REVERSIBLE RINGS
P. M. COHN
1. Introduction
Commutative rings form a very special subclass of rings, which shows quite
different behaviour from the general case. For example, in a (non-trivial) commutative
ring, the absence of zero-divisors is sufficient as well as necessary for the existence of
a field of fractions, whereas for general rings, another infinite set of conditions is
needed to characterize subrings of skew fields. This suggests the study of a class of
rings which includes all commutative rings as well as all integral domains : reversible
rings, where a ring is called reersible if ab l 0 implies ba l 0. It turns out that this
condition helps to simplify other ring conditions, as we shall see in Section 2, although
most of these results are at a somewhat superficial level. We therefore introduce a
more technical notion, full reversibility, in Section 3, and show that this is the precise
condition for the least matrix ideal to be proper and consist entirely of non-full
matrices. Further, we show in Section 4 that a fully reversible ring is embeddable in
a skew field if and only if it is an integral domain.
In what follows, all rings are associative, with a unit element 1 which is preserved
by ring homomorphisms, inherited by subrings and acts unitally on modules.
I am grateful to V. de O. Ferreira for his comments, in particular the suggestion
of using the notion of unit-stable rings.
2. Reersibility
As stated in the Introduction, a ring R is called reersible if ab l 0 implies ba l 0.
It is clear that such a ring is prime if and only if it is an integral domain, for an
integral domain is clearly a reversible prime ring and, conversely, if R is prime and
reversible, suppose that ab l 0, then abx l 0 for all x ? R, hence bxa l 0 by
reversibility, and now primeness shows that a l 0 or b l 0. Similarly, R is reduced
(that is, x# l 0 implies x l 0) if and only if it is reversible and semiprime. For suppose
that R is reduced : if ab l 0, then (ba)# l baba l 0, hence ba l 0, which shows R to
be reversible, and it is clearly also semiprime. Conversely, if R is reversible and semi-
prime, then the same argument as before shows that R is reduced. Thus we obtain
the following.
T 2.1. Let R be any ring. Then :
(i) R is an integral domain if and only if it is prime and reersible ;
(ii) R is reduced if and only if it is semiprime and reersible.
Received 2 September 1998.
1991 Mathematics Subject Classification 16U80.
Bull. London Math. Soc. 31 (1999) 641–648
14692120, 1999, 6, Downloaded from [Link] by Egyptian National Sti. Network (Enstinet), Wiley Online Library on [28/01/2026]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
642 . .
The following conjecture, raised by G. Ko$ the in the 1930s, is still open (compare
[4, p. 417]).
K$ ’ . If a ring has a non-zero nil right ideal, then it has a non-
zero nil ideal.
For reversible rings, it is easy to obtain an affirmative solution.
T 2.2. In any reersible ring, eery nil right ideal generates a two-sided nil
ideal, and the set of all nilpotent elements is a nil ideal.
Proof. Let R be a reversible ring. If x ? R is such that xr l 0, then for any c ? R
and any s r, xscxr−s l 0, hence xscxr−sd l 0 for any d. By induction, it follows that
if xr l 0, then xr"c xr# … xrn cn l 0, for any ci ? R and any ri such that r j…jrn l r.
" "
Hence if ; is a nil right ideal, then for any a ? ;, an l 0 for some n, thus for any x,
(xa)n l 0 and so R; is a nil ideal. Now let H be the set of all nilpotent elements of R, and
suppose that x, y ? H, say xr l ys l 0. Consider (xjy)r+s−" : this is a sum of products
of x and y, each with rjsk1 factors, so each term has either at least r factors x or
at least s factors y, and so vanishes, hence xjy ? H. Further, if xr l 0, then (bxc)r has
r factors x and so is 0, and this shows H to be an ideal, a nil ideal by definition.
We recall that a nilradical in any ring R is a nil ideal . such that the quotient R\.
is semiprime [4, p. 415]). In general, a ring may have more than one nilradical, but
from Theorems 2.2 and 2.1 we see that in a reversible ring, the set of all nilpotent
elements forms the unique nilradical.
3. Full reersibility
We recall from [2] that a matrix A over a ring R is said to be full if it is square,
say nin, and cannot be written as a product of an nir by an rin matrix, for r n.
