Systems of Equations and Linear Transformations

Overview

Systems of linear equations and linear transformations are usually taught as separate subjects: one is a computational technique, the other a structural theory. They are the same subject viewed from two positions. Every $m \times n$ coefficient matrix $A$ defines a linear transformation $T\colon \mathbb{C}^n \to \mathbb{C}^m$ by $T(\vec{x}) = A\vec{x}$ , and once that identification is made, every question about the system $A\vec{x} = \vec{b}$ becomes a question about $T$ .

The dictionary is short and exact. The kernel of $T$ is the null space of $A$ . The range of $T$ is the column space of $A$ . The solution set of $A\vec{x} = \vec{b}$ is the pre-image $T^{-1}(\vec{b})$ . From those three correspondences everything else follows: the system is consistent exactly when $\vec{b}$ lies in the range; when it is consistent, its solution set is a translate of the kernel by any single particular solution; and the solution is unique exactly when the kernel is trivial.

The reformulation is not decoration. It replaces case analysis with structure. Instead of three separate possibilities for a solution set — empty, a single vector, an infinite family — there is one statement: the pre-image is either empty or a coset of the kernel, and the dimension of the kernel is the number of free parameters. The three cases are the values that one integer can take, together with a single membership test.

It also makes dimension arguments available before any arithmetic. A system with more unknowns than equations induces a map from a larger space to a smaller one, so its nullity is at least the difference and it cannot have a unique solution: either no solution or infinitely many. A square system whose induced map has trivial kernel has an invertible transformation behind it, and the coefficient matrix is nonsingular, so every right-hand side has exactly one solution. Both conclusions come from counting dimensions, not from reducing a matrix.

Definition

Transformation Induced by a Matrix

Given an $m \times n$ matrix $A$ , the function $T\colon \mathbb{C}^n \to \mathbb{C}^m$ defined by $T(\vec{x}) = A\vec{x}$ is a linear transformation. Linearity is inherited directly from the distributive and scalar properties of the matrix-vector product: $A(\vec{x} + \vec{y}) = A\vec{x} + A\vec{y}$ and $A(\alpha\vec{x}) = \alpha (A\vec{x})$ .

The domain dimension is the number of columns of $A$ — the number of unknowns. The codomain dimension is the number of rows — the number of equations.

Pre-Image of a Vector

For a linear transformation $T\colon U \to V$ and a vector $\vec{v} \in V$ , the pre-image of $\vec{v}$ is the set of all inputs mapped to it:
$T^{-1}(\vec{v}) = \{\, \vec{u} \in U \mid T(\vec{u}) = \vec{v} \,\}$ .
The pre-image is defined whether or not $T$ is invertible, and it may be empty.

The notation $T^{-1}(\vec{v})$ denotes a set. It coincides with the value of the inverse map only when $T$ happens to be invertible, in which case that set is a singleton.

Solution Set as a Pre-Image

The solution set of the system $A\vec{x} = \vec{b}$ is exactly the pre-image of $\vec{b}$ under the induced transformation:
$\mathcal{S} = T^{-1}(\vec{b})$ .
When it is non-empty, it has the form $\vec{w} + \mathcal{K}(T)$ for any particular solution $\vec{w}$ , and its number of free parameters equals the nullity of $T$ .

Concepts

The dictionary between the two languages

Fix an $m \times n$ matrix $A$ and its induced transformation $T$ . The kernel is the set of $\vec{x}$ with $A\vec{x} = \vec{0}$ , which is the definition of the null space, so $\mathcal{K}(T) = \mathcal{N}(A)$ . The range is the set of all $A\vec{x}$ , and since $A\vec{x}$ is the linear combination of the columns of $A$ with weights taken from $\vec{x}$ , the range is the span of the columns: $\mathcal{R}(T) = \mathcal{C}(A)$ .

Those two identifications propagate. The nullity of $T$ is the nullity of $A$ ; the rank of $T$ is the rank of $A$ ; and the rank-nullity identity for transformations, applied to the domain $\mathbb{C}^n$ , reproduces the statement that the rank and nullity of a matrix sum to its column count. Nothing new has to be proved on the matrix side once the transformation side is in place.

