The Left Null Space

Q: What does a vector in the left null space actually tell me?

Two things at once. Read against the rows of $A$, its entries are the scalars in a linear combination of rows that cancels — a statement that those equations are not independent. Read against a right-hand side, it gives the scalar test $\vec{y}^{\,t}\vec{b} = 0$ that any consistent $\vec{b}$ must satisfy.

Q: When is the left null space trivial?

Exactly when $A$ has full row rank, $r = m$. Then no combination of rows cancels, no zero rows appear on reduction, and $A\vec{x} = \vec{b}$ is consistent for every possible $\vec{b}$ — equivalently the column space is all of $\mathbb{C}^m$. This is a common design goal for a well-posed system and a design failure for a self-checking instrument.

Overview

Three subsets of a matrix appear early and repeatedly: the null space, the column space and the row space. The fourth completes the set. The left null space is defined as the null space of the transpose, which places it in $\mathbb{C}^m$ for an $m \times n$ matrix — the same ambient space as the column space, and one size removed from the null space and the row space.

The name is explained by rewriting the defining condition. If $A^{t}\vec{y} = \vec{0}$ then transposing both sides gives $\vec{y}^{\,t} A = \vec{0}^{\,t}$ : the vector acts on $A$ from the left, and its entries serve as the scalars in a linear combination of the rows of $A$ that produces a row of zeros. Every element of the left null space is therefore a relation of linear dependence among the rows.

That reading gives the subspace immediate operational content. When a matrix row-reduces to a form with zero rows, each zero row was manufactured by some cumulative sequence of row operations, and the net scalars applied to the original rows constitute a vector in $\mathcal{L}(A)$ . The left null space is the complete inventory of such recipes. Its dimension is $m - r$ , exactly the number of zero rows that appear.

In application the same vectors do a second job. Because $\vec{y}^{\,t} A = \vec{0}^{\,t}$ , applying $\vec{y}^{\,t}$ to the system $A\vec{x} = \vec{b}$ produces $\vec{y}^{\,t}\vec{b} = 0$ regardless of $\vec{x}$ . Each basis vector of the left null space is thus a linear constraint that any consistent right-hand side must satisfy, and violation of one is a direct, computable signal of inconsistency. In estimation and diagnostics these are known as parity relations, and they are the mathematical basis of analytical redundancy.

Definition

Left Null Space

LNS

Let $A$ be an $m \times n$ matrix. The left null space of $A$ is $\mathcal{L}(A) = \mathcal{N}\!\left(A^{t}\right) \subseteq \mathbb{C}^m$ , the set of all vectors $\vec{y}$ of size $m$ satisfying $A^{t}\vec{y} = \vec{0}$ .

Equivalently, $\mathcal{L}(A) = \{\, \vec{y} \in \mathbb{C}^m \mid \vec{y}^{\,t} A = \vec{0}^{\,t} \,\}$ . The two formulations are related by transposition and the reversal rule $(A^{t}\vec{y})^{t} = \vec{y}^{\,t} A$ .

Row-dependence reading

RDR

A vector $\vec{y} \in \mathcal{L}(A)$ has entries $y_1, \ldots, y_m$ that satisfy $y_1 \vec{r}_1 + y_2 \vec{r}_2 + \cdots + y_m \vec{r}_m = \vec{0}$ , where $\vec{r}_i$ denotes row $i$ of $A$ . Non-zero elements of the left null space are therefore in one-to-one correspondence with non-trivial relations of linear dependence among the rows of $A$ .

Consistency condition

If $\vec{y} \in \mathcal{L}(A)$ and the system $A\vec{x} = \vec{b}$ has any solution, then $\vec{y}^{\,t}\vec{b} = 0$ . A basis for $\mathcal{L}(A)$ therefore supplies a complete set of $m - r$ independent scalar tests that a right-hand side must pass for the system to be consistent.

Concepts

Why the vectors act from the left

Start from the defining equation $A^{t}\vec{y} = \vec{0}$ and transpose it. Using the reversal rule for a transposed product and the fact that transposing twice returns the original matrix, the left-hand side becomes $\vec{y}^{\,t} A$ and the right-hand side becomes the zero row vector. So $\vec{y}^{\,t} A = \vec{0}^{\,t}$ : the vector multiplies $A$ on the left, whereas an element of the ordinary null space multiplies $A$ on the right. The name is a description of where the vector sits in the product, not a claim about any left-handed property of the matrix.

