Matrices and Linear Transformations

Overview

The matrix-vector product was introduced as a piece of notation for compressing a system of equations. Reading it as a function instead changes what it is for. The expression $\vec{x} \mapsto A\vec{x}$ takes a vector of $n$ components and returns a vector of $m$ components, and the two properties of matrix multiplication that distribute over vector addition and commute with scalar multiplication are exactly the two axioms of linearity. So every matrix is a linear transformation in disguise.

The converse is the substantial result, and it is proved by construction rather than by an existence argument. Suppose $T\colon \mathbb{C}^n \to \mathbb{C}^m$ is known only to be linear, with no formula given. Evaluate it at each standard unit vector $\vec{e}_i$ and collect the $n$ resulting vectors as the columns of a matrix $A$ . Then $T(\vec{x}) = A\vec{x}$ for every input, because writing $\vec{x}$ as a combination of the $\vec{e}_i$ and pushing $T$ through that combination reproduces the matrix-vector product exactly.

The engineering statement of this theorem is that a linear system is completely characterised by its unit responses. Excite one input at a time, record the whole output vector, and stack those records side by side: that array predicts the response to every possible input. Influence coefficient matrices in structural analysis, sensitivity matrices in process control, Jacobians in robotics and scattering matrices in RF engineering are all this construction under different names.

Because the representing matrix is unique, no information is lost in either direction. Structural questions about the transformation — whether it is injective, what its range is, how it behaves on a subspace — become questions about null spaces, column spaces and rank of the matrix, and the whole apparatus of row reduction becomes available. The correspondence is extended to arbitrary finite-dimensional vector spaces later, by first choosing bases and converting vectors to coordinate columns.

Definition

Matrix Representation of a Transformation on Column Vectors

MLTCV

For a linear transformation $T\colon \mathbb{C}^n \to \mathbb{C}^m$ , the matrix representation of $T$ is the $m \times n$ matrix whose $i$ -th column is $T(\vec{e}_i)$ , the image of the $i$ -th standard unit vector. It is the unique matrix satisfying $T(\vec{x}) = A\vec{x}$ for all $\vec{x} \in \mathbb{C}^n$ .

The matrix depends on the standard basis being used for both spaces. Representations with respect to other bases exist and are related to this one by change-of-basis matrices.

Standard Unit Vectors

SUV

The vectors $\vec{e}_1, \vec{e}_2, \ldots, \vec{e}_n$ of $\mathbb{C}^n$ , where $\vec{e}_i$ has a $1$ in position $i$ and zeros elsewhere. They are the columns of the identity matrix $I_n$ , and they form a basis of $\mathbb{C}^n$ , so every vector has the unique expansion $\vec{x} = \sum_{i=1}^{n} [\vec{x}]_i \, \vec{e}_i$ .

In engineering terms, $\vec{e}_i$ is the input that excites channel $i$ with unit magnitude and leaves every other channel at zero.

Influence Coefficient

The entry $[A]_{ji}$ of a matrix representation, read as the contribution to output $j$ per unit of input $i$ . Column $i$ collects all influence coefficients for a single input; row $j$ collects all contributions to a single output. This reading is the working interpretation of a matrix representation in structural, thermal and process engineering.

Concepts

Every matrix builds a linear transformation

Fix an $m \times n$ matrix $A$ and define $T(\vec{x}) = A\vec{x}$ . Additivity follows from the distributive property of matrix multiplication over vector addition, $A(\vec{x} + \vec{y}) = A\vec{x} + A\vec{y}$ , and homogeneity from $A(\alpha\vec{x}) = \alpha(A\vec{x})$ . Both are established facts about the product, so no separate verification is needed. The consequence is a factory: any array of numbers, with no conditions attached, immediately yields a linear transformation from $\mathbb{C}^n$ to $\mathbb{C}^m$ .

The output is a linear combination of the columns

Expanding the matrix-vector product gives $A\vec{x} = [\vec{x}]_1 \vec{A}_1 + [\vec{x}]_2 \vec{A}_2 + \cdots + [\vec{x}]_n \vec{A}_n$ , where $\vec{A}_i$ is column $i$ . The transformation therefore uses the components of the input as weights on a fixed set of $n$ output vectors. This is the single most useful way to picture the map: the columns are the available output directions, the input chooses how much of each, and the reachable set is exactly the span of the columns — the column space.

