Linear Transformations | KEVOS® Mathematics

Overview

Vector spaces come equipped with exactly two operations: addition of vectors and multiplication by a scalar. A function between two vector spaces is linear when it is compatible with both. That is a deliberately minimal requirement, and it is also the strongest one that can be imposed without immediately trivialising the function. Everything in the theory of linear transformations follows from these two lines.

The practical reading is superposition. If a system's response to input $\vec{u}_1$ is known and its response to $\vec{u}_2$ is known, then linearity fixes its response to any combination $\alpha\vec{u}_1 + \beta\vec{u}_2$ with no further experiment. This is why engineering disciplines invest so heavily in establishing that a model is linear over an operating range: linearity converts an unbounded family of load cases, input signals or excitations into a finite basis of tests plus arithmetic.

A linear transformation carries a domain $U$ and a codomain $V$ , and these need not be the same space or even the same kind of space. Differentiation maps polynomials to polynomials, the trace maps square matrices to scalars, the matrix-vector product maps $\mathbb{C}^n$ to $\mathbb{C}^m$ , and a Laplace transform maps functions to functions. Keeping the two spaces distinct in the notation matters, because the plus sign on the left of the additivity condition is an operation in $U$ while the plus sign on the right is an operation in $V$ , and those may be entirely different constructions.

The most common modelling error the definition guards against is the constant offset. A function such as $\vec{x} \mapsto A\vec{x} + \vec{c}$ with $\vec{c} \ne \vec{0}$ looks linear, plots as a straight line, and is called linear in everyday speech — but it violates both conditions and superposition fails for it. The theorem that every linear transformation sends the zero vector to the zero vector turns this into a one-line test that costs nothing to apply.

Definition

Linear Transformation

A linear transformation $T\colon U \to V$ is a function from a vector space $U$ , the domain, to a vector space $V$ , the codomain, satisfying two conditions.

Additivity. $T(\vec{u}_1 + \vec{u}_2) = T(\vec{u}_1) + T(\vec{u}_2)$ for all $\vec{u}_1, \vec{u}_2 \in U$ .
Homogeneity. $T(\alpha\vec{u}) = \alpha T(\vec{u})$ for all $\vec{u} \in U$ and all scalars $\alpha \in \mathbb{C}$ .

Both conditions are universally quantified: a single input pair or a single scalar for which either fails is enough to disqualify the function.

The addition and scalar multiplication on the left-hand sides are the operations of $U$ ; those on the right-hand sides are the operations of $V$ . When $U$ and $V$ are different spaces these are different operations that happen to share a symbol.

Linear Operator

A linear transformation whose domain and codomain are the same vector space, $T\colon V \to V$ . The restriction matters because only then can $T$ be composed with itself, so powers $T^{2}, T^{3}, \ldots$ , eigenvalues and invariant subspaces are defined. Every square matrix defines a linear operator on $\mathbb{C}^n$ .

Affine Map

AFM

A function of the form $F(\vec{u}) = T(\vec{u}) + \vec{c}$ , where $T$ is a linear transformation and $\vec{c}$ is a fixed vector in the codomain. An affine map is linear if and only if $\vec{c} = \vec{0}$ . Affine maps preserve straight lines and ratios along them but not the origin, and they do not obey superposition. Sensor offsets, datum shifts, bias currents and reference-temperature corrections all introduce affine rather than linear behaviour.

Concepts

Two conditions, one combined test

Additivity and homogeneity can be merged into a single statement: $T$ is linear if and only if $T(\alpha\vec{u}_1 + \beta\vec{u}_2) = \alpha T(\vec{u}_1) + \beta T(\vec{u}_2)$ for all vectors and all scalars. Taking $\alpha = \beta = 1$ recovers additivity and taking $\beta = 0$ recovers homogeneity, so the combined form is equivalent. In verification work the combined form saves one pass over the algebra; in proof work the two conditions are usually applied separately, since each is invoked at a different point.

