Invertibility Criteria for Linear Transformations

Overview

Invertibility is the property that lets a transformation be undone. Given $T\colon U \to V$ , the question is whether some companion function $S\colon V \to U$ returns every input to itself after a round trip, in both directions. Stated that way it looks like a search over an unbounded space of candidate functions. The central result of this topic removes the search entirely: $T$ is invertible if and only if it is injective and surjective, two properties that are decided by finite computations on the kernel and the range.

The two halves of the criterion correspond to the two ways a candidate inverse can fail. If $T$ misses some vector $\vec{v}$ of the codomain, no value can be assigned to $S(\vec{v})$ without contradicting the requirement $T(S(\vec{v})) = \vec{v}$ ; the putative inverse is undefined somewhere. If $T$ sends two distinct inputs to the same output, then $S$ would have to take two values at that output; the putative inverse is multiply defined. Surjectivity forbids the first failure, injectivity the second, and together they force every pre-image to be a set containing exactly one vector — which is precisely a recipe for a function.

Two structural bonuses come free. The inverse of a linear transformation is never merely a function: it inherits linearity from $T$ , so $T^{-1}$ is itself a linear transformation, and it is itself invertible with $(T^{-1})^{-1} = T$ . Invertibility is also preserved by composition, and the inverse of a composition reverses the order of the factors, exactly as it does for matrix products. Those closure properties are what allow long chains of coordinate changes, calibrations and representations to be assembled and then unwound.

For an engineer the criterion is the licence to speak of the solution rather than a solution. An invertible model map means measurements determine states uniquely and every conceivable measurement is achievable; a non-invertible one means either unobservable directions or unreachable outputs, and the criterion says exactly which of the two defects you have. The same equivalence, applied to a matrix-induced transformation on $\mathbb{C}^n$ , reproduces the familiar statement that a square matrix is invertible exactly when it is nonsingular.

Definition

Identity Linear Transformation

IDLT

For a vector space $W$ , the identity linear transformation $I_W \colon W \to W$ is defined by $I_W(\vec{w}) = \vec{w}$ for every $\vec{w} \in W$ . It is linear, injective and surjective, and it is the neutral element for composition: $I_W \circ F = F$ and $G \circ I_W = G$ whenever the compositions are defined.

The subscript matters. $I_U$ and $I_V$ act on different spaces, and the definition of invertibility uses both.

Invertible Linear Transformation

IVLT

A linear transformation $T\colon U \to V$ is invertible if there exists a function $S\colon V \to U$ satisfying both composition identities

$S \circ T = I_U$ — every input of $T$ is recovered, and
$T \circ S = I_V$ — every output of $T$ is reproduced.

The function

S

is then called the inverse of

T

and written

T^{-1}

. The inverse, when it exists, is unique.

The definition asks only for a function. That $T^{-1}$ turns out to be a linear transformation is a theorem, not an assumption.

Invertibility Criterion

ILTIS

Let $T\colon U \to V$ be a linear transformation. Then $T$ is invertible if and only if $T$ is both injective and surjective. Equivalently, in terms of the two subspaces that measure those properties, $T$ is invertible exactly when $\mathcal{K}(T) = \{\vec{0}\}$ and $\mathcal{R}(T) = V$ .

Concepts

Both composition identities are part of the definition

The definition demands $S \circ T = I_U$ and $T \circ S = I_V$ . Neither alone is sufficient in general. A map satisfying only $S \circ T = I_U$ has a left inverse and is necessarily injective; a map satisfying only $T \circ S = I_V$ has a right inverse and is necessarily surjective. Between finite-dimensional spaces of equal dimension the two conditions imply each other, which is why the distinction is invisible in most matrix work.

It becomes visible the moment dimensions differ or the spaces are infinite-dimensional. The right-shift operator on infinite sequences is injective with a left inverse but is not surjective, so it has no genuine inverse. Keeping both identities in the definition is what makes invertibility a symmetric relationship between $T$ and $T^{-1}$ rather than a one-way repair.

