Vector Form of Solution Sets | KEVOS® Mathematics

Q: Why does the fixed vector have zeros in the free positions?

Because it is defined as the solution obtained by setting every free variable to zero. That choice makes the 0/1 pattern in the direction vectors work out cleanly, so that assigning a value to a free variable produces a solution vector with exactly that value in the corresponding entry.

Q: Are the direction vectors unique?

As produced by this construction, yes: they are determined by the reduced row-echelon form, which is itself unique. As a description of the null space, no — any basis of $\mathcal{N}(A)$ would serve, and an SVD-based routine will return a different, orthonormal one describing the same set.

Q: What happens if the system has a unique solution?

Then $r = n$, there are no free variables, no direction vectors are constructed and the solution set is the single vector $\vec{c}$. The theorem remains true but adds nothing, since the constants column of the reduced augmented matrix is already the answer.

Q: Why is the solution set not a subspace?

Because it does not contain the zero vector unless $\vec{b} = \vec{0}$, and it is not closed under addition: the sum of two solutions of $A\vec{x} = \vec{b}$ satisfies $A\vec{x} = 2\vec{b}$. It is an affine translate of the null space, which is a subspace.

Q: Can I use this to check whether a proposed vector is a solution?

You can, but substitution into the original system is simpler and stronger. To use the vector form, subtract $\vec{c}$ and test whether the remainder lies in the span of the direction vectors, which is itself a linear system. Direct substitution costs one matrix-vector product.

Q: How does this relate to the general solution of a linear differential equation?

It is the same structure: one particular solution plus the general solution of the associated homogeneous problem. The vectors here play the role of the homogeneous basis functions, and the free variables play the role of the arbitrary constants.

Q: Should I reduce the coefficient matrix or the augmented matrix for a homogeneous system?

Either works. The final column stays zero under every row operation, so carrying it is pure overhead; reduce the coefficient matrix alone and remember that $\vec{c} = \vec{0}$. Just do not then mistake column $n$ for the constants column.

Overview

Row reduction answers whether a system has solutions and how many degrees of freedom the solution set carries, but it stops short of describing the set itself. When the solution set is infinite, an explicit description looks impossible: there is no way to list infinitely many vectors. The vector form closes that gap. It exhibits the entire set as one fixed vector plus a linear combination of $n - r$ further vectors, with the free variables serving as the coefficients.

The pay-off is that an infinite object becomes finite data. A system of five equations in seven unknowns with rank four has an infinite solution set, yet four vectors of size seven — twenty-eight numbers — describe it completely and unambiguously, provided the convention for combining them is agreed. That is a compression from the uncountable to the tabular, and it is what makes solution sets storable, transmissible and comparable.

The construction is entirely mechanical. Every entry of every vector is either a $0$ , a $1$ , an entry of the final column of the reduced augmented matrix, or the negative of an entry in a non-pivot column. Nothing is chosen and nothing is guessed, which is why the procedure can be implemented in code once the reduction has been performed and consistency confirmed. In practical terms it completes the conversion of equation solving from a craft into an algorithm.

Structurally, the vector form separates the two independent contributions to a solution. The fixed vector is one particular solution; the vectors scaled by free variables span the null space of the coefficient matrix and describe every direction in which a solution may be perturbed without leaving the solution set. That separation is the geometric statement that a consistent solution set is a translate of a subspace, and it recurs verbatim in the pre-image of a linear transformation and in the general solution of a linear differential equation.

Definition

Vector Form of the Solution Set

VFSS

Let $[\, A \mid \vec{b} \,]$ be the augmented matrix of a consistent system of $m$ equations in $n$ unknowns, and let $B$ be its reduced row-echelon form with $r$ non-zero rows, pivot column indices $D = \{d_1, \ldots, d_r\}$ and non-pivot variable column indices $F = \{f_1, \ldots, f_{n-r}\}$ . Define vectors $\vec{c}$ and $\vec{u}_1, \ldots, \vec{u}_{n-r}$ of size $n$ by

$[\vec{c}]_i = 0$ if $i \in F$ , and $[\vec{c}]_i = [B]_{k,n+1}$ if $i = d_k \in D$ ;
$[\vec{u}_j]_i = 1$ if $i = f_j$ , $[\vec{u}_j]_i = 0$ if $i \in F$ with $i \ne f_j$ , and $[\vec{u}_j]_i = -[B]_{k,f_j}$ if $i = d_k \in D$ .