Every transformation on column vectors comes from a matrix

For the converse, define $\vec{A}_i = T(\vec{e}_i)$ for $1 \le i \le n$ and set $A = [\, \vec{A}_1 \mid \cdots \mid \vec{A}_n \,]$ . Given any $\vec{x}$ , write it as $\sum_i [\vec{x}]_i \vec{e}_i$ , then apply $T$ and push it through the combination using additivity and homogeneity: $T(\vec{x}) = \sum_i [\vec{x}]_i T(\vec{e}_i) = \sum_i [\vec{x}]_i \vec{A}_i = A\vec{x}$ . This is the first argument in which the full strength of both axioms is used on a transformation supplied only as a hypothesis, and the proof is an algorithm: $n$ evaluations produce the matrix.

Uniqueness of the representing matrix

If $A\vec{x} = B\vec{x}$ for every $\vec{x}$ , then taking $\vec{x} = \vec{e}_i$ gives $\vec{A}_i = \vec{B}_i$ for each $i$ , so $A = B$ . The representation is therefore unique, which is what licenses the phrase the matrix of a transformation. It also gives the cheapest equality test for two linear maps: compare their action on $n$ basis vectors rather than attempting to compare them on all of $\mathbb{C}^n$ .

The size dictionary

An $m \times n$ matrix corresponds to a transformation $\mathbb{C}^n \to \mathbb{C}^m$ : the column count is the dimension of the domain and the row count is the dimension of the codomain. The reversal trips people up constantly. The mnemonic is that $A\vec{x}$ must be defined, so $\vec{x}$ has as many components as $A$ has columns, while the output inherits the row count. Composition then reads naturally: $S \circ T$ requires the middle dimensions to agree, which is precisely the condition for the matrix product to exist.

Structural questions transfer both ways

Because the correspondence is exact, every property of a transformation has a matrix counterpart and vice versa. The set of inputs mapped to zero is the null space of $A$ ; the set of achievable outputs is the column space of $A$ ; injectivity is equivalent to a trivial null space, hence to $A$ having full column rank; surjectivity is equivalent to the column space filling $\mathbb{C}^m$ , hence to full row rank. Row reduction, an algorithm defined for arrays of numbers, therefore answers questions about functions.

Extending the correspondence to abstract spaces

The theorem as stated needs domain and codomain to be spaces of column vectors, because the proof uses the standard unit vectors. For a transformation between arbitrary finite-dimensional vector spaces — polynomials to matrices, say — the repair is to fix a basis for each space and replace every vector by its coordinate column. The transformation then becomes a map between column vector spaces and acquires a matrix in the same way, though the matrix now depends on the two chosen bases as well as on $T$ .

Procedure: build the matrix of a linear transformation

Confirm linearity and the spacesEstablish that

T

is linear and that the domain is

\mathbb{C}^n

and the codomain

\mathbb{C}^m

. Without linearity the construction below is meaningless; without column vector spaces it needs a choice of bases first.

Evaluate on each standard unit vectorCompute

T(\vec{e}_1), T(\vec{e}_2), \ldots, T(\vec{e}_n)

. Each is a vector in

\mathbb{C}^m

. Experimentally, this is exciting one input channel at a time at unit magnitude.

Stack the results as columnsForm

A = \left[\, T(\vec{e}_1) \mid T(\vec{e}_2) \mid \cdots \mid T(\vec{e}_n) \,\right]

, an

m \times n

array. No rearrangement or normalisation is applied.

Verify on an independent inputChoose a vector with no zero components, evaluate

T

from its defining formula, and compare with

A\vec{x}

. Agreement on a generic vector catches transcription and ordering errors.

Read the structure from the matrixRow-reduce

A

to obtain rank, null space and column space, which answer injectivity, surjectivity and reachability for the transformation itself.

Equations

Matrix defines a transformation

EQ-MLT-01

T : ℂ^{n} \to ℂ^{m}, T (\vec{x}) = A \vec{x}

For any $m \times n$ matrix $A$ this function is linear, with no conditions on the entries of $A$ .