Consistency is a membership question

The system $A\vec{x} = \vec{b}$ has a solution precisely when some input maps to $\vec{b}$ , that is, when $\vec{b} \in \mathcal{R}(T)$ . Translated back, this says the system is consistent exactly when $\vec{b}$ lies in the column space of $A$ — when $\vec{b}$ can be written as a linear combination of the columns. The augmented-matrix pivot test familiar from row reduction is simply the computational realisation of that membership check.

This reframing separates two properties that are easy to confuse. Consistency depends on the pair $(A, \vec{b})$ , since it asks where $\vec{b}$ sits relative to the range. Uniqueness depends on $A$ alone, since it asks only whether the kernel is trivial. A single matrix can therefore give an inconsistent system for one right-hand side and an infinite solution set for another, and the transformation view makes clear why those are independent phenomena.

Every non-empty pre-image is a translate of the kernel

Suppose $\vec{w}$ satisfies $A\vec{w} = \vec{b}$ . If $\vec{z}$ lies in the kernel then $A(\vec{w} + \vec{z}) = \vec{b} + \vec{0} = \vec{b}$ , so every kernel vector added to $\vec{w}$ gives another solution. Conversely if $\vec{x}$ is any solution then $A(\vec{x} - \vec{w}) = \vec{0}$ , so $\vec{x} - \vec{w}$ lies in the kernel. The two directions together give $\mathcal{S} = \vec{w} + \mathcal{K}(T)$ exactly.

This is the structural content of the particular-plus-homogeneous decomposition. It also explains a point that troubles many practitioners: the solution set of an inhomogeneous system is not a subspace, because it does not contain the zero vector unless $\vec{b} = \vec{0}$ . It is a coset — a subspace shifted off the origin — and its dimension as an affine object is the nullity.

Dimension counting predicts the outcome

Consider a system of three equations in four unknowns. The induced map goes from $\mathbb{C}^4$ to $\mathbb{C}^3$ , so its rank is at most three and the rank-nullity identity forces $n(T) \ge 4 - 3 = 1$ . The kernel is non-trivial before a single entry of $A$ is inspected. Consequently the system cannot have a unique solution: it is either inconsistent, or consistent with an infinite solution set of at least one free parameter.

The same counting delivers the reverse conclusion. For a square system with $n$ equations in $n$ unknowns, a trivial kernel gives $n(T) = 0$ and therefore $r(T) = n$ , so $T$ is injective and surjective, hence invertible. Every right-hand side then has exactly one solution, and the coefficient matrix is nonsingular. A single kernel computation has settled solvability for all right-hand sides simultaneously.

The invertible case and why an explicit inverse is still the wrong tool

When $T$ is invertible the unique solution can be written $\vec{x} = T^{-1}(\vec{b}) = A^{-1}\vec{b}$ . That formula is theoretically decisive and computationally poor. Forming $A^{-1}$ costs roughly three times as much as an LU factorisation and applies more rounding error to the result than a factor-and-solve sequence does.

The transformation view suggests the better pattern directly. The map $T$ is fixed and only the target $\vec{b}$ varies, so the expensive object — the factorisation that encodes $T$ — should be computed once and reused for each new right-hand side. That is exactly what a factor-once, solve-many workflow does, and it is why production libraries expose solve and factorisation objects rather than encouraging matrix inversion.

Every theorem about systems has a transformation analogue

The correspondence runs in both directions and covers the whole first half of the subject. The statement that a consistent system with a free variable has infinitely many solutions is the statement that a non-empty pre-image under a map with non-trivial kernel is infinite. The statement that a homogeneous system always has the trivial solution is the statement that $T(\vec{0}) = \vec{0}$ . The statement that a nonsingular coefficient matrix gives a unique solution for every right-hand side is the statement that an invertible transformation has singleton pre-images.

The practical value is that the transformation statements survive the move to abstract spaces. Once a question about polynomials, matrices or functions has been posed as a linear transformation, the entire apparatus built for systems of equations applies without modification, and the coefficient array reappears only when a basis is chosen and a numerical answer is required.

From a system of equations to a structural verdict

Attach the transformationRead

A

as the map

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

from

\mathbb{C}^n

\mathbb{C}^m

, where

n

is the number of unknowns and

m

the number of equations.

Count dimensions firstIf

n > m

then

n(T) \ge n - m > 0

, so a unique solution is already impossible. If

n < m

the map cannot be surjective, so some right-hand sides are unreachable.

