Matrix Units, Full Idempotents and Matrix Ring Recognition

Q: Does recognition need any finiteness hypothesis on R?

No. The theorem is a statement about n^2 elements satisfying two identities, and its proof is a finite computation. No chain conditions, no dimension assumptions and no fullness hypothesis appear. What does matter is completeness: the diagonal units must sum to the identity of R.

Q: How do I know the coefficient ring is e_11Re_11 and not something else?

Because is constructed with entries in e_11Re_11 and is bijective onto M_n of that ring. Equivalently, e_11 is a full idempotent of M_n(S) with corner S, so the corner-ring theory identifies S uniquely up to isomorphism. The centraliser description gives the same ring in a coordinate-free way.

Q: Why is a complete orthogonal family of idempotents not enough?

Such a family gives the Peirce decomposition R = _i,j e_iRe_j, which is a *generalised* matrix ring: the diagonal entries are the possibly different rings e_iRe_i and the off-diagonal entries are bimodules. Turning it into an honest matrix ring requires all e_iR to be isomorphic, which is what supplies the exchanging elements e_ij.

Q: Is the quaternion example special to characteristic not 2?

Yes, as presented: the idempotents 12(1 i) require 2 to be invertible, and the standard presentation of quaternion algebras by i^2 = a, j^2 = b, ji = -ij is itself a characteristic-not-2 device. In characteristic 2 quaternion algebras are described by different relations and split by a different criterion.

Q: Does M_n(S) ≅ M_n(T) imply S ≅ T?

Not for arbitrary rings. It does when S and T are division rings, and more generally when they are basic, because then each is recovered as the endomorphism ring of a suitable indecomposable projective. In general the two rings are only Morita equivalent, and Morita equivalence does not imply isomorphism.

Executive Summary

Matrix rings are recognisable from the inside. If a ring $R$ contains $n^{2}$ elements $e_{i j}$ multiplying like matrix units, with $e_{11} + \dots + e_{n n} = 1$ , then $R$ is forced to be $M_{n} (S)$ for $S = e_{11} R e_{11}$ — the coefficient ring is not extra data, it is the corner at one of the units.

Lam's $(21.14)$ is the guiding computation: for $R = M_{n} (k)$ and $e = e_{11}$ one has $e_{11} r e_{11} = r_{11} e_{11}$ , so $e R e ≅ k$ , and $e$ is full because $e_{i 1} e_{11} e_{1 j} = e_{i j}$ . Everything about the pair $(R, e R e)$ — radicals, ideal lattices, Morita equivalence — specialises to familiar facts about matrix rings.

$n^{2}$ Matrix units needed

$e_{11} R e_{11}$ Recovered coefficient ring

$R e R = R$ Fullness of a matrix unit

$ℵ_{0}$ Matrix units from one failure of

a b = 1 \Rightarrow b a = 1

Overview

The standard matrix units of $M_{n} (k)$ are the matrices $e_{i j}$ with a single $1$ in position $(i, j)$ . Their entire multiplicative behaviour is captured by one identity, and it is that identity — not any reference to matrices — which does all the work.

e_{i j} e_{k l} = δ_{j k} e_{i l}, \sum_{i = 1}^{n} e_{i i} = 1 .

(MU)

A **complete set of $n \times n$ matrix units**. The Kronecker delta encodes composability; the sum condition says the family exhausts the identity.

This is the mechanism behind the last step of the Wedderburn–Artin theorem — a simple artinian ring is ${End}_{D} (V)$ , and choosing a basis of $V$ produces matrix units — and behind the splitting criteria for quaternion algebras. It also runs in reverse: Jacobson observed that a failure of Dedekind-finiteness manufactures an infinite family of matrix units, which is why rings with reasonable finiteness conditions cannot fail to be Dedekind-finite.

Two nearby pages carry the surrounding theory: Corner Rings supplies $rad (e R e) = e (rad R) e$ and the ideal correspondence, and Idempotents and Direct Decompositions of Modules supplies ${End}_{R} (e R) ≅ e R e$ .

Learning Objectives

Verify $e_{11} M_{n} (k) e_{11} ≅ k$ and that $e_{11}$ is a full idempotent.
Check that $(21.10)$ and $(21.11)$ specialise to $rad M_{n} (k) = M_{n} (rad k)$ and the classical ideal correspondence.
Prove that a complete set of $n \times n$ matrix units yields $R ≅ M_{n} (e_{11} R e_{11})$ .
Identify $e_{11} R e_{11}$ with the centraliser of the matrix units in $R$ .
Construct matrix units from $n$ pairwise isomorphic orthogonal idempotents summing to $1$ .
Reproduce Jacobson's infinite matrix units in a non-Dedekind-finite ring and deduce the finiteness criterion.

Definitions

Definition(MU)Matrix units

Let $R$ be a ring with identity and $I$ an index set. A family ${e_{i j}}_{i, j \in I}$ of elements of $R$ is a set of matrix units if $e_{i j} e_{k l} = δ_{j k} e_{i l}$ for all $i, j, k, l$ . It is complete if $I$ is finite and $\sum_{i \in I} e_{i i} = 1$ . Each $e_{i i}$ is then an idempotent, distinct diagonal units are orthogonal, and $e_{i j}$ implements an isomorphism $e_{j j} R ≅ e_{i i} R$ of right $R$ -modules.

