Cancellation of Modules over Semilocal Rings

Q: Why does part (1) require A to be projective when (20.11) requires no such thing?

Because the hypothesis has been moved. (20.11) assumes stable range one for End(A_R); (20.13)(1) assumes it only for R and must therefore reach A through a free module. Projectivity supplies the complement A' with A ⊕ A' ≅ R^n, and that is the only place it is used.

Q: Does invariant basis number really need a hypothesis?

Yes. End_k(V) for V of countably infinite dimension satisfies R ≅ R ⊕ R as right modules, so R^1 ≅ R^2. Leavitt algebras give examples with R^n ≅ R^m for prescribed n ≠ m. IBN is a theorem, not a convention.

Q: If stably free modules are free over semilocal rings, are all finitely generated projectives free?

No. Over M_2(k) the simple module is projective and not free, and over Z/12 the summand Z/4 is projective and not free. Only the *stably free* ones become free. Over a local ring, however, all finitely generated projectives are free, by Kaplansky's theorem.

Q: How does part (4) relate to Dedekind finiteness of R itself?

Part (4) with n = 1 is Dedekind finiteness of R. The content is the uniform statement over all n, that is, stable finiteness, which does not follow from Dedekind finiteness alone — Shepherdson gave a counterexample.

Executive Summary

Stable range one is a hypothesis on units; $(20.13)$ converts it into four statements about modules. Over a ring $R$ of left stable range one — for example any semilocal ring — a finitely generated projective module can be cancelled from any direct sum, free modules have well-defined rank, stably free modules are genuinely free, and one-sided inverses of matrices are two-sided.

All four follow from a single instance of the Cancellation Theorem: the case $A = R$ , where $End (R_{R}) ≅ R$ , so the hypothesis on the endomorphism ring is the hypothesis on $R$ itself. Everything else is bookkeeping with free summands.

4Consequences in

(20.13)

$A = R$ The only case actually needed

FreeWhat stably free becomes

$\neq$ K–SCancellation is not Krull–Schmidt

Overview

In K-theory one routinely proves statements only stably: after adding a free module. $K_{0} (R)$ remembers $P$ only up to the relation $P \oplus R^{m} ≅ Q \oplus R^{m}$ . Whether that stable information determines $P$ is exactly the cancellation question, and it is not automatic — over the coordinate ring of the real $2$ -sphere it fails.

P \oplus R^{m} ≅ Q \oplus R^{m} \overset{?}{⟹} P ≅ Q

(20.13a)

The stable-to-actual question. Left stable range one answers yes for finitely generated projective $P$ and $Q$ .

Lam's $(20.13)$ collects the affirmative answers available under that hypothesis. The list is worth memorising because each item is the standard hypothesis of some other theory: invariant basis number for anything involving matrix sizes, freeness of stably free modules for Serre-type problems, and stable finiteness for algorithms that invert matrices.

Learning Objectives

Reduce cancellation of a finitely generated projective to cancellation of $R$ itself.
Prove $(20.13) (1)$ using $End (R_{R}) ≅ R$ and the Cancellation Theorem.
Derive invariant basis number and identify why $R = 0$ must be excluded.
Prove that stably free implies free, handling both $s \geq r$ and $s < r$ .
Prove that $M_{n} (R)$ is Dedekind-finite by cancelling $R^{n}$ from $R^{n} \oplus \ker α$ .
Give an example where decomposition into indecomposables is not unique despite cancellation.

Definitions

Stably free: P is stably free if P direct sum R^m is free of finite rank for some m. Every stably free module is projective.
Invariant basis number (IBN): For R nonzero, R^n isomorphic to R^m implies n = m.
Stably finite: M_n(R) is Dedekind-finite for all n. Equivalent to: every epimorphism of right modules from R^n onto R^n is an isomorphism.
$V (R)$: The commutative monoid of isomorphism classes of finitely generated projective right R-modules under direct sum; K_0(R) is its group completion.

Throughout, modules are right $R$ -modules and $R$ has left stable range one; by Bass' Theorem this holds whenever $R$ is semilocal.