A ledger of row operations

Row operations replace the rows of a matrix with linear combinations of the original rows, and after a full reduction each row of the result is such a combination. When the reduced form has a zero row, the scalars that produced it form a vector satisfying $\vec{y}^{\,t} A = \vec{0}^{\,t}$ , hence an element of $\mathcal{L}(A)$ . Running the argument in reverse shows the correspondence is exact: the left null space contains precisely the net effects of all row-operation sequences that annihilate a row. This is why its dimension is $m - r$ , the number of zero rows produced.

Dimension and the rank connection

Applying the rank-nullity relation to $A^{t}$ , which has $m$ columns, gives $\dim \mathcal{L}(A) = m - \operatorname{rank}\!\left(A^{t}\right)$ . Since a matrix and its transpose share a rank, this is $m - r$ . A matrix with full row rank ( $r = m$ ) therefore has a trivial left null space: no combination of rows cancels, no zero rows appear on reduction, and every right-hand side gives a consistent system. A shortfall $m - r > 0$ counts the redundant equations exactly.

Parity relations and analytical redundancy

Model a measurement campaign as $\vec{b} = A\vec{x}$ , with $A$ the sensitivity matrix mapping $n$ physical states to $m$ channels. Any $\vec{y} \in \mathcal{L}(A)$ gives $\vec{y}^{\,t}\vec{b} = \vec{y}^{\,t} A \vec{x} = 0$ for every possible state, so the scalar $\vec{y}^{\,t}\vec{b}$ is identically zero on healthy data and depends on no unknown. Evaluating it on real readings produces a residual whose departure from zero indicates sensor fault, model error or noise. A basis for $\mathcal{L}(A)$ is a maximal independent set of such tests, and its size $m - r$ is the degree of analytical redundancy the instrumentation provides.

Relation to the column space

The left null space and the column space are both subspaces of $\mathbb{C}^m$ , with dimensions $m - r$ and $r$ . That complementarity is not an accident: $\vec{y} \in \mathcal{L}(A)$ holds exactly when $\vec{y}$ annihilates every column of $A$ , and hence every vector in $\mathcal{C}(A)$ , under the ordinary bilinear pairing $\vec{y}^{\,t}\vec{v}$ . The left null space is the set of linear functionals vanishing on the column space, which is precisely why its elements serve as consistency tests for $A\vec{x} = \vec{b}$ .

Computing it

The direct route is to transpose, reduce and read off a null-space basis in the usual way: the free columns of $\operatorname{RREF}\!\left(A^{t}\right)$ index the basis vectors, and each is obtained by setting one free variable to $1$ and the rest to $0$ . A second route avoids the transpose altogether. Augmenting $A$ with an $m \times m$ identity block and reducing the extended array records every row operation in the appended block; the rows of that block sitting alongside the zero rows of the reduced $A$ form a basis for $\mathcal{L}(A)$ directly.

Procedure: find a basis for the left null space

Transpose the matrixForm

A^{t}

, of size

n \times m

. The left null space of

A

is by definition the ordinary null space of this matrix.

Row-reduce $A^{t}$ Bring

A^{t}

to reduced row-echelon form. Record the pivot column indices and the free column indices; there are

m - r

free columns.

Build one basis vector per free columnSet the chosen free variable to

1

and the other free variables to

0

; the pivot rows then determine the remaining entries. Repeat for each free column.

Verify against the rows of $A$ Each basis vector's entries should annihilate the rows of

A

: check that

\vec{y}^{\,t} A = \vec{0}^{\,t}

by direct combination of the original rows.

Interpret the resultDimension

m - r

is the number of redundant equations. Each basis vector is one row dependency and one consistency test on the right-hand side.

Equations

Definition of the left null space

EQ-LNS-01

L (A) = N (A^{t}) = {\vec{y} \in ℂ^{m} ∣ A^{t} \vec{y} = \vec{0}}

For an $m \times n$ matrix $A$ . The result is a subspace of $\mathbb{C}^m$ , the same ambient space as the column space.

The left-acting form

EQ-LNS-02

The left-acting form

Obtained by transposing the defining equation. This is the form that explains the name and exposes the row-dependence reading.