Why injectivity and surjectivity are exactly what is needed

The proof of the criterion is worth carrying in memory because it explains the two failure modes rather than merely listing them. Assume $T$ is invertible. Injectivity follows by applying $T^{-1}$ to the equation $T(\vec{x}) = T(\vec{y})$ , which collapses to $\vec{x} = \vec{y}$ . Surjectivity follows by presenting, for any target $\vec{v} \in V$ , the explicit input $\vec{u} = T^{-1}(\vec{v})$ .

The converse is the constructive half. Assume $T$ is injective and surjective, and choose any $\vec{v} \in V$ . Surjectivity guarantees the pre-image $T^{-1}(\vec{v})$ is non-empty. The pre-image of a vector in the range is a coset $\vec{u} + \mathcal{K}(T)$ of the kernel, and injectivity forces the kernel to be trivial, so the pre-image collapses to the single vector $\{\vec{u}\}$ . Defining $S(\vec{v}) = \vec{u}$ is therefore legitimate — there is one candidate and it always exists — and the two composition identities follow immediately.

The inverse is automatically linear, and invertible in turn

Nothing in the definition of invertibility requires $S$ to respect addition or scalar multiplication, yet it must. For $\vec{x}, \vec{y} \in V$ , write each as $T$ applied to its own inverse image, push the sum through the linearity of $T$ , and cancel with $T^{-1}$ ; the result is $T^{-1}(\vec{x} + \vec{y}) = T^{-1}(\vec{x}) + T^{-1}(\vec{y})$ . The same manoeuvre with a scalar gives $T^{-1}(\alpha\vec{x}) = \alpha T^{-1}(\vec{x})$ .

Once $T^{-1}$ is known to be a linear transformation, the two defining identities can be read from its point of view, with $T$ playing the role of its companion. That says $T^{-1}$ is itself invertible and $(T^{-1})^{-1} = T$ . Invertibility is thus a symmetric property of an ordered pair of maps, and the operation of inverting is an involution on the set of invertible transformations.

Composition preserves invertibility and reverses order

If $T\colon U \to V$ and $S\colon V \to W$ are both invertible, so is $S \circ T$ . The clean argument routes through the criterion: composition of injective maps is injective, composition of surjective maps is surjective, and the criterion returns invertibility. The inverse can then be written down without further work, because $T^{-1} \circ S^{-1}$ satisfies both composition identities against $S \circ T$ .

The reversal of order in $(S \circ T)^{-1} = T^{-1} \circ S^{-1}$ is not an artefact of notation. Undoing a sequence of operations requires undoing the last one first, which is the same reason $(AB)^{-1} = B^{-1}A^{-1}$ for matrices and the same reason a garment put on last is removed first. Engineering pipelines that chain sensor calibration, unit conversion and basis change inherit this rule directly.

The equal-dimension shortcut

When $U$ and $V$ are finite-dimensional with $\dim(U) = \dim(V)$ , one of the two properties implies the other and a single test decides invertibility. The mechanism is the rank-nullity identity $r(T) + n(T) = \dim(U)$ . If the kernel is trivial then $n(T) = 0$ , so $r(T) = \dim(U) = \dim(V)$ , and a subspace of $V$ with the full dimension of $V$ must be $V$ itself — surjectivity follows. The reverse implication runs the same argument backwards.

This is the workhorse in practice. For a transformation from a space to itself, or between two spaces already known to have matching dimension, computing the kernel alone settles invertibility. It also explains why a square matrix with trivial null space is automatically nonsingular and why no rectangular matrix ever induces an invertible transformation.

Reading the criterion through a matrix

Every $m \times n$ matrix $A$ induces the transformation $T(\vec{x}) = A\vec{x}$ from $\mathbb{C}^n$ to $\mathbb{C}^m$ , whose kernel is the null space $\mathcal{N}(A)$ and whose range is the column space $\mathcal{C}(A)$ . The criterion then reads: $T$ is invertible exactly when $\mathcal{N}(A) = \{\vec{0}\}$ and $\mathcal{C}(A) = \mathbb{C}^m$ . Both can only hold when $m = n$ and $\operatorname{rank}(A) = n$ , which is the definition of a nonsingular square matrix.