Compute the kernelSolve

A\vec{x} = \vec{0}

. Its dimension is the number of free parameters any solution set will carry.

Test membership of the rangeDecide whether

\vec{b}

lies in

\mathcal{C}(A)

. A pivot in the augmented column says it does not, and the pre-image is empty.

Find one particular solutionAny single

\vec{w}

with

A\vec{w} = \vec{b}

will do; the reduction that tested consistency produces one at no extra cost.

Report the cosetState the solution set as

\vec{w} + \mathcal{K}(T)

with an explicit kernel basis, rather than as a single vector.

Equations

Transformation induced by a coefficient matrix

EQ-SLELT-01

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

The domain dimension is the number of unknowns and the codomain dimension is the number of equations.

Kernel is the null space, range is the column space

EQ-SLELT-02

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

The two identifications that carry every structural result across between the matrix language and the transformation language.

Solution set as a pre-image

EQ-SLELT-03

Solution set as a pre-image

The pre-image is defined for any transformation and any target; it is empty exactly when the system is inconsistent.

Consistency criterion

EQ-SLELT-04

S \neq \emptyset \Leftrightarrow \vec{b} \in C (A)

The system has a solution exactly when the right-hand side is a linear combination of the columns of the coefficient matrix.

Structure of a non-empty solution set

EQ-SLELT-05

Structure of a non-empty solution set

One particular solution plus the whole kernel. The set is a coset, not a subspace, unless $\vec{b} = \vec{0}$ .

Rank-nullity for the induced map

EQ-SLELT-06

Rank-nullity for the induced map

Applied to the induced transformation, the rank-nullity identity reproduces the statement that the rank and nullity of a matrix sum to its column count.

Unique solution for every right-hand side

EQ-SLELT-07

Unique solution for every right-hand side

For a square coefficient matrix the three statements coincide. The formula is a statement of uniqueness, not a recommended computational route.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Coefficient matrix	The $m \times n$ array of coefficients defining both the system and the transformation	m x n matrix
$T$	Induced transformation	The map $\vec{x} \mapsto A\vec{x}$ attached to the coefficient matrix	linear map C^n to C^m
$\vec{b}$	Vector of constants	The right-hand side; the target whose pre-image is sought	vector in C^m
$\vec{w}$	Particular solution	Any single vector satisfying $A\vec{w} = \vec{b}$ , used to anchor the coset	vector in C^n
$S$	Solution set	All vectors satisfying the system; equal to the pre-image of $\vec{b}$	subset of C^n
$N (A)$	Null space	Solutions of the homogeneous system; the kernel of the induced map	subspace of C^n
$C (A)$	Column space	Span of the columns of $A$ ; the range of the induced map	subspace of C^m
$n$	Number of unknowns	Column count of $A$ and dimension of the domain	positive integer
$m$	Number of equations	Row count of $A$ and dimension of the codomain	positive integer

Worked Numerical Example

Problem statement

A pipe network has four adjustable flow settings and three measured junction balances. The relationship is linear with a fixed coefficient matrix. Analyse the induced transformation once, then use it to characterise the solution sets for two different measured balance vectors — one achievable and one not.

