Arxiv:2008.03803V2 [Math.RA] 8 Sep 2020 Osbesz Facvrof Cover a of Size Possible Pern N[Ao2 L a Re Oe F2 Orsodn St Corresponding a 2

Home , Noncommutative ring, Order (ring theory)

arXiv:2008.03803v2 [math.RA] 8 Sep 2020 osbesz facvrof cover a of size possible pern n[ao2 l a re oe f2 orsodn st corresponding A 2. of when power that, a order fact had the all for [Maro12] explanation in one appearing provides theorem This iial,n igi h no ftosbig,btcnb h no fth of union the be can but subrings, two of If group. union Sc four subrings. the of Klein is theorem ring the no onto a Similarly, surject by exist that Indeed, there groups However, subgroups. the precisely proper subgroups. are three proper of two to union of is the union courses the algebra abstract never introductory is in problem standard A rprsbig,wihne o aeaymlilctv ntee if a even subrings unit these multiplicative any of have collection not the need which subrings, proper teristic n[ao2 twssonpeieywihascaiernsaeteunio the are rings associative which precisely have shown i.e., was subrings, it [Maro12] In cover. utetin quotient a hoe 1.1. Theorem theorem. following eut osuytecase the study to results with h rbe fcasfigrnsta r no fagvnnme fs of number given a of If union a a to results. are reduces our that it outline rings that now is classifying approach of our problem of the aspect attractice One unit. ul akt oe of cover a to back pulls R sarn,te eawy have always we then ring, a is σ ( fpoe urnswoeuinis union whose subrings proper of nt oa igwoersdefil a rm order. num prime covering has the field compute unital residue we of whose way classification ring the a local Along obtain finite four. and is number, number covering covering finite given a of Abstract. R p = ) o oeprime some for R NRNSA NOSO ORSUBRINGS FOUR OF UNIONS AS RINGS ON n S ( sa soitv ig esythat say we ring, associative an is and n h set The ) h oeignme fa soitv ring associative an of number covering The n a oqoin in quotient no has and σ σ ( ( R R/I .Temi olo hsppri oetn oeo their of some extend to is paper this of goal main The 3. = ) σ R S R ( For . ) n < p ( R n ewrite We . n > ;w banacmlt lsicto o rings for classification complete a obtain we 4; = ) ) sfiie If finite. is n ighscvrn number covering has ring A . o vr rprquotient proper every for 1. cover ≥ σ Introduction R ( ,dfieteset the define 3, O COHEN JON R eetbihasrtg ocasf ntlrings unital classify to strategy a establish We . σ ) of ( ≤ R R R = ) S σ o uhan such For . ( ( ∈ R/I m ∞ S ) ( o any for o nideal an for ) if n R ) R then is S sntcvrbeo a ofinite no has or coverable not is ( coverable n R n < m ob hs ntlrings unital those be to ) R R R/I stemnmlnumber minimal the is let , sfiieadhscharac- has and finite is u rtrsl sthe is result first Our . I n . ic oe of cover a since , σ fadol fi has it if only and if ( fi steuinof union the is it if ra[c2] these [Sco26], orza R nt computation finite hwta group a that show n ruswihare which groups e fevery of ber ig whose rings eteminimal the be ) R ,terings the 3, = os ecall We does. big.We ubrings. tmn for atement e rmore or ree fthree of n R with R/I R 2 JON COHEN groups can be found in Theorem 5 of [Bhar09]. Our next result was the original motivation for the paper, and answers the question of which (unital) rings are the union of four subrings. Theorem 1.2. A unital ring R has σ(R)=4 if and only if σ(R) =6 3 and R has a quotient isomorphic to one the rings in examples 3.3, 3.4, 3.5, or 3.6,

We now outline the article’s contents. In the first section, we define the sets S(n) and establish some of their general properties, including the first result stated above. Here we also prove a useful lemma on covering numbers of products of rings. In the next two sections, we compute the covering numbers for all rings of order at most p3, and all finite local rings with residue field of prime order. We then begin the attack on S(4) in earnest, proving with relative ease the classification of those rings in S(4) of characteristic 3. In the penultimate and most technical section, we classify the rings in S(4) of characteristic 2. Doing this involves first obtaining the bounds 24 ≤ |R| ≤ 28, and then performing a somewhat elaborate case-by-case elimination to show that no rings in S(4) exist besides those already constructed. In the final section we suggest some further directions of research.

2. Notation

We write Zn for the ring of order n generated by 1, and Cn for the cyclic group of order n. If q is a prime power, then Fq denotes the ﬁeld with q elements. For a set X we write |X| for its cardinality. The letter p will always indicate a prime number. For the entirety of this paper, R indicates a unital ring unless otherwise stated, while Si and S denote subrings of R which are not required to have a multiplicative unit.