The abstract statement is nevertheless the stronger one, because it applies verbatim to spaces of polynomials, matrices, functions or any other finite-dimensional vector space, where no coefficient array is available until a basis has been fixed. Choosing a basis and testing the resulting matrix is a valid tactic, but the conclusion must be reported for the transformation, not for the array.

Decision path: is this linear transformation invertible?

Confirm linearity firstVerify

T(\vec{x} + \vec{y}) = T(\vec{x}) + T(\vec{y})

and

T(\alpha\vec{x}) = \alpha T(\vec{x})

. The criterion applies only to linear maps.

Compare dimensionsIf

\dim(U) \ne \dim(V)

, stop:

T

cannot be invertible. Unequal dimensions rule out injectivity or surjectivity outright.

Compute the kernelSolve

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

. A trivial kernel means

T

is injective; a non-trivial one means

T

is not, and the search ends.

Settle surjectivityWith equal finite dimensions, a trivial kernel already forces

\mathcal{R}(T) = V

. Otherwise show every

\vec{v} \in V

has a non-empty pre-image.

Construct the inverseSolve

T(\vec{u}) = \vec{v}

for a general

\vec{v}

. The unique solution, expressed as a formula in the entries of

\vec{v}

, is

T^{-1}

Verify both compositionsCheck

T^{-1} \circ T = I_U

and

T \circ T^{-1} = I_V

on a basis of each space. Linearity extends the check to all vectors.

Equations

Identity linear transformation

EQ-IVC-01

I_{W} : W \to W, I_{W} (\vec{w}) = \vec{w}

The do-nothing map on $W$ . It is the reference against which both composition identities are measured.

Defining identities for an inverse

EQ-IVC-02

S \circ T = I_{U} and T \circ S = I_{V}

Both must hold. The first says $S$ recovers every input of $T$ ; the second says $S$ supplies a pre-image for every element of $V$ .

Invertibility criterion

EQ-IVC-03

T invertible \Leftrightarrow T injective and T surjective

The master equivalence. It replaces an existence question about functions with two subspace computations.

Criterion in terms of kernel and range

EQ-IVC-04

T invertible \Leftrightarrow K (T) = {\vec{0}} and R (T) = V

The computational form. The kernel is the solution set of $T(\vec{u}) = \vec{0}$ ; the range is the set of achievable outputs.

Pre-image of a vector under an invertible map

EQ-IVC-05

Pre-image of a vector under an invertible map

Surjectivity makes the pre-image non-empty; a trivial kernel collapses it to one vector. That singleton is what allows $T^{-1}$ to be defined as a function.

Inverse of a composition

EQ-IVC-06

(S \circ T)^{- 1} = T^{- 1} \circ S^{- 1}

Valid whenever $T\colon U \to V$ and $S\colon V \to W$ are both invertible. The order of the factors reverses.

Inversion is an involution

EQ-IVC-07

(T^{- 1})^{- 1} = T

The inverse of an invertible linear transformation is itself an invertible linear transformation, and inverting twice returns the original map.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$T$	Linear transformation	The map under test, with domain $U$ and codomain $V$	linear map U to V
$T^{- 1}$	Inverse transformation	The unique map satisfying both composition identities with $T$	linear map V to U
$U$	Domain	The vector space of inputs to $T$	vector space over C
$V$	Codomain	The vector space in which outputs of $T$ are declared to live	vector space over C
$I_{U}$	Identity on the domain	The map leaving every vector of $U$ unchanged	linear map U to U
$K (T)$	Kernel	All inputs sent to the zero vector; a subspace of $U$	subspace of U
$R (T)$	Range	All achievable outputs; a subspace of $V$	subspace of V
$\circ$	Composition	Application of one transformation to the output of another, right factor first	binary operation
$n (T)$	Nullity	Dimension of the kernel; zero exactly when $T$ is injective	non-negative integer

Worked Numerical Example

Problem statement

Three quantities $a$ , $b$ , $c$ describe a mixing stage, and an instrument reports only the three pairwise sums $a+b$ , $b+c$ and $a+c$ . Model the instrument as a linear transformation from $P_2$ to $\mathbb{C}^3$ , decide whether the original quantities can always be recovered, and if so build the recovery map.

