Uniqueness of Reduced Row-Echelon Form

Q: Why does the proof work with rows when the rest of the subject works with columns?

Because row operations act on rows, and the object being characterised is the result of applying them. Expressing a row of one reduced form as a combination of the rows of the other is the direct algebraic content of row equivalence; the column viewpoint has no equally short route to the same statement.

Q: Does the proof depend on the matrix being square or on any rank hypothesis?

No. It applies to any $m \times n$ matrix, and it deliberately avoids rank, dimension and independence, all of which are later defined in terms of the reduced form. Using them here would be circular.

Q: Is the row-echelon form unique?

No. The pivot column positions are the same for every echelon form of a matrix, but the entries above the staircase are unconstrained, so many different echelon forms exist. Uniqueness requires the additional reduced conditions: each leading entry equal to one, and alone in its column.

Q: What exactly goes wrong in floating-point arithmetic?

The pivot decision becomes a threshold test on a rounded quantity. A structurally zero entry may appear as a tiny non-zero value and be selected as a pivot under one ordering but rejected under another, producing genuinely different pivot patterns. The theorem's hypothesis of exact arithmetic is simply not met.

Q: If the reduced form is unique, is the sequence of row operations unique too?

No, and this is the point of the theorem: infinitely many sequences reach the same destination. The matrix $J$ recording the composition is not unique either when the matrix has dependent rows.

Q: How is this used to test whether two matrices are row-equivalent?

Reduce both and compare entry by entry. Row equivalence holds precisely when the reduced forms are identical, which converts a question about an infinite family of operation sequences into a finite comparison.

Q: Does the theorem hold over any field?

Yes. The argument uses only field arithmetic and the definitions of row operations and reduced row-echelon form, so it applies verbatim over the rationals, the reals, the complex numbers and any finite field. It is a statement about exact arithmetic, not about a particular number system.

Overview

Elementary row operations are reversible, so they partition the set of $m \times n$ matrices into equivalence classes. Existence of a reduced row-echelon form is easy: the row-reducing algorithm terminates on any input and delivers a matrix satisfying the four conditions. What is not obvious is that the algorithm's output does not depend on the choices made along the way — which row is used as a pivot, in what order operations are applied, whether a scaling is performed early or late.

The uniqueness theorem settles that. If $B$ and $C$ are both row-equivalent to $A$ and both in reduced row-echelon form, then $B = C$ entry for entry. Combined with existence, this makes RREF a genuine canonical form: each row-equivalence class contains exactly one matrix in reduced form, and two matrices are row-equivalent precisely when their reduced forms are identical.

Everything downstream depends on this. Speaking of the rank of a matrix, the pivot columns, the free variables or the nullity presupposes that these quantities do not depend on the reduction path. Without uniqueness they would be properties of a computation rather than of a matrix, and no structural theorem built on them would be well posed.

The argument is worth studying for its technique as much as its conclusion. It works with the rows of a matrix rather than its columns — one of the few places in the subject where that is the natural view — and it uses only the definition of row equivalence, the definition of reduced row-echelon form and the notion of a linear combination. Nothing about rank, dimension or independence is invoked, which is precisely what keeps the reasoning free of circularity, since those ideas are later defined in terms of the object being characterised here.

Definition

Reduced Row-Echelon Form is Unique

RREFU

Let $A$ be an $m \times n$ matrix, and suppose $B$ and $C$ are $m \times n$ matrices that are each row-equivalent to $A$ and each in reduced row-echelon form. Then $B = C$ .

Together with the existence guaranteed by the row-reducing algorithm, this justifies the notation $\operatorname{RREF}(A)$ for a single well-defined matrix determined by $A$ alone.

Rows Regarded as Vectors

RRV

For a matrix $B$ , write $\vec{b}_i$ for the vector of size $n$ whose entries are the entries of row $i$ of $B$ , in order. Converting rows into column vectors this way lets vector equality, addition, scalar multiplication and linear combination be applied to rows without defining a separate row-vector algebra.

This device is used only where rows must be compared directly. Elsewhere in the subject a matrix is treated as a collection of columns.

