is linear, show that reordering the basis changes the stored column, and un-coordinatise a stored column back to a profile.

Confirm the set is a basis
Express each candidate basis vector in the reference basis $\{1, x, x^2\}$ and assemble the coefficient columns into a matrix. Its determinant is $1(1 \cdot 1 - 0 \cdot 1) - 0 + 1(1 \cdot 1 - 1 \cdot 0) = 2$ , which is non-zero, so the three polynomials are independent. Three independent vectors in a space of dimension three form a basis.
$M = [\begin{matrix} 1 & 0 & 1 \\ 1 & 1 & 0 \\ 0 & 1 & 1 \end{matrix}], \det (M) = 2 \neq 0$
Set up the defining equation for the first profile
Take $u = 6 + 9x + 5x^2$ and seek scalars $a_1, a_2, a_3$ with $u = a_1(1 + x) + a_2(x + x^2) + a_3(1 + x^2)$ . Expanding the right-hand side and collecting powers of $x$ gives the coefficients as sums of pairs.
$a_{1} + a_{3} = 6, a_{1} + a_{2} = 9, a_{2} + a_{3} = 5$
Solve and record the coordinate vector
Adding all three equations gives $2(a_1 + a_2 + a_3) = 20$ , so the total is $10$ . Subtracting each equation from that total isolates the missing unknown: $a_2 = 10 - 6 = 4$ , $a_3 = 10 - 9 = 1$ and $a_1 = 10 - 5 = 5$ .
$ρ_{B} (6 + 9 x + 5 x^{2}) = [\begin{matrix} 5 \\ 4 \\ 1 \end{matrix}]$
Verify by rebuilding the profile
Un-coordinatise the answer: $5(1 + x) + 4(x + x^2) + 1(1 + x^2)$ has constant term $5 + 1 = 6$ , linear term $5 + 4 = 9$ and quadratic term $4 + 1 = 5$ . The original profile is recovered exactly, confirming both the arithmetic and the claim that $\rho_B^{-1}$ is a plain linear combination.
Coordinatise a second profile
For $v = 2 + 3x + x^2$ the same three equations become $a_1 + a_3 = 2$ , $a_1 + a_2 = 3$ and $a_2 + a_3 = 1$ , whose total is $3$ . This gives $a_1 = 2$ , $a_2 = 1$ and $a_3 = 0$ , and rebuilding confirms $2(1 + x) + 1(x + x^2) = 2 + 3x + x^2$ .
$ρ_{B} (2 + 3 x + x^{2}) = [\begin{matrix} 2 \\ 1 \\ 0 \end{matrix}]$
Check linearity on the sum
The sum of the two profiles is $u + v = 8 + 12x + 6x^2$ . Coordinatising it directly gives the system $a_1 + a_3 = 8$ , $a_1 + a_2 = 12$ , $a_2 + a_3 = 6$ with total $13$ , hence the column with entries $7$ , $5$ and $1$ — exactly the sum of the two coordinate vectors already computed.
$ρ_{B} (u + v) = [\begin{matrix} 7 \\ 5 \\ 1 \end{matrix}] = [\begin{matrix} 5 \\ 4 \\ 1 \end{matrix}] + [\begin{matrix} 2 \\ 1 \\ 0 \end{matrix}]$
Show that the ordering matters
Relative to the reference basis $D = \{1, \ x, \ x^2\}$ the first profile needs no computation at all: its coordinate column is just its coefficient list. Reversing that basis to $E = \{x^2, \ x, \ 1\}$ leaves the same three numbers but in a different order, so the stored column differs even though the polynomial has not changed.
$ρ_{D} (u) = [\begin{matrix} 6 \\ 9 \\ 5 \end{matrix}], ρ_{E} (u) = [\begin{matrix} 5 \\ 9 \\ 6 \end{matrix}]$
Un-coordinatise an arbitrary stored column
A stored record holding the column with entries $-1$ , $2$ and $3$ is decoded by forming the corresponding combination: $-1(1 + x) + 2(x + x^2) + 3(1 + x^2)$ , whose constant term is $-1 + 3 = 2$ , linear term $-1 + 2 = 1$ and quadratic term $2 + 3 = 5$ .
$Un-coordinatise an arbitrary stored column$

Result

Three different columns — the entries $5, 4, 1$ and $6, 9, 5$ and $5, 9, 6$ — all describe the same polynomial. A stored coefficient record is therefore meaningless without the basis that produced it, and any file format or interface that transmits coordinates must transmit the ordered basis alongside them. The linearity check also shows the practical payoff: once profiles are stored as columns, adding and scaling them requires no polynomial arithmetic at all.