Neither condition implies the other

The two axioms are genuinely independent, so both must be checked. Over the real scalars, additivity alone permits pathological functions that are not homogeneous, and homogeneity alone permits functions such as $\vec{x} \mapsto \vec{x}$ scaled by a factor depending on direction, which fail additivity. In the practical setting the failure is usually blunt: $S(x_1, x_2) = (x_1 x_2, x_1)$ is homogeneous in neither sense — it fails additivity because $(x_1+y_1)(x_2+y_2) \ne x_1x_2 + y_1y_2$ — while a constant offset fails both. Test each condition and state which one broke.

Linear transformations take zero to zero

The proof is three lines and worth remembering. In $U$ the zero vector satisfies $\vec{0}_U = 0 \cdot \vec{0}_U$ , a consequence of the vector space axioms. Applying $T$ and using homogeneity with $\alpha = 0$ gives $T(\vec{0}_U) = 0 \cdot T(\vec{0}_U)$ , and the same vector space property in $V$ says any vector multiplied by the scalar zero is the zero vector of $V$ . Hence $T(\vec{0}_U) = \vec{0}_V$ . Note carefully that the two zero vectors are different objects living in different spaces.

The zero test as an instant rejection

Because $T(\vec{0}) = \vec{0}$ is a necessary condition, evaluating a candidate function at the zero vector is the cheapest possible screening test. If the result is anything other than the zero vector of the codomain the function is not linear and no further work is required. The test is not sufficient — $S(x_1,x_2) = (x_1x_2, 0)$ sends zero to zero and is still not linear — so a passing candidate must still be verified against both axioms.

Linearity is not a subspace test

The two conditions superficially resemble the closure conditions that identify a subspace, and the resemblance causes persistent confusion. A subspace test asks whether a set is closed under the two operations. A linearity test asks whether a function commutes with the two operations. The objects are different, the quantifiers are different, and the conclusions are different. What connects them is a later result: the kernel and the range of a linear transformation turn out to be subspaces, but that is a theorem, not a restatement.

The definition reaches far beyond column vectors

Restricting attention to maps between $\mathbb{C}^n$ and $\mathbb{C}^m$ hides the generality. Differentiation $p \mapsto p'$ from $P_n$ to $P_{n-1}$ is linear; so is definite integration from $P_n$ to $\mathbb{C}$ ; so is multiplication by a fixed polynomial, $p(x) \mapsto (x - 2)p(x)$ , from $P_4$ to $P_5$ ; so are the transpose and the trace on matrix spaces. Because they are all linear, every theorem proved from the two axioms applies to all of them at once. This economy is the reason the abstract definition is used rather than a matrix-specific one.

Why superposition is the engineering payoff

Repeated application of the two conditions extends them from two terms to any finite number: $T\left(\sum_i \alpha_i \vec{u}_i\right) = \sum_i \alpha_i T(\vec{u}_i)$ . If a set of inputs spans the domain, knowing $T$ on that set determines $T$ everywhere. In structural analysis this licenses combining unit load cases; in circuit analysis it licenses superposing independent sources; in signal processing it licenses decomposing an input into sinusoids and recombining the responses. Each of these is a direct corollary of the definition and holds for no other class of function.

Decision path: is this function a linear transformation?

Identify the domain and codomainName the vector spaces

U

and

V

and their operations. Confirm the function is well defined: exactly one output in

V

for each input in

U

Evaluate at the zero vectorIf

T(\vec{0}_U) \ne \vec{0}_V

, stop. The function is not linear, and the culprit is almost always a constant term.

Test additivity symbolicallyExpand

T(\vec{u}_1 + \vec{u}_2)

from the defining formula and rearrange to

T(\vec{u}_1) + T(\vec{u}_2)

. Products of components, powers and absolute values are the usual obstructions.

Test homogeneity symbolicallyExpand

T(\alpha\vec{u})

and factor

\alpha

out of every component. A term where

\alpha

appears squared, conjugated or not at all breaks the condition.

If a condition fails, produce a counterexampleExhibit specific vectors or a specific scalar for which the two sides differ numerically. A concrete counterexample is a complete disproof; a vague appeal to shape is not.