The matrix A is called stably full if A & I is full for unit matrices of any order. We
recall, further, that a ring has unbounded generating number (UGN) if the unit matrix
of any order is full. A ring will be called n-reersible if it is non-zero and for any two
square matrices A, B of the same order, at most n, AB is full whenever BA is full. Thus
a reversible ring is a ring that is 1-reversible ; if R is n-reversible for all n, then it is
called fully reersible. Clearly, a commutative ring is always 1-reversible, but it need
not be fully reversible, as the example in Section 4 of a commutative ring which is not
3-reversible shows. Further, a ring will be called unit-stable if it is non-zero and for
any square matrices A, B of the same order and any invertible matrix U, AUB is full
whenever AB is full. It is clear that a fully reversible ring is always unit-stable ; below,
we shall find that under quite mild conditions the converse also holds (Theorem 3.1).
To show the relevance of these notions for localization, we briefly recall the
concept of matrix ideal, referring to [2, Chapter 7] or [5, Chapter 4] for details. In
order to localize, that is, introduce fractions in a general ring, we need to invert
matrices rather than merely elements, and the matrices that become singular form
what is called a matrix ideal. To give a precise definition, we need to introduce notions
of addition and multiplication. The latter is straightforward : for any square matrices
A, B, we form the diagonal sum A & B l 0 A 0
0 B 1. For addition, we need the
14692120, 1999, 6, Downloaded from [Link] by Egyptian National Sti. Network (Enstinet), Wiley Online Library on [28/01/2026]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
643
determinantal sum : for two square matrices A l (A Ah), B l (B Ah) which differ
" "
only in the first column, the determinantal sum is defined as A ] B l (A jB Ah), and
" "
similarly for other columns or for rows. Now a matrix ideal in a ring R is defined
as a set of square matrices, including all non-full matrices, which admits
determinantal sums (where defined) and diagonal sums with any square matrix, and
such that A & I ? implies A ? . A matrix ideal is said to be prime if its complement
is non-empty and closed under diagonal sums. Given any homomorphism R K of
R into a skew field K, the set of all square matrices mapping to singular (that is, non-
invertible) matrices over K forms a prime matrix ideal , and if K is an epic R-field
(that is, generated by the image of R), then it can be reconstructed from R and :
we adjoin to R the entries of all formal inverses of the square matrices in the
complement of . This is a local ring whose residue-class field is isomorphic to K.
Moreover, all epic R-fields are obtained in this way from R, using prime matrix ideals.
As in the commutative case, one proves that given any matrix ideal and any set Σ
of square matrices closed under diagonal sums and disjoint from , there exists a
matrix ideal containing , disjoint from Σ and maximal subject to these con-
ditions ; moreover, is prime.
In any ring R we can form the least matrix ideal , defined as the matrix ideal
generated by all non-full matrices. If is proper (in the set of all square matrices),
then it is disjoint from the set of all unit matrices, hence there exist prime matrix
ideals, and so R has homomorphisms to a skew field. Clearly, this condition (that
be proper) is necessary as well as sufficient, and in the same way one can give
conditions for R to be embeddable in a skew field (compare [2, 7.5] and [5, 4.4]).
If the set of all non-full square matrices over R forms a prime matrix ideal, then
R has a skew field of fractions over which every full matrix over R becomes invertible.
Such a ring is called a Sylester domain (compare [2, 5.5]). Similarly, we can ask under
what conditions the set of all non-full square matrices forms a proper matrix ideal,
not necessarily prime. Such a ring need not be embeddable in a skew field, but the con-
ditions for embeddability take on a much simpler form, as we shall see in Section 4.
We begin with a characterization of these rings.
T 3.1. Let R be a non-zero ring. Then the following conditions are
equialent.
(a) R is fully reersible.
(b) R is unit-stable, and eery full matrix is stably full.
(c) The least matrix ideal of R is proper and consists entirely of non-full
matrices.
For the proof of this theorem, we shall need a form of the magic lemma (compare
[2, Lemma 7.9.1] and [5, Lemma 9.3.3]) and a lemma on stably full matrices.
L 3.2. If R is a ring which is reersible or unit-stable, then the determinantal
sum of any two non-full matrices oer R, when defined, is non-full.