Then the solution set is

\mathcal{S} = \{\, \vec{c} + x_{f_1}\vec{u}_1 + \cdots + x_{f_{n-r}}\vec{u}_{n-r} \mid x_{f_1}, \ldots, x_{f_{n-r}} \in \mathbb{C} \,\}

The hypothesis of consistency is essential. If column $n+1$ is a pivot column the set $F$ as defined above is not the free-variable index set and the construction is meaningless — the solution set is empty.

Free and Dependent Variables

IDV

After reduction, a variable whose column contains a leading one is dependent; there are $r$ of them, indexed by $D$ . A variable whose column contains no leading one is free; there are $n - r$ of them, indexed by $F$ . Free variables may be assigned arbitrarily, after which the dependent variables are determined.

Concepts

Why the free variables become the coefficients

Each non-zero row of the reduced augmented matrix reads $x_{d_k} + \sum_{j} [B]_{k,f_j} x_{f_j} = [B]_{k,n+1}$ , because the leading one in column $d_k$ is alone in its column. Rearranged, this expresses the dependent variable $x_{d_k}$ as a constant minus a linear expression in the free variables. Collecting the constants into $\vec{c}$ and the coefficient of $x_{f_j}$ into $\vec{u}_j$ produces exactly the stated vectors.

The sign reversal in $[\vec{u}_j]_{d_k} = -[B]_{k,f_j}$ is a direct consequence of moving the free-variable terms to the right-hand side. It is the single most frequent source of error in hand computation, and it is worth stating explicitly rather than absorbing into a memorised pattern.

The pattern of zeros and ones

In the rows indexed by $F$ , the vectors carry a rigid pattern: $\vec{c}$ has a zero, $\vec{u}_j$ has a one in position $f_j$ and zeros in all other free positions. That pattern is not cosmetic. It guarantees that setting $x_{f_j} = \alpha_j$ genuinely produces a solution vector whose $f_j$ entry is $\alpha_j$ , so the parametrisation is faithful.

It also has a structural consequence that is exploited later: no non-trivial combination of the $\vec{u}_j$ can vanish, because in position $f_j$ such a combination equals its own $j$ -th coefficient. The vectors are therefore automatically linearly independent, which is why they form a basis of the null space rather than merely a spanning set.

Fixed vector plus null space directions

The vector $\vec{c}$ is obtained by setting every free variable to zero, so it is one specific solution — a particular solution. The remaining vectors satisfy $A\vec{u}_j = \vec{0}$ : they solve the associated homogeneous system. The vector form therefore says that the solution set is one point translated by the whole null space, $\mathcal{S} = \vec{c} + \mathcal{N}(A)$ .

Geometrically the solution set of a consistent system is never a subspace unless $\vec{b} = \vec{0}$ , because it does not contain the origin. It is an affine translate of a subspace, and the vector form makes the translation vector and the subspace explicit and separate.

Homogeneous systems collapse the fixed vector

If the system is homogeneous, the final column of the augmented matrix is zero and remains zero under every row operation, so $[B]_{k,n+1} = 0$ for all $k$ and hence $\vec{c} = \vec{0}$ . The vector form degenerates to a pure linear combination of the $\vec{u}_j$ — that is, to a span.

This is why homogeneous systems are treated separately and why the null space is a subspace while a general solution set is not. In practice one may reduce the coefficient matrix alone for a homogeneous system, holding the column of zeros in mind rather than carrying it through the arithmetic.

Unique solutions are the degenerate case

When $r = n$ there are no free variables, the list of $\vec{u}_j$ is empty and the vector form reduces to $\mathcal{S} = \{\vec{c}\}$ . The theorem still applies, but it is heavy machinery for a trivial conclusion; in that case the constants column of the reduced augmented matrix already is the answer.

The value of the construction lies entirely in the under-determined case $r < n$ , which is the normal situation in redundant structures, over-parametrised models and any physically indeterminate system.

