Bases from Row Spaces | KEVOS® Mathematics

Overview

Subspaces arrive in practice as spans of whatever vectors happened to be collected: measured response patterns, candidate reactions, sampled configurations. Such a list is typically redundant, unordered and numerically awkward. The task is to replace it with a basis — an independent spanning set — without losing the subspace it describes.

Two facts about row spaces combine to solve this in a single step. First, row-equivalent matrices have identical row spaces: each elementary row operation replaces the rows by linear combinations of themselves, and because the operations are reversible, no span is gained or lost. Second, the non-zero rows of a matrix already in reduced row-echelon form are linearly independent, because each carries a leading one in a column where every other row is zero. Together they say that reducing a matrix does not change its row space and delivers a basis for it at the same time.

The procedure follows immediately. Given a spanning set of column vectors, write them as the rows of a matrix, reduce to reduced row-echelon form, and read the non-zero rows back as column vectors. Those vectors are a basis for the original span, and their number is the rank of the matrix, hence the dimension of the subspace. No independence check is required afterwards, and no verification that the smaller set still spans: both are guaranteed by the theorems.

The output has a property no other basis-extraction method offers. Because the reduced row-echelon form of a matrix is unique, the basis produced depends only on the subspace and not on the vectors supplied or the order they were listed in. Two entirely different spanning sets for the same subspace reduce to the same matrix. That makes the reduced form a canonical identifier for a subspace, and it converts the otherwise awkward question "do these two lists describe the same subspace?" into an entry-by-entry comparison of two small matrices.

Definition

Row Space of a Matrix

RSM

Let $A$ be an $m \times n$ matrix. The row space of $A$ , written $\mathcal{R}(A)$ , is the span of the set of its rows, each regarded as a vector of size $n$ . Equivalently, $\mathcal{R}(A) = \mathcal{C}(A^{t})$ : the row space of $A$ is the column space of its transpose.

Rows are written horizontally but manipulated as vectors, so transposing is the standard device for reusing every result already proved about column spaces.

Row Equivalence

Matrices $A$ and $B$ of the same size are row-equivalent, written $A \sim B$ , if one can be obtained from the other by a finite sequence of elementary row operations: swapping two rows, multiplying a row by a non-zero scalar, or adding a multiple of one row to another.

Every elementary row operation is reversible, which is exactly why row equivalence preserves the row space rather than merely shrinking it.

Reduced Row-Echelon Form

RREF

A matrix is in reduced row-echelon form when all zero rows lie below all non-zero rows; the leading entry of each non-zero row is a $1$ ; each leading one lies strictly to the right of the leading one above it; and each leading one is the only non-zero entry in its column. Every matrix is row-equivalent to exactly one matrix in this form.

Uniqueness is what makes the basis produced by this method canonical. Without it the output would depend on the sequence of operations chosen.

Concepts

Row operations preserve the row space

Theorem. If $A \sim B$ , then $\mathcal{R}(A) = \mathcal{R}(B)$ .

The argument is a containment in each direction. Every row of $B$ is produced from rows of $A$ by scaling and addition, so every row of $B$ lies in $\mathcal{R}(A)$ , and therefore so does every linear combination of them: $\mathcal{R}(B) \subseteq \mathcal{R}(A)$ . Because each elementary operation has an inverse of the same type, $B \sim A$ as well, and the identical argument gives the reverse containment. The two spans coincide.

The reversibility clause is doing all the work. A one-way construction that replaced rows by combinations of themselves could only shrink the span; it is the ability to undo each step that forces equality. This is the same property that makes row operations safe for solving systems, applied to a different object.

The non-zero rows of a reduced form are independent

Theorem. If $B$ is in reduced row-echelon form with $r$ non-zero rows, those rows are a basis for $\mathcal{R}(B)$ .

Spanning is immediate, since the zero rows contribute nothing to any linear combination. Independence follows from the structure of the leading ones. Suppose some combination of the non-zero rows is the zero vector, and consider the column holding the leading one of row $i$ . In that column, row $i$ has entry $1$ and every other row has entry $0$ , so the coefficient of row $i$ in the combination must be zero. The argument applies to each $i$ in turn, so only the trivial relation exists.