Pivot Column Index Set

PCIS

For a matrix in reduced row-echelon form with $r$ non-zero rows, $D = \{d_1, d_2, \ldots, d_r\}$ with $d_1 < d_2 < \cdots < d_r$ lists the indices of the columns containing leading ones. The theorem proves that $D$ , and then $r$ , and then every entry, are determined by the original matrix.

Concepts

Rows of one reduced form combine to give rows of the other

Row equivalence is symmetric because each elementary row operation is reversible. So if $B$ and $C$ are both row-equivalent to $A$ , they are row-equivalent to each other, and there is a finite sequence of row operations converting $B$ into $C$ . Every such operation replaces a row by a linear combination of rows, so composing them shows that each row of $C$ is a linear combination of the rows of $B$ , and symmetrically each row of $B$ is a linear combination of the rows of $C$ .

That single observation is the entire engine of the proof. Everything that follows is the disciplined examination of individual entries of such a vector equality, using the rigid pattern of zeros and ones that reduced row-echelon form imposes.

Step one: the first pivots coincide

Let $D = \{d_1, \ldots, d_r\}$ be the pivot indices of $B$ and $D' = \{d'_1, \ldots, d'_{r'}\}$ those of $C$ . Suppose, for contradiction, that $d_1 < d'_1$ . Write the first row of $B$ as a combination of the rows of $C$ and read entry $d_1$ of both sides.

On the left the entry is $1$ , because $d_1$ is the position of the leading one in row $1$ of $B$ . On the right every row of $C$ has a zero in position $d_1$ , since $d_1$ is strictly to the left of $C$ 's first pivot and all entries left of a first pivot are zero. A linear combination of zeros is zero, so $1 = 0$ — impossible. Hence $d_1 \ge d'_1$ , and the mirror-image argument gives $d_1 \le d'_1$ .

Step two: induction along the staircase

Assume $d_i = d'_i$ for $i \le k$ and suppose $d_{k+1} < d'_{k+1}$ . Write row $k+1$ of $B$ as a combination $\sum_i a_i \vec{c}_i$ of the rows of $C$ . Reading entries $d_1, \ldots, d_k$ forces $a_1 = \cdots = a_k = 0$ : the left side is zero there because a leading one in reduced form is alone in its column, while on the right only $\vec{c}_i$ contributes at position $d_i = d'_i$ , and it contributes $a_i$ .

Now read entry $d_{k+1}$ . The left side is $1$ . On the right the first $k$ rows are multiplied by zero, and every remaining row of $C$ has a zero in position $d_{k+1}$ because that position lies strictly left of $d'_{k+1}$ . The right side is therefore zero, a contradiction. So $d_{k+1} \ge d'_{k+1}$ , and by symmetry equality holds.

Step three: the ranks agree

Suppose $r < r'$ . The previous steps give $d_i = d'_i$ for $i \le r$ . Write row $r'$ of $C$ as a combination of the $r$ non-zero rows of $B$ . Reading entry $d_i$ shows $a_i = 0$ for every $i$ , because row $r'$ of $C$ has zeros in all of the earlier pivot positions while only $\vec{b}_i$ contributes there.

With every coefficient zero, row $r'$ of $C$ is the zero row — contradicting the fact that it is a non-zero row carrying a leading one. Hence $r \ge r'$ , and by symmetry $r = r'$ , so $D = D'$ exactly.

Step four: the rows themselves are equal

With the pivot structure matched, write row $k$ of $C$ as $\sum_{i=1}^{r} a_i \vec{b}_i$ . Reading entry $d_i$ of both sides gives $a_i = [\vec{c}_k]_{d_i}$ , which is $0$ for $i \ne k$ and $1$ for $i = k$ , since in reduced form the entries of row $k$ at the pivot positions are the $k$ -th column of an identity block.

The combination therefore collapses to $\vec{c}_k = \vec{b}_k$ . As $k$ was arbitrary and the remaining rows of both matrices are zero, every row agrees, so $B = C$ . Note where the reduced condition is doing the work: it is the requirement that each leading one is alone in its column that makes the pivot entries read off the coefficients directly.