Core Concepts

The reduction to $A = R$

The Cancellation Theorem $(20.11)$ asks that $End (A_{R})$ have left stable range one, which is a statement about $A$ . When $A = R$ the endomorphism ring is $R$ itself, via $ϕ \mapsto ϕ (1)$ , so the hypothesis is exactly " $R$ has left stable range one". Every part of $(20.13)$ is obtained by inflating a projective module to a free one and cancelling copies of $R$ one at a time.

R

has left stable range

1

R

cancels

R^{n}

cancelsf.g. projectives cancel

Why projectivity of $A$ is needed here

$(20.13) (1)$ places its hypothesis on $R$ , not on $End (A_{R})$ , and pays for that by requiring $A$ to be finitely generated projective — precisely so that $A$ has a complement $A'$ with $A \oplus A' ≅ R^{n}$ . Without projectivity there is no way to move from $A$ to a free module, and one must return to $(20.11)$ and verify the hypothesis on $End (A_{R})$ directly.

Route 1

Hypothesis on the module

$(20.11)$ : $End (A_{R})$ has stable range one. Works for arbitrary $A$ — no projectivity, no finite generation. Cost: the hypothesis is often hard to check.

Route 2

Hypothesis on the ring

$(20.13) (1)$ : $R$ has stable range one and $A$ is finitely generated projective. Cost: $A$ is restricted. Benefit: the hypothesis is a property of $R$ alone and is inherited by all such $A$ at once.

The monoid picture

Cancellation says the monoid $V (R)$ of finitely generated projectives is cancellative. A cancellative commutative monoid embeds in its group completion, so

P ≅ Q ⟺ [P] = [Q] in K_{0} (R),

(20.13b)

Valid for finitely generated projective $P, Q$ over a ring of left stable range one. Without cancellation, $K_{0}$ forgets genuine distinctions.

Key Results

Theorem(20.13)Consequences of stable range one

Let $R$ be a ring with left stable range one — for instance, any semilocal ring. Then:

If $A, B, C$ are right $R$ -modules with $A$ finitely generated projective, then $A \oplus B ≅ A \oplus C$ implies $B ≅ C$ .
$R$ has invariant basis number: for $R \neq 0$ and natural numbers $n, m$ , $R^{n} ≅ R^{m}$ as right $R$ -modules implies $n = m$ .
Every finitely generated stably free right $R$ -module is free.
$M_{n} (R)$ is Dedekind-finite for every $n \geq 1$ ; that is, $R$ is stably finite.

Proofof (1)

Since $A$ is finitely generated projective, choose $A'$ with $A \oplus A' ≅ R^{n}$ . Adding $A'$ to both sides of $A \oplus B ≅ A \oplus C$ gives

R^{n} \oplus B ≅ A \oplus A' \oplus B ≅ A \oplus A' \oplus C ≅ R^{n} \oplus C .

It therefore suffices to cancel one copy of $R$ at a time. For $A = R$ the endomorphism ring is $End (R_{R}) ≅ R$ , which has left stable range one by hypothesis, so the Cancellation Theorem $(20.11)$ applies and gives $R \oplus X ≅ R \oplus Y \Rightarrow X ≅ Y$ . Applying this $n$ times to the display yields $B ≅ C$ . $□$

Proofof (2)

Suppose $R^{n} ≅ R^{m}$ with $n \geq m$ . Cancelling $R^{m}$ by part (1) gives $R^{n - m} ≅ 0$ . If $n > m$ then $R^{n - m}$ has $R$ as a direct summand, so $R = 0$ . Hence for $R \neq 0$ we must have $n = m$ . $□$

Proofof (3)

Let $P$ be finitely generated with $P \oplus R^{r} ≅ R^{s}$ .

If $s \leq r$ , cancel $R^{s}$ from both sides using part (1): $P \oplus R^{r - s} ≅ 0$ , so $P = 0$ , which is free of rank $0$ (and if $r > s$ this also forces $R = 0$ , where everything is free).