Combining the two theorems gives the whole method: reduction does not change the row space, and the reduced form exhibits a basis for it directly. The number of non-zero rows is therefore both the rank of the matrix and the dimension of the subspace.

The method, and why it needs no verification

To find a basis for the span of a set $S$ of column vectors, write the vectors as the rows of a matrix $C$ , so that $\mathcal{R}(C) = \operatorname{span}(S)$ by construction. Reduce $C$ to reduced row-echelon form, discard the zero rows, and transpose the survivors back to column vectors.

What makes this efficient is not the arithmetic but the absence of follow-up work. Methods that trim a spanning set by removing redundant vectors require a subsequent independence check and an argument that the smaller set still spans. Here both properties are conclusions of the theorems, so the reduction is the entire computation. The dimension is read off at the same time as the number of surviving rows, with no separate rank calculation.

The basis is canonical — and that is a tool

Since every matrix has exactly one reduced row-echelon form, the basis this method produces is determined by the subspace alone. Reordering the input vectors, scaling them, or replacing them by an entirely different spanning set for the same subspace all yield the same reduced matrix.

This gives a decisive test for subspace equality. To decide whether $\operatorname{span}(S) = \operatorname{span}(T)$ , form a matrix from each set's vectors as rows, reduce both, and compare entry by entry. No containment arguments, no membership tests and no dimension counting are needed. In software the reduced form serves as a canonical key for a subspace, so subspaces can be deduplicated, indexed or hashed exactly as ordinary data would be.

The basis vectors are not the vectors you started with

The output rows are linear combinations of the input vectors, and in general none of them is an input vector. In the standard shape, each basis vector has a $1$ in one pivot position and $0$ in every other pivot position, and the entries in the non-pivot columns carry all the remaining information — frequently as fractions, even when the inputs are integers.

This is the method's one real cost, and it is a genuine trade-off rather than an inconvenience. When the basis vectors must retain physical meaning — a subset of the actual measured load cases, a subset of the actual candidate reactions — a method that selects from the original list is required instead. When what is wanted is a clean, comparable, minimal description of the subspace, the canonical form is the better answer.

Choosing between the row method and column selection

Two extraction methods are available for the span of a list of vectors, and they differ in what they preserve. Placing the vectors as columns and keeping the columns of the original matrix in the pivot positions returns a basis that is a subset of the input, preserving identity and interpretability but depending on the order in which the vectors were listed. Placing them as rows and keeping the non-zero rows of the reduced form returns a canonical basis that is independent of ordering but composed of new vectors.

A useful summary: use column selection when the answer must name which of the original items are independent, and the row method when the answer is the subspace itself. Both cost one row reduction, and both report the same dimension, so the choice is entirely about the form of the output.

Reducing a spanning set to a basis

Collect the spanning vectorsAny list whose span is the subspace of interest. Redundancy, ordering and scaling are all irrelevant to the outcome.

Write the vectors as rowsForm a matrix

C

whose row

i

is the

i

th vector. By construction

\mathcal{R}(C)

equals the span of the list.

Reduce to reduced row-echelon formApply Gauss-Jordan elimination. The row space is unchanged because every elementary row operation is reversible.

Discard the zero rowsThe surviving

r

rows are linearly independent and still span. They are a basis, and

r

is the dimension.

Transpose back to column vectorsRewrite each surviving row as a column vector if the ambient space is a column vector space.

Optionally compare with another subspaceTwo spanning sets describe the same subspace exactly when their reduced forms are identical entry for entry.

Equations

Row space of a matrix

EQ-BRS-01

R (A) = span ({A_{1}, A_{2}, \dots, A_{m}}) = C (A^{t})

Here $A_i$ denotes row $i$ of $A$ regarded as a vector. Transposing converts every column space result into a row space result.

Row equivalence preserves the row space

EQ-BRS-02

A \sim B \Rightarrow R (A) = R (B)

Each row of $B$ is a combination of rows of $A$ and, because the operations are reversible, each row of $A$ is a combination of rows of $B$ .

