Properties of Matrix Multiplication

Q: In what sense do square matrices form a ring?

For a fixed size $n$, the set $M_{nn}$ is closed under addition and multiplication, addition is commutative with identity $O$ and inverses, multiplication is associative with identity $I_n$, and multiplication distributes over addition. That is exactly the definition of a ring with identity. It is non-commutative, and it has zero divisors, so it is not an integral domain and not a field.

Overview

Once matrix multiplication is defined, the immediate question is which of the familiar rules of arithmetic still apply. The answer is neither "all" nor "none". Associativity, both distributive laws, the interaction with scalars and the behaviour of the zero and identity matrices all carry over intact. Commutativity and the cancellation law do not. There is no way to guess which is which, so every rule has to be stated with its hypotheses and proved.

The proofs share a single template. Matrix equality means equality of corresponding entries, so any claimed identity is reduced to a statement about a scalar, expanded with the entry formula $[AB]_{ij} = \sum_k [A]_{ik}[B]_{kj}$ , and then settled using the ordinary arithmetic of $\mathbb{C}$ . Associativity becomes an exchange in the order of a double summation; distributivity becomes distributivity of complex numbers; the transpose rule becomes an index swap. None of these arguments needs a general array written out with ellipses, and that restraint is what keeps them short enough to check.

Taken together the rules make the set of square matrices of a fixed size into a ring with identity: closed under addition and multiplication, associative, distributive, with a multiplicative identity $I_n$ and an additive identity $O$ . It is a non-commutative ring, and one with zero divisors — non-zero matrices whose product is zero — which is precisely why cancellation is unavailable. Recognising the structure tells an engineer in advance which algebraic manipulations are legitimate.

The rules governing transposition, conjugation and the inner product are the ones most often misapplied. Conjugation passes through a product unchanged, but transposition and adjunction reverse the order of the factors. That reversal is not a curiosity: it is forced by conformability, since the transpose of an $m \times p$ product must be $p \times m$ and only the reversed order of the transposed factors has that shape. It reappears everywhere from the chain rule of backpropagation to the transformation of element stiffness matrices.

Definition

Matrix Multiplication and the Zero Matrix

MMZM

For an $m \times n$ matrix $A$ : $AO_{n \times p} = O_{m \times p}$ and $O_{p \times m}A = O_{p \times n}$ . Multiplying by a zero matrix annihilates, and the subscripts matter — the zero matrices on the two sides of each identity generally have different sizes.

The converse fails badly. $AB = O$ does not imply that $A$ or $B$ is the zero matrix; non-zero matrices with zero product are called zero divisors and they exist for every size $n \ge 2$ .

Matrix Multiplication and the Identity Matrix

MMIM

For an $m \times n$ matrix $A$ : $AI_n = A$ and $I_mA = A$ . The identity matrix $I_n$ has $1$ on the diagonal and $0$ elsewhere, and this pair of identities is what earns it the name — it plays the role for matrix multiplication that the scalar $1$ plays for ordinary multiplication.

The two identity matrices are of different sizes unless $A$ is square. Writing $AI = IA = A$ without subscripts is harmless for square matrices and wrong for rectangular ones.

Distributivity and the Scalar Rule

MMDAA

For $A$ of size $m \times n$ , $B$ and $C$ of size $n \times p$ , $D$ of size $p \times s$ and a scalar $\alpha$ :

$A(B + C) = AB + AC$
$(B + C)D = BD + CD$
$\alpha(AB) = (\alpha A)B = A(\alpha B)$

Both distributive laws must be stated separately, because non-commutativity means neither implies the other.

Distributivity requires $B$ and $C$ to be the same size, since their sum must exist. The scalar may be moved freely into either factor or left outside.

Associativity of Matrix Multiplication

MMA

For $A$ of size $m \times n$ , $B$ of size $n \times p$ and $D$ of size $p \times s$ : $A(BD) = (AB)D$ . Both groupings are conformable and both produce an $m \times s$ matrix. Because the grouping is immaterial, a product of several matrices may be written without parentheses at all.