Row dependency

EQ-LNS-03

y_{1} {\vec{r}}_{1} + y_{2} {\vec{r}}_{2} + \dots + y_{m} {\vec{r}}_{m} = \vec{0}

Written with $\vec{r}_i$ for the $i$ -th row of $A$ . The entries of a left null space vector are exactly the scalars in a vanishing combination of the rows.

Dimension

EQ-LNS-04

\dim L (A) = m - r, r = rank (A)

Equal to the number of zero rows in the reduced row-echelon form of $A$ . It vanishes precisely when $A$ has full row rank.

Consistency test

EQ-LNS-05

A \vec{x} = \vec{b} consistent \Rightarrow {\vec{y}}^{t} \vec{b} = 0 \forall \vec{y} \in L (A)

Applying $\vec{y}^{\,t}$ to both sides eliminates $\vec{x}$ entirely. Taken over a basis of $\mathcal{L}(A)$ , these $m - r$ conditions are also sufficient for consistency.

Parity residual

EQ-LNS-06

ρ = {\vec{y}}^{t} \vec{b}, ρ = 0 for all fault-free measurements

The scalar residual generated by a left null space vector. It is independent of the unknown state, so any non-zero value is attributable to fault, noise or model error.

Complementary dimensions in $\mathbb{C}^m$

EQ-LNS-07

Complementary dimensions in $\mathbb{C}^m$

The column space and left null space partition the dimension of the ambient space $\mathbb{C}^m$ ; the left null space annihilates the column space under the bilinear pairing.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Matrix	The $m \times n$ matrix under analysis	m x n matrix over C
$A^{t}$	Transpose	The $n \times m$ matrix whose ordinary null space is $\mathcal{L}(A)$	n x m matrix over C
$L (A)$	Left null space	All vectors annihilating $A$ from the left	subspace of C^m
$N (A)$	Null space	All vectors annihilating $A$ from the right; a subspace of $\mathbb{C}^n$	subspace of C^n
$\vec{y}$	Left null vector	A vector of scalars whose weighted combination of the rows of $A$ vanishes	vector in C^m
${\vec{r}}_{i}$	Row of $A$	The $i$ -th row of $A$ , treated as a vector of length $n$	vector in C^n
$r$	Rank	Number of non-zero rows in the reduced form of $A$	0 to min(m,n)
$m$	Row count	Number of rows of $A$ ; equivalently the number of measurement channels or equations	positive integer
$ρ$	Parity residual	The scalar $\vec{y}^{\,t}\vec{b}$ , identically zero on data consistent with the model	complex scalar

Worked Numerical Example

Problem statement

A condition-monitoring rig carries five transducers observing a three-parameter thermal state. Determine how much analytical redundancy the instrumentation provides, derive the parity relations, and confirm that they annihilate the sensitivity model.