If both hold, record the consequencesThe function sends zero to zero, obeys superposition over any finite linear combination, and — when domain and codomain are column vector spaces — is represented by a unique matrix.

Equations

Additivity condition

EQ-LT-01

T ({\vec{u}}_{1} + {\vec{u}}_{2}) = T ({\vec{u}}_{1}) + T ({\vec{u}}_{2}) \forall {\vec{u}}_{1}, {\vec{u}}_{2} \in U

The addition on the left is performed in $U$ ; the addition on the right is performed in $V$ . The condition says the two routes from a pair of inputs to a single output agree.

Homogeneity condition

EQ-LT-02

T (α \vec{u}) = α T (\vec{u}) \forall \vec{u} \in U, \forall α \in ℂ

Scaling the input scales the output by the same factor. Note that $\alpha$ ranges over all scalars, including zero and complex values.

Combined linearity test

EQ-LT-03

T (α {\vec{u}}_{1} + β {\vec{u}}_{2}) = α T ({\vec{u}}_{1}) + β T ({\vec{u}}_{2})

Equivalent to the two conditions together: set $\alpha = \beta = 1$ for additivity and $\beta = 0$ for homogeneity.

Linear transformations take zero to zero

EQ-LT-04

T ({\vec{0}}_{U}) = T (0 \cdot {\vec{0}}_{U}) = 0 \cdot T ({\vec{0}}_{U}) = {\vec{0}}_{V}

The proof and the statement in one line. The subscripts are essential: the input zero belongs to $U$ and the output zero belongs to $V$ .

Superposition over a finite combination

EQ-LT-05

T (\sum_{i}^{=} α_{i} {\vec{u}}_{i}) = \sum_{i}^{=} α_{i} T ({\vec{u}}_{i})

Obtained by induction from the two axioms. This is the form actually used in structural, circuit and signal analysis.

A linear transformation on column vectors

EQ-LT-06

T : ℂ^{3} \to ℂ^{2}, T ([\begin{matrix} x_{1} \\ x_{2} \\ x_{3} \end{matrix}]) = [\begin{matrix} 2 x_{1} + x_{3} \\ - 4 x_{2} \end{matrix}]

Every output component is a homogeneous linear expression in the input components — no constants, no products, no powers. That pattern is exactly what linearity looks like in coordinates.

An affine map that is not linear

EQ-LT-07

F (\vec{x}) = A \vec{x} + \vec{c}, F (\vec{0}) = \vec{c} \neq \vec{0}

The zero test fails immediately when $\vec{c} \ne \vec{0}$ . Both axioms fail as well: $F(\vec{x} + \vec{y})$ carries one copy of $\vec{c}$ while $F(\vec{x}) + F(\vec{y})$ carries two.

Linear transformations on function and polynomial spaces

EQ-LT-08

D (p) = p', M (p) = (x - 2) p, tr (A) = \sum_{i}^{=} [A]_{i i}

Differentiation, multiplication by a fixed polynomial and the matrix trace are all linear transformations, so every theorem derived from the two axioms applies to them unchanged.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$T$	Transformation	The function itself, considered as a whole object	map from U to V
$U$	Domain	The vector space the inputs are drawn from	vector space
$V$	Codomain	The vector space the outputs land in	vector space
$\vec{u}$	Input vector	A generic element of the domain	element of U
$T (\vec{u})$	Image of a vector	The output produced by $T$ from the input $\vec{u}$	element of V
$α$	Scalar	Multiplier used in the homogeneity condition	complex scalar
${\vec{0}}_{U}$	Zero vector of the domain	Additive identity of $U$ ; the input for the zero test	element of U
${\vec{0}}_{V}$	Zero vector of the codomain	Additive identity of $V$ ; the required output of the zero test	element of V
$\vec{c}$	Offset vector	Constant term that makes a map affine rather than linear	element of V

Worked Numerical Example

Problem statement

A three-element thermocouple stack feeds a conditioning stage that reports the mean temperature of the stack and the top-to-bottom difference. Establish whether the conditioning map is a linear transformation, then examine what happens when a fixed cold-junction offset is folded into the same block.