3. The set S(n)

Deﬁnition 3.1. For n ≥ 3, deﬁne S(n) to be the collection of (isomorphism classes of) unital rings R such that σ(R)= n < σ(R/I) for all proper ideals I of R. For R ∈ n S(n), with maximal subrings S1,...,Sn such that R = Si, call the corresponding iS=1 (n + 1)-tuple (R,S1,...,Sn) a good tuple.

n n If R ∈ S(n) and R = Si for proper subrings Si, the intersection S := Si does iS=1 iT=1 not contain any nonzero 2-sided ideal I of R. Otherwise we could replace R and Si by R/I and Si/I, respectively. Ours is a slightly stronger notion of a good tuple then that of [Maro12], which only requires that S not contain any proper 2-sided ideals. It also did not require the Si to be maximal, but that is automatic if n = 3. Example 3.2. The set S(3) was computed completely in [Maro12]. Its two elements 2 2 (we are requiring rings to have unit) are F2 × F2 and F2[x, y]/(x ,xy,y ) of order 4 and 8, respectively. These were distinguished by whether their maximal subrings all ONRINGSASUNIONSOFFOURSUBRINGS 3

contained 1R, and also by their radicals, which have orders 0 and 4, respectively. Note that both are commutative.

Example 3.3. Let R = F4 × F4. There are exactly four maximal subrings: F4 × F2,

F2 × F4, the subring {(t, t)}t∈F4 , and the subring {(0, 0), (1, 1), (x, x + 1), (x +1, x)}, where x ∈ F4 \ F2. Their union is R and no three form a cover, so R is coverable and σ(R)=4. Since no quotient of R is coverable, R ∈ S(4).

Example 3.4. Let R = M2(F2). There are exactly four maximal subrings: the unique copy of F4 inside R and the stabilizers of the three lines in F2 × F2. Their union is R and no three form a cover, so R is coverable and σ(R)=4. Since R has no proper 2-sided ideals, R ∈ S(4).

2 2 Example 3.5. Let R = F3[x, y]/(x ,xy,y ). There are exactly four maximal subrings: St := {a + bt : a, b ∈ F3} where t ∈{x, y, x + y, x +2y}. Their union is R and no three form a cover, so R is coverable and σ(R)=4. No quotient of R is coverable, e.g., by Proposition 5.1 below, so R ∈ S(4).

Example 3.6. Let R = T2(F3), the ring of upper triangular 2 × 2 matrices over the ﬁeld with three elements. There are exactly four maximal subrings, given by the matrices of the following forms a a b a + b a a b , , , . b a b a + b There union is R and no three of them form a cover, so σ(R)=4. The unique maximal proper quotient of T2(F3) is F3 × F3, which is not coverable by [Wer15]. Hence T2(F3) ∈ S(4).

Our study of S(n) begins with a useful lemma.

Lemma 3.7. Suppose R is a unital ring with σ(R) = n. Let S1,...,Sn be proper n maximal subrings such that R = S1 ∪···∪ Sn. If S := Si, then [R : S] ≤ n! iT=1

Proof. Since Si 6⊂ j=6 i Sj, the inequality [R : S] ≤ n! is a special case of Theorem 6 in [Bhar09]. S Theorem 3.8. The set S(n) is ﬁnite. If R ∈ S(n), then R is ﬁnite and has characteristic a prime p < n. If T is a unital ring then σ(T )= n if and only if T has a quotient in S(n) and has no quotient in S(m) for any m < n.

n Proof. Let (R,S1,...,Sn) be a good tuple, and S = Si. By Lemma 1 in [Lew67] i=1 and lemma 3.7, there is an ideal I of R contained in STwith 2 2 [R : I] < ([R : S]+1)([R:S]+1) ≤ (n!+1)(n!+1) . 2 Since S contains no nonzero ideals, |R| < (n!+1)(n!+1) , so R and S(n) are ﬁnite. 4 JON COHEN

For each prime p, let R(p) be the 2-sided ideal of R which is killed by some power of p. Since R is finite, R = R(p). We have σ(R) = minp{σ(R(p))} by Theorem Qp 2.2. of [Wer15]. Clearly R surjects onto each R(p), and σ(R) < σ(R/I) for all proper ideals I, so R = R(p) for some p. f Suppose R has characteristic p and pR 6⊂ Si. Then by maximality of Si, we have R = Si + pR. So we may write r = s + pr1, some r1 ∈ R. Then r1 = s1 + pr2, and f 2 r2 = s2 + pr3, and so on. Plugging back in and using p = 0 gives r = s + ps1 + p s2 + f−1 ··· + p sf−1. The right side lies in Si, which implies R = Si. This contradiction implies pR ⊂ Si for all i, hence pR ⊂ S, so pR = 0 and R has characteristic p. Let k = [R : S] > 1. The 2-sided ideal kR is contained in S, hence is zero. The characteristic of R is p, so p divides k, which is at most n!, and thus p ≤ n. In fact, p < n. This follows from a property of covering numbers for groups: if G is a finite non-cyclic p-group, then the minimum number of subgroups required to cover it is at least p + 1; see [Tom97]. The abelian p-group R/S is non-cyclic since it is covered by the proper subgroups S1/S,...,Sn/S. Since R/S is not a union of p subgroups, R is not a union of p subrings. Finally, let T be a unital ring with σ(T )= n. The above argument shows that T has a finite quotient T/I with σ(T/I) = n. If T/I ∈ S(n) then we’re done. Otherwise T/I has a proper quotient with covering number n. Since T/I is finite, we can iterate this process until we arrive at a minimal quotient with covering number n, and this is in S(n).

We will prove one more general result about the potential commutative rings inside the sets S(n). First we need a general lemma about covering numbers.

Lemma 3.9. Suppose that R1 and R2 are ﬁnite rings with no common nonzero quotient ring. Then σ(R1 × R2) = min(σ(R1), σ(R2)).