Basis from the reduced form

EQ-BRS-03

B = RREF (A) \Rightarrow {B_{1}, B_{2}, \dots, B_{r}} is a basis of R (A)

The $r$ non-zero rows of the reduced form span because the zero rows contribute nothing, and are independent because of the isolated leading ones.

Dimension read off directly

EQ-BRS-04

\dim (R (A)) = r = rank (A)

The number of surviving rows is simultaneously the size of the basis, the dimension of the span and the rank of the matrix.

Canonical test for subspace equality

EQ-BRS-05

span (S) = span (T) \Leftrightarrow RREF (C_{S}) = RREF (C_{T})

Where $C_S$ and $C_T$ hold the vectors of $S$ and $T$ as rows. Uniqueness of the reduced form makes this an entry-by-entry comparison.

Shape of the canonical basis

EQ-BRS-06

[\begin{matrix} 1 & 0 & * & * \\ 0 & 1 & * & * \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{matrix}]

A rank-two example with pivot columns $1$ and $2$ . Each basis row is a standard unit vector on the pivot columns and carries arbitrary entries elsewhere.

The basis produced in the worked example

EQ-BRS-07

B = {[\begin{matrix} 1 \\ 0 \\ \frac{1}{7} \\ \frac{11}{7} \end{matrix}], [\begin{matrix} 0 \\ 1 \\ - \frac{11}{7} \\ - \frac{2}{7} \end{matrix}]}

Fractions appear even though every input vector had integer entries, which is characteristic of the method.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Matrix	A general $m \times n$ matrix whose row space is under study	m x n matrix over C
$C$	Stacked matrix	The matrix formed by writing the spanning vectors as rows	matrix over C
$R (A)$	Row space	The span of the rows of $A$ , a subspace of $\mathbb{C}^n$	subspace of C^n
$A^{t}$	Transpose	The matrix with rows and columns interchanged, so that $\mathcal{R}(A) = \mathcal{C}(A^{t})$	n x m matrix
$S$	Spanning set	The original list of vectors, possibly redundant and unordered	finite subset of C^n
$B_{i}$	Basis row	A non-zero row of the reduced form, read back as a basis vector	vector in C^n
$r$	Rank	Number of non-zero rows of the reduced form; the dimension of the span	0 to min(m,n)
$D$	Pivot column set	Indices of the columns holding leading ones, which fix the shape of the canonical basis	subset of 1..n
$\sim$	Row equivalence	Relation between matrices connected by elementary row operations	equivalence relation on matrices

Worked Numerical Example

Problem statement

A structural test rig records the response of four strain channels to each of four separately applied load patterns, giving four response vectors in $\mathbb{C}^4$ . The engineer needs the dimension of the achievable response subspace, a clean basis for it, and a way to confirm that a second test campaign explored the same subspace.