If $s > r$ , cancel $R^{r}$ instead: $P ≅ R^{s - r}$ , which is free. $□$

Proofof (4)

Let $α, β \in M_{n} (R)$ with $α β = I_{n}$ . Viewing $R^{n}$ as a right $R$ -module of column vectors, left multiplication by a matrix is a right $R$ -linear endomorphism, so $α$ and $β$ define maps $R^{n} \to R^{n}$ .

From $α β = I_{n}$ , the map $α$ is surjective and is split by $β$ , which is therefore injective. Hence

R^{n} ≅ im β \oplus \ker α ≅ R^{n} \oplus \ker α .

Cancelling $R^{n}$ by part (1) gives $\ker α = 0$ . So $α$ is injective as well as surjective, hence an automorphism of $R^{n}$ , hence an invertible matrix. Multiplying $α β = I_{n}$ on the left by $α^{- 1}$ and on the right by $α$ gives $β α = I_{n}$ . $□$

CorollarySemilocal rings

Every semilocal ring satisfies (1)–(4), by Bass' Theorem $(20.9)$ . In particular every finite-dimensional algebra over a field, every finite ring, every left or right artinian ring and every local ring has invariant basis number, is stably finite, and has all stably free modules free.

RemarkEquivalent forms of stable finiteness

For a fixed $n$ , the following are equivalent: $M_{n} (R)$ is Dedekind-finite; $R^{n} ≅ R^{n} \oplus M$ forces $M = 0$ ; every right module epimorphism $R^{n} \to R^{n}$ is an isomorphism. A ring satisfying these for the given $n$ is called *weakly $n$ -finite*; satisfying them for all $n$ is stable finiteness.

RemarkCancellation does not give uniqueness of decompositions

None of (1)–(4) implies the Krull–Schmidt property. One may have $A \oplus B ≅ C \oplus D$ with all four modules finitely generated and indecomposable, $A not ≅ C$ , $A not ≅ D$ , $B not ≅ C$ , $B not ≅ D$ . There is no common summand to cancel, so the theorems above simply do not engage. Swan's example, over a commutative noetherian local domain, is exactly of this shape.

Proof Techniques and Method

How these proofs work, and which move to reuse.

1. Inflate to freeA finitely generated projective

A

has a complement

A'

with

A \oplus A' ≅ R^{n}

. Adding

A'

to both sides of an isomorphism converts a statement about

A

into one about

R^{n}

2. Cancel one $R$ at a time

End (R_{R}) ≅ R

, so

(20.11)

applies with the hypothesis on

R

. Induct on

n

3. Push degenerate cases to $R = 0$ Statements like

R^{n - m} ≅ 0

are not contradictions but conclusions: they force the ring to vanish. Handle the zero ring explicitly rather than assuming it away.

4. Convert matrices into module mapsA one-sided inverse in

M_{n} (R)

is a splitting of

R^{n} \to R^{n}

. Cancellation of

R^{n}

then kills the kernel and upgrades the inverse to a two-sided one.

Worked Example

Stably free equals free over $ℤ / 12$

$R = ℤ / 12 ≅ ℤ / 4 \times ℤ / 3$ is semilocal, so $(20.13)$ applies. Set $e = (1, 0)$ , the idempotent corresponding to the first factor, and $P = e R ≅ ℤ / 4$ . Then $P$ is projective and not free — it has $4$ elements while $R^{n}$ has $12^{n}$ — but it is not stably free either, consistently with (3): $P \oplus R^{m}$ has $4 \cdot 12^{m}$ elements, while $R^{n}$ has $12^{n}$ , and $4 \cdot 12^{m} = 12^{n}$ has no solution in non-negative integers.

A counting check of (2) in the same ring: $| R^{n} | = 12^{n}$ determines $n$ , so IBN is visible directly. Over a general semilocal ring no such counting is available and $(20.13) (2)$ is doing real work.

Cancellation without Krull–Schmidt: $ℤ [\sqrt{- 5}]$

Let $S = ℤ [\sqrt{- 5}]$ , a Dedekind domain of class number $2$ , and let $I = (2, 1 + \sqrt{- 5})$ .