Proof. By Theorem 2.2. of [Wer15], it suﬃces to show that every maximal subring of R := R1 ×R2 has the form M1 ×R2 or R1 ×M2, where Mi ⊂ Ri is a maximal subring. So let T be a maximal subring of R. If we write p1 : R → R1 for the projection map −1 onto the ﬁrst factor, then clearly T ⊂ p1 (p1(T )) = p1(T ) × R2. If p1(T ) =6 R1, then maximality of T forces T = p1(T ) × R2, and p1(T ) must be maximal since T is. Thus we may assume p1(T ) = R1, and similarly that p2(T ) = R2, where p2 is projection onto the second factor.

Now deﬁne I1 := {r ∈ R1 : (r, 0) ∈ T } and I2 := {r ∈ R2 : (0,r) ∈ T }. Suppose ′ ′ r ∈ I1 and r1 ∈ R1. Since p1(T ) = R1, there exists some r ∈ R2 with (r1,r ) ∈ T . ′ Then T ∋ (r, 0)(r1,r )=(rr1, 0) and similarly T ∋ (r1r, 0). Thus I1 is a 2-sided ideal of R1. If I1 = R1, then T ⊃ R1 × 0, which implies T = R1 × I2, and then I2 must be maximal since T is. So we may assume that I1 is a proper ideal of R1, and similarly that I2 is a proper ideal of R2. Note that I1 × I2 ⊂ T , and we may view T/(I1 × I2) as a subring of R/I1 × R/I2. ONRINGSASUNIONSOFFOURSUBRINGS 5

Define f : R1/I1 → R2/I2 by the condition (r1 +I1, f(r1 +I1)) ∈ T/(I1 ×I2). Similarly define g : R2/I2 → R1/I1 by the condition (g(r2 + I2),r2 + I2) ∈ T/(I1 × I2). It is routine to check that f and g are well-defined ring homomorphisms and are inverse ∼ to one another, so R1/I1 = R2/I2. This contradicts our initial assumption, and since all other avenues led to T having the desired form, this completes the proof. Corollary 3.10. Let R ∈ S(n) be a commutative ring. Then all local factors of R have the same residue field.

Proof. Since R is commutative and finite, it is a product of local rings. By grouping the factors in R, we may write R = R1 ×··· Rk where each Ri is a product of local rings with residue field Fqi , with qi =6 qj if i =6 j. Since quotients of local rings are local with the same residue field, we conclude that Ri and Rj have no common quotients if i =6 j. By Lemma 3.9, σ(R) = mini(σ(Ri)). Since R surjects onto each Ri, the definition of S(n) implies that R = Ri for some index i.

We now consider the question of which rings are unions of four, but no fewer, proper subrings. Thus we wish to determine the set S(4). By Theorem 3.8, if R ∈ S(4) then R has order pt for p ∈{2, 3} and t ≥ 1. The rest of this paper is devoted to proving the following theorem. Theorem 3.11. Let R ∈ S(4). a) If |R| =3t then R is isomorphic to the ring in example 3.5 or 3.6. b) If |R| =2t, then R is isomorphic to the ring in example 3.3 or 3.4.

4. Covering numbers for unital rings of order p3

We will compute covering numbers of all unital rings of order p3, using the classiﬁca- tion given in [Ant82]. First, we prove a lemma about rings of order p2.

2 Lemma 4.1. Among unital rings of order p , only F2 × F2 is coverable.

2 2 Proof. For a prime p, the four unital rings of order p are Zp2 , Fp2 , Fp[t]/(t ), and Fp ×Fp. The ﬁrst of these is not coverable since it is generated by 1. The second is not coverable since it is generated by any element not in Fp. The third is not coverable since it is generated by 1 + t. The last is not coverable if p> 2 since it is generated by (1, 2). It is easy to see that F2 × F2 is coverable and σ(F2 × F2) = 3. Corollary 4.2. If R is a coverable unital ring of order p3 that doesn’t surject onto F2 × F2, then R ∈ S(σ(R)). Lemma 4.3. The following is a complete list of coverable unital rings of order p3, with associated covering number σ:

I) T2(Fp), the ring of upper triangular 2 × 2 matrices over Fp, with σ = p +1 6 JON COHEN

2 2 II) Fp[x, y]/(x ,xy,y ), with σ = p +1

III) F3 × F3 × F3, with σ =6

IV) F2 × F2 × F2, with σ =3 2 V) F2 × F2[t]/(t ), with σ =3

VI) F2 × Z4, with σ =3

3 Proof. The ring T2(Fp) is the unique noncommutative unital ring of order p , for any prime p, by [Eld68]. It was shown in [Wer19] that σ(T2(Fp)) = p + 1. We now assume that R is commutative of order p3. Let σ = σ(R). 2 Suppose ﬁrst that R is decomposable, so R = Fp ×T where T has order p . If T = Fp2 , then R is not coverable by Lemma 3.9. If T = Fp × Fp then, by Theorems 3.5. and p 5.3. of [Wer15], R is coverable if and only if p ≤ 3, in which case σ = p + 2 . If R = Fp × Zp2 , then R is not coverable if p> 2 since it is generated by a = (1,2): we have a − ap(p−1) = (0, 1) and 2ap(p−1) − a = (1, 0). If p = 2, this is coverable since it 2 surjects onto F2 ×F2, so σ =3. If R = Fp ×Fp[t]/(t ), it is not coverable if p> 2 since p p+1 p 2p it is generated by a = (1, t − 1): we have a − a = (0, t) and 2 (a + a ) = (1, 0), which together with a give all of R. If p = 2 this is coverable since it surjects to F2 × F2, so σ = 3.