Output as a linear combination of columns

EQ-MLT-02

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

The components of the input are the weights and the columns of $A$ are the output directions. The reachable set is the span of the columns.

Construction of the matrix from unit responses

EQ-MLT-03

A = [T ({\vec{e}}_{1}) ∣ T ({\vec{e}}_{2}) ∣ \dots ∣ T ({\vec{e}}_{n})], {\vec{A}}_{i} = T ({\vec{e}}_{i})

The columns of the representing matrix are the images of the standard unit vectors. Experimentally, column $i$ is the recorded response to a unit excitation of input $i$ .

Proof chain for the representation theorem

EQ-MLT-04

T (\vec{x}) = T (\sum_{i}^{=} [\vec{x}]_{i} {\vec{e}}_{i}) = \sum_{i}^{=} [\vec{x}]_{i} T ({\vec{e}}_{i}) = \sum_{i}^{=} [\vec{x}]_{i} {\vec{A}}_{i} = A \vec{x}

Expand the input in the standard basis, push $T$ through the combination using both axioms, substitute the definition of the columns, and recognise the matrix-vector product.

Uniqueness of the representation

EQ-MLT-05

A \vec{x} = B \vec{x} \forall \vec{x} \in ℂ^{n} \Leftrightarrow A = B

Testing on the $n$ standard unit vectors is enough, since $A\vec{e}_i$ is column $i$ of $A$ . Two linear transformations are equal precisely when they agree on a basis.

Size dictionary

EQ-MLT-06

A \in M_{m n} \Leftrightarrow T : ℂ^{n} \to ℂ^{m}

The column count of the matrix is the dimension of the domain; the row count is the dimension of the codomain. The order is reversed relative to the arrow notation.

Structural correspondence

EQ-MLT-07

K (T) = N (A), R (T) = C (A)

The kernel of the transformation is the null space of its matrix and the range is the column space. Rank and nullity therefore answer surjectivity and injectivity questions about $T$ .

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$T$	Linear transformation	The function under study, from $\mathbb{C}^n$ to $\mathbb{C}^m$	linear map
$A$	Representing matrix	The unique matrix with $T(\vec{x}) = A\vec{x}$	m x n complex matrix
${\vec{A}}_{i}$	Column of the matrix	The image $T(\vec{e}_i)$ of the $i$ -th standard unit vector	vector in C^m
${\vec{e}}_{i}$	Standard unit vector	Column $i$ of $I_n$ ; a unit excitation of input channel $i$	vector in C^n
$n$	Domain dimension	Number of inputs; the column count of $A$	positive integer
$m$	Codomain dimension	Number of outputs; the row count of $A$	positive integer
$[\vec{x}]_{i}$	Component of the input	The $i$ -th entry of $\vec{x}$ , used as a weight on column $i$	complex scalar
$[A]_{j i}$	Influence coefficient	Contribution to output $j$ per unit of input $i$	complex scalar
$C (A)$	Column space	Span of the columns of $A$ ; the set of achievable outputs	subspace of C^m

Worked Numerical Example

Problem statement

A three-zone heater controls a plate instrumented with four thermocouples. Steady-state temperature rise at each thermocouple is linear in the three zone powers. Determine the influence coefficient matrix from unit-power tests, verify it against an independent operating point, and read off what the matrix says about reachability.