$I$ is not principal: an element of norm $2$ would need $a^{2} + 5 b^{2} = 2$ , which has no integer solutions.
$I^{2} = (2)$ : expanding, $I^{2} = (4, 2 + 2 \sqrt{- 5}, - 4 + 2 \sqrt{- 5}) = (2) \cdot (2, 1 + \sqrt{- 5}, - 2 + \sqrt{- 5})$ , and the second factor contains $(1 + \sqrt{- 5}) - (- 2 + \sqrt{- 5}) = 3$ as well as $2$ , hence is all of $S$ .
Steinitz' theorem for Dedekind domains gives $I \oplus J ≅ S \oplus I J$ for fractional ideals, so $I \oplus I ≅ S \oplus I^{2} ≅ S \oplus S$ .

I \oplus I ≅ S \oplus S, I not ≅ S .

(20.13c)

Two decompositions into indecomposables with no summand in common. Rank-one torsion-free modules over a Dedekind domain are indecomposable.

Cancellation is not violated: there is no common summand to cancel, and indeed finitely generated projectives over a Dedekind domain do cancel, by Steinitz' classification. What fails is uniqueness of the decomposition. This is the precise sense in which $(20.13)$ is weaker than a Krull–Schmidt theorem.

A ring where (4) fails

If $R = {End}_{k} (V)$ with $\dim_{k} V$ countably infinite then $V ≅ V \oplus V$ , so $R ≅ M_{2} (R)$ and $R_{R} ≅ R_{R} \oplus R_{R}$ . Here IBN fails, $R$ is not Dedekind-finite, and no part of $(20.13)$ survives — as it must be, since $R$ does not have stable range one.

Comparison and Classification

Cancellation results and the hypotheses they need
Result	Hypothesis on $R$	Hypothesis on $A$	Hypothesis on $B, C$
$(20.11)$	none	$End (A_{R})$ has left stable range one	none
$(20.12)$	module-finite over commutative noetherian semilocal $k$	finitely generated	none
$(20.13) (1)$	left stable range one	finitely generated projective	none
Finite length	none	$A$ has a composition series	none
Steinitz (Dedekind domains)	Dedekind domain	finitely generated projective	finitely generated

The trade is always the same: weaken the hypothesis on the ring and you must strengthen it on the module, and conversely. Only $(20.11)$ and $(20.13) (1)$ allow $B$ and $C$ to be completely arbitrary.

What holds where
Ring	IBN	Stably free = free	Stably finite
Semilocal	yes	yes	yes
Left stable range one	yes	yes	yes
Commutative nonzero	yes	no	yes
$ℤ [x_{1}, \dots, x_{d}]$	yes	yes (Quillen–Suslin)	yes
$ℝ [x, y, z] / (x^{2} + y^{2} + z^{2} - 1)$	yes	no	yes
${End}_{k} (V)$ , $\dim_{k} V$ infinite	no	no	no

Commutative rings always have IBN and are stably finite, but stably free modules over them need not be free — the sphere example is the standard witness. So $(20.13) (3)$ is strictly stronger than commutativity.

Relationship Map

Left stable range one — the hypothesis of $(20.13)$
- Cancellation of f.g. projectives — $(20.13) (1)$
  - $V (R)$ is a cancellative monoid
  - $[P] = [Q]$ in $K_{0} (R)$ iff $P ≅ Q$
  - Stably free modules are free $(20.13) (3)$
- Matrix consequences — $(20.13) (2)$ and $(20.13) (4)$
  - Invariant basis number
  - $M_{n} (R)$ Dedekind-finite for all $n$
  - Every epimorphism $R^{n} \to R^{n}$ is an isomorphism

Upstream, the hypothesis is supplied by Bass' Theorem for semilocal rings. Downstream it does not reach Krull–Schmidt, which needs indecomposables with local endomorphism rings — a genuinely different condition treated on the Krull–Schmidt page.