Assume now that R is indecomposable. If the additive group of R is Cp3 , then R is not coverable since no proper subring contains 1. If the additive group of R is 2 2 2 Cp × Cp2 , then R = Zp2 [x]/(px, f(x)) where f(x) ∈ {x , x − p, x − kp} and k is a nonsquare modulo p (only if p > 2). In all cases, R is not coverable since 2 σ(R) = σ(R/pR) = σ(Fp[x]/(x )) = ∞; see the proof of Theorem 3.8 for the ﬁrst 3 equality. If the additive group of R is Cp then R is one of three rings. The ring R = Fp3 is not coverable since it is generated by any element not in Fp. Next, 3 p2 R = Fp[x]/(x ) is not coverable since it is generated by a =1+ x: we have a = 1 p2 2 2 and a − a = x. Last, R = Fp[x, y]/(x ,xy,y ) is coverable with σ = p + 1, by example 6.1. in [Wer15]. Corollary 4.4. If |R| = 27 and R ∈ S(4) then R is isomorphic to the ring in Example 3.5 or 3.6.

2 2 For future reference, we observe that if R surjects onto T2(Fp) or onto Fp[x, y]/(x ,xy,y ), and if |R| is a power of p, then σ(R) = p + 1. To see this, ﬁrst note that clearly σ(R) ≤ p+1. By Theorem 3.8, there exists a quotient T of R, with T ∈ S(σ(R)); since T has order a power of p, Theorem 3.8 shows p < σ(T )= σ(R). Thus σ(R)= p + 1.

5. Covering numbers of finite local rings with residue field Fp

If R is a ﬁnite local ring with radical J, then the characteristic of R is pk for some 2 prime p, and J ⊃ pR. So the R-module J/J is an Fp-vector space. ONRINGSASUNIONSOFFOURSUBRINGS 7

Proposition 5.1. Let R be a ﬁnite commutative local ring with radical J and residue 2 2 ﬁeld Fp. Let n = dimFp J/J . If n ≤ 1, or if n = 2 and p 6∈ J , then R is not 2 2 coverable. Otherwise R is coverable, surjects onto Fp[x, y]/(x ,xy,y ), and σ(R) = 2 2 p +1. In particular, if R ∈ S(d) then d = p +1 and R =∼ Fp[x, y]/(x ,xy,y ).

Proof. If n = 0, then J = J 2 so J = 0. A finite local commutative ring with trivial radical is a field, and fields are not coverable. 2 Suppose n ≤ 2, with p 6∈ J if n = 2. Since R/J = Fp, we have R = Z · 1R + J. By 2 Nakayama’s Lemma, we can lift an Fp-basis for J/J to a minimal generating set for J. For r ∈ J, let [r] denote the image of r inside J/J 2. So we can choose r so that either {[r]} or {[p], [r]} is a basis for J/J 2, depending on n. In either case, J =(r, p) and R = Z · 1R +(r). m Since R is finite, J is nilpotent and r = 0 for some m ≥ 2. Let a = 1R + r and × × t t t = |R |. Since r is nilpotent, a ∈ R , a =1R, a − a = r, and a generates R. 2 Now assume n ≥ 2. Let a1,...,an be an Fp-basis for J/J . By Nakayama’s Lemma, and since J/J 2 is the radical of R/J 2, these can be chosen so that they are also a minimal generating set for J/J 2 as an R/J 2-module. In R/J 2 we have the relations 2 aiaj = 0 and pai = 0 for all 1 ≤ i, j ≤ n. Since the characteristic of R/J is either p or p2, we have shown that there is a surjective ring homomorphism 2 Zp2 [x1,...,xn]/In → R/J

given by xi 7→ ai, where In is the ideal generated by the elements pxi and xixj, n 1 ≤ i, j ≤ n. An element in the kernel, being a non-unit, is of the form pk + cixi iP=1 where 0 ≤ k ≤ p − 1 and 0 ≤ ci ≤ p − 1. If more than one of the cixi were nonzero, 2 or if k = 0, then we would obtain in R/J a linear dependence among the ai, which is absurd. Scaling by units, this leaves elements of the form p + cxi for 0 ≤ c ≤ p − 1 and 1 ≤ i ≤ n. If two such elements (for diﬀerent c or diﬀerent i) were in the kernel, we would again create a linear dependence among the ai. So the kernel is a principal 2 ideal, generated by 0, p, or p + cxi for some 1 ≤ c ≤ p − 1 and 1 ≤ i ≤ n. Thus R/J is isomorphic to one of Zp2 [x1,...,xn]/In

Zp[x1,...,xn]/In

Zp2 [x1,...,xn−1]/In−1. 2 2 2 If n ≥ 3, all three of these surject onto Fp[x, y]/(x ,y , xy), so R/J does. If n = 2 and p ∈ J 2, then R/J 2 has characteristic p and the third case does not arise, so we 2 2 can argue as before. The surjection R → Fp[x, y]/(x ,xy,y ) together with the fact that |R| is a power of p imply the other claims immediately.