Attach the transformation
Write the coefficient matrix and read it as a map from the four-dimensional space of flow settings to the three-dimensional space of junction balances. Note the dimensions before any computation: four unknowns and three equations.
$A = [\begin{matrix} 1 & 2 & 1 & 3 \\ 2 & 4 & 3 & 8 \\ 1 & 2 & 2 & 5 \end{matrix}], T : ℂ^{4} \to ℂ^{3}, T (\vec{x}) = A \vec{x}$
Predict from dimensions alone
The domain has dimension four and the codomain three, so $r(T) \le 3$ and the rank-nullity identity gives $n(T) \ge 4 - 3 = 1$ . The kernel is non-trivial, so $T$ is not injective and no right-hand side can produce a unique solution. Whatever the measurements turn out to be, the answer will be either no setting at all or an infinite family of settings.
Compute the kernel
Row-reduce $A$ . Subtracting twice row one from row two gives $\begin{bmatrix} 0 & 0 & 1 & 2 \end{bmatrix}$ , and subtracting row one from row three gives the same, so the third row clears to zero. Clearing above the second pivot leaves two non-zero rows with pivots in columns one and three, so $x_2$ and $x_4$ are free.
$RREF (A) = [\begin{matrix} 1 & 2 & 0 & 1 \\ 0 & 0 & 1 & 2 \\ 0 & 0 & 0 & 0 \end{matrix}]$
Write a kernel basis and record the two numbers
Setting each free variable to one in turn gives two independent kernel vectors. Direct substitution confirms both: for the first, $A$ applied to it gives $\begin{bmatrix} -2+2 & -4+4 & -2+2 \end{bmatrix}^{t} = \vec{0}$ , and for the second, $\begin{bmatrix} -1-2+3 & -2-6+8 & -1-4+5 \end{bmatrix}^{t} = \vec{0}$ .
$K (T) = span ({[\begin{matrix} - 2 \\ 1 \\ 0 \\ 0 \end{matrix}], [\begin{matrix} - 1 \\ 0 \\ - 2 \\ 1 \end{matrix}]}), n (T) = 2, r (T) = 2$
Note that the map is not surjective either
The rank is two but the codomain has dimension three, so $\mathcal{R}(T) = \mathcal{C}(A)$ is a two-dimensional plane inside $\mathbb{C}^3$ . Balance vectors off that plane correspond to no setting whatsoever. This is a property of the network, established once, and it applies to every measurement that will ever be taken.
Achievable measurement: characterise the full solution set
Take $\vec{b} = \begin{bmatrix} 2 & 5 & 3 \end{bmatrix}^{t}$ . Reducing the augmented matrix leaves the constants as $1$ in each pivot row and $0$ in the zero row, so no pivot appears in the augmented column and $\vec{b} \in \mathcal{C}(A)$ . Setting both free variables to zero reads off a particular solution, and the full pre-image is that vector translated by the kernel.
$\vec{w} = [\begin{matrix} 1 \\ 0 \\ 1 \\ 0 \end{matrix}], S = \vec{w} + K (T)$
Verify the particular solution and the coset
Direct multiplication gives $A\vec{w} = \begin{bmatrix} 1+1 & 2+3 & 1+2 \end{bmatrix}^{t} = \begin{bmatrix} 2 & 5 & 3 \end{bmatrix}^{t}$ , as required. Adding any multiple of either kernel vector changes the setting but not the measured balances, so the two kernel directions are exactly the recirculation patterns the instrumentation cannot see.
Unachievable measurement: an empty pre-image
Now take $\vec{d} = \begin{bmatrix} 2 & 5 & 4 \end{bmatrix}^{t}$ . The same row operations give constants $1$ and $2$ in the two pivot rows, so the third row becomes $\begin{bmatrix} 0 & 0 & 0 & 0 \end{bmatrix}$ with constant $1$ — the contradiction $0 = 1$ . Hence $\vec{d} \notin \mathcal{C}(A)$ and the pre-image is empty.
$Unachievable measurement: an empty pre-image$

Result

One analysis of the transformation answers both measurement cases and every future one. The nullity of two says the instrumentation leaves a two-parameter family of settings indistinguishable, so control authority exceeds observability by two degrees of freedom. The rank of two against a codomain of dimension three says a whole direction of balance readings is physically unreachable: a measurement such as the second one cannot arise from any legitimate setting and therefore indicates a sensor fault or a leak, not an unsolvable control problem.

Applications & Industry Use

Power systems

Network solvability and the reference bus

The linearised nodal admittance relation between bus angles and injected power has a kernel spanned by a uniform shift of all angles, since only differences are physical. That one-dimensional kernel is why a reference bus must be fixed before the system becomes uniquely solvable, and the transformation view identifies the constraint as the minimum needed to make the induced map injective.

Transport & traffic engineering

Link flows consistent with observed counts

Counting stations observe a subset of links, and the map from link flows to observed counts is typically far from injective. Its kernel enumerates the circulation patterns invisible to the sensor set, and the size of that kernel is a direct measure of how much additional instrumentation is required before flows can be identified rather than merely constrained.

Chemical process engineering

Species balances and independent reactions

An element balance across a reacting system produces a homogeneous system whose kernel contains the admissible reaction extents. The nullity is the number of independent reactions, and consistency of an inhomogeneous balance against measured feed and product streams is precisely the membership test for the measured vector in the column space.

Medical imaging

Tomographic reconstruction as an underdetermined system

The projection operator maps an image to a set of line integrals. With fewer projections than pixels the induced map has a large kernel, so infinitely many images match the data exactly. Reconstruction algorithms select one element of that coset, and the kernel is the formal description of what the additional prior information is being used to decide.