State the transformation
Encode the three quantities as the coefficients of a polynomial of degree at most two and the three readings as a column vector. The instrument is then the map
$T : P_{2} \to ℂ^{3}, T (a + b x + c x^{2}) = [\begin{matrix} a + b \\ b + c \\ a + c \end{matrix}]$
Compute the kernel
A polynomial lies in $\mathcal{K}(T)$ when all three readings vanish, giving $a + b = 0$ , $b + c = 0$ and $a + c = 0$ . The first two force $a = -b$ and $c = -b$ ; substituting into the third gives $-2b = 0$ , so $b = 0$ and then $a = c = 0$ .
$K (T) = {0}, n (T) = 0$
Conclude injectivity
A linear transformation is injective exactly when its kernel is trivial. The kernel here contains only the zero polynomial, so $T$ is injective: no two distinct mixtures can produce the same three readings.
Solve for a general target and obtain surjectivity
Given an arbitrary reading vector with entries $p$ , $q$ , $r$ , solve $a + b = p$ , $b + c = q$ , $a + c = r$ . Adding all three equations gives $2(a + b + c) = p + q + r$ . Subtracting each original equation in turn isolates one unknown, and the result exists for every $p$ , $q$ , $r$ , so every reading vector is achievable and $\mathcal{R}(T) = \mathbb{C}^3$ .
$a = \frac{p - q + r}{2}, b = \frac{p + q - r}{2}, c = \frac{- p + q + r}{2}$
Apply the criterion and name the inverse
$T$ is injective and surjective, so by the invertibility criterion it is invertible. The formulas of the previous step define the inverse explicitly, and by the linearity theorem this recovery map is itself a linear transformation.
$T^{- 1} ([\begin{matrix} p \\ q \\ r \end{matrix}]) = \frac{p - q + r}{2} + \frac{p + q - r}{2} x + \frac{- p + q + r}{2} x^{2}$
Verify both composition identities
Round trip one: apply $T$ to $a + bx + cx^2$ , then $T^{-1}$ . The first entry of the recovered polynomial is $\frac{(a+b) - (b+c) + (a+c)}{2} = \frac{2a}{2} = a$ , and the other two coefficients recover $b$ and $c$ in the same way, so $T^{-1} \circ T = I_{P_2}$ . Round trip two: apply $T$ to the formula above. Its first entry is $\frac{p-q+r}{2} + \frac{p+q-r}{2} = p$ , and similarly the remaining entries return $q$ and $r$ , so $T \circ T^{-1} = I_{\mathbb{C}^3}$ .
Chain a second invertible stage
Suppose the readings pass through a running-sum accumulator $R\left(\begin{bmatrix} p & q & r \end{bmatrix}^{t}\right) = \begin{bmatrix} p & p+q & p+q+r \end{bmatrix}^{t}$ , which is invertible by successive differencing. Then $R \circ T$ is invertible, and its inverse reverses the order of the stages: undo the accumulator first, then undo the instrument.
$(R \circ T)^{- 1} = T^{- 1} \circ R^{- 1}$
Numerical check of the chain
Take the mixture $1 + x + x^2$ . Then $T$ gives $\begin{bmatrix} 2 & 2 & 2 \end{bmatrix}^{t}$ and the accumulator gives $\begin{bmatrix} 2 & 4 & 6 \end{bmatrix}^{t}$ . Differencing returns $\begin{bmatrix} 2 & 2 & 2 \end{bmatrix}^{t}$ , and the inverse formula returns $a = b = c = 1$ , recovering the original mixture exactly.