State the measured spanning set
The four recorded response vectors span the subspace $V$ of all responses achievable by combining the applied load patterns. Nothing is known in advance about whether they are independent.
$S = {[\begin{matrix} 3 \\ - 1 \\ 2 \\ 5 \end{matrix}], [\begin{matrix} 1 \\ 2 \\ - 3 \\ 1 \end{matrix}], [\begin{matrix} 5 \\ 3 \\ - 4 \\ 7 \end{matrix}], [\begin{matrix} 2 \\ - 3 \\ 5 \\ 4 \end{matrix}]}$
Stack the vectors as rows
Form the matrix $C$ whose rows are the four vectors. By construction $\mathcal{R}(C) = \operatorname{span}(S) = V$ , so a basis for the row space of $C$ is a basis for the achievable response subspace.
$C = [\begin{matrix} 3 & - 1 & 2 & 5 \\ 1 & 2 & - 3 & 1 \\ 5 & 3 & - 4 & 7 \\ 2 & - 3 & 5 & 4 \end{matrix}]$
Row-reduce
Swap rows $1$ and $2$ to obtain a leading entry of $1$ , clear column $1$ below it, scale the new second row by $\tfrac{1}{7}$ , then clear column $2$ above and below. Two rows reduce to zero.
$Row-reduce$
Read off the basis and the dimension
Two non-zero rows survive, so $r = 2$ and the achievable response subspace is two-dimensional inside a four-dimensional measurement space. The two surviving rows, written as column vectors, are a basis for $V$ . No independence check and no re-verification of spanning are needed: both are conclusions of the theorems.
$B = {[\begin{matrix} 1 \\ 0 \\ \frac{1}{7} \\ \frac{11}{7} \end{matrix}], [\begin{matrix} 0 \\ 1 \\ - \frac{11}{7} \\ - \frac{2}{7} \end{matrix}]}$
Identify the redundancy explicitly
Since $r = 2$ but four vectors were measured, two of the load patterns produced no new information. Direct inspection confirms the dependencies: the third response is the first plus twice the second, and the fourth is the first minus the second. Two of the four test runs were, in this respect, repeats.
$u_{3} = u_{1} + 2 u_{2}, u_{4} = u_{1} - u_{2}$
Note that the basis is not part of the original data
Neither basis vector is one of the measured responses, and both carry fractions although every measurement was an integer. That is the characteristic signature of this method: it returns the tidiest description of the subspace, not a selection from the observations. If the deliverable had to name which two test runs were independent, the vectors would instead be placed as columns and the pivot columns of the original matrix selected.
Compare a second campaign using canonicity
A later campaign records two responses, $(4, 1, -1, 6)$ and $(1, -5, 8, 3)$ . Stacking these two as rows and reducing gives the identical reduced matrix, so the two campaigns explored exactly the same subspace despite sharing no measurement. Uniqueness of the reduced form is what licenses this conclusion from an entry-by-entry comparison alone.
$[\begin{matrix} 4 & 1 & - 1 & 6 \\ 1 & - 5 & 8 & 3 \end{matrix}] ⟶ [\begin{matrix} 1 & 0 & \frac{1}{7} & \frac{11}{7} \\ 0 & 1 & - \frac{11}{7} & - \frac{2}{7} \end{matrix}]$

Result

The achievable response subspace has dimension two, so only two of the four load patterns carried independent information and the test programme could have been halved. The canonical basis produced by the reduction describes that subspace in a form that depends on nothing but the subspace itself, which is what allows the second campaign — two different loads, two different responses, no shared measurement — to be certified as covering identical ground by comparing two small matrices. Where the report must instead name the two independent test runs, the same data placed as columns and reduced identifies them by pivot position.

Applications & Industry Use

Experimental mechanics

Pruning a redundant test matrix

Measured response vectors from a load or excitation programme are almost always partly redundant. Reducing the stacked response matrix gives the dimension of the achievable response space immediately, revealing how many test runs contributed independent information and bounding how much of the programme could be dropped without losing coverage.

Chemical reaction engineering

Independent reactions from a proposed mechanism

A proposed mechanism lists more elementary steps than are independent. Writing each stoichiometric coefficient vector as a row and reducing gives the number of independent reactions and a canonical basis for the reaction space, which fixes how many extents of reaction a process model needs and exposes steps that are combinations of others.

Coding theory

Systematic form of a generator matrix

A linear block code is the row space of its generator matrix, so two generator matrices define the same code exactly when they reduce to the same form. Row-reducing to systematic form produces the canonical basis of the code, which is why published codes are specified in that form and why encoder implementations can be compared without re-deriving the code.

Control systems

Comparing reachable subspaces across designs

The reachable subspace of a linear system is the span of the columns of its controllability matrix. Transposing and reducing gives a canonical description, so two candidate actuator layouts can be certified as reaching exactly the same set of states — or shown to differ — by comparing two reduced matrices rather than by testing individual target states.

Computational geometry & CAD

Detecting coincident affine subspaces

Planes, lines and higher-dimensional flats extracted from geometry are often duplicated by tolerance or by construction history. Reducing the direction vectors of each to canonical form gives a key on which duplicates can be matched exactly, which is more robust than pairwise angle comparisons and independent of how each flat was originally parameterised.