Applications & Industry Use

Structural dynamics

Modal coordinates

Expressing a displacement field in the basis of mode shapes replaces a coupled system of equations of motion with a set of independent single-degree-of-freedom equations. The modal amplitude vector is precisely $\rho_B$ applied to the displacement, with $B$ the ordered set of mode shapes, and truncating the expansion to the first few modes is a deliberate approximation made in coordinate space.

Finite element analysis

Nodal degrees of freedom as coefficients

Within an element, the displacement field is a linear combination of shape functions, and the nodal degree-of-freedom vector is its coordinate vector relative to the shape function basis. Assembly, constraint application and post-processing all operate on those coefficients, and the mapping back to a continuous field is exactly un-coordinatisation.

Signal processing

Transform coefficients

A discrete Fourier, cosine or wavelet transform is coordinatisation relative to a specific ordered basis of the signal space. The forward transform is $\rho_B$ and the inverse transform is $\rho_B^{-1}$ . The choice of basis is the entire engineering decision: compression works because the coefficient vector is sparse in a well-chosen basis and dense in a poor one.

Computational chemistry

Basis set expansion of orbitals

A molecular orbital is represented by its coefficients relative to a chosen set of atom-centred functions. Published results quote the basis set by name because the coefficients are meaningless without it, and comparing two calculations requires either the same basis or an explicit change of representation.

Geospatial information

Coordinates relative to a reference frame

A position is reported as a triple of numbers only after a datum, axis order and unit convention have been fixed. That specification is the ordered basis, and the widespread failure to transmit it — latitude and longitude exchanged, or the wrong datum assumed — is the practical form of the ordering hazard described on this page.

Computer graphics

Local, object and world frames

The same geometric point carries different coordinate triples in model space, world space and camera space, each corresponding to a different ordered basis. Rendering pipelines are chains of coordinatisations and un-coordinatisations, and the correctness of the pipeline depends on tracking which frame each stored triple belongs to.

Design Considerations

Transmit the basis with the coefficients

A coordinate vector is a description relative to a convention. Storing coefficients without the ordered basis that produced them creates data that cannot be interpreted later and cannot be exchanged safely. Treat the basis specification — vectors, order and normalisation — as part of the record, not as documentation.

Choose the basis for conditioning, not just for convenience

All bases of an $n$ -dimensional space are equivalent in theory; they are not equivalent numerically. Coordinatising a polynomial relative to the monomials on a wide interval involves a basis matrix whose condition number grows explosively with degree, so small changes in the function cause large changes in the coefficients. Orthogonal polynomial bases exist to avoid exactly this.

Exploit orthonormality when it is available

For a general basis, coordinatising requires solving an $n \times n$ system. For an orthonormal basis under an available inner product, each coefficient is a single inner product, so the whole coordinate vector costs $O(n^2)$ with no factorisation and no conditioning penalty. When an inner product exists, orthonormalising the basis once usually repays itself immediately.

Factor the basis matrix once

When many vectors are to be coordinatised relative to the same basis, the basis matrix does not change. Factorise it once and reuse the factors, reducing each subsequent coordinatisation from $O(n^3)$ to $O(n^2)$ . Re-solving from scratch for each vector is the most common inefficiency in coordinate-handling code.

Keep the vector and its representation distinct in the model

The polynomial does not change when the basis does; only its column changes. Code and documentation that conflate the two produce bugs that are hard to trace, because two columns that disagree may describe the same object. Name variables to indicate which basis a stored column belongs to, and convert explicitly rather than implicitly.

Validate the basis before trusting any coefficients

If the supposed basis is dependent, the coordinatisation system is singular and the coefficients are either non-unique or non-existent. Check independence — by a determinant, a rank computation or a factorisation that reports singularity — before accepting a set as a basis, particularly when it has been generated numerically rather than constructed.

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	Establishes the conventions for indexed families, component notation and the bracketed entry form $\left[ \vec{v} \right]_i$ used to state the definition of the coordinate map.
`ISO 19111`	Geographic information — Referencing by coordinates	The clearest standardised statement of the principle on this page: coordinates are meaningless without the coordinate reference system that defines them, and axis order and units are part of that definition.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Sets the precision within which the coordinatisation system is solved, and therefore the accuracy of a round trip through $\rho_B$ and $\rho_B^{-1}$ for an ill-conditioned basis.
`ISO 10303-42`	Product data representation and exchange: Geometric and topological representation	Fixes the basis conventions for spline and polynomial geometry in exchange files, so that a coefficient list transferred between systems is decoded against the basis it was encoded with.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Provides the semantic encoding for subscripted maps and column vectors on this page, keeping the basis subscript on $\rho_B$ machine-readable rather than typographic decoration.