Concepts

The entry-by-entry proof template

Every property on this page is established the same way. To prove $X = Y$ for matrix expressions, show $[X]_{ij} = [Y]_{ij}$ for all valid $i$ and $j$ , then invoke the definition of matrix equality. Expanding with the entry formula turns matrices into finite sums of complex numbers, at which point the field axioms take over. Associativity, for example, becomes $\sum_k \sum_\ell [A]_{ik}[B]_{k\ell}[D]_{\ell j}$ evaluated in two different orders, and finite sums may be reordered freely. The discipline of working with a single entry rather than a whole array is what keeps these proofs verifiable.

What associativity buys

Three consequences follow immediately. First, a chain $A B C D$ needs no parentheses, so notation stays clean. Second, matrix powers $A^{k}$ are unambiguous, which is what makes matrix polynomials, exponentials and eigenvalue theory possible. Third — and most practically — the grouping can be chosen for cost. Multiplying an $m \times n$ by an $n \times p$ by a $p \times 1$ costs $mnp + mp$ multiplications if grouped from the left and $np + mn$ from the right; for a large chain the optimal parenthesisation is found by dynamic programming and can change the run time by orders of magnitude.

Bilinearity

The two distributive laws together with the scalar rule say exactly that matrix multiplication is a bilinear map: linear in its left argument when the right is held fixed, and linear in its right argument when the left is held fixed. This is the precise sense in which multiplication respects the vector space structure of matrices, and it is what allows a product to be expanded, factored and differentiated in the ordinary way, provided the order of factors is never disturbed.

Conjugation, transposition and the order-reversal rule

$(AB)^{t} = B^{t}A^{t}$ , and since conjugation passes straight through a product, $(AB)^{*} = B^{*}A^{*}$ as well. The reversal is forced before any arithmetic is done: if $A B$ is $m \times p$ then $(AB)^{t}$ is $p \times m$ , and the only conformable way to build that from $A^{t}$ (of size $n \times m$ ) and $B^{t}$ (of size $p \times n$ ) is $B^{t}A^{t}$ . The entry proof then confirms the numbers agree. The same reversal governs inverses, $(AB)^{-1} = B^{-1}A^{-1}$ , and it is the reason gradients propagate backwards through a network of linear layers.

$\overline{AB} = \overline{A}\ \overline{B}$ , with no reversal of order. Conjugation is an entry-wise operation and complex conjugation is multiplicative and additive, so it commutes with every step of the entry formula. Combining this with the transpose rule gives the adjoint rule, and it explains the asymmetry: the reversal in $(AB)^{*} = B^{*}A^{*}$ comes entirely from the transposition, never from the conjugation.

The inner product is a matrix product

Regarding $\vec{u}, \vec{v} \in \mathbb{C}^m$ as $m \times 1$ matrices, $\langle \vec{u}, \vec{v} \rangle = \vec{u}^{\,t}\overline{\vec{v}} = \vec{v}^{*}\vec{u}$ , a $1 \times 1$ matrix identified with its lone entry. This is more than a notational trick: it means every theorem about matrix products applies to inner products for free, and it produces the adjoint relation $\langle A\vec{u}, \vec{v} \rangle = \langle \vec{u}, A^{*}\vec{v} \rangle$ by a single application of the reversal rule. Gram matrices, normal equations and projection formulas all descend from this identification.

What does not hold, and why it matters

Commutativity fails: $A B$ and $B A$ may differ, may have different sizes, or only one may exist. Cancellation fails: $AB = AC$ does not give $B = C$ , because $A(B - C) = O$ can occur with $B \ne C$ . Zero divisors exist, so $AB = O$ tells you nothing about $A$ or $B$ individually. And $(AB)^{-1}$ is $B^{-1}A^{-1}$ , never $A^{-1}B^{-1}$ . Each of these is a live source of error in derivations, and the discipline is simple: never perform a manipulation that is not licensed by a stated rule.