semilocalstable range

1

cancel f.g. projectivesstably free = freeIBN

Applications and Industry Use

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

Algebraic K-theory. $(20.13) (1)$ is what makes $K_{0}$ of a semilocal ring a faithful invariant of finitely generated projectives rather than merely a stable one.
Serre's problem and its relatives. "Stably free implies free" is the conclusion Quillen and Suslin obtained for polynomial rings over a field by other means; over semilocal rings it is elementary, which is why local-global arguments reduce hard cases to semilocal ones.
Integral and modular representation theory. Over a $p$ -adic order, cancellation of lattices holds by $(20.12)$ and $(20.13)$ , so genus theory and local-global counting of lattices are well posed.
Computer algebra and matrix algorithms. Stable finiteness guarantees that a one-sided matrix inverse over the coefficient ring is a genuine inverse, so routines that verify $α β = I$ need not separately check $β α = I$ . Over rings without the property this shortcut is unsound.
Systems and coding over rings. Free modules of well-defined rank are what make "a code of length $n$ " meaningful. IBN failure would make the length of a free code ambiguous; $(20.13) (2)$ rules that out for the finite chain rings used in practice.

Computational Notes

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

Deciding whether a given finitely generated projective module over a semilocal ring is free reduces, by $(20.13) (3)$ , to deciding whether it is stably free — a question about a class in $K_{0}$ rather than about explicit bases.
Over a finite-dimensional algebra, projective modules are computed as $e A$ for idempotents $e$ produced by lifting a Wedderburn decomposition of $A / rad A$ ; cancellation makes the resulting list of indecomposable projectives well defined up to isomorphism.
Matrix inversion over a semilocal coefficient ring may be verified one-sidedly, by $(20.13) (4)$ . Computer algebra systems that check both sides do so for generality, not necessity.
There is no algorithm to decide stable range one for an arbitrary finitely presented ring; in practice the property is certified structurally, by exhibiting semilocality or unit-regularity.

Failure Modes and Common Mistakes

Best Practices

State which of $(20.11)$ , $(20.12)$ and $(20.13) (1)$ you are using; they have different hypotheses and different reach.
When the ring is semilocal, quote Bass' Theorem once and then work with stable range one — it is the property the proofs actually consume.
Check finitely generated projective, not just finitely generated, before applying $(20.13) (1)$ .
If uniqueness of a decomposition is what you need, look for local endomorphism rings and a Krull–Schmidt theorem, not for cancellation.
For a concrete ring, test IBN and stable finiteness by counting or by exhibiting semilocality before assuming them.

Quick Reference

Hypothesis

R

has left stable range one; guaranteed for semilocal

R

(20.9)

$(20.13) (1)$

A

f.g. projective,

B, C

arbitrary:

A \oplus B ≅ A \oplus C \Rightarrow B ≅ C

$(20.13) (2)$ IBN:

R^{n} ≅ R^{m} \Rightarrow n = m

, for

R \neq 0

$(20.13) (3)$

P \oplus R^{r} ≅ R^{s} \Rightarrow P ≅ R^{s - r}

when

s \geq r

; otherwise

P = 0

$(20.13) (4)$

M_{n} (R)

is Dedekind-finite for all

n

α β = I_{n} \Rightarrow β α = I_{n}

Not impliedKrull–Schmidt uniqueness of decompositions.

Proof of each part in one line
Part	Argument
(1)	Add $A'$ with $A \oplus A' ≅ R^{n}$ ; cancel $R$ $n$ times using $End (R_{R}) ≅ R$ in $(20.11)$ .
(2)	Cancel $R^{m}$ from $R^{n} ≅ R^{m}$ ; a nonzero free module of positive rank forces $R \neq 0$ .
(3)	Cancel $R^{\min (r, s)}$ from $P \oplus R^{r} ≅ R^{s}$ .
(4)	$α β = I_{n}$ splits $R^{n} ↠ R^{n}$ ; cancel $R^{n}$ to kill $\ker α$ .

Frequently Asked Questions

Why does part (1) require

A

to be projective when

(20.11)

requires no such thing?

Because the hypothesis has been moved. $(20.11)$ assumes stable range one for $End (A_{R})$ ; $(20.13) (1)$ assumes it only for $R$ and must therefore reach $A$ through a free module. Projectivity supplies the complement $A'$ with $A \oplus A' ≅ R^{n}$ , and that is the only place it is used.