Proof. We have C l A ]B, where A is non-full, say A l PQ, where P is
ni(nk1) and Q is (nk1)in. We write Q l (Q Qh), where Q is a column and Qh
" "
is square of order nk1 ; then the first column of A is PQ . Denoting the first column
"
of B by B , we have
"
B l (B PQh) l (B P) (1 & Qh). (1)
" "
14692120, 1999, 6, Downloaded from [Link] by Egyptian National Sti. Network (Enstinet), Wiley Online Library on [28/01/2026]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
644 . .
It follows that
C l (A jB PQh) l (PQ jB P) (1 & Qh)
" " " "
l (B P)
" 0
1 0 1 0
Q 1 0 Qh
"
10 . 1
Briefly, B l ST and C l SUT, where S, T, U are square matrices and U is invertible.
Suppose that R is unit-stable and B l ST is full ; then C l SUT is also full, by
hypothesis. Conversely, when C is full, then so is B, and this shows the determinantal
sum of non-full matrices to be non-full. If R is reversible and ST is full, then so is TS,
hence also TSU and SUT, and again this argument can be reversed.
We note that for the above proof it is enough to assume that R is unit-stable for
unitriangular matrices.
L 3.3. Oer a fully reersible ring, eery full matrix is stably full.
Proof. Let R be a fully reversible ring, and suppose that A is a full matrix of
order n over R such that A & 1 is non-full. Then there exist square matrices P, Q of
order n, a row p l ( p , p , … , pn) and a column q l (q , q , … , qn)T such that
" # " #
0 1 0 1
A 0
0 1
l
P
p
(Q q).
We have Pq l 0, so PQh is not full, where Qh is the matrix with first column q and the
rest of the first row zero, while the unit matrix of order nk1 is in the bottom right-
hand corner, for then PQh has its first column zero and so is not full. By elemen-
tary column transformations, we can reduce the first column of Qh to the form
(q , 0, … , 0)T and so find that P(q & I ) is not full, and therefore neither is (q & I )P ; it
" " "
follows that (p q & I )P is not full. We claim that the same is true of ( pi qi & I )P ; for if
" "
T is the matrix obtained from the unit matrix of order n by interchanging the 1st and ith
rows, then PT:Tq l 0. Here Tq is q with the first and ith components interchanged.
The previous argument now shows that ( pi qi & I )PT is not full, so the same is true
of ( pi qi & I )P. If we now form the determinantal sum (with respect to the first row)
of these n matrices, bearing in mind that pi qi l 1, we find by Lemma 3.2 that P is
not full, but this contradicts the fact that A l PQ is full, and it proves the assertion.
Proof of Theorem 3.1. Assume that R is fully reversible ; if AB is full, then so is
BA, hence also BAU, for any invertible matrix U, and therefore AUB is full, so R is
unit-stable. Moreover, every full matrix is stably full, by Lemma 3.3. If R is unit-
stable and full matrices are stably full, then by Lemma 3.2, the set of all non-full
matrices is closed under determinantal sums ; moreover, I @ , and if A & I ? , then
A ? , because full matrices are stably full. It follows that is a proper matrix ideal.
Finally, assume that the set of all non-full matrices is a proper matrix ideal, and
assume that AB ? . Then we have the following series of transformations within :
AB ,- 0AB0 0I1 ,- 0ABA 0I1 ,- 0A0 kB
I
,-1 0
0
A
kB
0 1 0
,-
0 A
kB 0
. 1
Here we have formed determinantal sums with non-full matrices at steps 2 and 4, and
a permutation of rows and of columns at step 5. By reversing our steps, we now find
that BA ? , so R is fully reversible.
14692120, 1999, 6, Downloaded from [Link] by Egyptian National Sti. Network (Enstinet), Wiley Online Library on [28/01/2026]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
645
For a Sylvester domain, the non-full matrices form a prime matrix ideal, so every
Sylvester domain is fully reversible. These rings are included in the class of all Hermite
rings, where a ring is called Hermite if every matrix with a right inverse can be
completed to a square invertible matrix. For such rings, the condition of full
reversibility can still be weakened.
T 3.4. Any unit-stable Hermite ring is fully reersible.
Proof. By Theorem 3.1, it is enough to show that over a Hermite ring, any full
matrix is stably full, and this follows from Proposition 5.6.2 of [2]. For a direct proof,
we take an nin matrix A and suppose that A is full but A & 1 is not full ; thus we
have
0 1 0 1
A 0
0 1
l
P
p
(Q q),
where P, Q are square, p is a row and q is a column. Since pq l 1 and R is Hermite,
p, q form the first row and column, respectively, of matrices Ph, Qh that are mutually
inverse, so that
I
A 0 M K
0 1 0 l
J
M 0 I L
P
Ph 01
(Q Qh).