The description is canonical, the parametrisation is not

Because the reduced row-echelon form of a matrix is unique, the vectors $\vec{c}$ and $\vec{u}_j$ produced by this procedure are uniquely determined by the original augmented matrix. Two engineers following the recipe on the same system obtain identical vectors, which makes the output suitable as a canonical record.

What is not canonical is the parametrisation of the set as a mathematical object: the same solution set can be described by choosing a different particular solution or a different basis for the null space. When comparing two descriptions, compare the sets they generate, not the vectors themselves.

Three-step construction of the vector form

Reduce and check consistencyRow-reduce

[\, A \mid \vec{b} \,]

B

. If column

n+1

carries a leading one, stop: the solution set is empty.

Identify $D$ and $F$ List the pivot column indices

D

(dependent variables) and the non-pivot variable indices

F

(free variables). Record

r

and

n - r

Step 1 — lay out the templateWrite

\vec{x} = \vec{c} + x_{f_1}\vec{u}_1 + \cdots + x_{f_{n-r}}\vec{u}_{n-r}

with

1 + (n-r)

empty vectors of size

n

Step 2 — fill the free rowsIn each row indexed by

F

, put

0

\vec{c}

1

\vec{u}_j

if that row is

f_j

, and

0

in every other

\vec{u}

Step 3 — fill the dependent rowsFor each pivot row

k

, solve for

x_{d_k}

and copy the constant into

\vec{c}

and the negated free-variable coefficients into the

\vec{u}_j

Spot-checkChoose specific values for the free variables, form the combination and substitute the result into the original equations.

Equations

Vector form of the solution set

EQ-VFSS-01

S = {\vec{c} + x_{f_{1}} {\vec{u}}_{1} + x_{f_{2}} {\vec{u}}_{2} + \dots + x_{f_{n - r}} {\vec{u}}_{n - r} ∣ x_{f_{1}}, \dots, x_{f_{n - r}} \in ℂ}

Every solution is obtained exactly once by a choice of the $n-r$ free variables. The description uses $1 + (n-r)$ vectors of size $n$ .

Entries of the fixed vector

EQ-VFSS-02

[\vec{c}]_{i} = {\begin{cases} 0 & i \in F \\ [B]_{k, n + 1} & i = d_{k} \in D \end{cases}

The particular solution obtained by setting every free variable to zero. Its dependent entries are the constants column of the reduced augmented matrix.

Entries of the direction vectors

EQ-VFSS-03

[{\vec{u}}_{j}]_{i} = {\begin{cases} 1 & i = f_{j} \\ 0 & i \in F, i \neq f_{j} \\ - [B]_{k, f_{j}} & i = d_{k} \in D \end{cases}

The negation in the third case comes from transposing the free-variable terms across the equals sign. These vectors span the null space of $A$ .

Pivot row solved for its dependent variable

EQ-VFSS-04

x_{d_{k}} = [B]_{k, n + 1} - [B]_{k, f_{1}} x_{f_{1}} - [B]_{k, f_{2}} x_{f_{2}} - \dots - [B]_{k, f_{n - r}} x_{f_{n - r}}

Row $k$ of the reduced augmented matrix, rearranged. No back-substitution is required because the leading one is alone in its column.

Affine structure of the solution set

EQ-VFSS-05

S = \vec{c} + N (A) = {\vec{c} + \vec{z} ∣ \vec{z} \in N (A)}

A consistent solution set is a translate of the null space by any one particular solution.

Degrees of freedom

EQ-VFSS-06

\dim N (A) = n - r, | S | = 1 \Leftrightarrow r = n

The number of direction vectors equals the nullity. A unique solution is the case where that count is zero.