State the conditioning map
The three raw channel readings form the input vector; the two reported quantities form the output. Both are column vector spaces, so the domain is $\mathbb{C}^3$ and the codomain is $\mathbb{C}^2$ .
$T : ℂ^{3} \to ℂ^{2}, T ([\begin{matrix} x_{1} \\ x_{2} \\ x_{3} \end{matrix}]) = [\begin{matrix} \frac{1}{3} (x_{1} + x_{2} + x_{3}) \\ x_{3} - x_{1} \end{matrix}]$
Apply the zero test first
With all three channels reading zero the reported mean is zero and the reported difference is zero, so $T(\vec{0}) = \vec{0}$ . The candidate survives the screening test, which means nothing yet but costs nothing.
$T ([\begin{matrix} 0 \\ 0 \\ 0 \end{matrix}]) = [\begin{matrix} 0 \\ 0 \end{matrix}]$
Test additivity
Take two arbitrary input vectors $\vec{x}$ and $\vec{y}$ . Add first, then transform; the arithmetic separates cleanly into the two individual outputs because every expression involved is a sum of scaled components.
$T (\vec{x} + \vec{y}) = [\begin{matrix} \frac{1}{3} (x_{1} + y_{1} + x_{2} + y_{2} + x_{3} + y_{3}) \\ (x_{3} + y_{3}) - (x_{1} + y_{1}) \end{matrix}] = [\begin{matrix} \frac{1}{3} (x_{1} + x_{2} + x_{3}) \\ x_{3} - x_{1} \end{matrix}] + [\begin{matrix} \frac{1}{3} (y_{1} + y_{2} + y_{3}) \\ y_{3} - y_{1} \end{matrix}] = T (\vec{x}) + T (\vec{y})$
Test homogeneity
Now scale the input by an arbitrary scalar $\alpha$ . Every component of the output is a sum of terms each containing exactly one input component, so $\alpha$ factors straight out.
$T (α \vec{x}) = [\begin{matrix} \frac{1}{3} (α x_{1} + α x_{2} + α x_{3}) \\ α x_{3} - α x_{1} \end{matrix}] = α [\begin{matrix} \frac{1}{3} (x_{1} + x_{2} + x_{3}) \\ x_{3} - x_{1} \end{matrix}] = α T (\vec{x})$
Conclude and record the matrix
Both conditions hold for all inputs and all scalars, so $T$ is a linear transformation. Because domain and codomain are column vector spaces, $T$ is realised by a single matrix acting on the input.
$T (\vec{x}) = A \vec{x}, A = [\begin{matrix} \frac{1}{3} & \frac{1}{3} & \frac{1}{3} \\ - 1 & 0 & 1 \end{matrix}]$
Fold in a cold-junction offset
Suppose the same block also adds a fixed reference temperature of $20$ units to the reported mean, to correct for the cold junction. Call the modified map $S$ . The zero test now fails immediately, so $S$ is not a linear transformation, and no further checking is needed.
$Fold in a cold-junction offset$
Exhibit the failure numerically
A counterexample makes the failure concrete. Take $\alpha = 2$ and the input $\vec{x} = (30, 30, 30)^{t}$ . Doubling the physical temperatures does not double the reported mean, because the offset is added once in one case and twice in the other. Superposition of two load cases would go wrong in exactly the same way.
$S (2 \vec{x}) = [\begin{matrix} 80 \\ 0 \end{matrix}], 2 S (\vec{x}) = 2 [\begin{matrix} 50 \\ 0 \end{matrix}] = [\begin{matrix} 100 \\ 0 \end{matrix}]$
Separate the offset from the linear part
The repair is structural, not cosmetic. Write $S(\vec{x}) = T(\vec{x}) + \vec{c}$ with $\vec{c} = (20, 0)^{t}$ , an affine map with linear part $T$ . Superposition, calibration by unit responses and matrix representation all apply to $T$ ; the offset is carried separately and added once at the end of the chain.
$S (\vec{x}) = A \vec{x} + \vec{c}, S (\vec{x}) - S (\vec{y}) = A (\vec{x} - \vec{y})$

Result

The conditioning map itself is linear, so the whole stack can be characterised by three unit-channel tests and every other input predicted by arithmetic. The offset-compensated block is affine, not linear: it must never be fed to a superposition argument directly. The last identity shows the standard workaround — differences of affine outputs are linear in the differences of inputs, which is why calibration procedures work with deltas rather than absolute readings.