On the right, multiply the first factor by Qh on the right and the second factor by Ph
on the left. Bearing in mind that PhQh l QhPh l I, we obtain
IA 0 M K PQh I0 Pd K 0
l0
I 1 0 1
1 0
0 1 0 (PhQ I) l 1 0 ,
JM 0 I L J0 I L Qd 0 I
where Pd is ni(nk1), Qd is (nk1)in, and the zeros in the top left-hand corners
of the factors on the right arise from the form of the left-hand side. We now have
A l PdQd, showing that A is not full.
4. Properties of localization
Let R be any ring. For a matrix ideal in R, we define its radical N as the set
of all matrices A such that a diagonal sum of the form A & … & A with sufficiently
many terms lies in . It can be shown that N is just the intersection of all prime
matrix ideals containing (Theorem 7.4.4 of [2]). Thus, in any ring R we have
N l λ,
where λ runs over all the prime matrix ideals containing the least matrix ideal .
Write Kλ l R\λ for the epic R-field defined by λ, consider the natural
homomorphism } : R Kκ, and denote its kernel by ;. We claim that
; l N ER l (N ), (2)
"
where the suffix 1 indicates 1i1 matrices. If x} l 0, then x lies in each λ, so x ? N .
Conversely, if x ? N , then x maps to 0 in all the Kλ, so x ? ; ; this proves (2). In any
14692120, 1999, 6, Downloaded from [Link] by Egyptian National Sti. Network (Enstinet), Wiley Online Library on [28/01/2026]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
646 . .
reversible ring, the set of all nilpotent elements is the unique nilradical H, as we saw
in Section 2, and it is clear that H 7 ; ; here, equality holds when R is fully reversible,
for if x ? ;, then x} l 0, so xI ? , which means that xI is non-full, hence so is xn,
that is, x is nilpotent. Consider the quotient R* l R\H, for any fully reversible ring R.
This is still fully reversible, and it is reduced.
Any direct product of skew fields is fully reversible. For if R l Kλ with
projections pλ, then a square matrix A over R is non-full if and only if Apλ is singular
for all λ, and a product AB over any skew field is singular if and only if BA is singular.
Let R be any fully reversible reduced ring. If λ is the family of all prime matrix
ideals of R and Kλ l R\λ as before, then, as we have seen, R is embedded in Kλ.
Thus we obtain the following result.
T 4.1. Let R be a fully reersible ring and H its nilradical. Then R\H is a
subring of a direct product of skew fields ; in particular, any fully reersible ring has a
homomorphism into a skew field.
This result is similar to a theorem of Andrunakievich and Ryabukhin in [1]. They
showed that a (non-zero) reduced ring is a subdirect product of integral domains.
If we apply Theorem 4.1 to semiprime rings and remember Theorem 3.1, then we
obtain the following.
C 4.2. Any fully reersible semiprime ring is a subring of a direct
product of skew fields. In particular, any such ring has a homomorphism into a skew
field.
Theorem 4.1 can be made more precise, as follows. Let R be any ring and a set
of full matrices over R closed under diagonal sums. If, further,
E l6,
then, as we saw, there is a prime matrix ideal disjoint from , and over the
corresponding epic R-field, the members of become invertible. Of course, this
conclusion does not require full reversibility.
Even for an integral domain R, it may happen that the set of all diagonal
matrices with no zero on the main diagonal meets the least matrix ideal . This is
the case for any integral domain not embeddable in a skew field (compare [2, 2.11]
and [5, 5.7]). But if R is fully reversible and prime, then is disjoint from , and
we obtain an embedding in a skew field.
C 4.3. A fully reersible ring is embeddable in a skew field if and only
if it is an integral domain or, equialently, a prime ring.
Proof. The condition is clearly necessary. Conversely, assume that R is a prime.
By Theorem 2.1 it is an integral domain, and by Theorem 4.1 it is a subring of a direct
product of fields. We now apply Proposition 1.2.3 of [5] (see also Corollary 1.2.5),
which states that a ring R with a family of homomorphisms fλ to fields is embeddable
in a field provided that for any finite subset X of R B o0q there is an fλ which maps no
member of X to 0. When R is an integral domain, we can replace the family X by the
product, and it follows that a subring of a direct product of fields which is an integral
domain is embeddable in a field, as we had to show.