Computer-aided design

Geometric constraint solving

A sketch with dimensional constraints yields a linear system in the incremental coordinates. A non-trivial kernel signals an under-constrained sketch whose remaining degrees of freedom are exactly the kernel basis; an inconsistent right-hand side signals over-constrained geometry that no adjustment can satisfy. The two defects require different fixes, and the transformation view separates them cleanly.

Economics

Input-output production planning

A Leontief model relates gross output to final demand through a coefficient matrix. Solvability for every demand vector is invertibility of the induced transformation, and where the map is not invertible, the unreachable part of the codomain describes demand patterns that the production structure cannot deliver at any activity level.

Design Considerations

Separate the property of $A$ from the property of the pair

Uniqueness is decided by the kernel and therefore by $A$ alone; consistency is decided by membership of $\vec{b}$ in the range and therefore by the pair. Analyse the matrix once, and treat each new right-hand side as a cheap membership test against a structure already computed. Conflating the two leads to repeated full analyses when only the target has changed.

Factor once, solve many

The transformation is fixed and only the target varies in most engineering settings — load cases, time steps, demand scenarios. Compute a factorisation of $A$ once and apply it to each right-hand side, at a cost of $O(n^2)$ per solve against $O(n^3)$ for the factorisation. Repeating a full solve for every target is the most common avoidable expense in linear modelling code.

Never form the inverse to solve a system

The identity $\vec{x} = A^{-1}\vec{b}$ is a statement about uniqueness, not an algorithm. Explicit inversion costs about three times a factorisation, requires more storage, and yields a less accurate answer than triangular solves. Reserve an explicit inverse for the rare cases where the entries themselves are the deliverable, such as a covariance matrix.

Report the solution set, not a solution

When the nullity is positive, returning a single vector conceals the ambiguity. Report a particular solution together with an explicit kernel basis, so downstream users can see how many degrees of freedom remain and in which directions. Least-squares and pseudo-inverse routines silently select the minimum-norm element of the coset, which is a defensible choice only if it is stated.

Decide what to do when the target is outside the range

An empty pre-image is a modelling result, not a failure. Either the data is inconsistent with the model, or the model is missing a degree of freedom. Approximating with a least-squares solution replaces the exact question with a different one; that substitution should be a deliberate decision recorded in the analysis, not a side effect of calling a routine that never reports inconsistency.

Treat structural and numerical rank as different quantities

The rank implied by the sparsity pattern of $A$ — which columns can in principle be independent — bounds but does not determine the numerical rank. A network model can be structurally full rank and numerically near-deficient because of extreme parameter ratios. Check both, and report the smallest singular value alongside the rank whenever the model is derived from physical data.

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 for the matrix-vector product, set-builder descriptions of solution sets, and the distinction between the inverse map and the pre-image of a set, all of which appear on this page.
`BLAS Level 2`	Basic Linear Algebra Subprograms, matrix-vector operations	The routine `xGEMV` is the induced transformation in executable form: evaluating $T(\vec{x})$ is precisely one general matrix-vector product, and its performance characteristics set the cost of every forward evaluation.
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Provides the solve paths that correspond to each structural case: `xGESV` for a square nonsingular system, `xGELSD` for rank-deficient or overdetermined cases where the target lies outside the range and a least-squares surrogate is used.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Determines whether a residual small enough to be rounding noise is reported as consistency. Any range-membership test on measured data must compare the residual against a tolerance derived from this arithmetic rather than against zero.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the augmented matrices and set-builder expressions here as structured markup, keeping the distinction between a solution set and a solution vector machine-readable.