Why echelon form is not unique

Drop the requirement that each leading one is alone in its column and the argument collapses at step four. Entries above a pivot are then unconstrained, so a matrix has many row-echelon forms: scaling and adding rows produces different entries above the staircase while preserving the staircase itself. What survives is the pivot position set $D$ , which the first three steps establish without using the reduced condition at all.

This explains the division of labour in practice. LU factorisation is not unique as a set of entries but reveals the same pivot structure; RREF pins down the entries as well, at extra arithmetic cost, and is what makes entry-by-entry comparison of two matrices a valid test of row equivalence.

Structure of the uniqueness proof

Set upAssume

B

and

C

are both in RREF and both row-equivalent to

A

; convert rows to vectors and note each row of one is a combination of the rows of the other.

First pivotAssume

d_1 < d'_1

, read entry

d_1

of a vector equality, obtain

1 = 0

. Conclude

d_1 = d'_1

by symmetry.

Inductive stepAssuming

d_i = d'_i

for

i \le k

, kill the first

k

coefficients using earlier pivot positions, then read entry

d_{k+1}

to reach a contradiction.

Equal ranksAssume

r < r'

and show row

r'

C

must be zero, contradicting its leading one. Conclude

r = r'

and

D = D'

Equal rowsExpand row

k

C

over the rows of

B

; the pivot entries force all coefficients to zero except

a_k = 1

, giving

\vec{c}_k = \vec{b}_k

ConcludeAll rows agree, so

B = C

. RREF is therefore a canonical form for row equivalence.

Equations

Uniqueness statement

EQ-RREFU-01

B \sim A, C \sim A, B, C in RREF \Rightarrow B = C

Two reduced row-echelon forms of the same matrix are identical entry for entry, whatever sequence of row operations produced them.

A row of one matrix over the rows of the other

EQ-RREFU-02

{\vec{c}}_{k} = a_{1} {\vec{b}}_{1} + a_{2} {\vec{b}}_{2} + \dots + a_{m} {\vec{b}}_{m}

Row equivalence makes every row of one matrix a linear combination of the rows of the other. Individual entries of this equality drive the whole proof.

Pivot entries of a reduced form

EQ-RREFU-03

[B]_{i, d_{j}} = {\begin{cases} 1 & i = j \\ 0 & i \neq j \end{cases}

In reduced row-echelon form the pivot columns form an identity block. Reading a vector equality at position $d_j$ therefore isolates a single coefficient.

Contradiction at the first pivot

EQ-RREFU-04

1 = [{\vec{b}}_{1}]_{d_{1}} = \sum_{i}^{=} a_{i} [{\vec{c}}_{i}]_{d_{1}} = \sum_{i}^{=} a_{i} \cdot 0 = 0

The impossible equality obtained from assuming $d_1 < d'_1$ . Every row of $C$ vanishes to the left of its first pivot.

Coefficient extraction at a known pivot

EQ-RREFU-05

a_{j} = \sum_{i}^{=} a_{i} [{\vec{b}}_{i}]_{d_{j}} = [{\vec{c}}_{k}]_{d_{j}}

Reading the expansion at pivot position $d_j$ recovers coefficient $a_j$ directly, which forces all coefficients except the $k$ -th to vanish.

Canonical form property

EQ-RREFU-06

A \sim A' \Leftrightarrow RREF (A) = RREF (A')

The operational payoff: a question about an infinite equivalence class reduces to comparing two specific matrices.

Well-defined invariants

EQ-RREFU-07

Well-defined invariants

Because the reduced form is unique, these quantities are properties of the matrix rather than of a particular reduction.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Original matrix	The matrix whose reduced form is in question	m x n matrix
$B$	First reduced form	A matrix in RREF row-equivalent to $A$	m x n matrix
$C$	Second reduced form	Another matrix in RREF row-equivalent to $A$	m x n matrix
${\vec{b}}_{i}$	Row of B as a vector	Row $i$ of $B$ regarded as a vector of size $n$	vector in C^n
${\vec{c}}_{k}$	Row of C as a vector	Row $k$ of $C$ regarded as a vector of size $n$	vector in C^n
$d_{k}$	Pivot index in B	Column of the leading one in row $k$ of $B$	1 to n
$d'_{k}$	Pivot index in C	Column of the leading one in row $k$ of $C$	1 to n
$Non-zero row count of B$	Non-zero row count of B	Number of leading ones in the first reduced form	0 to min(m,n)
$a_{i}$	Expansion coefficient	Scalar multiplying a row in the linear combination	complex scalar