Applications & Industry Use

Structural engineering

Unit load cases and load combinations

Within the elastic range, the map from applied loads to deflections and member forces is linear. An analyst therefore computes the response to a small set of unit load cases and forms every design combination — dead, live, wind, seismic, with their partial factors — as a scaled sum. Every code-prescribed load combination rule is an application of superposition, and every geometric or material non-linearity invalidates it.

Electrical engineering

Superposition of independent sources

In a network of linear resistors, capacitors and inductors, the map from source values to branch currents is linear. The superposition method analyses one source at a time with the others suppressed and adds the results. The technique fails the moment a diode, a saturating core or any offset-generating element appears, which is exactly the affine-versus-linear distinction in circuit form.

Signal processing & communications

Linear time-invariant systems

A filter is linear when its response to a sum of inputs is the sum of its responses. Combined with time invariance, this is what makes convolution and the frequency-domain description valid: an arbitrary input is decomposed into elementary signals, each is transformed independently, and the outputs are recombined. Amplifier clipping and quantisation are the practical linearity violations that limit the model.

Control systems

Linearisation about an operating point

Physical plants are rarely linear globally. The standard remedy is to expand the dynamics about an equilibrium and retain the first-order term, producing a linear model in the deviation variables. That deviation framing is precisely the affine-to-linear repair: the operating point is the offset, and the linear transformation acts on departures from it.

Computer graphics & robotics

Rotations, scalings and the need for homogeneous coordinates

Rotation, scaling, shear and projection are linear and compose by matrix multiplication. Translation is not linear — it moves the origin — which is why graphics and robotics pipelines embed points in homogeneous coordinates, adding one dimension so that an affine map in $\mathbb{R}^3$ becomes a genuinely linear map in $\mathbb{R}^4$ and the whole chain can be composed as a single matrix product.

Chemical & process engineering

Component balances and blending

Steady-state mass balances for a blending or separation unit express outlet compositions as linear functions of the inlet flows. Linearity is what allows a recipe to be scaled and two feed streams to be combined by adding their individual contributions, and it underlies the linear programming formulations used for refinery and blend optimisation.

Design Considerations

Check both conditions, not just the plausible one

Additivity and homogeneity are independent, and a function can satisfy one while failing the other. In verification work the temptation is to test additivity, see it succeed, and stop. Test both, and when the scalars are complex remember that homogeneity must hold for complex $\alpha$ — a map involving conjugation is additive and real-homogeneous but not linear over $\mathbb{C}$ .

Be explicit about which space each operation belongs to

When the domain is a polynomial space and the codomain a matrix space, the two plus signs in the additivity condition denote entirely different operations. Writing the verification with the spaces named prevents the common error of applying a codomain identity to a domain object, and it is the discipline that makes abstract proofs reliable.

Treat offsets as a modelling decision, not an inconvenience

Bias, tare, datum and reference values are real and must appear somewhere. The engineering choice is where. Absorbing them into the transformation destroys linearity and invalidates every superposition argument downstream; carrying them as a separate additive term keeps the linear machinery intact and makes the offset auditable.

State the range over which linearity is claimed

Physical systems are linear only over an interval: below yield, below saturation, below the small-angle limit, within the dynamic range of the converter. A linearity claim without a stated validity range is not usable. Record the range with the model and check operating points against it before applying superposition.

Prefer a basis of tests to an enumeration of cases

If a system is linear and the domain has dimension $n$ , then $n$ well-chosen experiments determine the entire response. Spending the test budget on a spanning set rather than on an arbitrary collection of realistic cases yields complete coverage; the standard unit vector basis is usually the cheapest choice because each test excites one input alone.