The following corollary, which will not be used in this paper, addresses a special case of a question of Werner; see [Wer15]. 8 JON COHEN

Corollary 5.2. Let R be a ﬁnite commutative ring, and let R1,...,Rt be the local factors of R. Assume all Ri have residue ﬁeld Fp. 2 2 a) If any Ri is coverable, then R surjects onto Fp[x, y]/(x ,xy,y ) and σ(R)= p +1.

b) Suppose all Ri are not coverable. Then R is coverable if and only if t ≥ p, in which p case p +1 ≤ σ(R) ≤ p + 2 Proof. a) This is immediate from the previous proposition. b) The inequality p +1 ≤ σ(R) is automatic if σ(R)= ∞ and follows from Theorem p 3.8 if R is coverable. If t ≥ p then R surjects onto Fp , which is coverable with covering p p number p + 2 by [Wer15]. Thus R is coverable and σ(R) ≤ p + 2 . So assume t < p. From the proof of Proposition 5.1, each Ri is a quotient of p−1 mi mi Zpki [ti]/(ti ) for some ki, mi ≥ 1. If the ring T := Zpki [ti]/(ti ) is not cover- i=1 able then neither is its quotient R. Let M be the subringQ generated by

x = (1, 2,...,p − 1)+(t1, t2,...,tp−1). × × f We claim M = T . Let f = |T |. Since each i+ti ∈ Ri , we have x = (1, 1,..., 1), so f f f x−x = (0, 1, 2,...,p−2)+(t1, t2,...,tp−1), and (x−x ) = (0, 1, 1,..., 1). Subtract- ing this from xf gives (1, 0, 0 ..., 0) ∈ M. This shows that M ∋ (1, 0,..., 0)(x−xf )= m1 (t1, 0,..., 0), and so M ⊃ Zpk1 [t1]/(t1 ) × 0 ×···× 0. Iterating this argument yields M = T , so T is not coverable. Proposition 5.3. Let R be a ﬁnite noncommutative local ring with residue ﬁeld 2 2 Fp. Then R surjects onto Fp[x, y]/(x ,xy,y ) and σ(R) = p + 1. In particular, R 6∈ S(p + 1).

Proof. Let J be the radical of R. Since xy − yx ∈ J 2 for x, y ∈ J, and since R = 2 Z · 1R + J, we see that R/J is commutative. Using the criteria of Proposition 5.1, we will show the commutative local ring R/J 2 is coverable, and thus obtain a surjection 2 2 2 R → R/J → Fp[x, y]/(x ,xy,y ) which implies the rest. By Lemma 2.2 of [Cor I], 2 2 2 dimFp J/J > 1. Suppose dimFp J/J = 2 and p 6∈ J . Then there exists x ∈ J such 2 that x and p provide an Fp-basis for J/J . By Lemma 2.1 of [Cor I], J =(p, x). But then R = Z · 1R + J = Z · 1R +(x), so R is commutative.

6. Rings in S(4) of order 3t

4 Let (R,S1,S2,S3,S4) be a good tuple, with |R| a power of 3. Recall S = Si. Since iT=1 [R : S] = [R : Si][Si : S] ≤ 4! = 24 by Lemma 3.7, and [R : S] is a power of 3, we have [R : Si] = [Si : S] = 3 for every 1 ≤ i ≤ 4. In particular, 3R ⊂ S, so 3R = 0. Note that Si ∩ Sj = S whenever i =6 j, since Si ∩ Sj is a proper subring of Si, but S is a maximal subring of Si for each i. Suppose there are 2 indices i =6 j with ONRINGSASUNIONSOFFOURSUBRINGS 9

1R 6∈ Si ∪Sj. Then Si and Sj are both maximal ideals of R by Lemma 2.1. of [Wer18], and S = Si ∩ Sj is an ideal, hence is zero. But S = 0 implies |R| = 9, contradicting Lemma 4.1. Thus 1R ∈ S. Let S1/S = {0, x1, 2x1} and S2/S = {0, x2, 2x2}. We 3 index so that S3/S = {0, x1 + x2, 2x1 + 2x2}. Since x1 + 2x2 6∈ Si, we have iS=1 S4/S = {0, x1 +2x2, 2x1 + x2}. For r ∈ R \ S, let IR(r) := {s ∈ S : sr ∈ S} and IL(r) := {s ∈ S : rs ∈ S}. These are additive subgroups of S.

Lemma 6.1. The sets IR(r) and IL(r) are independent of r.

Proof. Let r = ax1 + bx2 be one of our eight nontrivial coset representatives. We will show that the sets IR(r) are all equal. Clearly IR(r) = IR(2r). Suppose s ∈ IR(x1). If s 6∈ IR(x2) then sx2 = kx2 + s2 for some s2 ∈ S, 1 ≤ k ≤ 2. But then s(x1 + x2)= sx1 +s2 +kx2 is a nontrivial coset representative of S2/S, hence cannot be in S3, which contradicts S3 being a ring. Similarly, if s 6∈ IR(x1+x2), then s(x1+x2)= s3+kx1+kx2 for some s3 ∈ S, 1 ≤ k ≤ 2. But then s(2x1 + x2) = sx1 + s3 + kx1 + kx2 lies in S3 instead of S4. Finally, if s 6∈ IR(2x1 + x2), so that s(2x1 + x2) = s4 +2kx1 + kx2, for some s4 ∈ S and 1 ≤ k ≤ 2, then sx2 = sx1 + s4 +2kx1 + kx2, which lies in S4 instead of S2. Now apply symmetry.