Worked Numerical Example

Problem statement

A sensor calibration matrix is reduced independently by two analysts. One pivots on the first row as written; the other swaps rows first to avoid fractions. Track both reductions and confirm that the results coincide, then locate the step in the proof that each observation illustrates.

The matrix
Both analysts start from the same $3 \times 4$ array. Note that row $3$ is the sum of rows $1$ and $2$ , so a zero row is inevitable, but neither analyst assumes this.
$A = [\begin{matrix} 2 & 4 & 1 & 3 \\ 1 & 2 & 0 & 1 \\ 3 & 6 & 2 & 5 \end{matrix}]$
Route one: scale the first row
Analyst one keeps the row order and scales: $R_1 \to \tfrac{1}{2}R_1$ , then $R_2 \to R_2 - R_1$ and $R_3 \to R_3 - 3R_1$ . Fractions appear immediately.
$[\begin{matrix} 1 & 2 & \frac{1}{2} & \frac{3}{2} \\ 0 & 0 & - \frac{1}{2} & - \frac{1}{2} \\ 0 & 0 & \frac{1}{2} & \frac{1}{2} \end{matrix}]$
Route one: clear the second pivot
Column $3$ supplies the next pivot. Apply $R_2 \to -2R_2$ , then $R_1 \to R_1 - \tfrac{1}{2}R_2$ and $R_3 \to R_3 - \tfrac{1}{2}R_2$ .
$[\begin{matrix} 1 & 2 & 0 & 1 \\ 0 & 0 & 1 & 1 \\ 0 & 0 & 0 & 0 \end{matrix}]$
Route two: swap first
Analyst two swaps $R_1 \leftrightarrow R_2$ to obtain a leading entry of $1$ , then applies $R_2 \to R_2 - 2R_1$ and $R_3 \to R_3 - 3R_1$ . No fractions arise at any point.
$[\begin{matrix} 1 & 2 & 0 & 1 \\ 0 & 0 & 1 & 1 \\ 0 & 0 & 2 & 2 \end{matrix}]$
Route two: finish
A single operation $R_3 \to R_3 - 2R_2$ completes the reduction. Row $1$ already has a zero in column $3$ , so no back-clearing is needed.
$[\begin{matrix} 1 & 2 & 0 & 1 \\ 0 & 0 & 1 & 1 \\ 0 & 0 & 0 & 0 \end{matrix}]$
Compare the outcomes
The two reduced forms are identical, although one route used seven operations with fractions and the other five with integers only. Pivot columns are $D = D' = \{1, 3\}$ and $r = r' = 2$ , exactly as the first three steps of the proof require.
See the argument at work
Take row $1$ of the second analyst's result and expand it over the rows of the first analyst's result: $\vec{c}_1 = a_1\vec{b}_1 + a_2\vec{b}_2 + a_3\vec{b}_3$ . Reading entry $d_1 = 1$ gives $a_1 = 1$ ; reading entry $d_2 = 3$ gives $a_2 = 0$ ; row $3$ is zero so $a_3$ is irrelevant. The combination collapses to $\vec{c}_1 = \vec{b}_1$ , which is step four in miniature.
${\vec{c}}_{1} = 1 \cdot {\vec{b}}_{1} + 0 \cdot {\vec{b}}_{2} = [\begin{matrix} 1 \\ 2 \\ 0 \\ 1 \end{matrix}]$
Contrast with the echelon forms
The intermediate matrices reached by the two analysts before back-clearing are genuinely different — one has $\tfrac{1}{2}$ and $\tfrac{3}{2}$ in row $1$ , the other has $0$ and $1$ — yet both are in row-echelon form with the same pivot columns. This is the concrete demonstration that echelon form is not unique while reduced echelon form is.