Homogeneous specialisation

EQ-VFSS-07

\vec{b} = \vec{0} \Rightarrow \vec{c} = \vec{0}, S = span ({{\vec{u}}_{1}, \dots, {\vec{u}}_{n - r}})

For a homogeneous system the fixed vector vanishes and the solution set is exactly a span, hence a subspace.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Coefficient matrix	The $m \times n$ matrix of the system	m x n matrix
$B$	Reduced augmented matrix	Reduced row-echelon form of $[\, A \mid \vec{b} \,]$	m x (n+1) matrix
$\vec{c}$	Fixed vector	The particular solution obtained by zeroing all free variables	vector in C^n
${\vec{u}}_{j}$	Direction vector	The vector scaled by free variable $x_{f_j}$ ; an element of the null space	vector in C^n
$r$	Rank	Number of non-zero rows of the reduced form	0 to min(m,n)
$D$	Pivot index set	Column indices carrying a leading one; the dependent variables	subset of 1..n
$F$	Free index set	Variable column indices with no leading one	subset of 1..n
$x_{f_{j}}$	Free variable	An unconstrained parameter serving as a coefficient	complex scalar
$S$	Solution set	The set of all vectors satisfying the system	subset of C^n

Worked Numerical Example

Problem statement

A three-loop pipe network is described by three flow-balance equations in four branch flows. The system is redundant. Produce a complete description of every admissible flow state in vector form, and use it to generate two specific states.

State the system
The balance equations are $x_1 + 2x_2 + x_3 + 2x_4 = 3$ , $2x_1 + 4x_2 - x_3 - 5x_4 = 12$ and $3x_1 + 6x_2 + 2x_3 + 3x_4 = 11$ . Assemble the augmented matrix.
$[\begin{matrix} 1 & 2 & 1 & 2 & 3 \\ 2 & 4 & - 1 & - 5 & 12 \\ 3 & 6 & 2 & 3 & 11 \end{matrix}]$
Row-reduce and check consistency
Gauss-Jordan elimination gives the reduced form below. Column $5$ carries no leading one, so the system is consistent. There are $r = 2$ non-zero rows, $D = \{1, 3\}$ and $F = \{2, 4\}$ , so $x_2$ and $x_4$ are free and $n - r = 4 - 2 = 2$ .
$Row-reduce and check consistency$
Step 1 — lay out the template
Three vectors of size four are required: the fixed vector $\vec{c}$ and one direction vector for each free variable, scaled by $x_2$ and $x_4$ respectively.
$Step 1 — lay out the template$
Step 2 — fill rows 2 and 4
Rows $2$ and $4$ are the free rows. Put $0$ in $\vec{c}$ ; put $1$ in $\vec{u}_1$ at row $2$ and $0$ at row $4$ ; put $0$ in $\vec{u}_2$ at row $2$ and $1$ at row $4$ . Rows $1$ and $3$ remain to be determined.
$Step 2 — fill rows 2 and 4$
Step 3 — fill the dependent rows
Row $1$ of the reduced matrix reads $x_1 + 2x_2 - x_4 = 5$ , so $x_1 = 5 - 2x_2 + x_4$ : place $5$ in $\vec{c}$ , $-2$ in $\vec{u}_1$ and $+1$ in $\vec{u}_2$ . Row $2$ reads $x_3 + 3x_4 = -2$ , so $x_3 = -2 - 3x_4$ : place $-2$ in $\vec{c}$ , $0$ in $\vec{u}_1$ and $-3$ in $\vec{u}_2$ .
$Step 3 — fill the dependent rows$
Generate a specific solution
Choosing $x_2 = 1$ and $x_4 = 2$ gives a concrete flow state without any further elimination.
$\vec{x} = [\begin{matrix} 5 \\ 0 \\ - 2 \\ 0 \end{matrix}] + 1 [\begin{matrix} - 2 \\ 1 \\ 0 \\ 0 \end{matrix}] + 2 [\begin{matrix} 1 \\ 0 \\ - 3 \\ 1 \end{matrix}] = [\begin{matrix} 5 \\ 1 \\ - 8 \\ 2 \end{matrix}]$
Verify against the original equations
Substituting $(5,\, 1,\, -8,\, 2)$ into the untouched original system gives $5 + 2 - 8 + 4 = 3$ , $10 + 4 + 8 - 10 = 12$ and $15 + 6 - 16 + 6 = 11$ . All three hold, so the reduction and the vector form are both confirmed.
Recover the fixed vector as a solution
Setting both free variables to zero gives $\vec{c} = (5,\, 0,\, -2,\, 0)$ , which also satisfies the original equations: $5 - 2 = 3$ , $10 + 2 = 12$ and $15 - 4 = 11$ . The fixed vector is genuinely a particular solution, not a bookkeeping artefact.