Write the sensitivity matrix
Row $i$ holds the sensitivity of transducer $i$ to the three state parameters. There are $m = 5$ channels and $n = 3$ states.
$A = [\begin{matrix} 1 & 2 & - 1 \\ 2 & - 1 & 3 \\ 0 & 1 & 1 \\ 3 & 1 & 2 \\ 4 & - 3 & 5 \end{matrix}]$
Transpose
The left null space of $A$ is the ordinary null space of $A^{t}$ , so form the $3 \times 5$ transpose. Its columns are the original rows of $A$ .
$A^{t} = [\begin{matrix} 1 & 2 & 0 & 3 & 4 \\ 2 & - 1 & 1 & 1 & - 3 \\ - 1 & 3 & 1 & 2 & 5 \end{matrix}]$
Reduce
Bringing $A^{t}$ to reduced row-echelon form gives leading ones in columns $1$ , $2$ and $3$ . So $\operatorname{rank}\!\left(A^{t}\right) = 3 = r$ , columns $4$ and $5$ are free, and $\dim \mathcal{L}(A) = m - r = 5 - 3 = 2$ .
$[\begin{matrix} 1 & 0 & 0 & 1 & 0 \\ 0 & 1 & 0 & 1 & 2 \\ 0 & 0 & 1 & 0 & - 1 \end{matrix}]$
Build the basis vectors
Set $y_4 = 1$ , $y_5 = 0$ : the three pivot equations give $y_1 = -1$ , $y_2 = -1$ , $y_3 = 0$ . Then set $y_4 = 0$ , $y_5 = 1$ : they give $y_1 = 0$ , $y_2 = -2$ , $y_3 = 1$ .
$L (A) = span ({[\begin{matrix} - 1 \\ - 1 \\ 0 \\ 1 \\ 0 \end{matrix}], [\begin{matrix} 0 \\ - 2 \\ 1 \\ 0 \\ 1 \end{matrix}]})$
Verify the first relation against the rows
The first basis vector asserts that row 4 equals the sum of rows 1 and 2. Checking directly: $(1,2,-1) + (2,-1,3) = (3,1,2)$ , which is row 4. So $-\vec{r}_1 - \vec{r}_2 + \vec{r}_4 = \vec{0}$ , and the vector is genuinely in $\mathcal{L}(A)$ .
$- {\vec{r}}_{1} - {\vec{r}}_{2} + {\vec{r}}_{4} = - (1, 2, - 1) - (2, - 1, 3) + (3, 1, 2) = (0, 0, 0)$
Verify the second relation
The second basis vector asserts $-2\vec{r}_2 + \vec{r}_3 + \vec{r}_5 = \vec{0}$ . Evaluating gives $-(4,-2,6) + (0,1,1) + (4,-3,5) = (0,0,0)$ , confirming the relation and hence the second parity check.
State the parity relations
Writing $b_i$ for the reading of transducer $i$ , each basis vector converts into a scalar test on the measurement vector. Both hold identically for any thermal state consistent with the model, so both are pure functions of the data.
$ρ_{1} = - b_{1} - b_{2} + b_{4}, ρ_{2} = - 2 b_{2} + b_{3} + b_{5}$

Result

The rig has $\dim \mathcal{L}(A) = 2$ degrees of analytical redundancy: five channels observe only three independent states, so two independent scalar combinations of the readings must vanish for any physically realisable state. Monitoring $\rho_1$ and $\rho_2$ detects a fault in any single channel without knowing the thermal state at all, and the pattern of which residuals move localises the fault. A rig with $m = r$ would have a trivial left null space and no self-checking capability whatsoever.

Applications & Industry Use

Aerospace and avionics

Sensor fault detection by parity space

Redundant inertial and air-data sensors are related to a smaller set of physical states by a fixed geometry matrix. The left null space of that matrix generates the parity vector whose components are zero in fault-free operation. Threshold logic on these residuals isolates a failed unit without any estimate of the vehicle state, which is the standard basis for analytical redundancy in flight-critical monitoring.

Process and chemical engineering

Data reconciliation and gross error detection

Flow, composition and energy balances around a plant impose linear constraints on measured streams. The left null space of the balance matrix supplies the independent balance closures that measured data must satisfy. Systematic non-closure of a specific combination points to a biased instrument or an unrecorded stream, and quantifies how much redundancy the measurement layout provides before any reconciliation is attempted.

Electrical power systems

State estimation observability and bad data

In power system state estimation the measurement Jacobian relates meter readings to bus states. Elements of its left null space form residual tests that are independent of the true state, and the dimension of that subspace determines how many bad measurements can in principle be identified. A trivial left null space means no measurement can be cross-checked at all.

Structural engineering

Redundant equilibrium equations

Assembling nodal equilibrium equations for a structure often yields more equations than independent conditions, because global equilibrium is implied by the local ones. The left null space enumerates those implied combinations exactly and thereby identifies the equilibrium checks that carry no new information but serve as valuable arithmetic verification of an assembled model.

Metrology and calibration

Closure conditions in a measurement network

Levelling loops, interferometric chains and comparison schemes all produce measurements whose sums around a closed path must vanish. Those closure conditions are precisely the left null space of the design matrix. Their number sets the redundancy of the network, and their evaluated values are the misclosures used to assess measurement quality before adjustment.

Systems biology

Conserved moieties in a reaction network

For a stoichiometric matrix with reactions as columns and species as rows, the left null space consists of weightings of species concentrations that remain constant under every reaction. These conservation relations — total phosphate, total enzyme, total cofactor — are read directly from a basis and constrain any dynamic simulation of the network.