Result

The instrument is invertible, so the three pairwise sums carry the same information as the three original quantities: the design loses nothing. Had the third channel reported $a + b + c$ rather than $a + c$ , the kernel would still be trivial, but a channel reporting a repeat of $a+b$ would give a one-dimensional kernel and a two-dimensional range — an unobservable mixture direction and an unreachable reading, the two distinct defects the criterion separates.

Applications & Industry Use

Communications & signal processing

Analysis and synthesis transform pairs

Transform coding relies on an analysis map having an exact synthesis partner. Requiring the analysis operator to be injective and surjective on the signal space is precisely the requirement of perfect reconstruction; loss of surjectivity means some signals cannot be represented, loss of injectivity means distinct signals share a code word and cannot be told apart at the decoder.

Control systems

State transformations and observer design

Moving a state-space model into modal or controllable canonical form uses an invertible change of state variable. Invertibility guarantees the transformed model carries identical dynamics and that any control law designed in the new coordinates can be mapped back. A non-invertible change of variable silently discards state directions and produces a controller that cannot be implemented.

Colour science & imaging

Round-trip colour space conversion

Conversion between linear RGB and CIE XYZ is a linear map between three-dimensional spaces. Its invertibility is what allows an image to be moved into a working space for processing and returned without residual error. Where a conversion is deliberately non-invertible — a projection to greyscale, for instance — the lost dimension is exactly the kernel.

Cryptography

Diffusion layers in block ciphers

The linear mixing stage of a block cipher must be invertible so that decryption exists at all. Designers therefore select mixing maps over a finite field with trivial kernel, and the reversal rule for compositions is what lets a multi-round cipher be inverted by running the inverse rounds in reverse order.

Structural & mechanical engineering

Unique response to applied load

For a properly restrained structure the map from nodal displacements to nodal forces is invertible, so each load case produces exactly one displacement field. A rigid-body mechanism left in the model shows up as a non-trivial kernel, and the criterion identifies it as a failure of injectivity rather than a numerical accident.

Metrology & instrumentation

Sensor fusion and calibration inversion

A calibration matrix maps physical quantities to instrument channels. Invertibility of that map decides whether the physical state can be reconstructed from the readings. Testing the criterion at design time, before hardware is committed, distinguishes a redundant channel set from one that leaves a physical direction unobservable.

Design Considerations

Do not accept a one-sided inverse without checking dimensions

A left inverse proves injectivity and a right inverse proves surjectivity, but neither on its own proves invertibility. Only when the domain and codomain are finite-dimensional of equal dimension does one side imply the other. In infinite-dimensional settings such as sequence or function spaces, one-sided inverses exist routinely for maps that are not invertible, and code written on the assumption that they are will fail on the missing side.

Equal dimension is necessary but never sufficient

Matching dimensions rule nothing in. A map from $\mathbb{C}^3$ to $\mathbb{C}^3$ that collapses a plane onto a line has equal dimensions and is not invertible. Use the dimension comparison only as a fast rejection test, then compute the kernel to decide the remaining cases.

Test the transformation, report the transformation

Fixing bases and inspecting the resulting matrix is a legitimate way to decide invertibility, because the matrix is nonsingular exactly when the transformation is invertible. The conclusion, however, belongs to the map, not to the array; a different pair of bases produces a different matrix with the same verdict. State which bases were used whenever a matrix is offered as evidence.

Invertible is not the same as numerically safe

The criterion is a yes-or-no statement about exact arithmetic. A transformation can be invertible and still be so ill-conditioned that recovering an input from a measured output amplifies noise beyond usefulness. Pair the algebraic verdict with a condition number, or with the ratio of largest to smallest singular value of a matrix representation, before promising that an inverse is usable.

Prefer solving to inverting

Even when $T^{-1}$ exists, forming it explicitly is rarely the cheapest way to recover an input. Solving $T(\vec{u}) = \vec{v}$ directly for the particular $\vec{v}$ in hand is faster and better conditioned than building the full inverse and applying it. Build the inverse only when it will be reused across many right-hand sides or when the formula itself is the deliverable.