Full idempotent: $e$ with $R e R = R$ . Equivalently $1$ is a finite sum $\sum a_{i} e b_{i}$ .
Centraliser of the matrix units: $C = {r \in R : r e_{i j} = e_{i j} r for all i, j}$ , a subring of $R$ containing $1$ .
Dedekind-finite: $a b = 1 \Rightarrow b a = 1$ . Also called von Neumann finite or directly finite.
$M_{n} (S)$: The ring of $n \times n$ matrices over $S$ . This collection writes $M_{n} (S)$ , never $𝕄_{n} (S)$ .
Quaternion algebra: For a field $k$ of characteristic not $2$ and $a, b \in k^{\times}$ , the $4$ -dimensional $k$ -algebra with basis $1, i, j, i j$ subject to $i^{2} = a$ , $j^{2} = b$ , $j i = - i j$ .

Matrix units are elements of a ring, not matrices. Nothing in the definition mentions matrices; the recognition theorem is what produces them.

Core Concepts

The corner at a matrix unit

In $R = M_{n} (k)$ take $e = e_{11}$ . For $r = (r_{i j})$ , left multiplication by $e_{11}$ keeps row $1$ and right multiplication keeps column $1$ , so $e_{11} r e_{11} = r_{11} e_{11}$ . Hence $e R e = k e_{11}$ , a ring isomorphic to $k$ with $e_{11} \leftrightarrow 1$ . The coefficient ring is not lost when we form the matrix ring; it is sitting in the corner.

The idempotent $e_{11}$ is full, because $e_{i 1} e_{11} e_{1 j} = e_{i j}$ puts every matrix unit — and hence every matrix — inside $R e_{11} R$ . By the corner-ring theory this means $M_{n} (k)$ and $k$ have the same ideal lattice and are Morita equivalent, which is the structural reason matrix rings are so well behaved.

Recognition: the units determine the ring

Suppose only that $R$ contains a complete set ${e_{i j}}$ of $n \times n$ matrix units. Write $S = e_{11} R e_{11}$ . The map that sends $r$ to the matrix of its components along the units is a ring isomorphism onto $M_{n} (S)$ :

ϕ : R ⟶ M_{n} (S), ϕ (r)_{i j} = e_{1 i} r e_{j 1} \in e_{11} R e_{11} .

(REC)

Recognition. The matrix multiplication rule appears when $1 = \sum_{k} e_{k k}$ is inserted between $r$ and $r'$ .

The mechanism is the Peirce decomposition $R = ⨁_{i, j} e_{i i} R e_{j j}$ together with the fact that the units make all the pieces $e_{i i} R e_{j j}$ isomorphic to a single one, namely $S$ , by $x \mapsto e_{1 i} x e_{j 1}$ .

Where matrix units come from

In practice one rarely hands over $n^{2}$ elements. The usable form of recognition starts from a complete orthogonal family $1 = e_{1} + \dots + e_{n}$ of idempotents that are pairwise isomorphic, meaning $e_{i} R ≅ e_{j} R$ for all $i, j$ . Isomorphisms of these modules correspond to elements $a_{i} \in e_{1} R e_{i}$ and $b_{i} \in e_{i} R e_{1}$ with $a_{i} b_{i} = e_{1}$ and $b_{i} a_{i} = e_{i}$ , and setting $e_{i j} = b_{i} a_{j}$ manufactures the matrix units.

Key Results

Example(21.14)The guiding example

Let $k$ be a ring and $R = M_{n} (k)$ , and let $e = e_{11}$ be the matrix unit. Then $e r e = r_{11} e$ for $r = (r_{i j})$ , so $e R e ≅ k$ as rings. The idempotent $e$ is full. Consequently:

$(21.10)$ gives $rad (k) ≅ e (rad M_{n} (k)) e$ , consistent with the identity $rad M_{n} (k) = M_{n} (rad k)$ ;
$(21.11) (2)$ gives a multiplication-preserving bijection between the ideals of $k$ and those of $M_{n} (k)$ , namely $𝔞 \mapsto M_{n} (𝔞)$ — the correspondence already recorded in Lam's $(3.1)$ .

Theorem(REC)Matrix ring recognition

Let $R$ be a ring with identity containing a complete set of $n \times n$ matrix units ${e_{i j}}_{1 \leq i, j \leq n}$ , that is, $e_{i j} e_{k l} = δ_{j k} e_{i l}$ and $\sum_{i} e_{i i} = 1$ . Put $S = e_{11} R e_{11}$ , a ring with identity $e_{11}$ . Then

ϕ : R ⟶ M_{n} (S), ϕ (r) = (e_{1 i} r e_{j 1})_{1 \leq i, j \leq n}

is a ring isomorphism. Moreover $S$ is isomorphic to the centraliser $C = {r \in R : r e_{i j} = e_{i j} r \forall i, j}$ , via $c \mapsto e_{11} c e_{11}$ .

Proof

Well defined. $e_{11} (e_{1 i} r e_{j 1}) e_{11} = e_{1 i} r e_{j 1}$ because $e_{11} e_{1 i} = e_{1 i}$ and $e_{j 1} e_{11} = e_{j 1}$ ; so each entry lies in $S$ . Additivity is clear.