Decision path: is this manipulation legitimate?

Are all products conformable?Write the sizes of every factor first. Many invalid manipulations are caught at this stage because the rewritten expression does not typecheck.

Does the step change the order of factors?If yes, it is illegal unless justified by a specific rule such as the transpose reversal or a proven commuting pair. Non-commutativity is the default.

Does it only regroup, expand or move a scalar?Associativity, the two distributive laws and the scalar rule permit these freely. No hypothesis beyond conformability is required.

Does it cancel a common factor?Not permitted. Cancellation requires the factor to be invertible, in which case multiply both sides by its inverse on the correct side, and state which side.

Does it involve a transpose, conjugate or inverse of a product?Conjugation preserves order. Transposition, adjunction and inversion all reverse it. Check the sizes of the rewritten expression as a confirmation.

Equations

Zero matrix annihilates

EQ-PMM-01

A O_{n \times p} = O_{m \times p}, O_{p \times m} A = O_{p \times n}

Valid for any $m \times n$ matrix $A$ . The zero matrices on the two sides of each equation are of different sizes, which the subscripts record.

Identity matrix acts trivially

EQ-PMM-02

A I_{n} = A, I_{m} A = A

For an $m \times n$ matrix. The two identities have different sizes unless $A$ is square; this is the property that names the matrix.

Distributive laws

EQ-PMM-03

A (B + C) = A B + A C, (B + C) D = B D + C D

Both must be stated. Non-commutativity means the left-hand and right-hand versions are logically independent, though both are true.

Scalars commute with products

EQ-PMM-04

α (A B) = (α A) B = A (α B), α \in ℂ

A scalar may be extracted from a product or pushed into either factor. Scalars are the one thing that commutes with everything.

Associativity

EQ-PMM-05

A (B D) = (A B) D

For $A$ of size $m \times n$ , $B$ of size $n \times p$ and $D$ of size $p \times s$ . The common value is $\sum_k \sum_\ell [A]_{ik}[B]_{k\ell}[D]_{\ell j}$ , summed in either order.

Conjugate, transpose and adjoint of a product

EQ-PMM-06

\bar{A B} = \bar{A} \bar{B}, (A B)^{t} = B^{t} A^{t}, (A B)^{*} = B^{*} A^{*}

Conjugation preserves the order; transposition and adjunction reverse it. The reversal is required for conformability alone, before any entries are compared.

Inner product as a matrix product

EQ-PMM-07

⟨ \vec{u}, \vec{v} ⟩ = {\vec{u}}^{t} \bar{\vec{v}} = {\vec{v}}^{*} \vec{u}, ⟨ A \vec{u}, \vec{v} ⟩ = ⟨ \vec{u}, A^{*} \vec{v} ⟩

Treating vectors as single-column matrices identifies the inner product with a $1 \times 1$ product, and the adjoint relation follows from the order-reversal rule.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Left factor	Matrix appearing first in a product	m x n matrix
$B, C$	Middle factors	Matrices of matching size, so that $B + C$ exists in the distributive laws	n x p matrices
$D$	Right factor	Matrix completing a three-term chain	p x s matrix
$α$	Scalar	Multiplier that may be moved into either factor of a product	element of C
$I_{n}$	Identity matrix	Square matrix with ones on the diagonal, the multiplicative identity of $M_{nn}$	n x n matrix
$O_{m \times p}$	Zero matrix	Matrix of all zeros, annihilator under multiplication and identity under addition	m x p matrix
$A^{t}$	Transpose	Rows and columns exchanged; reverses factor order in a product	n x m matrix
$A^{*}$	Adjoint	Conjugate transpose $\overline{A}^{\,t}$ ; reverses factor order and conjugates entries	n x m matrix
$M_{m n}$	Matrix space	The set of all $m \times n$ matrices; $M_{nn}$ is a ring under these operations	vector space over C

Worked Numerical Example

Problem statement

A measurement chain applies three linear stages to a raw reading: a $4$ -channel calibration vector is combined by a $3 \times 4$ sensor-fusion stage, then reduced by a $2 \times 3$ output stage. Verify associativity numerically, count the cost of both groupings, and confirm the transpose reversal rule on the same data.