14692120, 1999, 6, Downloaded from [Link] by Egyptian National Sti. Network (Enstinet), Wiley Online Library on [28/01/2026]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
647
We note that a fully reversible Hermite ring need not be embeddable in a skew
field : for example, the direct product of two fields is Hermite, as is easily verified,
and it is fully reversible, as we saw earlier. This also provides an example of a fully
reversible Hermite ring which is not a Sylvester domain.
In connection with Corollary 4.3, we note that a subring of a skew field need not
be fully reversible. By Theorem 3.1, it will be enough to find a subring of a skew field
which is not unit-stable. For example, consider the polynomial ring in three variables
over a field, k[x, y, z], which is a subring of the rational function field k(x, y, z). We
recall from [3] that this ring has two full 3i3 matrices whose product is non-full :
I 0 z ky K Ix 0 0 K I z ky K I ky 1 0 K.
kz l
L J kz L
0 x 0 1 0 0 x
J y kx 0 L J0 0 1 L J kx 0
0 1
However, by interposing a suitable invertible matrix, we obtain the matrix
I 0 z ky K I0 0 1 K Ix 0 0 K I kxy z 0 K
kz 0 x 0 1 0 0 1 0 l x# 0 kz , (3)
J y kx 0 L J1 0 0 L J0 0 1 L J 0 kx y L
which we claim is full, by a proof along the lines of [3]. Thus assume that the matrix
on the right is not full. Denoting it by C, we have C l PQ, where P is 3i2 and Q is
2i3. Let Qh be the 2i2 matrix consisting of the last two columns of Q. Then
I z 0 K
PQh l 0 kz . (4)
J kx y L
By omitting a row on the right, we obtain a product of two 2i2 matrices, and taking
determinants, we find that det Qh must be a factor of z#, yz, xz ; hence Qh has the
determinant 1 or z (except for a scalar factor). Suppose that det Qh l z, and put
Qh l 0ac bd1 .
The vector
(kx y) Qh−" l z−"(kx y) 0kcd kb
a 1
has components in k[x, y, z], so cyjdx and ayjbx are divisible by z. As in [3], we
reach a contradiction, so det Qh l 1 and Qh is invertible. We now multiply P on the
right by Qh and Q on the left by Qh−" ; this replaces Qh by the unit matrix, while P
reduces to the matrix on the right of (4). Thus the equation C l PQ takes the form
I kxy z 0 K I z 0 K Iq 1 0 K.
x# kz l kz
L J qh L
0 0
J 0 kx y L J kx y
0 1
This leads to the equations qz lkxy, qhz lkx#, xqkyqh l 0, but these equations
have no solution for q, qh, since, for example, xy is not divisible by z. So we have
reached a contradiction, and this shows the matrix (3) to be full, as claimed.
14692120, 1999, 6, Downloaded from [Link] by Egyptian National Sti. Network (Enstinet), Wiley Online Library on [28/01/2026]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
648 . .
In the same way, it can be shown that if
I0 0 1 K I 0 z ky K
Al 0 1 0 , B l kz 0 x ,
Jx 0 0 L J y kx 0 L
then AB is full but BA is non-full. By Theorem 3.1, there are full matrices, such as
the above matrix B, which lie in the least matrix ideal ; it would be interesting to
have an explicit representation of this fact.
References
1. V. A. A and Y. M. R, ‘ Rings without nilpotent elements, and completely
prime ideals ’ (Russian), Dokl. Akad. Nauk SSSR 180 (1968) 9–11.
2. P. M. C, Free rings and their relations, London Math. Soc. Monographs 19 (2nd edn, Academic
Press, London, 1985).
3. P. M. C, ‘ Around Sylvester’s law of nullity ’, Math. Sci. 14 (1989) 73–83.
4. P. M. C, Algebra, Vol. 3 (2nd edn, John Wiley & Sons, Chichester, 1991).
5. P. M. C, ‘ Skew fields ’, Theory of general diision rings, Encyclopedia Math. Appl. 57 (Cambridge
University Press, 1995).
University College London
Gower Street
London WC1E 6BT