A single counterexample settles a negative claim

Proving linearity requires an argument covering all inputs and all scalars. Disproving it requires one concrete pair of vectors, or one vector and one scalar, for which the two sides differ. When a candidate is suspected to be non-linear, search for the counterexample first — it is far cheaper than a general argument and constitutes a complete proof.

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	Specifies the notation for functions and mappings, including the arrow form $T\colon U \to V$ and the convention that the map name is italic while the operation symbols are upright.
`IEC 60050-131`	International Electrotechnical Vocabulary — Circuit theory	Defines linearity for electrical networks and the conditions under which the superposition theorem may be applied, the engineering statement of the two axioms on this page.
`ISO 10303 (STEP), Part 42`	Geometric and topological representation	Represents geometric transformations for CAD interchange, distinguishing the linear part of a placement from its translation exactly as this page distinguishes linear from affine.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the function application and mapping notation semantically, so that $T(\vec{u})$ is exposed as an application rather than as juxtaposed symbols.
`W3C WCAG 2.1 AA`	Web Content Accessibility Guidelines	Requires the algebraic statements of additivity and homogeneity to be available as structured text to assistive technology rather than as rendered 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
Real scalars $\mathbb{R}$	Mechanical, structural and thermal models where every quantity is a physical magnitude.	Matches the physics directly and halves storage relative to complex, but linear maps over $\mathbb{R}$ may have no eigenvalues at all, so rotation-like behaviour must be handled with $2\times 2$ blocks.
Complex scalars $\mathbb{C}$	Phasor analysis of AC networks, frequency-domain signal processing and any setting where a spectrum must be complete.	Guarantees that characteristic polynomials factor completely and that eigenvalues exist, at double the storage and roughly four times the arithmetic per multiplication.
Exact rational arithmetic	Verifying linearity symbolically, or deriving an exact matrix representation from a defining formula.	Removes any doubt about whether a residual is a round-off artefact or a genuine non-linearity, but is impractical for large numerical work.
IEEE 754 binary64 evaluation	Numerically testing a black-box system for linearity by comparing $T(\vec{x} + \vec{y})$ against $T(\vec{x}) + T(\vec{y})$ .	The comparison must use a relative tolerance, since exact equality will not hold even for a genuinely linear implementation; choosing the tolerance is the whole difficulty of the test.
Fixed-point arithmetic	Embedded conditioning and control on hardware without a floating-point unit.	Deterministic and cheap, but saturation and truncation make the implemented map non-linear near the range limits even when the mathematical model is linear — the discrepancy must be budgeted.
Homogeneous coordinates	Geometry pipelines that must compose translations with rotations and scalings as a single chain.	Turns an affine map into a linear one by adding a dimension, so the whole transformation stack composes by matrix multiplication; costs one extra row and column and requires a perspective divide on output.

Manufacturing Notes

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

Verifying linearity by hand

Work symbolically with generic vectors, never with specific numbers, since agreement on particular inputs proves nothing. Expand the left-hand side from the defining formula, then reassemble the right-hand side, justifying each rearrangement by a vector space property. If the algebra stalls, the stalling point usually identifies the offending term: a constant, a product of components, a power or an absolute value.

Numerically probing a black-box implementation

For a system available only as code or hardware, evaluate the residuals $\|T(\vec{x} + \vec{y}) - T(\vec{x}) - T(\vec{y})\|$ and $\|T(\alpha\vec{x}) - \alpha T(\vec{x})\|$ over randomised inputs, normalised by $\|T(\vec{x})\|$ . Sweep the amplitude: a residual that stays at round-off level and then rises sharply above a threshold amplitude is the signature of saturation, which is the most common practical non-linearity.

Cost of characterisation

A linear transformation on an $n$ -dimensional domain is completely determined by its values on a basis, so characterisation costs $n$ evaluations regardless of how the transformation is implemented internally. Using the standard unit vectors makes each evaluation excite exactly one input, which both simplifies instrumentation and yields the matrix representation directly as the collected outputs.