Multiplicative. Insert $1 = \sum_{k} e_{k k} = \sum_{k} e_{k 1} e_{1 k}$ between $r$ and $r'$ :

ϕ (r r')_{i j} = e_{1 i} r r' e_{j 1} = \sum_{k} (e_{1 i} r e_{k 1}) (e_{1 k} r' e_{j 1}) = \sum_{k} ϕ (r)_{i k} ϕ (r')_{k j},

which is the matrix product. Also $ϕ (1)_{i j} = e_{1 i} e_{j 1} = δ_{i j} e_{11}$ , the identity matrix of $M_{n} (S)$ .

Injective. If $ϕ (r) = 0$ then $e_{1 i} r e_{j 1} = 0$ for all $i, j$ , hence $e_{i i} r e_{j j} = e_{i 1} (e_{1 i} r e_{j 1}) e_{1 j} = 0$ ; summing over $i$ and $j$ gives $r = 1 \cdot r \cdot 1 = \sum_{i, j} e_{i i} r e_{j j} = 0$ .

Surjective. Given $(a_{i j}) \in M_{n} (S)$ set $r = \sum_{i, j} e_{i 1} a_{i j} e_{1 j}$ . Since $e_{1 k} e_{i 1} = δ_{k i} e_{11}$ and $e_{1 j} e_{l 1} = δ_{j l} e_{11}$ , and $e_{11} a_{k l} e_{11} = a_{k l}$ , we get $e_{1 k} r e_{l 1} = a_{k l}$ , so $ϕ (r) = (a_{i j})$ .

The centraliser. Define $ψ : S \to R$ by $ψ (s) = \sum_{i} e_{i 1} s e_{1 i}$ . Then $ψ (s) e_{k l} = e_{k 1} s e_{1 l} = e_{k l} ψ (s)$ , so $ψ (s) \in C$ . Conversely $c \mapsto e_{11} c e_{11}$ maps $C$ into $S$ , and the two are mutually inverse: $e_{11} ψ (s) e_{11} = s$ , while for $c \in C$ one has $ψ (e_{11} c e_{11}) = \sum_{i} e_{i 1} e_{11} c e_{1 i} = \sum_{i} e_{i i} c = c$ , using $c e_{1 i} = e_{1 i} c$ . Multiplicativity of $c \mapsto e_{11} c e_{11}$ follows from $e_{11} c' = c' e_{11}$ for $c' \in C$ .

Corollary(REC')Recognition from isomorphic idempotents

Let $R$ be a ring with identity and let $e_{1}, \dots, e_{n}$ be pairwise orthogonal nonzero idempotents with $e_{1} + \dots + e_{n} = 1$ . If $e_{i} R ≅ e_{1} R$ as right $R$ -modules for every $i$ , then $R ≅ M_{n} (e_{1} R e_{1})$ .

Proof

Fix isomorphisms $e_{i} R ≅ e_{1} R$ . By the isomorphism ${Hom}_{R} (e_{i} R, e_{1} R) ≅ e_{1} R e_{i}$ of $(21.6)$ and its inverse, a mutually inverse pair of homomorphisms corresponds to elements $a_{i} \in e_{1} R e_{i}$ and $b_{i} \in e_{i} R e_{1}$ with $a_{i} b_{i} = e_{1}$ and $b_{i} a_{i} = e_{i}$ ; take $a_{1} = b_{1} = e_{1}$ .

Set $e_{i j} = b_{i} a_{j}$ . Since $a_{j} \in e_{1} R e_{j}$ and $b_{k} \in e_{k} R e_{1}$ , the product $a_{j} b_{k}$ lies in $e_{1} R e_{j} e_{k} R e_{1}$ , which is $0$ for $j \neq k$ and contains $a_{j} b_{j} = e_{1}$ for $j = k$ . Therefore

e_{i j} e_{k l} = b_{i} (a_{j} b_{k}) a_{l} = δ_{j k} b_{i} e_{1} a_{l} = δ_{j k} b_{i} a_{l} = δ_{j k} e_{i l},

using $e_{1} a_{l} = a_{l}$ . Also $e_{i i} = b_{i} a_{i} = e_{i}$ , so $\sum_{i} e_{i i} = 1$ . The family is a complete set of $n \times n$ matrix units, and the theorem gives $R ≅ M_{n} (e_{11} R e_{11}) = M_{n} (e_{1} R e_{1})$ .

Construction(21.26)Jacobson's infinite matrix units

Let $R$ be a ring that is not Dedekind-finite: there exist $a, b \in R$ with $a b = 1$ and $e : = b a \neq 1$ . Then $e$ is an idempotent, and for $i, j \geq 0$ the elements

e_{i j} = b^{i} (1 - e) a^{j}

form an infinite set of matrix units: $e_{i j} e_{k l} = δ_{j k} e_{i l}$ , and every $e_{i j}$ is nonzero. In particular ${e_{i i}}_{i \geq 0}$ is an infinite family of nonzero pairwise orthogonal idempotents, and $⨁_{i \geq 0} e_{i i} R$ is an infinite direct sum of nonzero right ideals of $R$ .