State the transformation
Let $\vec{x}$ hold the three zone powers and $T(\vec{x})$ the four temperature rises. The physical model, established from the governing equations, gives each output as a weighted sum of the inputs with no constant term — the transformation is linear.
$T : ℂ^{3} \to ℂ^{4}, T ([\begin{matrix} x_{1} \\ x_{2} \\ x_{3} \end{matrix}]) = [\begin{matrix} 2 x_{1} + x_{2} - x_{3} \\ - x_{1} + 4 x_{3} \\ 3 x_{2} + 2 x_{3} \\ x_{1} - x_{2} + x_{3} \end{matrix}]$
Run the first unit test
Drive zone one at unit power with zones two and three at zero. The recorded vector of four temperature rises is the first column of the influence matrix. Note that thermocouple two falls, which the sign of the second entry records faithfully.
$T ({\vec{e}}_{1}) = T ([\begin{matrix} 1 \\ 0 \\ 0 \end{matrix}]) = [\begin{matrix} 2 \\ - 1 \\ 0 \\ 1 \end{matrix}]$
Run the remaining unit tests
Repeat for zones two and three. Three experiments, each exciting a single input, produce three output vectors. Nothing more is required: linearity guarantees that these determine every other response.
$T ({\vec{e}}_{2}) = [\begin{matrix} 1 \\ 0 \\ 3 \\ - 1 \end{matrix}], T ({\vec{e}}_{3}) = [\begin{matrix} - 1 \\ 4 \\ 2 \\ 1 \end{matrix}]$
Assemble the influence matrix
Stack the three recorded vectors as columns, in input order. The result is $4 \times 3$ : four rows because there are four outputs, three columns because there are three inputs.
$A = [T ({\vec{e}}_{1}) ∣ T ({\vec{e}}_{2}) ∣ T ({\vec{e}}_{3})] = [\begin{matrix} 2 & 1 & - 1 \\ - 1 & 0 & 4 \\ 0 & 3 & 2 \\ 1 & - 1 & 1 \end{matrix}]$
Verify against an independent operating point
Take a setting with no zero components, $\vec{z} = (2, -3, 1)^{t}$ . Evaluate the transformation from its defining formula: $2(2) + (-3) - 1 = 0$ , $-2 + 4 = 2$ , $3(-3) + 2 = -7$ , $2 + 3 + 1 = 6$ . Now compute the matrix-vector product and compare.
$T (\vec{z}) = [\begin{matrix} 0 \\ 2 \\ - 7 \\ 6 \end{matrix}] = A \vec{z} = 2 [\begin{matrix} 2 \\ - 1 \\ 0 \\ 1 \end{matrix}] - 3 [\begin{matrix} 1 \\ 0 \\ 3 \\ - 1 \end{matrix}] + 1 [\begin{matrix} - 1 \\ 4 \\ 2 \\ 1 \end{matrix}]$
Read the structure by row reduction
Reducing $A$ gives three pivot columns, so the rank is $3$ . Full column rank means the null space is trivial, hence distinct power settings always produce distinct temperature patterns: the transformation is injective and the settings are recoverable from measurements.
$A ⟶ [\begin{matrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\ 0 & 0 & 0 \end{matrix}], rank (A) = 3, nullity (A) = 0$
Interpret the reachable set
The column space is a three-dimensional subspace of $\mathbb{C}^4$ , so the achievable temperature patterns form a proper subspace: with three actuators and four sensors, most four-component target profiles cannot be realised exactly. Deciding whether a specific target is achievable is a consistency question for $A\vec{x} = \vec{b}$ , answered by row reduction of the augmented matrix.

Result

Three unit-power experiments completely characterise a system with an unbounded number of possible operating points. The influence matrix is $4 \times 3$ with rank $3$ : the plate is fully observable in the sense that no two distinct power settings produce the same temperature pattern, but it is under-actuated, since only a three-dimensional set of four-component profiles can be achieved. Any target outside that subspace requires either a fourth heater zone or a least-squares compromise.

Applications & Industry Use

Structural engineering

Flexibility and stiffness influence coefficients

Applying a unit load at one degree of freedom and recording deflections at all degrees of freedom produces one column of the flexibility matrix. Assembling all columns yields the matrix representation of the load-to-deflection transformation, and inverting it gives the stiffness matrix used in the direct stiffness method.

Process control

Steady-state gain matrices for multivariable plants

Step-testing each manipulated variable while holding the others constant produces one column of the plant gain matrix. That matrix drives controller pairing decisions, relative gain array analysis and the assessment of whether the number of independent control handles is sufficient for the number of controlled outputs.

Robotics

The Jacobian as a local matrix representation

The map from joint velocities to end-effector velocity is linear at each configuration, and its matrix is the Jacobian. Its columns are the end-effector velocities produced by unit motion of each joint, so rank deficiency of the Jacobian identifies a singular configuration in which some task-space direction becomes momentarily unreachable.

Computer graphics

Transformation matrices from images of the axes

A rotation, scaling or shear is specified by where it sends the coordinate axes, and those images are exactly the columns of its matrix. This is why building a change-of-frame matrix in a graphics or CAD system amounts to writing down the three basis directions of the new frame as columns.