Result

Every admissible flow state is $(5,\, 0,\, -2,\, 0)$ plus any combination of the two directions $(-2,\, 1,\, 0,\, 0)$ and $(1,\, 0,\, -3,\, 1)$ . The network is indeterminate to degree two: two independent circulating flow patterns satisfy the balances with no net supply, and twelve numbers describe an infinite family of admissible states exactly.

Applications & Industry Use

Civil & structural engineering

Self-stress states in indeterminate structures

For a statically indeterminate frame, the vector form separates one equilibrium force distribution from the space of self-stress states that carry no external load. Designers use the direction vectors to explore prestress options while remaining in equilibrium by construction.

Water & process networks

Admissible flow distributions

Continuity equations in a looped pipe or duct network are under-determined. The fixed vector gives one feasible distribution and the direction vectors give the independent loop circulations, which is exactly the parametrisation that a Hardy Cross or gradient-based hydraulic solver refines.

Electrical engineering

Branch currents with redundant loops

Kirchhoff current equations for a network with more branches than independent nodes yield an infinite solution family before the constitutive laws are imposed. The vector form provides the parametrisation over which the resistive or reactive constraints are subsequently minimised.

Robotics

Null-space motion of a redundant manipulator

For a manipulator with more joints than task dimensions, the joint velocities achieving a commanded end-effector twist form exactly this affine set. The fixed vector is the minimum-effort command and the direction vectors are self-motions used for obstacle and joint-limit avoidance.

Chemical engineering

Independent reactions in a network

Stoichiometric balances usually admit a family of solutions. The direction vectors are the independent reaction extents, and their count is the number of reactions that must be measured or specified before the composition is pinned down.

Data science

Non-identifiable regression parameters

When a design matrix is rank-deficient, the set of parameter vectors fitting the data exactly is an affine family. Writing it in vector form shows which parameter combinations are estimable and which directions the data cannot distinguish, guiding reparametrisation or regularisation.

Design Considerations

Confirm consistency before constructing anything

The construction presumes a consistent system. If column $n+1$ is a pivot column, the index set $F$ silently acquires the value $n+1$ and a mechanical application of the recipe produces a plausible-looking but entirely spurious answer. Test consistency first, as a separate gate.

Choose the parametrisation to suit the downstream use

The free variables are a legitimate but arbitrary set of parameters. If a physical quantity such as a particular member force or loop flow is the natural control variable, reorder the columns before reduction so that variable becomes free. The solution set is unchanged; the description becomes far more interpretable.

The fixed vector is not canonical, only convenient

Setting the free variables to zero is one choice among infinitely many particular solutions. For physical reporting, a minimum-norm particular solution is often preferable because it does not privilege an arbitrary variable ordering. Both describe the same set.

Store the vectors, not a sampled solution list

It is tempting to record a handful of representative solutions. Do not: the vector form is smaller, exact and complete, whereas any finite sample is none of those. Interfaces exchanging solution sets should exchange $\vec{c}$ and the $\vec{u}_j$ with the variable ordering that defines them.

In floating point the rank decision drives everything

The number of direction vectors is $n - r$ , and $r$ is a discontinuous function of the matrix entries. A near-dependence misclassified as a dependence adds a spurious direction vector; misclassified the other way, a genuine degree of freedom disappears. Report the tolerance used alongside the answer.

Guard against apparent over-determination

Zero rows in the reduced form mean redundant equations, not error. A model that produces many zero rows may be duplicating physics — for example applying both node and loop laws where only one set is independent — which is worth investigating even though the mathematics is untroubled.