Design Considerations

Distinguish carefully from the ordinary null space

For a non-square matrix the two live in different spaces: $\mathcal{N}(A) \subseteq \mathbb{C}^n$ and $\mathcal{L}(A) \subseteq \mathbb{C}^m$ , with dimensions $n - r$ and $m - r$ . They answer different questions — the null space describes ambiguity in the unknowns, the left null space describes redundancy in the equations. Confusing them is a dimensional error that will usually surface as a shape mismatch, but not always.

Use the extended reduction to avoid transposing

Materialising $A^{t}$ and reducing it duplicates work if the reduction of $A$ is needed anyway. Augmenting $A$ with an identity block and reducing once yields both the reduced form of $A$ and, in the appended block alongside the zero rows, a basis for $\mathcal{L}(A)$ . For any workflow that wants more than one of the four subsets, this is the economical route.

Complex data requires a choice of transpose

Over the complex numbers, two candidate operations exist: the plain transpose $A^{t}$ and the conjugate transpose $A^{*}$ . The definition here uses the plain transpose, giving $\vec{y}^{\,t} A = \vec{0}^{\,t}$ . Numerical libraries and much of the applied literature default to the conjugate transpose, which yields a different subspace unless $A$ is real. State which convention is in force before comparing results across sources.

A basis is not unique — choose one that suits the task

Any invertible recombination of a left null space basis is another basis, and different choices give different-looking parity relations. For fault isolation, a basis engineered so that each residual involves few channels gives a clean and interpretable fault signature. For numerical robustness, an orthonormal basis from an SVD is preferable. The subspace is fixed; the presentation is a design decision.

Redundancy in principle is not redundancy in practice

A non-trivial left null space guarantees that exact model data satisfies the parity relations, but says nothing about noise amplification. A residual whose generating vector is nearly orthogonal to the useful signal direction, or whose entries span many orders of magnitude, will be dominated by measurement noise. Assess each candidate residual for signal-to-noise before committing it to a monitoring scheme.

Full row rank means no self-checking

When $r = m$ the left null space is trivial, every right-hand side gives a consistent system, and no combination of measurements can be cross-validated. If the design intent is a self-checking instrumentation layout, the row rank must be deliberately kept below the channel count; verifying $m - r \ge 1$ is a legitimate design acceptance criterion.

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	Standardises the transpose and conjugate-transpose notation whose distinction determines which of two candidate subspaces the phrase 'left null space' denotes.
`ISO 5725-1`	Accuracy (trueness and precision) of measurement methods and results	Provides the vocabulary of trueness and precision against which parity residuals derived from the left null space are interpreted as bias or as random scatter.
`IEC 61508`	Functional safety of electrical/electronic/programmable electronic safety-related systems	Diagnostic coverage claims for redundant sensing arrangements rest on analytical redundancy; the dimension of the left null space bounds how many independent cross-checks a layout can support.
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Left null space bases in production are obtained from the trailing left singular vectors returned by `xGESVD` or from a rank-revealing QR, not from an explicit reduced form.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the row-vector products and matrix layouts on this page semantically, preserving the distinction between $\vec{y}^{\,t} A$ and $A\vec{y}$ for assistive technology.

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	Structural or symbolic analysis where the exact dimension $m - r$ and integer parity relations are the deliverable.	Yields unambiguous relations with integer coefficients that read cleanly as conservation laws, but is impractical for large or measured matrices.
IEEE 754 binary64 with SVD	Measured sensitivity matrices in monitoring, estimation or reconciliation.	The trailing left singular vectors give a well-conditioned orthonormal basis with a quantified singular-value gap, at higher cost than elimination and with no integer structure.
Rank-revealing QR of $A^{t}$	A left null space basis is needed repeatedly for a fixed geometry, and the cost of a full SVD is not justified.	Cheaper than SVD and numerically sound, but the rank decision is less robust for matrices with a shallow singular-value gap.
Integer arithmetic over $\mathbb{Z}$	Stoichiometric and incidence matrices, where the entries are small integers and conservation relations should be reported in integer form.	Preserves the physical interpretability of conserved moieties and loop sums; requires care with coefficient growth and with normalising the sign and scale of each basis vector.
Fixed-point arithmetic	Parity residuals evaluated online in an embedded monitor with no floating-point unit.	Deterministic timing and small footprint, but the basis coefficients must be pre-scaled offline and the residual thresholds set with quantisation error explicitly budgeted.
Sparse storage with a sparse QR	Large network models where the sensitivity or incidence matrix is very sparse.	Retains sparsity in the factors with a good ordering, but the computed basis vectors are typically dense, so the memory saving applies to the factorisation rather than the result.