RF & microwave engineering

Scattering parameter matrices

An $S$ -parameter matrix records the wave leaving each port per unit wave incident at one port, with all other ports matched — a column-by-column construction from unit excitations. The resulting matrix represents the linear network exactly and composes with other networks by matrix operations.

Analytical chemistry

Calibration matrices for multi-analyte instruments

A spectrometer measuring several wavelengths for a mixture of several analytes has a response linear in the concentrations. Running a pure standard of each analyte gives one column of the calibration matrix; inverting or least-squares solving with that matrix converts a measured spectrum back into concentrations.

Design Considerations

Unit tests are the cheapest complete characterisation

For a genuinely linear system, $n$ single-channel experiments determine the response to every input, so a test programme that exercises arbitrary realistic combinations wastes budget and leaves gaps. Design the test matrix to span the input space; the standard unit vectors are the simplest spanning choice and each isolates one influence path.

Unit excitation may be experimentally impractical

Some systems cannot be driven one input at a time — a pump cannot deliver negative flow, a joint may not move independently. When that happens, excite any set of $n$ independent input vectors, collect the responses as the columns of $B$ , and recover the standard representation as $A = B P^{-1}$ , where $P$ has the chosen excitation vectors as its columns. The construction is unchanged; only the basis differs.

Watch the row-column orientation

An $m \times n$ matrix represents a map from $\mathbb{C}^n$ to $\mathbb{C}^m$ , so the dimensions appear in the opposite order to the arrow notation. Storing unit responses as rows instead of columns produces the transpose, which silently computes a different transformation. Verify orientation on a generic vector before the matrix enters any downstream calculation.

Rank determines what the system can and cannot do

Full column rank means inputs are recoverable from outputs; full row rank means every output is achievable. With $m > n$ the map cannot be surjective, so exact tracking of an arbitrary target is impossible and a least-squares formulation is required. With $n > m$ the map cannot be injective, so a null space of input combinations produces no output at all — useful for secondary objectives, dangerous if unmodelled.

Rebuild the representation when the operating point moves

A matrix derived by linearising a non-linear plant is valid only near the point at which it was measured. Influence coefficients for a structure change once yielding begins; a robot Jacobian changes at every configuration. Record the operating point alongside the matrix and re-identify it whenever the point moves outside the validated envelope.

Measurement noise makes an identified matrix approximate

Columns obtained experimentally carry error, so an identified matrix should be reported with uncertainty and its rank determined with a tolerance rather than by exact-zero tests. Repeating each unit test and averaging, or fitting all excitations simultaneously by least squares, gives a better-conditioned estimate than a single test per channel.

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 bracket notation for matrices, the italic convention for the map name $T$ , and the indexing order in which $[A]_{ji}$ names the entry in row $j$ and column $i$ .
`BLAS Level 2`	Basic Linear Algebra Subprograms, matrix-vector operations	Standardises `xGEMV`, the matrix-vector product that evaluates a linear transformation; the leading-dimension and transpose flags in its interface are exactly the orientation questions discussed here.
`IEEE 370-2020`	Electrical characterization of printed circuit board and related interconnects at frequencies up to 50 GHz	Governs the measurement and validation of scattering parameter matrices, an industrial instance of building a matrix representation column by column from single-port excitations.
`ISO 10303 (STEP), Part 42`	Geometric and topological representation	Encodes geometric transformations for CAD exchange by their action on the coordinate axes, which is the matrix-from-unit-vectors construction expressed as a data model.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Provides the semantic encoding of the matrix and column notation on this page so that structure, not just appearance, is available to 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
Dense row-major or column-major storage	Small to moderate influence matrices from instrumentation, robotics or control identification.	Simple and cache-friendly for matrix-vector products, but the storage order must be declared consistently or the transpose is computed silently; memory grows as $m n$ .
Sparse compressed storage (CSR or CSC)	Large finite element or network influence matrices where each output depends on only a few inputs.	Reduces storage and product cost to the number of non-zeros, but CSC favours column access, which is the natural direction for the unit-response construction, while CSR favours row access.
Matrix-free operator	The action of $T$ is available as a simulation or a physical experiment but the matrix is too large to store.	Needs only $O(m + n)$ memory and suits Krylov solvers, but explicit structural questions such as rank and null space require iterative estimation rather than row reduction.
IEEE 754 binary64 entries	Identified matrices from measurement, where entries carry physical units and limited significant figures.	Ample precision relative to measurement noise, but rank must be judged with a tolerance derived from the noise level, not from machine epsilon.
Exact rational entries	Analytically derived transformations, symbolic verification and textbook-scale structural analysis.	Gives an exact rank and an exact null space basis, but is unavailable for identified matrices and impractical beyond modest sizes.
Fixed-point entries	Embedded implementation of an identified gain matrix in a controller without floating-point hardware.	Deterministic timing and low cost, but requires per-column scaling analysis to avoid overflow, and the quantised matrix represents a slightly different transformation than the identified one.