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
Orthonormal basis with binary64 coefficients	The space carries an inner product and repeated coordinatisation is required, as in modal or spectral work.	Coefficients come from inner products at $O(n^2)$ cost with perfect conditioning, but the orthonormal basis must be constructed first and may be dense where the original basis was sparse.
General basis with a stored LU factorisation	The basis is fixed by the application and many vectors must be coordinatised against it.	Amortises the $O(n^3)$ factorisation over all subsequent solves, but the factors must be kept alongside the basis and invalidated whenever the basis changes.
Exact rational coefficients	Small symbolic spaces where the coordinate vector is the answer, as in verifying that a proposed set is a basis.	Gives an unambiguous result and detects a singular basis matrix decisively, but is impractical once dimensions reach the hundreds.
Orthogonal polynomial basis	Representing functions on an interval where the monomial basis would be ill-conditioned.	Condition numbers stay modest even at high degree and coefficients decay meaningfully, at the cost of extra machinery for evaluation and for conversion to the monomial form other tools expect.
Sparse coefficient storage	A basis chosen so that typical vectors have few significant coefficients, as in wavelet compression.	Storage proportional to the significant coefficients only, but the basis must be transmitted or agreed in advance and thresholding introduces a controlled approximation.
Fixed-point coefficients	Embedded systems storing modal or transform coefficients under tight memory and power budgets.	Compact and fast without a floating-point unit, but the coefficient dynamic range must be analysed against the chosen basis in advance, since a poorly scaled basis will saturate.

Manufacturing Notes

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

Reducing an abstract representation to a numerical system

Choose a convenient reference basis of the abstract space — monomials for polynomials, single-entry matrices for $M_{mn}$ — and coordinatise both the target and each vector of the working basis against it. Assemble the working basis columns into a matrix and solve against the target column. This two-stage route turns any abstract coordinatisation into a standard linear solve.

Cost of the two directions

Un-coordinatising is one linear combination: $n$ scalar multiples and $n-1$ additions of vectors, or $O(n^2)$ scalar operations when the vectors are themselves columns of length $n$ . Coordinatising against a general basis requires a solve at $O(n^3)$ for the first vector and $O(n^2)$ for each subsequent one if the factorisation is retained. Against an orthonormal basis, coordinatising drops to $O(n^2)$ with no factorisation.

Verification by round trip

The cheapest and most effective check is to un-coordinatise the computed column and compare with the original vector. It exercises the whole computation, requires no second method, and in exact arithmetic must reproduce the input exactly. In floating point, compare the residual against a tolerance scaled by the norm of the input and the condition number of the basis matrix.

Library support

There is no single library function named for this operation, because it is a solve. In NumPy the coordinate vector is numpy.linalg.solve(M, w) with $M$ holding the basis as columns, and scipy.linalg.lu_factor with lu_solve gives the factor-once pattern. SymPy provides exact coordinatisation for symbolic bases, and for orthonormal bases the coefficients are simply a matrix-vector product with the conjugate transpose.

Detecting a bad basis at run time

A basis matrix that is singular or nearly so signals that the supposed basis is dependent or badly scaled. Rather than testing the determinant, which is not scale-aware, obtain a reciprocal condition estimate from the factorisation and reject the basis when it falls below a documented threshold. This catches numerically generated bases that have drifted towards dependence.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Storing coefficients without the basis	high	A coordinate vector is written to a file or passed across an interface with no record of the ordered basis used.	Make the basis specification part of the data structure and the file format. Reject any incoming coefficient record that does not identify its basis.
Treating the basis as an unordered set	high	Two implementations use the same basis vectors in different orders and exchange columns that appear compatible.	Define the basis as an ordered list, and validate an interface with a test vector whose coefficients are all distinct so a permutation is immediately visible.
Using a spanning set that is dependent	high	A generating set is mistaken for a basis, so the coordinatisation system is singular and the coefficients are not unique.	Verify independence before use. Numerically, obtain a rank or reciprocal condition estimate from a factorisation of the basis matrix.
Ill-conditioned basis at high degree	high	Coordinatising relative to monomials on a wide interval, where the basis matrix condition number grows very rapidly with dimension.	Use an orthogonal polynomial basis, or map the interval to a standard one before representation, and monitor the condition number of the basis matrix.
Comparing coordinate vectors from different bases	medium	Two columns describing the same vector under different bases are compared entry by entry and judged different.	Convert to a common basis before comparing, or compare the un-coordinatised vectors directly. Tag every stored column with its basis identifier.
Assuming coordinates respect operations the space does not have	medium	Multiplying coordinate vectors entry by entry and expecting the result to represent a product of the original vectors.	The coordinate map preserves addition and scalar multiplication only. Any other operation must be represented separately and explicitly.
Rescaling a basis vector silently	medium	Normalising the basis after coefficients have been computed, so old and new columns disagree by scale factors.	Treat normalisation as part of the basis definition. Recompute or rescale all stored coefficients whenever the basis is changed in any way.
Assuming a standard basis makes the step free	low	Working in $\mathbb{C}^n$ and forgetting that the coordinate column coincides with the vector only for the standard basis in its usual order.	State the basis even when it is the standard one, so that a later change of basis does not silently invalidate stored results.