Define the three factors
$A$ is the $2 \times 3$ output stage, $B$ the $3 \times 4$ fusion stage and $\vec{d}$ the $4 \times 1$ raw reading. The chain $AB\vec{d}$ is conformable as $(2 \times 3)(3 \times 4)(4 \times 1)$ and produces a $2 \times 1$ result whichever way it is grouped.
$A = [\begin{matrix} 1 & 2 & - 1 \\ 0 & 3 & 4 \end{matrix}], B = [\begin{matrix} 2 & 0 & 1 & - 1 \\ 1 & 3 & - 2 & 0 \\ 0 & 1 & 4 & 2 \end{matrix}], \vec{d} = [\begin{matrix} 1 \\ - 2 \\ 3 \\ 1 \end{matrix}]$
Group from the left: form the composite stage first
Computing $A B$ collapses the two stages into a single $2 \times 4$ operator. Entry $(1,1)$ , for instance, is $(1)(2) + (2)(1) + (-1)(0) = 4$ .
$A B = [\begin{matrix} 4 & 5 & - 7 & - 3 \\ 3 & 13 & 10 & 8 \end{matrix}]$
Apply the composite to the reading
Row $1$ of $A B$ against $\vec{d}$ gives $4 - 10 - 21 - 3 = -30$ ; row $2$ gives $3 - 26 + 30 + 8 = 15$ .
$(A B) \vec{d} = [\begin{matrix} - 30 \\ 15 \end{matrix}]$
Group from the right instead
Now push the reading through the fusion stage first, then through the output stage. $B\vec{d}$ is a $3 \times 1$ intermediate: $2 + 0 + 3 - 1 = 4$ , $1 - 6 - 6 + 0 = -11$ and $0 - 2 + 12 + 2 = 12$ .
$B \vec{d} = [\begin{matrix} 4 \\ - 11 \\ 12 \end{matrix}], A (B \vec{d}) = [\begin{matrix} 4 - 22 - 12 \\ 0 - 33 + 48 \end{matrix}] = [\begin{matrix} - 30 \\ 15 \end{matrix}]$
Compare the cost of the two groupings
Left grouping: $2 \times 3 \times 4 = 24$ multiplications to form $A B$ , then $2 \times 4 \times 1 = 8$ to apply it, giving $32$ . Right grouping: $3 \times 4 \times 1 = 12$ for $B\vec{d}$ , then $2 \times 3 \times 1 = 6$ , giving $18$ . Associativity guarantees identical answers; only the work differs, here by a factor of $1.8$ , and by far more at realistic sizes.
$32 multiplications versus 18 multiplications$
Verify the transpose reversal
$A B$ is $2 \times 4$ , so $(AB)^{t}$ is $4 \times 2$ . The only conformable product of the transposed factors is $B^{t}A^{t}$ , of size $(4 \times 3)(3 \times 2)$ . Computing both gives the same array, entry for entry.
$(A B)^{t} = [\begin{matrix} 4 & 3 \\ 5 & 13 \\ - 7 & 10 \\ - 3 & 8 \end{matrix}] = B^{t} A^{t}$
Confirm that the wrong order is not even defined
The erroneous form $A^{t}B^{t}$ would be $(3 \times 2)(4 \times 3)$ , which is not conformable at all. Checking sizes therefore catches this particular mistake before any arithmetic, which is why the size check belongs at the start of every manipulation rather than the end.
Note what is not being claimed
Nothing here shows $AB = BA$ ; indeed $B A$ does not exist, since $B$ has four columns and $A$ has two rows. Associativity concerns grouping only, never the order of the factors, and the two are routinely confused.