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 set-builder and vector notation used to state the solution set, including the conditional bar and the entry notation $[\vec{c}]_i$ .
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Rank-deficient systems are handled by `xGELSD` and `xGELSY`, which return a minimum-norm particular solution rather than the pivot-based fixed vector; the null-space directions must be recovered separately from the SVD or from `xGEQP3`.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Determines the rounding behaviour of the reduction that produces every entry of the fixed and direction vectors, and hence the reliability of the rank on which their count depends.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the parametrised solution set so that the vectors and their coefficients remain semantically distinguishable to search and assistive technology.
`W3C WCAG 2.1 AA`	Web Content Accessibility Guidelines	Requires the column vectors to be exposed as structured mathematics rather than images, keeping multi-row displays navigable by screen reader and legible under magnification.

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	Small integer systems where the direction vectors are to be reported symbolically or compared for equality.	Entries are exact and the rank is unambiguous, but coefficients can grow large during reduction and performance degrades sharply with size.
IEEE 754 binary64	Engineering systems assembled from measured coefficients.	Fast and adequate for generating solutions, but the count of direction vectors becomes a tolerance decision and near-zero pivots must be handled explicitly.
Singular value decomposition instead of RREF	Numerically sensitive rank determination, or when an orthonormal null-space basis is wanted.	Far more robust and gives orthonormal directions, at roughly an order of magnitude more arithmetic and with directions that no longer carry the 0/1 pattern.
Sparse storage with fill-reducing ordering	Large network or finite element systems where the coefficient matrix is mostly zero.	Preserves tractability during factorisation, but the direction vectors themselves are typically dense, so the output can be much larger than the input.
Fixed-point arithmetic	Embedded solvers regenerating solutions on-line from a precomputed vector form.	Cheap and deterministic for evaluating the combination, provided the vectors were produced off-line in higher precision; unsuitable for the reduction itself.
Finite field arithmetic	Coding theory and cryptanalysis, where solution sets over a finite field are enumerated exactly.	Exact, fast and gives a genuinely finite solution set of size $q^{n-r}$ , but the structure over a finite field need not mirror the structure over the rationals.

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 construction

Once the reduction is done the construction is free: each of the $1 + (n-r)$ vectors of size $n$ is filled by copying or negating entries already present in $B$ , with no arithmetic beyond sign changes. Total cost is $O(n(n-r))$ data movement. The expense is entirely in reaching the reduced form.

Hand procedure and the discipline of order

Complete Step 2 for every vector before beginning Step 3, and within Step 3 handle one dependent variable at a time, writing all of its entries across all vectors before moving on. Working vector by vector instead of variable by variable is the reliable route to a transposed sign.

Library behaviour

SymPy's linsolve returns a parametrised solution equivalent to this form, and Matrix.nullspace() returns the direction vectors with the same 0/1 pattern. NumPy has no direct equivalent: numpy.linalg.lstsq gives a minimum-norm particular solution and scipy.linalg.null_space gives an orthonormal null-space basis from the SVD. The two conventions describe the same set with different vectors.

Verification technique

Two checks are worth running. First, confirm $A\vec{c} = \vec{b}$ against the original matrix. Second, confirm $A\vec{u}_j = \vec{0}$ for each direction vector. Together these certify the whole set, because any member is $\vec{c}$ plus a combination of the $\vec{u}_j$ . Checking a few random members instead is weaker and slower.

Generating solutions on demand

Stored as a matrix $U$ whose columns are the $\vec{u}_j$ , the family becomes $\vec{x} = \vec{c} + U\vec{t}$ for a free parameter vector $\vec{t}$ . Generating a solution is then one gemv call, which is the right structure for embedded or real-time use where a solution must be produced each control cycle.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Applying the construction to an inconsistent system	high	Skipping the consistency gate, so a leading one in the final column is never noticed.	Test whether column $n+1$ is a pivot column immediately after reduction and before any vectors are written.
Sign error on the direction vectors	high	Copying $[B]_{k,f_j}$ into $\vec{u}_j$ instead of $-[B]_{k,f_j}$ .	Always write the pivot row as an explicit equation solved for its dependent variable before transcribing, and verify with $A\vec{u}_j = \vec{0}$ .
Counting the augmented column as a free variable	high	Including index $n+1$ in $F$ , producing one direction vector too many.	Restrict $F$ to indices $1$ through $n$ ; the constants column is never a variable.
Using the echelon form rather than the reduced form	medium	Stopping after forward elimination, so entries above the leading ones are non-zero.	Complete the back-clearing; the construction depends on each leading one being alone in its column so that no back-substitution is needed.
Wrong number of direction vectors	medium	Counting zero rows instead of non-pivot variable columns, which agree only when $m = n$ .	Compute $n - r$ from the number of unknowns and the rank, never from the number of zero rows.
Losing the variable ordering	medium	Reporting the vectors without recording which entry corresponds to which physical unknown, especially after a column permutation.	Transmit the variable name ordering alongside the vectors as part of the same record.
Treating the fixed vector as unique	low	Assuming that a differently obtained particular solution indicates an error.	Compare descriptions by checking that the difference of the two fixed vectors lies in the span of the direction vectors.
Spurious degrees of freedom in floating point	high	A tiny but non-zero pivot rejected by the tolerance, inflating $n - r$ .	Use a scaled tolerance relative to the matrix norm, or determine rank from singular values and report the numerical rank explicitly.