Does invariant basis number really need a hypothesis?

Yes. ${End}_{k} (V)$ for $V$ of countably infinite dimension satisfies $R ≅ R \oplus R$ as right modules, so $R^{1} ≅ R^{2}$ . Leavitt algebras give examples with $R^{n} ≅ R^{m}$ for prescribed $n \neq m$ . IBN is a theorem, not a convention.

If stably free modules are free over semilocal rings, are all finitely generated projectives free?

No. Over $M_{2} (k)$ the simple module is projective and not free, and over $ℤ / 12$ the summand $ℤ / 4$ is projective and not free. Only the stably free ones become free. Over a local ring, however, all finitely generated projectives are free, by Kaplansky's theorem.

How does part (4) relate to Dedekind finiteness of

R

itself?

Part (4) with $n = 1$ is Dedekind finiteness of $R$ . The content is the uniform statement over all $n$ , that is, stable finiteness, which does not follow from Dedekind finiteness alone — Shepherdson gave a counterexample.

Why does Krull–Schmidt not follow?

Cancellation compares two decompositions that already share a summand. Krull–Schmidt asserts that any two decompositions into indecomposables match up, and needs each indecomposable to have a local endomorphism ring. The ideal example over $ℤ [\sqrt{- 5}]$ has indecomposables with non-local endomorphism rings and fails uniqueness while cancellation still holds.

Is the converse of

(20.13)

true — does cancellation imply stable range one?

No. Cancellation of finitely generated projectives holds over every Dedekind domain by Steinitz' theorem, yet $ℤ$ has stable range two. The implications in $(20.13)$ are one-directional.

References

T. Y. Lam, A First Course in Noncommutative Rings, Graduate Texts in Mathematics 131, Springer-Verlag, 1991, §20, (20.13) and Exercises 20.2–20.8, pp. 316–317.
H. Bass, “K-theory and stable algebra”, Publications Mathématiques de l'IHÉS 22 (1964), 5–60.
R. G. Swan, Algebraic K-Theory, Lecture Notes in Mathematics 76, Springer-Verlag, 1968.
P. M. Cohn, Free Rings and Their Relations, 2nd edition, Academic Press, 1985, for invariant basis number and weakly finite rings.
T. Y. Lam, Serre's Problem on Projective Modules, Springer Monographs in Mathematics, Springer-Verlag, 2006.
A. Facchini, Module Theory: Endomorphism Rings and Direct Sum Decompositions in Some Classes of Modules, Progress in Mathematics 167, Birkhäuser, 1998.

AI Suggested Questions

Prove Steinitz' theorem that $I \oplus J ≅ R \oplus I J$ for fractional ideals over a Dedekind domain.
Construct Swan's example of finitely generated indecomposable modules over a commutative noetherian local domain with non-unique direct-sum decomposition.
Show that the Leavitt algebra $L (m, n)$ fails invariant basis number with $R^{m} ≅ R^{n}$ .
Give the details of Kaplansky's theorem that projective modules over a local ring are free.
Explain why the tangent module of the real $2$ -sphere is stably free but not free, and identify the stable range obstruction.
For which rings is $V (R)$ cancellative but $R$ of stable range greater than one?
How do cancellation and Krull–Schmidt interact for modules of finite length over an arbitrary ring?

01Executive Summary

02Overview

03Learning Objectives

04Definitions

05Core Concepts

The reduction to A=R

Why projectivity of A is needed here

Hypothesis on the module

Hypothesis on the ring

The monoid picture

06Key Results

07Proof Techniques and Method

08Worked Example

Stably free equals free over ℤ/12

Cancellation without Krull–Schmidt: ℤ[−5]

A ring where (4) fails

09Comparison and Classification

10Relationship Map

11Applications and Industry Use

12Computational Notes

13Failure Modes and Common Mistakes

14Best Practices

15Quick Reference

16Frequently Asked Questions

17Related KEVOS Topics

18References

19AI Suggested Questions

Continue learning