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 binary64 storage	Small to moderate systems where every coefficient is potentially non-zero, such as calibration or constraint blocks.	Simplest and fastest per entry with mature library support, but memory grows as the square of the dimension and sparsity in the model is wasted.
Sparse compressed storage with a direct solver	Network, finite element and process models where each equation involves only a few unknowns.	Handles dimensions far beyond dense limits, but fill-in during factorisation depends heavily on the ordering, and the kernel basis it produces is rarely sparse.
Sparse storage with an iterative solver	Very large systems where even a sparse factorisation is uneconomic and an approximate solution suffices.	Memory is proportional to the number of non-zeros only, but convergence depends on conditioning and the method returns one solution without revealing the kernel.
Exact rational arithmetic	Determining the structural verdict — rank, nullity, consistency — for a small symbolic or integer model.	Gives an unambiguous kernel dimension with no tolerance to defend, but coefficient growth makes it unusable beyond modest sizes.
Extended precision for the residual check	Deciding consistency when the residual is close to the rounding level of the working precision.	Makes a marginal membership test decisive at low incremental cost, but only if the residual is computed from the original data rather than from reduced quantities.
Fixed-point arithmetic	Embedded solvers applying a precomputed factorisation to new right-hand sides in real time.	Deterministic timing without a floating-point unit, but ranges must be analysed offline and any near-singularity in the model becomes an overflow rather than a warning.

Manufacturing Notes

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

One reduction answers three questions

Row-reducing the augmented matrix $[\, A \mid \vec{b} \,]$ simultaneously tests consistency, produces a particular solution and exposes the free columns that index a kernel basis. There is no need for separate passes. The pivot columns give the rank, the non-pivot variable columns give the nullity, and a pivot in the final column gives an empty pre-image.

Cost of the analysis versus cost of a solve

Reducing an $m \times n$ augmented matrix costs about $O(m n \min(m,n))$ operations, dominated by the structural analysis. Once a factorisation of $A$ exists, each additional right-hand side costs only two triangular solves, about $2n^2$ operations. When many targets share one coefficient matrix, more than 90% of the total work belongs to the one-off analysis.

Verifying against the original data

Check a claimed solution by computing the residual $A\vec{w} - \vec{b}$ from the original matrix and right-hand side, never from the reduced forms. An arithmetic slip during reduction produces a reduced system that is internally consistent but no longer equivalent to the input, and only a residual against the original data will detect it.

Library behaviour differs by structural case

NumPy's numpy.linalg.solve handles only square nonsingular systems and raises otherwise; numpy.linalg.lstsq accepts any shape and silently returns a least-squares answer whether or not the system is consistent, together with a rank and residual that should always be inspected. SciPy's scipy.linalg.null_space supplies the kernel basis needed to report the full coset. SymPy's linsolve returns a parametrised solution set directly, which matches the structural description most closely.

Presenting the answer

The natural output format is a particular solution plus a list of kernel basis vectors, each labelled with the free variable that generated it. This makes the number of degrees of freedom visible, allows any specific solution to be reconstructed by choosing parameters, and communicates the ambiguity to engineers who will otherwise assume that a returned vector is the only answer.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Testing consistency by reducing the coefficient matrix alone	high	The constants column is dropped, discarding exactly the information that decides membership of the range.	Always reduce the augmented matrix, and check for a pivot in the final column before interpreting anything else.
Reporting one solution when the pre-image is infinite	high	A solver returns a single vector and the positive nullity is not inspected.	Report a particular solution together with a kernel basis, and treat any solver that cannot supply the nullity as unsuitable for structural work.
Confusing the kernel with the solution set	medium	Treating the inhomogeneous solution set as a subspace and expecting it to contain the zero vector.	Remember that $\mathcal{S} = \vec{w} + \mathcal{K}(T)$ is a coset. It is a subspace only in the homogeneous case $\vec{b} = \vec{0}$ .
Assuming a square system has a unique solution	high	Equating equation count with unknown count and concluding solvability without testing the kernel.	Squareness only makes injectivity and surjectivity equivalent. Compute the kernel, or a condition estimate, before promising a unique answer.
Silently substituting least squares for a solution	high	Calling a least-squares routine on an inconsistent system and reporting the output as though the equations were satisfied.	Inspect the returned residual. A non-negligible residual means the target lies outside the range and the answer solves a different problem from the one posed.
Forming the inverse to solve a system	medium	Translating $\vec{x} = A^{-1}\vec{b}$ literally into code.	Use a factorisation and triangular solves. Reserve explicit inversion for cases where the entries of the inverse are themselves the required output.
Using the row count as the domain dimension	medium	Applying the rank-nullity identity with the number of equations instead of the number of unknowns.	The domain of $T(\vec{x}) = A\vec{x}$ is $\mathbb{C}^n$ with $n$ the column count. Rank and nullity sum to the number of unknowns.
Treating an empty pre-image as a numerical failure	medium	Interpreting inconsistency as a solver problem and increasing tolerances until an answer appears.	An empty pre-image is a valid structural result. Investigate whether the data or the model is at fault before relaxing any tolerance.