Proof

First $e^{2} = b (a b) a = b a = e$ . Next $a (1 - e) = a - a b a = a - a = 0$ and $(1 - e) b = b - b a b = b - b = 0$ , and $a^{m} b^{m} = 1$ for all $m \geq 0$ since $a b = 1$ .

Compute $e_{i j} e_{k l} = b^{i} (1 - e) a^{j} b^{k} (1 - e) a^{l}$ . If $j > k$ then $a^{j} b^{k} = a^{j - k}$ and $a^{j - k} (1 - e) = 0$ because $a (1 - e) = 0$ . If $k > j$ then $a^{j} b^{k} = b^{k - j}$ and $(1 - e) b^{k - j} = 0$ because $(1 - e) b = 0$ . If $j = k$ then $a^{j} b^{j} = 1$ and, since $1 - e$ is idempotent, the product collapses to $b^{i} (1 - e) a^{l} = e_{i l}$ .

Nonvanishing: if $b^{i} (1 - e) a^{j} = 0$ then multiplying on the left by $a^{i}$ and on the right by $b^{j}$ gives $a^{i} b^{i} (1 - e) a^{j} b^{j} = 1 - e = 0$ , i.e. $e = 1$ , contrary to hypothesis.

Directness of $\sum_{i} e_{i i} R$ : if $\sum_{i} x_{i} = 0$ is a finite sum with $x_{i} \in e_{i i} R$ , then applying $e_{j j}$ on the left kills every term but the $j$ -th, since $e_{j j} e_{i i} = δ_{i j} e_{i i}$ , giving $x_{j} = 0$ .

Corollary(21.27)A finiteness criterion for Dedekind-finiteness

Let $R$ be a ring such that $\bar{R} = R / rad R$ contains no infinite direct sum of nonzero right ideals — for instance, $\bar{R}$ right noetherian. Then $R$ is Dedekind-finite.

Proof

If $\bar{R}$ were not Dedekind-finite, $(21.26)$ would produce inside $\bar{R}$ an infinite family of nonzero pairwise orthogonal idempotents and hence an infinite direct sum of nonzero right ideals, contrary to hypothesis. So $\bar{R}$ is Dedekind-finite.

Now let $a b = 1$ in $R$ . Passing to $\bar{R}$ gives $\bar{a} \bar{b} = 1$ , hence $\bar{b} \bar{a} = 1$ , i.e. $b a \in 1 + rad R \subseteq U (R)$ . Choose $u \in R$ with $(b a) u = 1$ . Multiplying on the left by $a$ and using $a b = 1$ gives $a = a (b a u) = (a b) (a u) = a u$ . Multiplying $a = a u$ on the left by $b$ gives $b a = b (a u) = (b a) u = 1$ .

Proof Techniques and Method

How these proofs work, and which move to reuse.

Move 1

Insert the resolution of identity

$1 = \sum_{k} e_{k k} = \sum_{k} e_{k 1} e_{1 k}$ turned the product $r r'$ into a matrix product. Every recognition proof turns on this substitution and nothing else.

Move 2

Transport across the units

$x \mapsto e_{1 i} x e_{j 1}$ carries $e_{i i} R e_{j j}$ bijectively to $S$ . Once every Peirce block has been identified with a single ring, the generalised matrix ring becomes an honest one.

Move 3

Amplify a defect

One relation $a b = 1$ , $b a \neq 1$ was amplified by the powers $b^{i} (1 - b a) a^{j}$ into infinitely many independent idempotents. Turning a single failure into an infinite family is the standard way to prove that a finiteness condition forbids the failure.

Move 3 is worth isolating as a proof pattern: to show a finiteness hypothesis $H$ implies a cancellation property $P$ , construct from any failure of $P$ an infinite configuration that $H$ forbids. The same shape appears in proofs that right noetherian rings are Dedekind-finite and that surjective endomorphisms of noetherian modules are injective.

Note also which hypotheses are load-bearing. Recognition needs the units to be complete; without $\sum e_{i i} = 1$ one gets only that $R$ contains a matrix subring, not that $R$ is one. Jacobson's construction, by contrast, deliberately produces an incomplete family — the diagonal sum is an infinite orthogonal family that cannot reach $1$ .

Worked Example

Splitting a quaternion algebra

Let $k$ be a field with $char k \neq 2$ , let $b \in k^{\times}$ , and let $A$ be the quaternion algebra with $k$ -basis $1, i, j, i j$ and relations $i^{2} = 1$ , $j^{2} = b$ , $j i = - i j$ . We show $A ≅ M_{2} (k)$ by exhibiting matrix units. Put

e_{11} = \frac{1}{2} (1 + i), e_{22} = \frac{1}{2} (1 - i), e_{21} = j e_{11}, e_{12} = b^{- 1} e_{11} j .

(E.1)

Since $i^{2} = 1$ , $e_{11}^{2} = \frac{1}{4} (1 + 2 i + i^{2}) = \frac{1}{2} (1 + i) = e_{11}$ , and similarly $e_{22}^{2} = e_{22}$ ; also $e_{11} e_{22} = \frac{1}{4} (1 - i^{2}) = 0 = e_{22} e_{11}$ and $e_{11} + e_{22} = 1$ . The anticommutation $j i = - i j$ gives the key exchange relations

j e_{11} = \frac{1}{2} (j + j i) = \frac{1}{2} (j - i j) = e_{22} j, j e_{22} = e_{11} j .