Result

Both groupings return $\begin{bmatrix} -30 & 15 \end{bmatrix}^{t}$ , as associativity requires, but the right-to-left evaluation uses $18$ multiplications against $32$ . The engineering rule follows directly: when a chain of operators ends in a vector, apply the factors to the vector rather than forming the composite operator — unless the composite will be reused across many readings, in which case forming it once amortises the cost. The transpose of the two-stage chain is the reversed product of the transposed stages, which is the identity that propagates sensitivities backwards through the same chain.

Applications & Industry Use

Robotics

Kinematic chains and frame composition

The pose of an end effector is the product of the homogeneous transforms of each joint. Associativity lets sub-chains be precomputed and cached, while non-commutativity is the formal statement that rotating then translating differs from translating then rotating — a distinction that shows up as a physical error, not a numerical one.

Structural engineering

Element stiffness transformation

A local element stiffness matrix is mapped into global coordinates by the congruence $K = T^{t}kT$ . The order-reversal rule is what guarantees this expression is conformable and symmetric: $\left(T^{t}kT\right)^{t} = T^{t}k^{t}T$ , which equals the original whenever $k$ is symmetric.

Machine learning

Backpropagation through linear layers

A forward pass composes weight matrices in one order; the gradient computation composes their transposes in the reverse order, which is exactly the rule $(AB)^{t} = B^{t}A^{t}$ applied along the chain. Frameworks implement the backward pass as a transposed GEMM precisely because of this identity.

Control systems

State transition over multiple steps

For a discrete-time system the state after $k$ steps is $A^{k}\vec{x}_0$ , and the power is well defined only because multiplication is associative. Time-varying systems compose their step matrices in strict order, and reversing that order gives a different and physically wrong trajectory.

Wireless communications

Cascaded MIMO channels and precoders

An effective channel is the product of precoder, propagation and combiner matrices. Associativity permits the receiver to fold the combiner into the channel estimate, while the adjoint rule gives the matched filter $H^{*}$ as the natural receive processing, with $\langle H\vec{x}, \vec{y} \rangle = \langle \vec{x}, H^{*}\vec{y} \rangle$ making the relationship precise.

Computational statistics

Gram matrices and normal equations

The cross-product matrix $X^{t}X$ of a design matrix is symmetric because transposing it reverses the order and returns the same expression. That symmetry halves the storage and permits a Cholesky factorisation, and it is a direct consequence of the reversal rule rather than an accident of the data.

Design Considerations

Choose the parenthesisation before you compute

Associativity means grouping is free to choose and never free of consequence. For a chain of several matrices the optimal ordering is found by a standard dynamic programming algorithm in $O(k^{3})$ time for $k$ factors, which is negligible against the products themselves. Libraries that evaluate expressions lazily — Eigen, NumPy's multi_dot, most tensor frameworks — do this automatically; hand-written loops do not.

Never rely on commutativity, even when it happens to hold

Some specific pairs commute: a matrix with its own powers, with the identity, with any scalar multiple of the identity, and diagonal matrices with each other. Building a derivation on such a coincidence is fragile, because a later change of data or generalisation of the code silently invalidates it. State the commuting hypothesis explicitly if the argument depends on it.

Cancel only through an inverse, and record the side

From $AB = AC$ nothing follows unless $A$ is invertible, and then only by multiplying on the left by $A^{-1}$ . Left and right cancellation are different operations with different requirements. Where $A$ is rectangular or singular, the correct tool is the null space of $A$ : $B - C$ must have every column in $\mathcal{N}(A)$ .

Use transpose flags rather than materialising transposes

Every BLAS multiply accepts transpose and conjugate-transpose flags on each operand, so $B^{t}A^{t}$ is computed without ever forming $B^{t}$ or $A^{t}$ in memory. Explicitly transposing a large matrix costs a full pass over memory with poor locality and doubles the footprint, for no benefit.