Lemma 6.2. The additive groups S/IR(r) and S/IL(r) have order 3.

Proof. We can assume that r is one of our eight nontrivial coset representatives. Since 1R ∈ S \ IR(r), we have [S : IR(r)] > 1. Suppose that s1,s2,s1 − s2 ∈ S \ IR(r). ′ ′ We claim s1 + s2 ∈ IR(r). Write s1r = s1 + k1r, s2r = s2 + k2r with 1 ≤ ki ≤ 2. ′ ′ Then (s1 − s2)r = s1 − s2 + (k1 − k2)r must have k1 − k2 ∈ {1, 2}. But then ′ ′ (s1 + s2)r = s1 + s2 +(k1 + k2)r ∈ S since k1 + k2 = 0.

Lemma 6.3. We have IR(r) ∩ IL(r)=0.

Proof. Let t ∈ IR(r) ∩ IL(r). We will show that z1tz2 ∈ S for zi varying among any of our eight coset representatives. Suppose ﬁrst that z1 and z2 belong to diﬀerent Si. Since z1t ∈ S, z1tz2 lies in the same Si as z2 does. Since tz2 ∈ S, z1tz2 lies in the same Si as z1 does. Since the intersection of two distinct Si is S, we have z1tz2 ∈ S. It remains only to show that x1tx1 and x2tx2 lie in S. We have x1tx1 = s + bx1 where 0 ≤ b ≤ 2. So x1t(x1 + x2)= s + x1tx2 + bx1. The left side lies in S3, the right side in S1 since x1tx2 ∈ S, thus b = 0. The case x2tx2 is similar. Therefore RtR is an ideal of R contained in S. This forces RtR = 0, so t = 0.

Since 3 ≤|S| = [S : IR(r) ∩ IL(r)] ≤ [S : IR(r)][S : IL(r)] = 9, we see that |R| = 27 or |R| = 81. We already showed what was possible if |R| = 27 in Corollary 4.4. The next result completes the proof of part a) of Theorem 3.11 Proposition 6.4. If R ∈ S(4) then |R|= 6 81. 10 JON COHEN

Proof. Suppose |R| = 81 and R ∈ S(4). We have |S| = 9 and |Si| =27 for 1 ≤ i ≤ 4. If R was commutative, then IR(r) would be a 2-sided ideal in S, hence would be zero. But this would force |S| = 3, so R is noncommutative. Let J = J(R) denote the Jacobson radical of R. If J = 0 then R is semisimple, so by order considerations and noncommutativity R = M2(F3). But σ(M2(F3)) = 7 =6 4 by [Wer18]. If |J| = 27, then R is a local ring with residue field F3. By Proposition 5.3, R 6∈ S(4). This leaves |J|∈{3, 9}. Noncommutative unital rings of order p4, for p a prime, have been classified in [Derr94]; we make use of their list for R having characteristic p and |J|∈{p,p2}. There are six such rings: five with |J| =32, and one with |J| = 3.

The unique noncommutative ring R of order 81 with |J| = 3 is R = F3 × T2(F3), which clearly surjects onto T2(F3). So σ(R) = 4 but R 6∈ S(4). ∼ The unique noncommutative ring R of order 81 with R/J = F9 is given by matrices of the form a b 0 a3 where a, b ∈ F9. The only nonzero 2-sided ideal of R is J, which is deﬁned by a = 0. Let S1 be the index-3 subring deﬁned by the condition a ∈ F3. Suppose M was another index-3 subring of R. Since M 6⊂ S1, M contains a matrix X as above with a ∈ F9 \ F3. Then [R : M] = 3 implies there exists 0 =6 Y ∈ M ∩ J. But any such X and Y together generate R. So all other maximal subrings S2,...,Sσ(R) have order at most 9. They all contain 1R since they would otherwise be contained in J, hence in S1. Thus |Si| =9for2 ≤ i ≤ σ(R) and Si ∩ Sj = {0, 1, 2} for all 1 ≤ i =6 j ≤ σ(R). Since |R \ S1| = 54 and each other Si contributes 6 additional elements, we see that σ(R) = 10. ∼ If R is noncommutative of order 81 and R/J = F3 ×F3, then R is one of the following four rings: a b a a b c a b c  a  b a  ,  a  ,  d  ,  d   c d   d   d           c d           where a,b,c,d ∈ F3. In each case, let B be the matrix with a = c = d = 0 and b = 1. The F3-span of B is a two-sided ideal, and R/(B) is noncommutative, so ∼ R/(B) = T2(F3). Thus σ(R) = 4 but R 6∈ S(4).

7. Rings in S(4) of order 2t

Now we commence the systematic study of the 2-rings in S(4), which presents more diﬃculties than the 3-rings.