Library and API conventions

Numerical libraries do not represent a linear transformation as a first-class object; they represent it as a matrix, or as a callable that computes matrix-vector products. Matrix-free interfaces such as scipy.sparse.linalg.LinearOperator take the second view and require the caller to supply the action of $T$ and of its adjoint, which is enough for every Krylov method. Symbolic systems keep the functional form and can verify the axioms directly.

Guarding a linearity assumption in production

Where a downstream algorithm relies on superposition, add an inexpensive runtime assertion: evaluate the additivity residual on one randomly chosen pair per batch and raise an alarm when it exceeds tolerance. This catches the case where an upstream change introduces clipping, a saturating limiter or a newly added offset, which would otherwise corrupt results silently.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Calling an affine map linear	high	Everyday usage treats any straight-line relationship as linear, so a constant offset is not recognised as disqualifying.	Apply the zero test to every candidate. If a constant term is genuinely present, write the map as linear plus offset and keep the offset separate.
Verifying with specific numbers only	high	Checking additivity on one or two chosen vectors and concluding the general condition holds.	Verification must be symbolic and cover all inputs and all scalars. Numerical checks can only ever disprove linearity, never establish it.
Testing only one of the two conditions	medium	Assuming additivity implies homogeneity, or the converse. Neither implication holds.	Check both explicitly, and state which condition was verified in any written argument.
Confusing linearity with the subspace test	medium	The closure conditions for a subspace resemble the two axioms, so a set property is applied to a function.	Ask what object is being tested. A set is tested for closure; a function is tested for commuting with the operations.
Ignoring complex scalars	medium	A map involving complex conjugation satisfies additivity and homogeneity for real $\alpha$ but fails for complex $\alpha$ .	State the scalar field explicitly. Over $\mathbb{C}$ , homogeneity must be checked with a genuinely complex scalar.
Applying superposition outside the validated range	high	A model linear below yield, below saturation or within a small-angle approximation is used at an operating point beyond that limit.	Record the validity range with the model and check every operating point against it before combining responses.
Mismatched domain and codomain in a composition	medium	Composing $S$ after $T$ when the codomain of $T$ is not the domain of $S$ , so the composite is undefined.	Write every transformation with its spaces attached, $T\colon U \to V$ , and confirm the middle space matches before composing.
Assuming a function is linear because it looks linear in coordinates	low	An expression such as $x_1x_2$ or $\|x_1\|$ is mistaken for a first-order term at a glance.	Require each output component to be a sum of input components each multiplied by a constant, with no products of unknowns, no powers and no constant term.

FAQs

Why is a function with a constant term not linear?

Because it violates both axioms and, most visibly, sends the zero vector to the offset rather than to zero. Concretely, $F(\vec{x} + \vec{y})$ contains the offset once while $F(\vec{x}) + F(\vec{y})$ contains it twice, so the two sides differ by exactly $\vec{c}$ . Such a map is called affine, and superposition arguments do not apply to it.

Does

T(\vec{0}) = \vec{0}

prove that a function is linear?

No, it is necessary but not sufficient. The map $S(x_1, x_2) = (x_1x_2, 0)$ sends the zero vector to the zero vector yet fails additivity, since $(x_1+y_1)(x_2+y_2)$ is not $x_1x_2 + y_1y_2$ . Use the zero test as a fast rejection, then verify both axioms symbolically for any candidate that survives.

Do the domain and codomain have to be the same kind of space?

No. A linear transformation may map polynomials to matrices, matrices to scalars, or $\mathbb{C}^4$ to $\mathbb{C}^2$ . The only requirements are that both are vector spaces over the same scalar field and that the two axioms hold, with each operation interpreted in its own space.

What is the difference between a linear transformation and a linear operator?

A linear operator is a linear transformation whose domain and codomain coincide. The restriction matters because only then can the map be composed with itself, which is what makes powers, eigenvalues, invariant subspaces and matrix exponentials meaningful. Every square matrix is a linear operator on $\mathbb{C}^n$ ; a rectangular matrix is a transformation but not an operator.

How does this relate to the superposition principle in engineering?