Data engineering

Canonical keys for linear feature relationships

When derived features are linear combinations of raw ones, the set of relationships is a subspace. Reducing it to canonical form yields a stable identifier that can be stored, hashed and compared across pipeline versions, so a change in feature engineering that alters the span is detected automatically rather than by inspection.

Design Considerations

Choose the method by what the output must be

The row method returns a canonical basis composed of new vectors; column selection returns a subset of the original vectors but depends on their order. Decide before computing which the deliverable requires. A report that must identify which measurements were independent needs column selection; a report that must describe the subspace needs the row method.

Exploit canonicity deliberately

The reduced form is a fingerprint of the subspace. Storing it as a key allows subspaces to be deduplicated, indexed and compared with ordinary equality, which is far cheaper and more reliable than mutual containment testing. This is available only because the reduced row-echelon form is unique, so it does not transfer to bases obtained by other routes.

Expect fractions and plan the arithmetic

Integer input almost always yields rational output, because the leading entries are normalised to one. In exact arithmetic this is harmless but can produce large numerators and denominators on bigger problems; clearing denominators afterwards gives a tidier basis but destroys canonicity, so the two goals cannot both be met.

Do not use this method on noisy measured data

The pivot pattern is discontinuous in the entries, so an arbitrarily small perturbation can change the rank and hence the entire output. Where the vectors come from measurement, use a singular value decomposition with an explicit tolerance to determine the dimension and an orthonormal basis, and reserve row reduction for exact or symbolic data.

Confirm which space you are working in

The row space of $A$ is a subspace of $\mathbb{C}^n$ and the column space is a subspace of $\mathbb{C}^m$ ; they generally have nothing in common beyond their shared dimension. Transposing at the wrong moment produces a basis for the wrong subspace, and the error is easy to miss because the rank agrees either way.

Use the same reduction to extract everything at once

A single reduction of the stacked matrix delivers the basis, the dimension, the rank and the pivot column set. If the dependencies among the original vectors are also wanted, reduce the transposed arrangement instead and read the non-pivot columns, which express each redundant vector in terms of the retained ones. Deciding which of the two arrangements is needed before starting avoids doing both.

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 transpose, for spans and for matrix rows and columns, so that a statement about $\mathcal{R}(A)$ cannot be confused with one about $\mathcal{C}(A)$ across documents.
`CCSDS 131.0-B`	TM Synchronization and Channel Coding	Specifies block codes whose codewords form a subspace over a finite field. A generator matrix in systematic form is precisely the row-reduced basis of that subspace, which is what makes independently published generator matrices comparable and encoders interoperable.
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Supplies the numerically defensible alternatives for measured data: `xGESVD` for a tolerance-based rank and an orthonormal basis, and `xGEQP3` for a rank-revealing factorisation when the basis must be selected from the input vectors.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Defines the arithmetic in which the pivot decisions of this method become tolerance judgements, and therefore the reason a floating-point reduced form must not be used as a canonical subspace key.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the reduced matrices and their fractional entries as structured markup, keeping the exact rational values readable and machine-processable rather than rendered as images.

Material Selection

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

Representation and precision selection
Representation	Select when	Trade-off
Exact rational arithmetic	The default for this method, since normalising leading entries to one introduces fractions and canonicity depends on exactness.	Guarantees a correct pivot pattern and a genuinely canonical basis, but numerators and denominators can grow substantially on larger matrices.
Arbitrary-precision integers with fraction-free elimination	Large integer input where rational blow-up must be controlled during the elimination.	Bounds intermediate growth, but delivers an echelon rather than a fully reduced form, so a final normalisation is still required to recover canonicity.
IEEE 754 binary64 with a rank tolerance	Exploratory work on large computed matrices where an approximate dimension is sufficient.	Fast and memory-efficient, but the pivot pattern is not robust to rounding and the result must not be used as a canonical identifier.
Singular value decomposition with an orthonormal basis	Measured, estimated or noisy vectors, where rank must be decided against a stated tolerance.	Gives a numerically stable basis and a quantitative measure of near-dependence, at several times the cost and with no canonical form.
Finite field arithmetic	Coding theory and cryptographic constructions, where the subspace is a code over a finite field.	Exact, fast and free of coefficient growth, and the systematic form is the standard published representation; the pivot pattern may differ from the pattern over the rationals.
Sparse row storage	Very large spanning sets from network or mesh models where each vector has few non-zero entries.	Saves storage on input, but reduction causes fill-in that typically destroys sparsity, so a sparse QR or an iterative subspace method is usually preferable.