Manufacturing Notes

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

Cost of building and applying the representation

Constructing the matrix costs $n$ evaluations of the transformation, which may be $n$ physical experiments or $n$ simulation runs. Once built, applying it costs $2mn$ floating-point operations per input vector via a single matrix-vector product, independent of how expensive the original evaluation was. For a system evaluated many times, identifying the matrix once is almost always the cheaper route.

Building the matrix by hand from a formula

Set all inputs but one to zero and read the surviving terms; that gives a column directly, with no algebraic rearrangement. The alternative is to group the defining formula by input variable and factor the coefficient vectors out — the same answer reached by collecting rather than by substitution. Cross-check the two methods on one column when the formula is complicated.

Verification strategy

Testing on standard unit vectors only reproduces the construction and cannot detect an error. Verify on a vector with all components non-zero and mutually distinct, so that a transposed matrix, a swapped column pair or a sign error all produce a visible discrepancy. In floating point, compare $\|T(\vec{x}) - A\vec{x}\|$ against $\|A\| \, \|\vec{x}\|$ times the unit round-off.

Library conventions and pitfalls

NumPy's A @ x and MATLAB's A*x both implement the matrix-vector product with the column convention used here. NumPy broadcasting will silently accept a one-dimensional array of the wrong length in some reduction contexts, and MATLAB will interpret a row vector argument as requiring the transposed product; both errors surface as a wrong answer rather than an exception. Assert the shape of $A$ against $(m, n)$ before use.

Identification by least squares rather than one test per channel

Given a set of $k \ge n$ excitation vectors collected as the columns of $X$ and their measured responses as the columns of $Y$ , the best-fit representation solves $AX \approx Y$ , giving $A = YX^{+}$ with the pseudoinverse. Redundant excitations average out measurement noise and the condition number of $X$ quantifies how well the chosen test set spans the input space.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Transposing the representation	high	Recording unit responses as rows instead of columns, producing $A^{t}$ and a transformation with the domain and codomain exchanged.	Check the shape against $m \times n$ and verify on a generic vector, not on a standard unit vector, before using the matrix.
Building a matrix for a non-linear map	high	Applying the unit-response construction to a function that has not been verified linear, most often one with a constant offset.	Establish linearity first, including the zero test. For an affine map, identify the linear part from response differences and carry the offset separately.
Verifying only on the standard unit vectors	medium	The construction guarantees agreement on the $\vec{e}_i$ by definition, so such a check is circular and detects nothing.	Verify on an input whose components are all non-zero and mutually different.
Confusing the domain and codomain dimensions	medium	Reading an $m \times n$ matrix as a map from $\mathbb{C}^m$ to $\mathbb{C}^n$ because the arrow notation lists domain first.	Anchor on the requirement that $A\vec{x}$ be defined: the input length matches the column count, the output length matches the row count.
Column order not matching input order	medium	Unit tests performed or filed in a different order from the input variable numbering.	Label each recorded response vector with its input index at the time of the experiment and assemble strictly in index order.
Reusing an identified matrix outside its operating envelope	high	A linearised or measured representation is applied at a configuration or load level well away from where it was obtained.	Store the operating point and validity range as metadata with the matrix and re-identify whenever the point moves outside it.
Declaring rank by exact-zero tests on measured data	medium	A column of an identified matrix that should be dependent is only nearly dependent because of noise.	Determine rank from singular values against a tolerance derived from the measurement uncertainty, not from row reduction with exact comparisons.
Assuming a representation exists for abstract spaces without choosing bases	low	Applying the theorem to a map between polynomial or matrix spaces, where standard unit vectors are not defined.	Fix a basis for the domain and codomain, convert to coordinate vectors, and note that the resulting matrix depends on both bases.