Result

Both analysts obtain the same reduced row-echelon form, so both report $\operatorname{rank}(A) = 2$ , pivot columns $\{1, 3\}$ and one degree of freedom. Because the destination is independent of the route, the pivoting strategy can be chosen purely for numerical or arithmetic convenience — here, the row swap avoids fractions entirely without affecting the answer.

Applications & Industry Use

Software engineering

Regression testing of linear algebra code

Uniqueness supplies an exact expected output for a reduction routine. A test suite can assert equality against a stored reduced form rather than against a hand-computed sequence of operations, which makes the tests independent of pivoting strategy and of internal refactoring.

Computer algebra

Canonical representatives for equality testing

Symbolic systems decide whether two matrices are row-equivalent, or whether two systems of equations have the same solution set, by reducing both and comparing. Without a canonical form the question would require a search over operation sequences.

Coding theory

Standard form of a generator matrix

A linear block code is specified by a generator matrix, and many matrices generate the same code. Reducing to standard form gives a unique representative for a given coordinate ordering, which is how equivalent code specifications are recognised and catalogued.

Structural engineering

Reproducible indeterminacy reports

The degree of static indeterminacy is read from the pivot pattern of an equilibrium matrix. Uniqueness guarantees that two analysis packages using different elimination orders report the same degree, which is what makes the figure quotable in a design report.

Control engineering

Consistent controllability assessments

Rank conditions on controllability and observability matrices decide whether a design is feasible. Because the reduced form is unique, the verdict cannot depend on which tool performed the elimination — only on the tolerance applied to it.

Education technology

Automatic marking of reduction exercises

A grading system can accept any valid sequence of row operations and still check the final answer exactly, because every correct route ends at the same matrix. Intermediate steps can then be assessed for validity rather than for conformity to a model answer.

Design Considerations

Uniqueness is a statement about exact arithmetic

The theorem is about matrices over a field, with no rounding. In floating point the pivot pattern can genuinely differ between reduction orders, because a near-zero entry may be accepted as a pivot in one order and rejected in another. Do not use the theorem to justify comparing two floating-point reduced forms for exact equality.

Choose the pivot strategy freely, but choose it deliberately

Since the destination is fixed, pivoting can be selected on other grounds: a pivot of $1$ to keep exact arithmetic fraction-free, or the largest available magnitude to limit growth in floating point. These are different objectives and rarely coincide.

Do not permute columns

The theorem concerns row equivalence only. Column operations change which variable each column represents and give a different canonical form entirely — the Smith or the rank normal form. If a routine pivots on columns, its output is not the RREF of the original matrix.

Use the theorem to justify vocabulary, not to compute

Uniqueness is what licenses phrases such as the rank and the pivot columns. It does not make computing the reduced form worthwhile when only rank is needed; a rank-revealing factorisation is cheaper and numerically better behaved.

Record the reduced form, not the operation log

Since the destination is canonical and the route is not, archives and interchange formats should store the reduced matrix. An operation log is useful for audit but is not a stable identifier for the equivalence class.

Beware of the augmented column

The reduced form of $[\, A \mid \vec{b} \,]$ is unique as a matrix in its own right, and its leading columns need not agree with those of $\operatorname{RREF}(A)$ — column $n+1$ may itself be a pivot column. Uniqueness does not license transferring conclusions between the augmented and unaugmented reductions.

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 in the proof: subscripted indices, entry notation $[B]_{ij}$ , and the typographic distinction between a matrix and its rows regarded as vectors.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Explains why the theorem's exact-arithmetic hypothesis fails in machine computation: rounding makes the pivot decision, and hence the reduced form, path-dependent.
`ISO/IEC 14882 (C++)`	Programming languages — C++	Specifies the exact-arithmetic requirements for integral types and the implementation-defined behaviour of floating types, which determines whether a reduction routine can reproduce a canonical result bit for bit.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the matrices and vector equalities of the proof semantically, keeping index-level detail available to search and assistive technology.
`W3C WCAG 2.1 AA`	Web Content Accessibility Guidelines	Requires the multi-step algebraic argument to be presented as structured, navigable content rather than as rendered images of equations.