Manufacturing Notes

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

Two computational routes

The direct route transposes and reduces, costing roughly the same as reducing $A$ itself. The extended route augments $A$ with $I_m$ and reduces the $m \times (n + m)$ array once, extracting the basis from the rows of the appended block that sit alongside the zero rows of the reduced $A$ . The second route costs more per reduction but delivers the reduced form of $A$ and material for the other three subsets at the same time.

Doing it by hand

Transpose first and write the result down explicitly — attempting the reduction mentally on the untransposed matrix is the dominant source of error. Reduce, identify the free columns, and construct one basis vector per free column by setting that variable to $1$ and the other free variables to $0$ . Then check every vector by combining the original rows of $A$ ; this catches both transposition slips and arithmetic slips.

Library behaviour

SymPy's Matrix.nullspace() applied to A.T returns an exact basis. In SciPy, scipy.linalg.null_space(A.T) returns an orthonormal basis computed from the SVD with a documented tolerance; note that A.T is the plain transpose while A.conj().T is the adjoint, and the two differ for complex data. MATLAB's null(A.') behaves analogously.

Verification technique

Compute $Y^{t} A$ where $Y$ holds the candidate basis vectors as columns; the result should be the zero matrix to within tolerance. Independently confirm the count: the number of basis vectors must equal $m$ minus the rank obtained from a separate computation. Both checks are cheap relative to the derivation and catch the common transposition error immediately.

Conditioning of the residuals

When a left null space basis is used to generate monitoring residuals, normalise each basis vector to unit norm so that residual magnitudes are comparable across relations and thresholds can be set on a common scale. An orthonormal basis from the SVD does this automatically and additionally minimises the amplification of measurement noise into any single residual.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Computing $\mathcal{N}(A)$ instead of $\mathcal{N}(A^{t})$	high	Omitting the transpose, so the ordinary null space is returned in place of the left null space.	Check the size of the result: elements of $\mathcal{L}(A)$ have $m$ entries, elements of $\mathcal{N}(A)$ have $n$ . Verify $\vec{y}^{\,t} A = \vec{0}^{\,t}$ explicitly.
Silently using the conjugate transpose	high	A library defaulting to the adjoint $A^{*}$ on complex data, producing a different subspace from the one defined by $A^{t}$ .	State the convention in the specification and use the explicit plain-transpose operator; for real matrices the two agree, so the bug only surfaces on complex data.
Concluding no redundancy from a trivial computed basis	medium	A floating-point rank computation without tolerance reports full row rank on a matrix that is only numerically close to rank-deficient.	Inspect the singular values rather than counting pivots, and report the gap that justifies the chosen rank.
Treating any non-zero residual as a fault	medium	Applying the exact condition $\vec{y}^{\,t}\vec{b} = 0$ to noisy measurements with no threshold.	Propagate the measurement covariance through the residual to obtain its distribution under the healthy hypothesis, and set thresholds from that distribution rather than from zero.
Assuming a basis is canonical	low	Comparing left null space bases from two sources and concluding the subspaces differ because the vectors differ.	Compare subspaces, not vectors: stack both bases and check that the combined rank equals $m - r$ .
Ignoring scaling when interpreting residuals	medium	Using a raw basis whose entries span several orders of magnitude, so one channel dominates the residual and the others are effectively unmonitored.	Normalise the basis vectors, or derive them from an SVD, and inspect each vector's entries for near-zero coefficients before deploying the residual.
Confusing the left null space with a left inverse	low	Reading the word 'left' as referring to inversion rather than to the side on which the vector multiplies.	The defining relation is $\vec{y}^{\,t} A = \vec{0}^{\,t}$ . A left inverse satisfies $BA = I_n$ and is a different object entirely, existing only when $A$ has full column rank.
Expecting a non-trivial left null space from a full-row-rank matrix	low	Assuming that having more equations than unknowns guarantees redundancy relations.	Redundancy is governed by $m - r$ , not by $m - n$ . A tall matrix with full row rank ( $r = m$ ) has a trivial left null space.