FAQs

Why does the fixed vector have zeros in the free positions?

Because it is defined as the solution obtained by setting every free variable to zero. That choice makes the 0/1 pattern in the direction vectors work out cleanly, so that assigning a value to a free variable produces a solution vector with exactly that value in the corresponding entry.

Are the direction vectors unique?

As produced by this construction, yes: they are determined by the reduced row-echelon form, which is itself unique. As a description of the null space, no — any basis of $\mathcal{N}(A)$ would serve, and an SVD-based routine will return a different, orthonormal one describing the same set.

What happens if the system has a unique solution?

Then $r = n$ , there are no free variables, no direction vectors are constructed and the solution set is the single vector $\vec{c}$ . The theorem remains true but adds nothing, since the constants column of the reduced augmented matrix is already the answer.

Why is the solution set not a subspace?

Because it does not contain the zero vector unless $\vec{b} = \vec{0}$ , and it is not closed under addition: the sum of two solutions of $A\vec{x} = \vec{b}$ satisfies $A\vec{x} = 2\vec{b}$ . It is an affine translate of the null space, which is a subspace.

Can I use this to check whether a proposed vector is a solution?

You can, but substitution into the original system is simpler and stronger. To use the vector form, subtract $\vec{c}$ and test whether the remainder lies in the span of the direction vectors, which is itself a linear system. Direct substitution costs one matrix-vector product.

How does this relate to the general solution of a linear differential equation?

It is the same structure: one particular solution plus the general solution of the associated homogeneous problem. The vectors here play the role of the homogeneous basis functions, and the free variables play the role of the arbitrary constants.

Should I reduce the coefficient matrix or the augmented matrix for a homogeneous system?

Either works. The final column stays zero under every row operation, so carrying it is pure overhead; reduce the coefficient matrix alone and remember that $\vec{c} = \vec{0}$ . Just do not then mistake column $n$ for the constants column.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section LC, Subsection VFSS. Licensed under the GNU Free Documentation License v1.2.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013.
Anderson, E. et al. LAPACK Users' Guide, 3rd edition. Society for Industrial and Applied Mathematics, 1999.
IEEE 754-2019, IEEE Standard for Floating-Point Arithmetic. Institute of Electrical and Electronics Engineers.

AI Suggested Questions

Take a 4x7 augmented matrix, row-reduce it and produce the full vector form of the solution set with every step shown.
Show how permuting the columns before reduction changes which variables are free without changing the solution set.
Compare the pivot-based fixed vector with the minimum-norm particular solution returned by a least-squares routine on the same system.
Why are the direction vectors automatically linearly independent, and where exactly does the 0/1 pattern enter that argument?
Given two different vector-form descriptions of the same solution set, how do I verify computationally that they agree?
How should a numerical rank tolerance be chosen so that the count of direction vectors is trustworthy?

Related Calculators

Vector Form Solution Calculator

Row-reduce an augmented matrix and emit the fixed vector and direction vectors with the free-variable labelling.

Null Space Basis Calculator

Compute a basis for the null space of a matrix using the pivot pattern, with an optional orthonormal alternative.

Linear System Solver

Solve $A\vec{x} = \vec{b}$ , report consistency and rank, and present the solution set in vector form.