(E.2)

Multiplication by $j$ swaps the two idempotents; this is what makes $e_{11} A ≅ e_{22} A$ .

Hence $e_{21} = j e_{11} = e_{22} j e_{11} \in e_{22} A e_{11}$ and $e_{12} = b^{- 1} e_{11} j = b^{- 1} j e_{22} \in e_{11} A e_{22}$ . Now compute, using that $j^{2} = b$ is central:

e_{12} e_{21} = b^{- 1} e_{11} j j e_{11} = b^{- 1} \cdot b e_{11} = e_{11}, e_{21} e_{12} = b^{- 1} j e_{11} j = b^{- 1} j j e_{22} = e_{22} .

(E.3)

So ${e_{11}, e_{12}, e_{21}, e_{22}}$ is a complete set of $2 \times 2$ matrix units in $A$ . Recognition gives $A ≅ M_{2} (S)$ with $S = e_{11} A e_{11}$ , and comparing dimensions, $4 = \dim_{k} A = 4 \dim_{k} S$ , forces $\dim_{k} S = 1$ , i.e. $S = k$ . Therefore $A ≅ M_{2} (k)$ : a quaternion algebra containing a non-central square root of $1$ is split.

Recognition inside Wedderburn's theorem

Let $D$ be a division ring and $V$ a right $D$ -vector space of finite dimension $n$ . Choosing a basis $v_{1}, \dots, v_{n}$ of $V$ and letting $e_{i j} \in {End}_{D} (V)$ be the map sending $v_{j} \mapsto v_{i}$ and $v_{l} \mapsto 0$ for $l \neq j$ produces a complete set of $n \times n$ matrix units, with $e_{11} {End}_{D} (V) e_{11} ≅ D^{op}$ or $D$ depending on the side conventions in force. Recognition then delivers ${End}_{D} (V) ≅ M_{n} (D)$ — the final step of Wedderburn–Artin, obtained without any further structure theory.

A non-example: an incomplete family

In the ring $R$ of upper triangular $2 \times 2$ matrices over $k$ , the elements $e_{11}, e_{12}, e_{22}$ behave correctly but there is no $e_{21}$ , since the $(2, 1)$ entry is forced to be zero. No complete set of $2 \times 2$ matrix units exists in $R$ , and indeed $R$ is not a $2 \times 2$ matrix ring: $e_{11} R$ has dimension $2$ and $e_{22} R$ dimension $1$ , so the diagonal idempotents are not isomorphic.

Process and Workflow

Find a complete orthogonal familyDecompose

1 = e_{1} + \dots + e_{n}

into orthogonal idempotents, ideally primitive ones.

Test the summands for isomorphism

e_{i} R ≅ e_{j} R

iff there are

a \in e_{i} R e_{j}

b \in e_{j} R e_{i}

with

a b = e_{i}

b a = e_{j}

. Use

{Hom}_{R} (e_{i} R, e_{j} R) ≅ e_{j} R e_{i}

to search.

Build the matrix unitsWith

a_{i} \in e_{1} R e_{i}

b_{i} \in e_{i} R e_{1}

inverse to each other, set

e_{i j} = b_{i} a_{j}

Read off the coefficient ring

S = e_{11} R e_{11}

, and

R ≅ M_{n} (S)

. Identify

S

by dimension count or by recognising it directly.

Transport resultsRadicals, ideals, chain conditions and module categories now transfer between

R

and

S

by the matrix-ring dictionary.

Your ring has a complete orthogonal family of idempotents. Are they pairwise isomorphic?

Yes, all $n$ of themRecognition applies:

R ≅ M_{n} (e_{1} R e_{1})

. This is the best possible outcome and is what happens for simple artinian rings.

They fall into several isomorphism classes

R

is a generalised matrix ring, block-decomposed by class. Group the isomorphic ones to get

R ≅

a matrix ring over a smaller algebra only within each block; globally you get a basic algebra.

No two are isomorphic

R

is already basic. No matrix ring structure is available, and

R

is Morita equivalent to itself only.

You cannot decidePass to

R / rad R

, where isomorphism of idempotents is decidable by Wedderburn–Artin, then lift the conclusion if the radical permits idempotent lifting.

Comparison and Classification

The matrix-ring dictionary induced by $e = e_{11}$
Feature of $M_{n} (S)$	Corresponding feature of $S$	Reference
$rad M_{n} (S) = M_{n} (rad S)$	$rad S$	(21.10), (21.14)
Ideals $M_{n} (𝔞)$	Ideals $𝔞$ of $S$	(21.11)(2), (3.1)
Centre	$Z (M_{n} (S)) ≅ Z (S)$ via scalar matrices	direct computation
Simple	$M_{n} (S)$ simple iff $S$ simple	(21.13) plus fullness
Left noetherian / artinian	same for $S$	(21.13) plus fullness
Module category	equivalent to that of $S$	Morita, via the progenerator $e_{11} M_{n} (S)$
Commutativity	not preserved for $n \geq 2$	$M_{2} (k)$ is noncommutative for every $k \neq 0$

Which hypotheses each result needs
	$\sum e_{i i} = 1$	Units pairwise isomorphic	Fullness	Finiteness condition
$e R e ≅ k$ for $e = e_{11}$ in $M_{n} (k)$	no	no	no	no
Recognition $R ≅ M_{n} (S)$	yes	yes	no	no
Ideal correspondence for $e_{11}$	no	no	yes	no
Jacobson's infinite units	no	no	no	no
Dedekind-finiteness criterion	no	no	no	yes

Which hypotheses each result needs

The second row records that recognition needs both completeness and mutual isomorphism when starting from idempotents; if the matrix units are handed over directly, mutual isomorphism is automatic because $e_{i j}$ implements it.