Do not confuse mathematical associativity with floating-point associativity

$A(BD)$ and $(AB)D$ are equal as matrices but will generally differ in their last bits when computed in floating point, because floating-point addition is not associative and the two routes sum different intermediate quantities. Any test comparing two groupings must use a tolerance, and any claim of bitwise reproducibility must fix the evaluation order.

Exploit structure that the rules preserve

The product of two upper triangular matrices is upper triangular; the product of a matrix with its own adjoint is Hermitian; congruence $T^{*}kT$ preserves Hermitian symmetry and positive semi-definiteness. Recognising that a product is structured allows a specialised kernel and a specialised storage format, often halving both time and memory.

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
`BLAS Level 3`	Basic Linear Algebra Subprograms, matrix-matrix operations	The `xGEMM` signature $C \leftarrow \alpha \, \mathrm{op}(A)\,\mathrm{op}(B) + \beta C$ encodes the scalar rule and distributivity directly in its interface, and its transpose flags implement the order-reversal rule without materialising a transpose.
`ISO 80000-2`	Quantities and units — Part 2: Mathematics	Standardises the notation for the identity matrix, the transpose and the conjugate transpose, distinctions on which every identity in this page depends.
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Floating-point addition is not associative, so mathematically identical groupings of a matrix chain produce numerically different results; the standard defines the rounding that bounds the discrepancy.
`ISO/IEC 14882`	Programming languages — C++	Operator overloading and expression templates allow a library to see a whole product chain before evaluating it, which is what permits automatic reassociation into the cheapest grouping.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes products, transposes and adjoints with explicit grouping structure, so that the distinction between $A(BD)$ and $(AB)D$ survives into machine-readable form.

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 integer or rational arithmetic	Verifying an algebraic identity, testing a proof, or working with combinatorial and stoichiometric data.	The identities hold exactly, so a discrepancy is a genuine error rather than rounding; but entry growth in long chains is rapid and memory use unbounded.
IEEE 754 binary64	General engineering and scientific chains.	Identities hold to within a relative error of order $n\varepsilon$ ; equality tests between different groupings must use a tolerance rather than exact comparison.
IEEE 754 binary32 or reduced precision	Inference, graphics and real-time pipelines where throughput dominates.	Two to four times the throughput, but the gap between mathematically equal groupings widens, and long chains may need a wider accumulator to remain meaningful.
Finite field arithmetic	Coding theory, cryptography and combinatorial algorithms over $\mathbb{F}_2$ or a prime field.	All ring properties on this page hold verbatim and there is no rounding at all; but there is no notion of magnitude, so conditioning and tolerance have no meaning.
Structured storage: triangular, symmetric, banded	The operands or the product are known to be structured, as in stiffness assembly or Cholesky-based statistics.	Halves storage and roughly halves arithmetic through specialised kernels, at the cost of code paths that must be selected correctly — using a general kernel on packed storage produces silent nonsense.
Real versus complex scalars	The data carries phase, as in spectral, phasor or quantum computation.	The transpose must be replaced by the adjoint throughout or the adjoint relation for inner products fails; complex multiplication costs roughly four times its real counterpart per entry.

Manufacturing Notes

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

The proof pattern in practice

To verify any proposed identity, expand a single entry of each side with the entry formula, then justify each rewriting step by naming the scalar property used — distributivity in $\mathbb{C}$ , commutativity of complex multiplication, or exchange of finite summation order. If no such justification exists for a step, the identity is false, and constructing a $2 \times 2$ counterexample is usually a matter of minutes.

Matrix chain ordering

Given $k$ factors with dimensions $p_0 \times p_1, \ldots, p_{k-1} \times p_k$ , the minimum multiplication count over all parenthesisations is computed by dynamic programming in $O(k^{3})$ time and $O(k^{2})$ space. For chains that mix wide and tall factors the saving over a naive left-to-right evaluation is routinely a factor of ten or more, and it costs nothing at run time.