FAQs

Why are the columns of the matrix the images of the standard unit vectors?

Because $A\vec{e}_i$ selects column $i$ of $A$ exactly. So if $T(\vec{x}) = A\vec{x}$ is to hold for all inputs, it must hold at $\vec{x} = \vec{e}_i$ , which forces column $i$ to equal $T(\vec{e}_i)$ . The construction is not a convenient choice; it is the only possibility.

Is the matrix of a linear transformation unique?

With respect to the standard bases on both spaces, yes. If two matrices give the same output for every input then they agree on each standard unit vector, hence column by column. Change either basis and the matrix changes, which is why representations with respect to arbitrary bases carry explicit basis labels.

Does every linear transformation have a matrix?

Every linear transformation between finite-dimensional spaces does, once bases have been chosen for the domain and the codomain. When both spaces are column vector spaces the standard bases are implied and no choice is needed. Transformations on infinite-dimensional spaces, such as differentiation on all smooth functions, have no finite matrix.

How many experiments do I need to characterise a linear system?

Exactly as many as the dimension of the input space, provided the excitations are linearly independent. With $n$ independent excitations collected as the columns of $P$ and the responses as the columns of $B$ , the standard representation is $A = BP^{-1}$ ; if the excitations are the standard unit vectors then $P = I_n$ and $A = B$ directly.

What does rank deficiency of the representing matrix mean physically?

Deficient column rank means a non-trivial null space: some non-zero input combination produces no output at all, so inputs cannot be recovered from measurements. Deficient row rank means the achievable outputs form a proper subspace, so some target responses are unattainable. A robot Jacobian at a singular configuration exhibits both readings simultaneously.

Why does an

m \times n

matrix map

\mathbb{C}^n

\mathbb{C}^m

rather than the other way round?

Because the product $A\vec{x}$ requires the input to have as many components as $A$ has columns, and it produces one output component per row. The dimension order therefore reverses relative to the arrow $T\colon \mathbb{C}^n \to \mathbb{C}^m$ . This is a constant source of shape errors and is worth checking explicitly in code.

How does this correspondence extend to spaces of polynomials or matrices?

Choose a basis for the domain and one for the codomain, replace every vector by its coordinate column with respect to those bases, and the transformation becomes a map between column vector spaces with a matrix built exactly as above. The resulting matrix depends on both bases, and changing a basis conjugates or multiplies it by a change-of-basis matrix.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section LT, subsection MLT. Licensed under the GNU Free Documentation License v1.2.
Strang, G. Introduction to Linear Algebra, 5th edition. Wellesley-Cambridge Press, 2016. Chapter 8 on linear transformations and their matrices.
Axler, S. Linear Algebra Done Right, 3rd edition. Springer, 2015. Chapter 3, the matrix of a linear map.
Dongarra, J. J. et al. An Extended Set of FORTRAN Basic Linear Algebra Subprograms. ACM Transactions on Mathematical Software 14(1), 1988.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.

AI Suggested Questions

If I can only excite my system with three specific non-standard input vectors, how do I recover the standard matrix representation from the measured responses?
Show how the matrix of a rotation in the plane is obtained purely from the images of the two coordinate axes.
Why does a robot Jacobian lose rank at a singular configuration, and what does the null space of the Jacobian mean for joint motion?
Compare identifying an influence matrix by one unit test per channel against a least-squares fit over redundant excitations, in the presence of measurement noise.
Take a linear transformation from $P_3$ to $M_{22}$ and build its matrix with respect to explicitly chosen bases for both spaces.
Explain what changes in the matrix representation when I reorder the input channels, and why it is a permutation similarity.

Related Calculators

Transformation Matrix Builder

Evaluate a symbolic transformation on the standard unit vectors and assemble the representing matrix automatically.

Matrix-Vector Product Calculator

Compute $A\vec{x}$ and display it simultaneously as a linear combination of the columns of $A$ .

Matrix Rank & Nullity Calculator

Report rank, nullity and pivot columns to decide injectivity and surjectivity of the represented transformation.