Relationship Map

Complete orthogonal family

1 = \sum e_{i}

all

e_{i} R

isomorphiccomplete set of matrix units

R ≅ M_{n} (e_{1} R e_{1})

Matrix units in $R$ — $e_{i j} e_{k l} = δ_{j k} e_{i l}$
- complete, finite
  - $R ≅ M_{n} (e_{11} R e_{11})$
  - $e_{11}$ is a full idempotent
  - ideals and radicals correspond
- incomplete, infinite
  - infinite orthogonal family ${e_{i i}}$
  - infinite direct sum of right ideals
  - $R$ is not Dedekind-finite
- arising from
  - a basis of a vector space over a division ring
  - isomorphic primitive idempotents
  - a non-central square root of $1$ in a quaternion algebra

The two branches are exact opposites: a complete family certifies the best possible structure, an infinite one certifies the failure of a finiteness property. Both are read off the same multiplication rule.

Applications and Industry Use

Applications here means where this structure is used — inside mathematics and in the engineering and computing disciplines that consume it.

Structure theory

Wedderburn–Artin

The classification of semisimple rings as products of matrix rings over division rings is finished by recognition: a simple artinian ring acts faithfully on a vector space over a division ring, and a basis supplies the matrix units.

Arithmetic

Splitting central simple algebras

A central simple algebra is split — isomorphic to a matrix ring — exactly when it contains a complete set of matrix units of the right size. Local-global principles in the Brauer group are statements about when such units exist over completions.

Coding and cryptography

Algebras over finite fields

By Wedderburn's little theorem every finite division ring is a field, so every finite simple ring is $M_{n} (𝔽_{q})$ . Recognition is what turns an abstract presentation of such an algebra into explicit matrices usable by an implementation.

Operator algebras

Matrix units and UHF algebras

Uniformly hyperfinite $C^{*}$ -algebras are built as limits of full matrix algebras along embeddings of matrix unit systems; the recognition principle is what makes such a limit description meaningful.

In each case the value is the same: matrix units convert an abstract algebra into coordinates, and coordinates are what algorithms and explicit constructions need.

Design Considerations

Design considerations here means the choices made when modelling a problem with these algebraic structures.

Store the units, not the isomorphism. If your data structure records ${e_{i j}}$ , the isomorphism $R ≅ M_{n} (S)$ and its inverse are both one-line formulas. Recording only an abstract isomorphism loses that.
Choose which unit is the base point. $S = e_{11} R e_{11}$ depends on the choice of $e_{11}$ only up to isomorphism, but explicit coordinates change; fix the base point once.
Watch the opposite ring. With modules on one side and endomorphisms written on the other, ${End}_{D} (V)$ is $M_{n} (D)$ for one convention and $M_{n} (D^{op})$ for the other. State the convention before claiming which.
Do not over-refine. Splitting past primitive idempotents is impossible; splitting into non-isomorphic primitives destroys the matrix structure. Group the isomorphic ones deliberately.
Finiteness assumptions. Recognition is a purely finite statement about $n^{2}$ elements; it needs no chain conditions. Do not import artinian hypotheses you do not use.

Standards and Notation

Standards here covers notation, symbol and markup standards, and reference implementations, rather than material or design codes.

Matrix units

e_{i j}

; also

E_{i j}

ε_{i j}

in linear-algebra sources

Defining rule

e_{i j} e_{k l} = δ_{j k} e_{i l}

\sum_{i} e_{i i} = 1

Matrix ring

M_{n} (S)

here;

S^{n \times n}

{Mat}_{n} (S)

M_{n} (S)

elsewhere

Fullness

R e R = R

; also *

e

is a generating idempotent*

Dedekind-finiteAlso von Neumann finite, directly finite, inverse symmetric

MarkupPresentation MathML per ISO/IEC 40314; symbols per ISO 80000-2

SoftwareGAP MatrixAlgebra, IsomorphismMatrixAlgebra; Magma MatrixAlgebra, IsIsomorphic

Computational Notes

Computational notes cover algorithms, cost and library behaviour rather than manufacturing process.