Proposition 7.1. Suppose (R,S1,S2,S3,S4) is a good tuple with |R| a power of 2. 4 Then [R : Si]=8. iT=1 ONRINGSASUNIONSOFFOURSUBRINGS 11

4 Proof. Let S = Si. Since [R : S] = [R : Si][Si : S] ≤ 4! = 24 by Lemma 3.7, we i=1 have [R : S] ∈{4T, 8, 16}.

We cannot have [R : S] = 4, since the abelian group R/S has the Si/S as four distinct nontrivial proper subgroups, and no group of order 4 has that many nontrivial proper subgroups. 4 Suppose [R : S] = 16. By Theorem 3.8, R has characteristic 2, so R/S = C2 . Let the subgroups Si/S be ordered so that |Si/S| is nonincreasing. Since the intersection of four index-4 subgroups contains the identity, they cannot cover the whole group. Thus |S1/S| = 8. Any subgroup of order 4 will intersect S1/S nontrivially, hence the remaining Si/S cannot all be order 4 (or less), since they would not be able to 4 cover R/S. Thus |S2/S| = 8. We will show that the intersection Si/S must be i=1 T 4 nontrivial, contradicting the deﬁnition of S. Let w,x,y,z be a generating set for C2 . Without loss of generality, S1/S is generated by {w,x,y} and S2/S by {w, x, z}. The 4 elements not in S1/S ∪ S2/S are y + z, w + y + z, x + y + z, and w + x + y + z. A group containing any three of these will contain the fourth, so they must be split evenly in two between S3/S and S4/S. If y + z, w + y + z ∈ S3/S, then w ∈ S/S. If y + z, x + y + z ∈ S3/S, then x ∈ S/S. And if y + z, w + x + y + z ∈ S3/S, then w + x ∈ S/S. So y + z cannot be paired with anything, which is a contradiction.

4 Suppose (R,S1,S2,S3,S4) is a good tuple, with R a 2-ring and S = Si. For i=1 r ∈ R \ S, deﬁne the sets T

IR(r) := {s ∈ S : rs ∈ S}

IL(r) := {s ∈ S : sr ∈ S}

IR := IR(r) r\∈R

IL := IL(r) r\∈R

All four are additive subgroups of R. The set IR is a left R-ideal, and IL is a right R-ideal. Clearly IR(r)= IR(r+s) for any s ∈ S, so that in studying these sets we may assume r is one of the seven nontrivial elements of R/S. Our strategy is to bound |S| = [S : IR ∩ IL]|IR ∩ IL| ≤ [S : IR][S : IL]|IR ∩ IL| by bounding [S : IR], [S : IL], and |IR ∩ IL|. As a last observation, note that if T is a subring of R containing S and r, then the additive map s 7→ rs : S → T induces an injection S/IR(r) → T/S. Below we will use letters x, y, z to denote elements of R/S and any lift of them to R; the reader can check that this creates no issue. ∼ 3 Since R has characteristic 2 and [R : S] = 8, we have R/S = C2 . Assume the Si have been ordered so that |Si/S| is nonincreasing. If S1/S and S2/S did not both 12 JON COHEN have order 4, then we could not form a cover of R/S. So S1/S = {0, x, y, x + y} and S2/S = {0, x, z, x + z}, where x, y, z is some set of generators for R/S.

Suppose that S3/S and S4/S also had order 4. From the inclusion-exclusion principle, |Si/S ∪ Sj/S ∪ Sk/S| =6+ |Si/S ∩ Sj/S ∩ Sk/S|, where i, j, k are distinct. So for our cover to be irredundant, we must have Si ∩ Sj ∩ Sk = S whenever i, j, k are distinct. We can assume, without loss of generality, that S3/S = {0,y,z,y + z}. But now S4/S, which must contain x + y + z, cannot contain x, y, or z, as this would create a nontrivial triple intersection. Similarly S4/S cannot contain x + y, x + z, or y + z, as any of these would, with x + y + z, yield one of x or y or z. So |S4/S| = 2. There are two options for |S3/S|.

7.1. The case |S3/S| = 2. We write S3/S = {0,y + z} and S4/S = {0, x + y + z}. Since R/S3 and R/S4 have nonprime order, the maximal subrings S3 and S4 contain 1R by Lemma 2.1. of [Wer18]. So 1R ∈ S3 ∩ S4 = S, and in particular, |S| ≥ 2.

Lemma 7.2. We have IR(x) = IR(y + z) = IR(x + y + z) ⊃ IR(y) = IR(x + y) = IR(z)= IR(x + z)= IR.

Proof. Suppose that (x + y + z)s 6∈ S, so that (x + y + z)s = x + y + z + s′ for some s′ ∈ S. Then none of xs, ys, zs, (x+y)s, (x+z)s, (y +z)s can lie in S. For example, if ′ xs ∈ S, then (y + z)s = x + y + z +(s + xs). The left side lies in the subring S3, and the right side is in S4 \ S, which is absurd. The other cases are completely analogous, and rely only on the fact that Si ∩ S4 = S for i =6 4. Thus IR(r) ⊂ IR(x + y + z) for every r ∈ R \ S. Suppose that (y + z)s 6∈ S, so that (y + z)s = y + z + s′ for some s′ ∈ S. Then none of xs, ys, zs, (x + y)s, (x + z)s, (x + y + z)s can lie in S. For example, if xs ∈ S then ′ (x + y + z)s = y + z +(s + xs). The left side lies in S4 and the right side lies in S3 \ S, which is absurd. The other cases are completely analogous, and rely only on the fact that Si ∩ S3 = S for i =6 3. Thus IR(r) ⊂ IR(y + z) for every r ∈ R \ S. In particular, IR(x + y + z)= IR(y + z).