Verify on a basis, not on random samples

Because both $T$ and a candidate $S$ are linear, checking $S \circ T = I_U$ on a basis of $U$ and $T \circ S = I_V$ on a basis of $V$ establishes the identities for every vector. That is a finite, complete verification. Spot-checking a handful of arbitrary vectors is neither cheaper nor conclusive.

Standards & Codes

Notation, interchange and numerical standards that govern how this material is written down, stored and computed in production systems.

Applicable standards, conventions and reference implementations
Reference	Title	Relevance to this topic
`ISO 80000-2`	Quantities and units — Part 2: Mathematics	Fixes the notation used here for functions, composition and inverse, including the upright rendering of operator names such as $\dim$ and the superscript $-1$ convention for an inverse map rather than a reciprocal.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Governs the arithmetic in which a matrix representation of $T$ is tested. Exact-zero tests for a trivial kernel are unsafe under this arithmetic, which is why invertibility decisions in floating point require a stated tolerance.
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Supplies the practical invertibility test for a matrix representation: `xGETRF` factorises and `xGECON` estimates the reciprocal condition number, so the yes-or-no algebraic verdict arrives with a usability figure attached.
`ISO/IEC 18033-3`	Information technology — Security techniques — Encryption algorithms — Part 3: Block ciphers	Standardised block ciphers specify linear diffusion layers together with their inverses; the reversal rule for compositions is what defines the inverse cipher as the reverse-ordered chain of inverse rounds.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes every expression on this page as semantic markup, so composition symbols and inverse superscripts are machine-readable and available to assistive technology rather than baked into images.

Material Selection

For a mathematical topic, "material" is the numeric representation: the scalar field, storage format and precision the computation is built from.

Representation and precision selection
Representation	Select when	Trade-off
Exact rational arithmetic	Deciding invertibility of a small symbolic or integer-valued transformation where the verdict must be certain.	A zero pivot is genuinely zero, so the kernel computation is exact; intermediate fractions can grow rapidly and the method does not scale to large systems.
IEEE 754 binary64	Engineering transformations built from measured data, where the input already carries uncertainty.	Fast and universally supported, but invertibility becomes a tolerance decision and a nearly singular map is indistinguishable from a singular one without a condition estimate.
Finite field arithmetic	Cryptographic diffusion layers and error-control coding, where the transformation acts on symbols rather than magnitudes.	Exact, closed and free of conditioning concerns, but invertibility over one field says nothing about invertibility of the same coefficient pattern over another.
Real scalars only	Geometric and mechanical models where complex outputs would have no physical reading.	Halves storage and keeps every intermediate interpretable, at the cost of losing the eigen-structure arguments that require an algebraically closed field.
Symbolic representation of the map itself	The deliverable is a reusable formula for $T^{-1}$ rather than a numeric answer for one input.	Produces a closed-form recovery rule that documents itself and can be code-generated, but expression growth makes it impractical beyond a handful of dimensions.
Fixed-point arithmetic	Embedded targets that must apply an inverse calibration map in real time without a floating-point unit.	Deterministic timing and low cost, but the scaled inverse coefficients must be range-analysed in advance or the recovery saturates on legitimate inputs.

Manufacturing Notes

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

Cost of the decision versus cost of the inverse

For a transformation between spaces of dimension $n$ , deciding invertibility from a matrix representation costs one LU factorisation, about $\tfrac{2}{3}n^3$ operations. Forming the inverse explicitly costs roughly $2n^3$ — three times as much — because it amounts to solving $n$ systems. Applying an already-formed inverse costs $2n^2$ per input, the same as a single triangular solve pair, so the inverse only repays its construction cost after many reuses.

Constructing the inverse formula by hand

Write the general output vector with symbolic entries, set it equal to $T$ applied to a symbolic input, and solve the resulting system once. Because the system is the same for every target, one solution in symbols yields the whole inverse map. Reading the coefficients of that solution against the entries of the output gives the formula directly, and its linearity in those entries is a useful self-check.