Library behaviour

NumPy's multi_dot performs chain ordering automatically; the plain @ operator evaluates strictly left to right. Eigen builds an expression tree and can fuse a scalar and an addition into a single gemm call, implementing the scalar and distributive rules as a compile-time optimisation. MATLAB evaluates left to right and offers no automatic reassociation, so the parenthesisation written is the one executed.

Verification on random data

The cheapest test of an implementation is to assert the identities themselves: generate random conformable matrices and check $\|A(BD) - (AB)D\|$ , $\|(AB)^{t} - B^{t}A^{t}\|$ and $\|AI - A\|$ against a tolerance proportional to $\varepsilon$ times the norms involved. Failures at a level far above tolerance point to an index or conformability bug rather than to rounding.

Cost of the identities themselves

Nothing on this page is free at scale. Forming $A B$ costs $2mnp$ operations, and rewriting an expression to avoid forming a large intermediate is usually a bigger win than any low-level optimisation. Where an identity permits the same result to be obtained with a smaller intermediate — associativity around a vector, or a transpose flag instead of an explicit transpose — the algebra is the performance work.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Assuming commutativity	high	Rewriting $A B$ as $B A$ inside a derivation by analogy with scalar algebra.	Preserve factor order in every step. Where a swap is genuinely needed, use the transpose or adjoint rule, or prove that the specific pair commutes.
Writing $(AB)^{t} = A^{t}B^{t}$	high	Applying transposition factor by factor without reversing the order.	Check conformability of the rewritten expression: for non-square factors the wrong order is not even defined, which exposes the error immediately.
Writing $(AB)^{-1} = A^{-1}B^{-1}$	high	Distributing inversion over a product in the same erroneous way as transposition.	Use $(AB)^{-1} = B^{-1}A^{-1}$ and confirm by multiplying out: $ABB^{-1}A^{-1} = AIA^{-1} = I$ .
Cancelling a common factor	high	Deducing $B = C$ from $AB = AC$ without establishing that $A$ is invertible.	Multiply by $A^{-1}$ explicitly on the correct side, and only after invertibility is proved. Otherwise conclude only that the columns of $B - C$ lie in $\mathcal{N}(A)$ .
Concluding a factor is zero from a zero product	high	Treating $AB = O$ as implying $A = O$ or $B = O$ , which fails because matrix rings have zero divisors.	Recall the standard counterexample of two non-zero $2 \times 2$ matrices with zero product, and reason instead about column spaces and null spaces.
Mismatched sizes in the distributive law	medium	Applying $A(B + C) = AB + AC$ when $B$ and $C$ have different shapes, so their sum does not exist.	Confirm $B$ and $C$ are the same size before distributing; the rule requires the sum to be defined, not merely each product.
Confusing transpose with adjoint on complex data	high	Using $A^{t}$ where $A^{*}$ is required, so conjugates are omitted and the adjoint relation for inner products fails.	Use the adjoint throughout complex work and test the relation $\langle A\vec{u}, \vec{v} \rangle = \langle \vec{u}, A^{*}\vec{v} \rangle$ on random complex data.
Expecting bitwise equality between groupings	low	Comparing $A(BD)$ with $(AB)D$ using exact equality in a unit test.	Compare with a tolerance scaled by $\varepsilon$ and the norms of the operands; mathematical associativity does not survive floating-point rounding.

FAQs

Which familiar arithmetic rules survive for matrices and which do not?

Associativity, both distributive laws, the scalar rule and the identity and zero laws all survive. Commutativity and cancellation do not, and the product of two non-zero matrices can be zero. The practical rule is to perform no manipulation that is not licensed by a stated theorem, because intuition transferred from scalar algebra is unreliable here.

Why does transposition reverse the order of a product but conjugation does not?