For a finite-dimensional algebra $A$ over a field $k$ with $\dim_{k} A = N$ :

Once matrix units are known, computing $ϕ$ costs $n^{2}$ multiplications per element, so translating a basis of $A$ into matrices over $S$ costs $O (N n^{2})$ algebra multiplications.
The hard step is finding the units. The standard route is to compute $rad A$ , split $A / rad A$ into simple factors, find a primitive idempotent in each, and test the resulting summands for isomorphism using ${Hom}_{A} (e A, f A) ≅ f A e$ ; each test is a linear-algebra computation.
Over a finite field, explicitly splitting a simple algebra $A ≅ M_{n} (𝔽_{q})$ is equivalent to finding a zero divisor of the right kind; randomised algorithms succeed in expected polynomial time, and this is what implementations use.
Over $ℚ$ the analogous problem for a central simple algebra is much harder: producing an explicit splitting is at least as hard as factoring, so implementations either work over local completions or accept certificates.

Failure Modes and Common Mistakes

Do not assume $M_{n} (S) ≅ M_{m} (T)$ forces $n = m$ and $S ≅ T$ ; it does for $S, T$ division rings and for many other classes, but not in general — this is the failure of cancellation studied under Morita theory.
Do not assume every ring with a nontrivial idempotent is a matrix ring; the idempotent must sit in a complete family of mutually isomorphic ones.
Do not read $(21.26)$ as producing a matrix ring: the family is infinite and incomplete, and its conclusion is negative — a finiteness condition fails.
Do not forget that $(21.27)$ hypothesises the condition on $R / rad R$ , not on $R$ ; the reduction through the radical is part of the statement.

Best Practices

Verify the two defining identities on a candidate family before invoking recognition; one missing relation invalidates the whole conclusion.
Compute a dimension count as a check: $\dim_{k} R = n^{2} \dim_{k} S$ must hold exactly.
When claiming an algebra is split, exhibit the matrix units rather than asserting an abstract isomorphism — the units are the certificate.
Record explicitly whether your matrix units come from a basis, from isomorphic idempotents, or from an ad hoc construction; the provenance determines what else you may assume.

Historical Notes and Lessons Learned

1907Wedderburn's structure theoremWedderburn classifies finite-dimensional simple algebras as matrix algebras over division algebras. The passage from an abstract algebra to matrices is exactly the recognition step.
1927Artin's chain-condition versionArtin extends the theorem to rings with the descending chain condition, where the same idempotent bookkeeping applies without finite dimensionality.
1930sBrauer groups and splittingSplitting fields of central simple algebras are studied systematically; being split becomes the statement that a complete set of matrix units exists over an extension field.
1950sJacobson's finiteness observationThe construction of infinitely many orthogonal idempotents from a one-sided inverse pins down which rings can fail to be Dedekind-finite, and gives the criterion recorded as $(21.27)$ .
1958Morita theoryMorita's equivalence theorem explains recognition conceptually: $R ≅ M_{n} (S)$ says $R$ has a progenerator that is a direct sum of $n$ copies of a single module.

The lesson is that a coordinate system can be an existence statement. Matrix units look like notation but are a structure theorem in disguise: their existence is equivalent to the ring being a matrix ring, so producing them is the same problem as proving the classification.

Quick Reference

Matrix units

e_{i j} e_{k l} = δ_{j k} e_{i l}

Complete

e_{11} + \dots + e_{n n} = 1

Recognition

R ≅ M_{n} (e_{11} R e_{11})

r \mapsto (e_{1 i} r e_{j 1})

Coefficient ring

S = e_{11} R e_{11} ≅

centraliser of the units

Corner in $M_{n} (k)$

e_{11} M_{n} (k) e_{11} ≅ k

;

e_{11}

is full

Radical

rad M_{n} (k) = M_{n} (rad k)

From idempotents

1 = \sum e_{i}

orthogonal with all

e_{i} R

isomorphic

\Rightarrow R ≅ M_{n} (e_{1} R e_{1})

Infinite units

a b = 1 \neq b a

gives

e_{i j} = b^{i} (1 - b a) a^{j}

Reference numbers used on this page
Reference	Statement
(21.14)	$e_{11} M_{n} (k) e_{11} ≅ k$ ; $e_{11}$ full; consistency with (21.10), (21.11)
(3.1)	Ideals of $M_{n} (k)$ are exactly $M_{n} (𝔞)$ for ideals $𝔞$ of $k$
(21.26)	Jacobson's matrix units $b^{i} (1 - b a) a^{j}$ in a non-Dedekind-finite ring
(21.27)	No infinite direct sum of right ideals in $R / rad R$ implies $R$ Dedekind-finite

Frequently Asked Questions

Does recognition need any finiteness hypothesis on

R

No. The theorem is a statement about $n^{2}$ elements satisfying two identities, and its proof is a finite computation. No chain conditions, no dimension assumptions and no fullness hypothesis appear. What does matter is completeness: the diagonal units must sum to the identity of $R$ .

How do I know the coefficient ring is

e_{11} R e_{11}

and not something else?

Because $ϕ$ is constructed with entries in $e_{11} R e_{11}$ and is bijective onto $M_{n}$ of that ring. Equivalently, $e_{11}$ is a full idempotent of $M_{n} (S)$ with corner $S$ , so the corner-ring theory identifies $S$ uniquely up to isomorphism. The centraliser description gives the same ring in a coordinate-free way.

Why is a complete orthogonal family of idempotents not enough?