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 reduction

Reducing an $m \times n$ matrix to reduced row-echelon form costs $O(mn\min(m,n))$ operations, which for a spanning set of $m$ vectors in $\mathbb{C}^n$ with $m \le n$ is about $m^2n$ . Since the method requires exactly one reduction and no follow-up verification, this is also the total cost of obtaining the basis, the dimension and the rank together.

Doing it by hand

Order the rows so that a row with a leading entry of $1$ comes first; this avoids fractions during the elimination phase and confines them to the final normalisation. Work strictly left to right, one pivot column at a time, and clear both above and below before moving on. Clearing above is what distinguishes the reduced form from an echelon form and is essential here, since only the reduced form is unique.

Library behaviour

SymPy's Matrix.rref returns the reduced form and the pivot column indices in exact arithmetic, which is what this method requires. NumPy deliberately provides no such routine; the closest equivalents are scipy.linalg.orth, which returns an orthonormal basis for a column space from the SVD, and numpy.linalg.matrix_rank, which reports the dimension against a tolerance. Neither returns a canonical basis, and neither is a substitute when subspace comparison is the objective.

Verifying the result

Run two independent checks. Confirm every original vector is a combination of the basis vectors by solving for its coordinates and checking the residual; this catches an over-aggressive reduction. Then confirm the basis has exactly $r$ elements where $r$ is the independently computed rank; this catches a reduction that stopped early. Checking one direction alone will not detect a basis that spans too little or too much.

Using the reduced form as a key

In exact arithmetic the reduced form can be serialised — rational entries in lowest terms, row-major order — and used directly as a dictionary key or hash for the subspace. This makes deduplication and cross-version comparison of subspaces an $O(1)$ lookup rather than a pairwise containment test. The technique is valid only with exact arithmetic; a floating-point reduced form is not a stable key.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Confusing the row space with the column space	high	Reducing the matrix without transposing, or transposing when it was not needed, so the basis describes the wrong subspace.	Check the ambient dimension of the answer against the problem. The row space of an $m \times n$ matrix lives in $\mathbb{C}^n$ ; the column space lives in $\mathbb{C}^m$ .
Expecting the basis to consist of original vectors	high	Using the row method when the deliverable is a selection from the input list.	Use column placement with pivot column selection when the identity of the original vectors matters. Use the row method only when the subspace itself is the deliverable.
Using a floating-point reduced form as a canonical key	high	Assuming the uniqueness property survives rounding, so two computations of the same subspace produce comparable output.	Canonicity requires exact arithmetic. For measured data, compare subspaces with principal angles or with an SVD-based test at a stated tolerance.
Adding an unnecessary independence check	low	Not recognising that independence of the surviving rows is a conclusion of the theorem, and re-verifying it.	The non-zero rows of a reduced form are independent by construction of the leading ones. The reduction is the whole computation.
Clearing denominators to tidy the basis	medium	Scaling the rows to integers after reduction for presentation, then reusing the result as a canonical identifier.	Scaling produces a valid basis but not the canonical one. Keep the normalised form for comparison and generate a scaled version only for display.
Stopping at echelon form	medium	Performing forward elimination only, so entries above the leading ones remain non-zero.	An echelon form still yields a basis but is not unique, so canonicity and the subspace equality test are both lost. Complete the reduction by clearing above every pivot.
Reporting the dimension as the number of input vectors	medium	Assuming the supplied spanning set was independent.	The dimension is the number of non-zero rows after reduction, which may be far smaller. Compare the two counts explicitly and report the redundancy.
Applying the method to a set that does not span the intended subspace	medium	Reducing a partial or filtered list of vectors and reporting a basis for the subspace they happen to span.	Confirm that the input list spans the intended subspace before reducing. The method faithfully reproduces the span of whatever it is given, including a span that is too small.