Verification strategy

Apply the candidate inverse to the image of each basis vector of the domain and confirm the original basis vector returns; then apply the transformation to the inverse image of each basis vector of the codomain. Two finite sweeps of $\dim(U) + \dim(V)$ evaluations settle both identities completely. Verifying only one direction is the most common shortcut and the one that hides a missing surjectivity.

Library behaviour

Numerical libraries expose invertibility through solvers rather than through a boolean. NumPy's numpy.linalg.solve raises on a singular matrix, numpy.linalg.cond quantifies how close to singular it is, and numpy.linalg.inv exists mainly for expository code. Symbolic systems are different: SymPy's Matrix.inv works over the rationals and reports exact singularity, which is the correct tool when the transformation is defined by symbols rather than measurements.

Chained pipelines and inverse ordering

When several invertible stages are composed in software, store the stages in a list and invert by walking the list backwards applying each stage's inverse. Implementing the reversal by inverting a fused matrix product instead is numerically worse and destroys the ability to diagnose which stage is responsible for a conditioning problem.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Accepting a one-sided inverse	high	Only $S \circ T = I_U$ is verified, and $T \circ S = I_V$ is assumed to follow.	Check both compositions explicitly, or first establish that the domain and codomain are finite-dimensional of equal dimension, where one side does imply the other.
Inferring invertibility from equal dimensions	high	Treating $\dim(U) = \dim(V)$ as the criterion rather than as a necessary condition.	Use the dimension comparison only to reject; always compute the kernel before concluding that a map is invertible.
Forgetting the order reversal in a composition	medium	Writing $(S \circ T)^{-1} = S^{-1} \circ T^{-1}$ by analogy with the forward composition.	Check the domains: $(S \circ T)^{-1}$ must accept vectors from $W$ , and only $S^{-1}$ does. The correct order is $T^{-1} \circ S^{-1}$ .
Confusing the inverse map with a pointwise reciprocal	medium	Reading the superscript $-1$ in $T^{-1}$ as an exponent applied to output values.	Reserve $T^{-1}$ for the inverse transformation and write reciprocals explicitly. Note also that the pre-image notation $T^{-1}(\vec{v})$ is meaningful even for non-invertible $T$ , where it denotes a set.
Declaring invertibility from a floating-point determinant	high	Testing a computed determinant against exact zero on rounded data, where scaling alone can move the determinant by many orders of magnitude.	Use a reciprocal condition estimate or the smallest singular value of a matrix representation, both of which are scale-aware, instead of the determinant.
Reporting invertibility without naming the bases	low	A matrix representation is tested, and its nonsingularity is quoted as though it were a property of the array.	Record the domain and codomain bases alongside any matrix used as evidence, and state the conclusion as a property of the transformation.
Assuming an invertible model gives a usable reconstruction	high	Treating the algebraic verdict as a guarantee that measured outputs can be inverted in the presence of noise.	Report a condition number with the verdict, and specify the noise amplification factor that the inverse will apply to measurement error.
Skipping the linearity check	medium	Applying the criterion to a map that is bijective but not linear, such as an affine map with a non-zero offset.	Confirm additivity and scalar homogeneity before invoking the criterion; an affine map has an inverse, but it is not a linear transformation and the kernel test does not apply to it.

FAQs

Why is being injective not enough for a linear transformation to be invertible?

Injectivity guarantees that distinct inputs give distinct outputs, so no output is claimed by two inputs. It says nothing about outputs that are claimed by none. If some $\vec{v} \in V$ lies outside the range, there is no legitimate value for $T^{-1}(\vec{v})$ , and the inverse fails to be defined on all of $V$ . Only when the codomain happens to have the same finite dimension as the domain does injectivity force surjectivity as well.

Does the inverse of a linear transformation have to be linear?

It has to be, but that is a conclusion rather than an assumption. The definition asks only for a function satisfying both composition identities. Pushing sums and scalar multiples through $T$ and cancelling with $T^{-1}$ shows that $T^{-1}$ respects both operations, so it is automatically a linear transformation — and, being invertible with inverse $T$ , it sits in the same class as $T$ .