FAQs

Why is it called the left null space?

Because its vectors annihilate $A$ from the left. Transposing the defining equation $A^{t}\vec{y} = \vec{0}$ gives $\vec{y}^{\,t} A = \vec{0}^{\,t}$ , in which $\vec{y}^{\,t}$ appears on the left of $A$ . Ordinary null space vectors appear on the right, in $A\vec{x} = \vec{0}$ . The name records the position in the product and has nothing to do with left inverses.

How is the left null space different from the ordinary null space?

They describe different things and generally live in different spaces. For an $m \times n$ matrix, $\mathcal{N}(A) \subseteq \mathbb{C}^n$ has dimension $n - r$ and measures ambiguity in the unknowns; $\mathcal{L}(A) \subseteq \mathbb{C}^m$ has dimension $m - r$ and measures redundancy among the equations. They coincide only in special cases such as certain symmetric matrices.

What does a vector in the left null space actually tell me?

Two things at once. Read against the rows of $A$ , its entries are the scalars in a linear combination of rows that cancels — a statement that those equations are not independent. Read against a right-hand side, it gives the scalar test $\vec{y}^{\,t}\vec{b} = 0$ that any consistent $\vec{b}$ must satisfy.

When is the left null space trivial?

Exactly when $A$ has full row rank, $r = m$ . Then no combination of rows cancels, no zero rows appear on reduction, and $A\vec{x} = \vec{b}$ is consistent for every possible $\vec{b}$ — equivalently the column space is all of $\mathbb{C}^m$ . This is a common design goal for a well-posed system and a design failure for a self-checking instrument.

Can I find it without transposing the matrix?

Yes. Augment $A$ with an $m \times m$ identity block and row-reduce the extended array. The appended block records the cumulative row operations, and the rows of that block sitting alongside the zero rows of the reduced $A$ form a basis for the left null space directly. This is also the route that delivers all four subsets from one reduction.

Should I use the transpose or the conjugate transpose for complex matrices?

The definition here uses the plain transpose, which is what makes the identity $\vec{y}^{\,t} A = \vec{0}^{\,t}$ exact. Most numerical libraries default to the conjugate transpose, giving the orthogonal complement of the column space under the Hermitian inner product. For real matrices the two agree; for complex matrices state the convention explicitly.

How does the left null space relate to the column space?

Its elements are precisely the vectors annihilating every column of $A$ , and hence everything in $\mathcal{C}(A)$ , under the pairing $\vec{y}^{\,t}\vec{v}$ . The two subspaces sit in the same $\mathbb{C}^m$ with dimensions $r$ and $m - r$ summing to $m$ , which is why membership of $\vec{b}$ in the column space can be tested entirely through left null space conditions.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section FS, Subsection LNS. 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.
Gertler, J. Fault Detection and Diagnosis in Engineering Systems. Marcel Dekker, 1998.
Narasimhan, S. and Jordache, C. Data Reconciliation and Gross Error Detection. Gulf Publishing, 2000.
IEC 61508:2010, Functional safety of electrical/electronic/programmable electronic safety-related systems. International Electrotechnical Commission.

AI Suggested Questions

Given a 6x4 sensitivity matrix, derive a set of parity relations from its left null space and show how a single-channel bias appears in the residuals.
Explain the difference between the left null space defined with the plain transpose and the one defined with the conjugate transpose, using a complex 3x2 example.
Why does the left null space of a stoichiometric matrix correspond to conserved quantities, and how do I read a conserved moiety off a basis vector?
Show how augmenting a matrix with an identity block and reducing once yields a left null space basis without ever forming the transpose.
How should I choose among the infinitely many bases for a left null space when designing fault-isolation residuals?
For a matrix with a shallow singular-value gap, how do I decide whether a computed left null space vector is real or an artefact of rounding?

Related Calculators

Left Null Space Calculator

Compute a basis for $\mathcal{L}(A)$ with exact or floating-point arithmetic, reporting dimension $m - r$ and verifying $\vec{y}^{\,t} A = \vec{0}^{\,t}$ .

Parity Relation Generator

Turn a sensitivity matrix into a normalised set of measurement residuals, with noise propagation and suggested detection thresholds.

Matrix Rank & Nullity Calculator

Report rank, nullity and left-nullity for a matrix with a selectable rank tolerance.