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	Any computation that relies on uniqueness — canonical comparison, equivalence testing, symbolic rank.	Delivers exactly the matrix the theorem describes, but intermediate coefficients can grow rapidly and memory use is unbounded.
Arbitrary-precision integers with fraction-free elimination	Integer inputs where exactness is needed and rational growth must be controlled.	The Bareiss approach limits growth, but its natural output is an echelon form requiring a final normalisation pass to reach the canonical reduced form.
Finite field arithmetic modulo a prime	Coding theory, and rank or equivalence certification by modular methods.	Exact, fast and with operands of bounded size, but the canonical form over a finite field may differ from the one over the rationals for primes dividing a pivot.
IEEE 754 binary64	Numerical work where an approximate reduced form is acceptable and a tolerance is declared.	Fast, but the theorem does not apply: two orderings can yield different pivot sets, so exact comparison of results is invalid.
Interval arithmetic	Verifying that a pivot decision is justified despite input uncertainty.	Gives a rigorous bracket on each entry and can certify a pivot as non-zero, but intervals widen quickly and the result may be inconclusive.
Modular reduction with rational reconstruction	Large exact reductions where direct rational arithmetic is prohibitive.	Avoids coefficient explosion and parallelises well, at the cost of implementation complexity and the need to detect unlucky primes.

Manufacturing Notes

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

Where the cost lies

Uniqueness costs nothing to invoke but something to realise: the back-clearing phase that makes each leading one alone in its column adds roughly $\tfrac{1}{3}n^3$ operations to the $\tfrac{2}{3}n^3$ of forward elimination. That extra third is precisely what buys the canonical property, since it is exactly the freedom removed by the reduced condition.

Checking uniqueness empirically

A useful implementation test is to reduce the same exact matrix under several randomised pivot orderings and assert that all results agree. Over the rationals or a finite field this must hold; a failure is a bug. Over floating point the same test with a tolerance measures the sensitivity of the pivot pattern rather than the correctness of the code.

Reconstructing the transforming matrix

Although the reduced form is unique, the non-singular matrix $J$ with $JA = \operatorname{RREF}(A)$ is not, whenever $A$ has fewer than $m$ independent rows. Reducing $[\, A \mid I_m \,]$ produces one such $J$ , determined by the algorithm rather than by $A$ , so it must not be treated as canonical.

Library behaviour

SymPy's Matrix.rref() returns both the reduced matrix and the pivot tuple, and over exact entries the pair is reproducible across versions. Sage and Maxima behave similarly. NumPy provides no rref, and the standard justification is exactly the failure of the theorem's hypothesis under rounding: an output that looks canonical would imply an exactness the arithmetic cannot deliver.

Hand verification of a claimed uniqueness violation

When two hand reductions of the same matrix disagree, at least one is wrong. Locate the error by checking that both results are row-equivalent to the original — solve each against a common right-hand side, or verify that each row of the original is a combination of the rows of the claimed reduced form. The discrepancy is always an arithmetic slip, never a genuine second reduced form.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Assuming uniqueness for row-echelon form	high	Conflating the echelon and reduced echelon conditions, then expecting two forward eliminations to agree.	Only the fully reduced form is unique. Two echelon forms share the pivot pattern but may differ in every entry above the staircase.
Comparing floating-point reduced forms for exact equality	high	Applying an exact-arithmetic theorem to rounded computation.	Compare with a relative tolerance, or move the comparison to exact rational or modular arithmetic where the theorem genuinely applies.
Allowing column operations	high	Swapping columns to obtain a convenient pivot, by analogy with row swaps.	Restrict to the three elementary row operations. Column operations produce a different canonical form and destroy the correspondence with the original variables.
Circular justification of rank	medium	Proving uniqueness using rank or dimension, both of which are defined via the reduced form.	Keep the argument to row equivalence, the RREF conditions and linear combinations of rows, as the proof above does.
Assuming the transforming matrix is unique	medium	Treating the recorded product of elementary matrices as canonical because the reduced form is.	Uniqueness applies to the destination, not the path. Never use $J$ as an identifier for a matrix or its equivalence class.
Transferring the pivot set between augmented and unaugmented forms	medium	Assuming the pivot columns of $\operatorname{RREF}([\, A \mid \vec{b} \,])$ restricted to the first $n$ columns equal those of $\operatorname{RREF}(A)$ .	They do agree when the system is consistent, but an inconsistent system adds a pivot in the final column. Verify consistency before making the transfer.
Concluding uniqueness of the null space basis	low	Assuming that because the reduced form is unique, every object derived from it is the only possible one.	The basis produced from the pivot pattern is uniquely determined by that procedure, but the null space itself has infinitely many bases.
Mixing exact and inexact entries	medium	Reducing a matrix with some rational and some floating-point entries, so exactness is lost silently at the first contaminated pivot.	Promote the whole matrix to a single arithmetic type before reducing, and state which type the result is canonical in.