Is the inverse unique?

Yes. If $S$ and $S'$ both satisfy the two identities, then $S = S \circ I_V = S \circ (T \circ S') = (S \circ T) \circ S' = I_U \circ S' = S'$ . Uniqueness is what licenses the notation $T^{-1}$ and the phrase the inverse.

How does this criterion relate to a matrix being nonsingular?

Apply it to the transformation $T(\vec{x}) = A\vec{x}$ induced by a matrix $A$ . Injectivity is the statement $\mathcal{N}(A) = \{\vec{0}\}$ and surjectivity is $\mathcal{C}(A) = \mathbb{C}^m$ ; both can hold only for a square matrix of full rank, which is exactly a nonsingular matrix. The abstract criterion is the more general statement because it applies to spaces of polynomials, matrices and functions where no array exists until a basis is chosen.

If two spaces have the same dimension, is every linear transformation between them invertible?

No. Equal dimension is necessary for invertibility but says nothing about a particular map. The zero transformation from $\mathbb{C}^4$ to $\mathbb{C}^4$ has matching dimensions and a four-dimensional kernel. What equal dimension buys is that one test suffices: a trivial kernel then forces surjectivity, and a full range then forces injectivity.

Why does the inverse of a composition reverse the order of the factors?

Because undoing a sequence requires undoing the last operation first. If $S \circ T$ applies $T$ then $S$ , the reverse process must strip off $S$ before it can address $T$ , which is the composition $T^{-1} \circ S^{-1}$ . A domain check makes the same point: $(S \circ T)^{-1}$ takes inputs from $W$ , and only $S^{-1}$ accepts them.

Can a linear transformation between spaces of different dimensions ever be invertible?

Not in the finite-dimensional case. An invertible map is injective, which forces $\dim(U) \le \dim(V)$ , and surjective, which forces $\dim(U) \ge \dim(V)$ . The two inequalities leave only equality. This is also the reason invertible transformations are the exact mechanism by which two vector spaces can be declared structurally identical.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section IVLT. Licensed under the GNU Free Documentation License v1.2.
Halmos, P. R. Finite-Dimensional Vector Spaces, 2nd edition. Springer, 1974.
Axler, S. Linear Algebra Done Right, 3rd edition. Springer, 2015.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
Anderson, E. et al. LAPACK Users' Guide, 3rd edition. Society for Industrial and Applied Mathematics, 1999.
Higham, N. J. Accuracy and Stability of Numerical Algorithms, 2nd edition. SIAM, 2002.

AI Suggested Questions

Give me a linear transformation between infinite-dimensional sequence spaces that has a left inverse but no genuine inverse, and show exactly which composition identity fails.
For a transformation from $P_3$ to $M_{22}$ defined by four coefficient formulas, walk through the kernel computation and then build the inverse map in closed form.
How does the condition number of a matrix representation change the practical meaning of an invertibility verdict, and what threshold should trigger a redesign?
Show why an affine map has an inverse function even though it is not a linear transformation, and explain why the kernel test does not apply to it.
Compare the operation counts of deciding invertibility, forming the inverse, and solving a single system for a transformation of dimension 500.
Construct two invertible transformations whose composition is invertible, and verify the reversal rule for the inverse by evaluating both sides on a basis.

Related Calculators

Invertibility Tester

Enter a linear transformation by its action on a basis and receive kernel, range, and an invertible or not verdict with the deciding computation shown.

Inverse Transformation Builder

Solve $T(\vec{u}) = \vec{v}$ symbolically for a general target and return a closed-form formula for $T^{-1}$ .

Kernel & Range Calculator

Compute bases for $\mathcal{K}(T)$ and $\mathcal{R}(T)$ , with dimensions reported against the domain and codomain.

Condition Number Calculator

Estimate how much a matrix representation of an invertible map amplifies measurement error during reconstruction.