FAQs

Why is the solution set of an inhomogeneous system not a subspace?

Because it does not contain the zero vector unless the right-hand side is zero. It is a coset $\vec{w} + \mathcal{K}(T)$ : the kernel, which is a genuine subspace, translated by any particular solution. Sums of two solutions are not solutions, but the difference of two solutions always lies in the kernel.

What is the difference between the pre-image

T^{-1}(\vec{b})

and the inverse transformation?

The pre-image is a set and is defined for every linear transformation, including maps with no inverse; it may be empty or infinite. The inverse transformation is a function and exists only when $T$ is invertible. The notation is shared because when $T$ is invertible every pre-image is a single vector, and the two readings agree.

How does the transformation view explain the three possibilities for a solution set?

The pre-image of $\vec{b}$ is empty when $\vec{b}$ lies outside the range. When it is non-empty it equals $\vec{w} + \mathcal{K}(T)$ , which is a single vector if the kernel is trivial and an infinite family otherwise. The three cases are therefore one membership test combined with one integer, the nullity, rather than three separate phenomena.

Can a system have exactly two solutions?

No. If $\vec{x}$ and $\vec{y}$ are distinct solutions, their difference is a non-zero kernel vector, and adding any scalar multiple of that difference to $\vec{x}$ produces another solution. Over an infinite field of scalars this yields infinitely many. Linearity is what forbids intermediate cases.

Why does having more unknowns than equations guarantee a non-trivial kernel?

The induced map goes from $\mathbb{C}^n$ to $\mathbb{C}^m$ with $n > m$ . The rank cannot exceed $m$ , so the rank-nullity identity gives $n(T) = n - r(T) \ge n - m > 0$ . The consequence for the system is that it cannot have a unique solution: it is either inconsistent or has infinitely many solutions.

If the coefficient matrix is nonsingular, why not just compute the inverse?

Because inversion is a more expensive and less accurate route to the same answer. Forming $A^{-1}$ costs roughly three times an LU factorisation and then still requires a matrix-vector product per right-hand side, while the factorisation plus triangular solves is cheaper and has better backward error. The inverse is worth forming only when its entries are themselves required.

Does this correspondence work for abstract vector spaces too?

Yes, and that is the point of stating it in transformation language. A map between spaces of polynomials or matrices has a kernel, a range and pre-images with exactly the same structure, so the entire theory of solution sets transfers without modification. A coefficient array reappears only once bases are chosen and a numerical answer is wanted.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section IVLT, Subsection SLELT. Licensed under the GNU Free Documentation License v1.2.
Strang, G. Linear Algebra and Its Applications, 4th edition. Cengage, 2006.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013.
Davis, T. A. Direct Methods for Sparse Linear Systems. Society for Industrial and Applied Mathematics, 2006.
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.

AI Suggested Questions

Take a $5 \times 7$ coefficient matrix and show how one row reduction of the augmented matrix delivers consistency, a particular solution and a kernel basis at once.
Explain why the difference of any two solutions of $A\vec{x} = \vec{b}$ lies in the null space, and how that fact alone forces the three-case classification of solution sets.
For a DC power flow model, show that the kernel of the nodal map is spanned by the all-ones vector and explain what fixing a reference bus does to the induced transformation.
Compare the cost of solving twenty right-hand sides by repeated Gaussian elimination against a single LU factorisation with twenty triangular solve pairs.
When a least-squares routine returns an answer with a large residual, how should the result be reported so that readers understand the original system was inconsistent?
Reformulate the statement that a homogeneous system with more unknowns than equations has a non-trivial solution purely in terms of ranks and nullities of the induced map.

Related Calculators

Linear System Solver

Solve $A\vec{x} = \vec{b}$ and report consistency, rank, nullity and the full solution set as a particular solution plus a kernel basis.

Pre-Image Calculator

Compute $T^{-1}(\vec{b})$ for a matrix-induced transformation, returning the empty set or an explicit coset description.

Null Space Calculator

Produce a basis for $\mathcal{N}(A)$ with the free variables identified, giving the degrees of freedom in every solution set for that matrix.

Column Space Membership Tester

Decide whether a given right-hand side lies in $\mathcal{C}(A)$ and report the residual when it does not.