Conjugation acts entry by entry and leaves positions alone, so it commutes with every step of the entry formula. Transposition moves entries between rows and columns, and for the result to be conformable the factors must swap: the transpose of an $m \times p$ product is $p \times m$ , and only $B^{t}A^{t}$ has that shape. The adjoint reverses order because it contains a transposition.

AB = AC

, can I conclude that

B = C

No, unless $A$ is invertible. Cancellation fails because $A(B - C) = O$ can hold with $B \ne C$ whenever the columns of $B - C$ lie in the null space of $A$ . If $A$ is invertible, multiply on the left by $A^{-1}$ — and note that this is left cancellation only; the right-hand version requires the factor on the right.

Does associativity mean I can compute a chain in any order I like?

It means the result is the same for any grouping, not for any ordering of the factors. Rearranging which matrices are adjacent is illegal; changing where the parentheses sit is not only legal but is the standard optimisation, since the cost of different groupings can differ by orders of magnitude.

Why do two groupings of the same product give slightly different numbers in floating point?

Because floating-point addition is not associative, and the two groupings accumulate different intermediate sums with different rounding. Both results are correct within their error bounds, and the discrepancy is of order $\varepsilon$ times the norms involved. Any comparison must therefore use a tolerance.

In what sense do square matrices form a ring?

For a fixed size $n$ , the set $M_{nn}$ is closed under addition and multiplication, addition is commutative with identity $O$ and inverses, multiplication is associative with identity $I_n$ , and multiplication distributes over addition. That is exactly the definition of a ring with identity. It is non-commutative, and it has zero divisors, so it is not an integral domain and not a field.

How does the inner product fit into these rules?

Treat vectors as single-column matrices. Then $\langle \vec{u}, \vec{v} \rangle = \vec{u}^{\,t}\overline{\vec{v}} = \vec{v}^{*}\vec{u}$ is an ordinary matrix product with a $1 \times 1$ result. Every property of matrix products then applies to inner products automatically, and the adjoint relation $\langle A\vec{u}, \vec{v} \rangle = \langle \vec{u}, A^{*}\vec{v} \rangle$ is a single application of the order-reversal rule.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section MM, subsection PMM. Licensed under the GNU Free Documentation License v1.2.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.
Horn, R. A. and Johnson, C. R. Matrix Analysis, 2nd edition. Cambridge University Press, 2013.
Cormen, T. H., Leiserson, C. E., Rivest, R. L. and Stein, C. Introduction to Algorithms, 4th edition. MIT Press, 2022. Matrix-chain multiplication.
Dongarra, J. J., Du Croz, J., Duff, I. S. and Hammarling, S. A Set of Level 3 Basic Linear Algebra Subprograms. ACM Transactions on Mathematical Software, 1990.
IEEE 754-2019, IEEE Standard for Floating-Point Arithmetic. Institute of Electrical and Electronics Engineers.

AI Suggested Questions

Give me two non-zero 2x2 matrices whose product is the zero matrix, and explain why this makes cancellation impossible.
Prove associativity of matrix multiplication by exchanging the order of a double summation, naming the property used at each step.
Find the optimal parenthesisation for a chain of matrices with dimensions 30x35, 35x15, 15x5, 5x10 and 10x20, and give the multiplication counts.
Which pairs of matrices commute, and how would I characterise the set of matrices commuting with a given diagonal matrix?
Show how the transpose reversal rule produces the backward pass of a two-layer linear network from its forward pass.
How large can the difference between A(BD) and (AB)D become in binary64, and how should a unit test set its tolerance?

Related Calculators

Matrix Identity Verifier

Test associativity, distributivity, the transpose reversal and the adjoint rule on supplied or random matrices with a scaled tolerance.

Matrix Chain Order Optimiser

Compute the cheapest parenthesisation of a product chain from the operand dimensions and report the multiplication count for each grouping.

Commutator Calculator

Evaluate $AB - BA$ to measure how far two matrices are from commuting, with a relative measure against their norms.