Superposition is the two axioms extended by induction to any finite linear combination. If a system is linear, its response to a weighted sum of inputs is the same weighted sum of the individual responses, which is why load combinations, source-by-source circuit analysis and convolution are valid. Every one of those methods becomes invalid when the underlying map is not linear.

Is differentiation really a linear transformation?

Yes. The derivative of a sum is the sum of the derivatives and the derivative of a constant multiple is that multiple of the derivative — precisely the two axioms. Restricted to a polynomial space, differentiation maps $P_n$ to $P_{n-1}$ and can be written as a matrix once bases are chosen, which is how differential operators are handled numerically.

Why do graphics systems use homogeneous coordinates?

Because translation is not linear: it moves the origin, so $T(\vec{0}) \ne \vec{0}$ . Embedding a three-dimensional point as a four-component vector with a trailing $1$ makes translation the action of a $4 \times 4$ matrix, so an affine transformation in three dimensions becomes a genuinely linear transformation in four and the entire pipeline composes as one matrix product.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section LT, subsection LT. Licensed under the GNU Free Documentation License v1.2.
Axler, S. Linear Algebra Done Right, 3rd edition. Springer, 2015. Chapter 3 on linear maps.
Halmos, P. R. Finite-Dimensional Vector Spaces. Springer, 1974.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
IEC 60050-131:2002, International Electrotechnical Vocabulary — Part 131: Circuit theory. International Electrotechnical Commission.

AI Suggested Questions

Give me a function that satisfies additivity but not homogeneity, and explain why such examples are hard to write down explicitly over the reals.
Show step by step how homogeneous coordinates turn a translation in three dimensions into a linear transformation in four.
How would I design a numerical experiment to decide whether a black-box amplifier is linear, and what tolerance should the residuals be judged against?
Work through the linearisation of a non-linear plant about an equilibrium and identify precisely where the affine offset goes.
Which of the following are linear transformations on $M_{22}$ : transpose, trace, determinant, and multiplication by a fixed matrix? Justify each.
Explain why complex conjugation on $\mathbb{C}^n$ is additive and real-homogeneous but not a linear transformation over $\mathbb{C}$ .

Related Calculators

Linearity Checker

Test a symbolic map for additivity and homogeneity and report which condition fails, with a counterexample where one exists.

Linear Transformation Evaluator

Apply a transformation to any input vector and to arbitrary linear combinations, showing the superposition check side by side.

Affine Map Decomposer

Split a candidate map into its linear part and its constant offset and report the residual after removal.

Overview

Definition

Linear Transformation

Linear Operator

Affine Map

Concepts

Two conditions, one combined test

Neither condition implies the other

Linear transformations take zero to zero

The zero test as an instant rejection

Linearity is not a subspace test

The definition reaches far beyond column vectors

Why superposition is the engineering payoff

Decision path: is this function a linear transformation?

Equations

Variable Definitions

Worked Numerical Example

Problem statement

State the conditioning map

Apply the zero test first

Test additivity

Test homogeneity

Conclude and record the matrix

Fold in a cold-junction offset

Exhibit the failure numerically

Separate the offset from the linear part

Applications &amp; Industry Use

Unit load cases and load combinations

Superposition of independent sources

Linear time-invariant systems

Linearisation about an operating point

Rotations, scalings and the need for homogeneous coordinates

Component balances and blending

Design Considerations

Check both conditions, not just the plausible one

Be explicit about which space each operation belongs to

Treat offsets as a modelling decision, not an inconvenience

State the range over which linearity is claimed

Prefer a basis of tests to an enumeration of cases

A single counterexample settles a negative claim

Standards &amp; Codes

Material Selection

Manufacturing Notes

Verifying linearity by hand

Numerically probing a black-box implementation

Cost of characterisation

Library and API conventions

Guarding a linearity assumption in production

Failure Modes &amp; Common Mistakes

FAQs

Related Engineering Topics

References

AI Suggested Questions

Related Calculators

Continue learning

Applications & Industry Use

Standards & Codes

Failure Modes & Common Mistakes