Such a family gives the Peirce decomposition $R = ⨁_{i, j} e_{i} R e_{j}$ , which is a generalised matrix ring: the diagonal entries are the possibly different rings $e_{i} R e_{i}$ and the off-diagonal entries are bimodules. Turning it into an honest matrix ring requires all $e_{i} R$ to be isomorphic, which is what supplies the exchanging elements $e_{i j}$ .

What exactly does Jacobson's construction prove?

That failure of Dedekind-finiteness is never isolated: from a single pair with $a b = 1$ and $b a \neq 1$ one manufactures infinitely many nonzero orthogonal idempotents, hence an infinite direct sum of nonzero right ideals. Contrapositively, any ring whose semiprimitive quotient forbids such a direct sum — in particular any ring with $R / rad R$ right noetherian — is Dedekind-finite.

Is the quaternion example special to characteristic not

2

Yes, as presented: the idempotents $\frac{1}{2} (1 \pm i)$ require $2$ to be invertible, and the standard presentation of quaternion algebras by $i^{2} = a$ , $j^{2} = b$ , $j i = - i j$ is itself a characteristic-not- $2$ device. In characteristic $2$ quaternion algebras are described by different relations and split by a different criterion.

Does

M_{n} (S) ≅ M_{n} (T)

imply

S ≅ T

Not for arbitrary rings. It does when $S$ and $T$ are division rings, and more generally when they are basic, because then each is recovered as the endomorphism ring of a suitable indecomposable projective. In general the two rings are only Morita equivalent, and Morita equivalence does not imply isomorphism.

References

T. Y. Lam, A First Course in Noncommutative Rings, Graduate Texts in Mathematics 131, Springer-Verlag, 1991, §21, especially (21.14), (21.26) and (21.27), together with (3.1).
N. Jacobson, Structure of Rings, American Mathematical Society Colloquium Publications 37, revised edition, 1964, Chapters II–III.
R. S. Pierce, Associative Algebras, Graduate Texts in Mathematics 88, Springer-Verlag, 1982, Chapters 3 and 13.
I. N. Herstein, Noncommutative Rings, Carus Mathematical Monographs 15, Mathematical Association of America, 1968, Chapter 2.
F. W. Anderson and K. R. Fuller, Rings and Categories of Modules, 2nd edition, Graduate Texts in Mathematics 13, Springer-Verlag, 1992, §§17 and 22.

AI Suggested Questions

Prove that $M_{n} (S) ≅ M_{n} (T)$ implies $S ≅ T$ when $S$ and $T$ are basic algebras, and give a counterexample without that hypothesis.
Give an algorithm that, given structure constants for a simple algebra over a finite field, produces an explicit complete set of matrix units.
Show that a central simple algebra of degree $2$ over a field of characteristic not $2$ is split if and only if its norm form is isotropic.
What is the analogue of the recognition theorem for infinite sets of matrix units, and what replaces the completeness condition?
Prove that a right noetherian ring is Dedekind-finite directly, without passing through the radical.
How does the trace ideal $R e_{11} R$ detect fullness in a ring that is not a matrix ring?
Explain how systems of matrix units are used to build UHF and AF algebras as inductive limits.

Matrix Units and Full Idempotents

Executive Summary

Overview

Learning Objectives

Definitions

Core Concepts

The corner at a matrix unit

Recognition: the units determine the ring

Where matrix units come from

Key Results

Proof Techniques and Method

Insert the resolution of identity

Transport across the units

Amplify a defect

Worked Example

Splitting a quaternion algebra

Recognition inside Wedderburn's theorem

A non-example: an incomplete family

Process and Workflow

Comparison and Classification

Relationship Map

Applications and Industry Use

Wedderburn–Artin

Splitting central simple algebras

Algebras over finite fields

Matrix units and UHF algebras

Design Considerations

Standards and Notation

Computational Notes

Failure Modes and Common Mistakes

Best Practices

Historical Notes and Lessons Learned

Quick Reference

Frequently Asked Questions

References

AI Suggested Questions

Continue learning

01Executive Summary

02Overview

03Learning Objectives

04Definitions

05Core Concepts

The corner at a matrix unit

Recognition: the units determine the ring

Where matrix units come from

06Key Results

07Proof Techniques and Method

Insert the resolution of identity

Transport across the units

Amplify a defect

08Worked Example

Splitting a quaternion algebra

Recognition inside Wedderburn's theorem

A non-example: an incomplete family

09Process and Workflow

10Comparison and Classification

11Relationship Map

12Applications and Industry Use

Wedderburn–Artin

Splitting central simple algebras

Algebras over finite fields

Matrix units and UHF algebras

13Design Considerations

14Standards and Notation

15Computational Notes

16Failure Modes and Common Mistakes

17Best Practices

18Historical Notes and Lessons Learned

19Quick Reference

20Frequently Asked Questions

21Related KEVOS Topics

22References

23AI Suggested Questions

Continue learning

Executive Summary

Overview

Learning Objectives

Definitions

Core Concepts

Key Results

Proof Techniques and Method

Worked Example

Process and Workflow

Comparison and Classification

Relationship Map

Applications and Industry Use

Design Considerations

Standards and Notation

Computational Notes

Failure Modes and Common Mistakes

Best Practices

Historical Notes and Lessons Learned

Quick Reference

Frequently Asked Questions

Related KEVOS Topics

References

AI Suggested Questions