FAQs

Why does the definition require a basis rather than just a spanning set?

A spanning set guarantees that some expansion exists but not that it is unique, so the recipe could return several different columns for one vector and would not define a function. Linear independence removes the ambiguity. Conversely an independent set that fails to span leaves the map undefined on part of the space. Only a basis supplies both existence and uniqueness.

Does the coordinate vector change if I reorder the basis?

Yes. The $i$ th entry is by definition the weight on the $i$ th basis vector, so permuting the basis permutes the entries. The vector itself is unaffected. This is why a basis for representation purposes must be treated as an ordered list, and why coordinate exchange formats always fix a component order.

Why is un-coordinatising cheaper than coordinatising?

Rebuilding a vector from coefficients is a single linear combination, evaluated directly. Extracting coefficients requires solving a system whose coefficient matrix holds the basis. The asymmetry disappears for an orthonormal basis, where each coefficient is an inner product, which is one of the main reasons orthonormal bases are constructed.

Is the coordinate map the same thing as an isomorphism with

\mathbb{C}^n

Yes. Because $\rho_B$ is a linear transformation that is both injective and surjective, it is invertible, and the existence of an invertible linear transformation is exactly what makes two vector spaces isomorphic. Every choice of ordered basis therefore produces a different isomorphism between $V$ and $\mathbb{C}^n$ .

Can two different vectors have the same coordinate vector?

Not relative to the same basis. The coordinate map is injective because a vector whose coefficients all vanish must be the zero vector, so distinct vectors give distinct columns. Two different bases can of course assign the same column to different vectors, which is why comparisons must always be made within one basis.

How do I coordinatise a vector in an abstract space where there is no obvious arithmetic?

Introduce a convenient reference basis for the space, such as monomials for polynomials or single-entry matrices for a matrix space. Coordinatise both the target and each working basis vector against the reference basis, then solve the resulting numerical system. This reduces any abstract representation to a standard linear solve.

What goes wrong when the basis is badly conditioned?

The coordinatisation system becomes sensitive: small perturbations in the vector produce large changes in the coefficients, and a round trip through $\rho_B$ and $\rho_B^{-1}$ loses accuracy. The monomial basis on a wide interval is the standard example. The remedy is a better-conditioned basis, typically an orthogonal or orthonormal one, rather than higher precision.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section VR. Licensed under the GNU Free Documentation License v1.2.
Halmos, P. R. Finite-Dimensional Vector Spaces, 2nd edition. Springer, 1974.
Trefethen, L. N. Approximation Theory and Approximation Practice. Society for Industrial and Applied Mathematics, 2013.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
ISO 19111:2019, Geographic information — Referencing by coordinates. International Organization for Standardization.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013.

AI Suggested Questions

Show how the condition number of the monomial basis matrix on the interval from zero to ten grows with polynomial degree, and compare it with a Chebyshev basis.
Coordinatise a $3 \times 3$ symmetric matrix relative to a basis of six symmetric matrices and explain how the choice relates to Voigt notation.
Explain why the discrete Fourier transform is a vector representation map and identify the ordered basis it uses.
Given two bases of the same space, derive the matrix that converts a coordinate vector from one to the other and show it is nonsingular.
For an orthonormal basis, prove that each coordinate is an inner product and quantify the operation count saved compared with solving a system.
Construct a case where a numerically generated basis has drifted towards dependence, and show how a reciprocal condition estimate detects it before the coefficients become meaningless.

Related Calculators

Coordinate Vector Calculator

Compute $\rho_B(\vec{w})$ for a chosen ordered basis, with the solve shown and the reconstruction verified.

Basis Validator

Check whether a proposed set is an ordered basis by testing independence and reporting the condition number of the basis matrix.

Change-of-Basis Calculator

Convert a coordinate vector from one ordered basis to another and display the conversion matrix.

Polynomial Basis Converter

Translate polynomial coefficients between monomial, Chebyshev and Bernstein bases and compare conditioning.

Overview

Definition