FAQs

Why does the proof work with rows when the rest of the subject works with columns?

Because row operations act on rows, and the object being characterised is the result of applying them. Expressing a row of one reduced form as a combination of the rows of the other is the direct algebraic content of row equivalence; the column viewpoint has no equally short route to the same statement.

Does the proof depend on the matrix being square or on any rank hypothesis?

No. It applies to any $m \times n$ matrix, and it deliberately avoids rank, dimension and independence, all of which are later defined in terms of the reduced form. Using them here would be circular.

Is the row-echelon form unique?

No. The pivot column positions are the same for every echelon form of a matrix, but the entries above the staircase are unconstrained, so many different echelon forms exist. Uniqueness requires the additional reduced conditions: each leading entry equal to one, and alone in its column.

What exactly goes wrong in floating-point arithmetic?

The pivot decision becomes a threshold test on a rounded quantity. A structurally zero entry may appear as a tiny non-zero value and be selected as a pivot under one ordering but rejected under another, producing genuinely different pivot patterns. The theorem's hypothesis of exact arithmetic is simply not met.

If the reduced form is unique, is the sequence of row operations unique too?

No, and this is the point of the theorem: infinitely many sequences reach the same destination. The matrix $J$ recording the composition is not unique either when the matrix has dependent rows.

How is this used to test whether two matrices are row-equivalent?

Reduce both and compare entry by entry. Row equivalence holds precisely when the reduced forms are identical, which converts a question about an infinite family of operation sequences into a finite comparison.

Does the theorem hold over any field?

Yes. The argument uses only field arithmetic and the definitions of row operations and reduced row-echelon form, so it applies verbatim over the rationals, the reals, the complex numbers and any finite field. It is a statement about exact arithmetic, not about a particular number system.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section LC, Subsection URREF. Licensed under the GNU Free Documentation License v1.2.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
Yuster, T. The Reduced Row Echelon Form of a Matrix is Unique: A Simple Proof. Mathematics Magazine, 57(2), 1984.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013.
von zur Gathen, J. and Gerhard, J. Modern Computer Algebra, 3rd edition. Cambridge University Press, 2013.

AI Suggested Questions

Give a 4x5 matrix and reduce it by three deliberately different pivot orders, showing that all three reach the same reduced form.
Construct two different row-echelon forms of the same matrix to show concretely that the unreduced form is not unique.
Where precisely would the uniqueness proof break if leading entries were allowed to be values other than one?
Show a small floating-point matrix where two elimination orders produce different pivot patterns, and explain the tolerance issue.
Prove that two matrices have the same row space if and only if they have the same reduced row-echelon form.
Why is the matrix J with JA equal to the reduced form not unique, even though the reduced form is?

Related Calculators

RREF Calculator

Reduce a matrix to reduced row-echelon form with a full audit trail, in exact or floating-point arithmetic.

Row Equivalence Tester

Decide whether two matrices are row-equivalent by comparing their canonical reduced forms.

Pivot Pattern Explorer

Reduce a matrix under randomised pivot orderings and compare the resulting pivot sets and entries.