FAQs

Why does row reduction not change the row space?

Because every elementary row operation replaces the rows by linear combinations of themselves, so the new row space is contained in the old one, and every operation is reversible, so the containment also runs the other way. The two spans therefore coincide. Reversibility is the essential clause: a one-way construction could only shrink the span.

Why are the non-zero rows of a reduced form automatically independent?

Because of the isolated leading ones. In the column holding the leading one of row $i$ , that row has entry $1$ and every other row has entry $0$ , so any linear combination equal to the zero vector must have coefficient zero on row $i$ . Applying the argument to each row in turn leaves only the trivial relation.

Will the basis contain any of my original vectors?

Generally not. The output rows are linear combinations of the inputs, shaped so that each has a $1$ in one pivot column and zeros in the others, and fractional entries are common even for integer input. If the basis must be a subset of the original list, place the vectors as columns instead and select the columns of the original matrix in the pivot positions.

How does this give a test for whether two spans are equal?

Every matrix has exactly one reduced row-echelon form, so the reduced form of the stacked matrix depends only on the subspace, not on which spanning set was supplied or in what order. Reducing both lists and comparing the resulting matrices entry by entry therefore settles equality outright, with no containment arguments or membership tests.

Can I use this on measured data?

Not safely. The pivot pattern is a discontinuous function of the entries, so a small measurement error can change the rank and produce a completely different output. Use a singular value decomposition with an explicit tolerance to determine the dimension and to obtain an orthonormal basis, and reserve this method for exact or symbolic data.

How is this different from casting out vectors from a spanning set?

Casting out removes vectors that are combinations of earlier ones, so the result is a subset of the input and its content depends on the ordering. The row method returns a canonical basis of new vectors that does not depend on ordering at all. Both cost one row reduction and both report the same dimension; they differ only in the form of the output.

May I clear the fractions to make the basis tidier?

Yes for presentation, no for comparison. Scaling a basis vector by a non-zero scalar leaves it a basis of the same subspace, so the scaled set is perfectly valid. But it is no longer the unique reduced form, so the subspace equality test and any use of the output as a canonical key both stop working.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section B, Subsection BRS, with Theorems REMRS and BRS. Licensed under the GNU Free Documentation License v1.2.
Strang, G. Linear Algebra and Its Applications, 4th edition. Cengage Learning, 2006.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013.
MacWilliams, F. J. and Sloane, N. J. A. The Theory of Error-Correcting Codes. North-Holland, 1977.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
Cohen, H. A Course in Computational Algebraic Number Theory. Springer, 1993.

AI Suggested Questions

Take a set of six vectors in $\mathbb{C}^5$ and produce both a canonical row-space basis and a subset basis from the original vectors, then explain when each output is the right deliverable.
Show why the reduced row-echelon form can be used as a hash key for a subspace in exact arithmetic but not in floating point.
Demonstrate a rank change caused by a perturbation of $10^{-12}$ in one entry, and show how an SVD-based tolerance avoids it.
Explain how a linear block code's generator matrix in systematic form relates to the canonical basis of the code, and what changes if the code is over a field with four elements.
Compare principal angles between subspaces with the reduced-form equality test as methods for deciding whether two measured subspaces agree.
How does fraction-free Bareiss elimination change the coefficient growth of this method on a large integer matrix, and what has to be done afterwards to recover the canonical form?

Related Calculators

Row Space Basis Calculator

Stack a list of vectors as rows, reduce exactly, and return the canonical basis, the dimension and the pivot column set.

Subspace Equality Tester

Compare two spanning sets by reducing both to canonical form and reporting whether they describe the same subspace.

RREF Calculator

Reduce any matrix to reduced row-echelon form with a full audit trail of the row operations applied.

Matrix Rank & Nullity Calculator

Compute rank, nullity and the pivot column set with a selectable numerical tolerance.