Suppose s ∈ IR(y + z) = IR(x + y + z), so that (x + y + z)s ∈ S and (y + z)s ∈ S. We see immediately that xs ∈ S, and so IR(x)= IR(y + z)= IR(x + y + z).

If s ∈ IR(y) ⊂ IR(x) we have ys,xs ∈ S so (x + y)s ∈ S. Symmetry implies IR(y) ⊂ IR(x + y), and an identical argument shows IR(z)= IR(x + z). Similarly, if x ∈ IR(y) ⊂ IR(y + z) we have ys ∈ S and (y + z) ∈ S, which immediately yields zs ∈ S, so IR(y) ⊂ IR(z). By symmetry IR(y) = IR(z). Since IR is the intersection of all IR(r), there is nothing left to show.

Lemma 7.3. We have 2 ≤ [S : IR] ≤ 4 and 2 ≤ [S : IL] ≤ 4.

Proof. Since 1R ∈ S \ IR, we know S/IR is nontrivial. Since IR = IR(y), we have an injection [s] 7→ [ys]: S/IR → S1/S. The case for IL is symmetric.

Lemma 7.4. We have |IR ∩ IL| ≤ 2. ONRINGSASUNIONSOFFOURSUBRINGS 13

Proof. Let t ∈ IR ∩ IL. We have (y + z)tx, (y + z)ty, (y + z)tz, xt(y + z), yt(y + z), and zt(y + z) all lying in S since they lie in S3 ∩ Si = S for i =1, 2. The same holds if we replace all occurences of y + z with x + y + z, since S4 ∩ Si = S for i = 1, 2. From this we can conclude that (y + z)t(y + z) ∈ S and that (x + y + z)t(x + y + z)= xtx + xt(y + z)+(y + z)tx +(y + z)t(y + z) ∈ S. Thus xtx ∈ S. Since (x + y + z)ty and (y + z)ty lie in S, their sum xty ∈ S. Similarly ytx ∈ S. These then force ztx, xtz ∈ S. The remaining elements to consider are yty,ztz,zty, and ytz. From the above we see that yty, zty, ztz, and ytz are all congruent to one another modulo S. So yty ∈ S iﬀ RtR ⊂ S iﬀ t = 0. Since ytz ∈ S1 ∩ S2, yty ∈ S1 ∩ S2. Suppose we have 0 =6 a, b ∈ IR ∩ IL. If yay,yby 6∈ S, then since S1/S ∩ S2/S has order 2, we have y(a + b)y ∈ S.

Corollary 7.5. If (R,S1,S2,S3,S4) is a good tuple with [S1 : S] = [S2 : S]=4 and t [S3 : S] = [S4 : S]=2, then R has order 2 with 4 ≤ t ≤ 8. If R is commutative, then 4 ≤ t ≤ 5.

5 Proof. We have 2 ≤ |S| = [S : IR ∩ IL]|IR ∩ IL| ≤ [S : IR][S : IL] ∗ 2 ≤ 2 , and so 4 8 2 ≤ |R| ≤ 2 . If R is commutative then IR = IL is a 2-sided ideal in S, forcing IR =0. So 2 ≤|S| = [S : IR] ≤ 4 and 16 ≤|R| ≤ 32.

7.2. The case |S3/S| =4. Since R/S4 has nonprime order, the maximal subring S4 must contain 1R. In fact, 1R must be contained in at least two of the other Si as well. For if, say, 1R 6∈ S1 ∪ S2, then S1 and S2 would be maximal 2-sided ideals, and ∼ then the injection R/(S1 ∩ S2) → R/S1 × R/S2 = F2 × F2 must be surjective since [R : S1 ∩ S2] = 4, which forces σ(R) = 3. But since S4 ∩ Si = S for 1 ≤ i ≤ 3, we deduce that 1R ∈ S.

Having written S1/S = {0, x, y, x+y} and S2/S = {0, x, z, x+z}, we consider what is possible for S3/S and S4/S. Either S4/S = {0, x+y+z} or S4/S = {0,y+z}; without loss of generality, we assume S4/S = {0, x + y + z}. Then S3/S = {0,y,z,y + z} or {0, x + y, x + z,y + z}, since there are 3 subgroups of order 4 containing y + z and the third would contain S4/S. Without loss of generality, we can take S3/S to be the former; the other choice is obtained by replacing y and z by x + y and x + z, which doesn’t change S1,S2, or S4, nor the sizes of intersections of any number of the Si.

Lemma 7.6. We have IR(r) ⊂ IR(x + y + z) for r ∈ R \ S. Also IR(x)= IR(y + z), IR(y)= IR(x + z), and IR(z)= IR(x + y).

′ ′ Proof. Suppose s 6∈ IR(x + y + z), so that (x + y + z)s = x + y + z + s for some s ∈ S. Then we cannot have xs, ys, zs, (x + y)s, (x + z)s, or (y + z)s in S. For example if we had xs ∈ S then we would obtain (y + z)s = x + y + z +(s′ + xs), with the left side in S3 and the right in S4 \ S, which is absurd. So IR(r) ⊂ IR(x + y + z).

Now if s ∈ IR(x) ⊂ IR(x + y + z). Then clearly s ∈ IR(y + z), and by symmetry IR(x)= IR(y + z). Similarly IR(y)= IR(x + z) and IR(z)= IR(x + y). 14 JON COHEN

Lemma 7.7. The additive groups S/IR(r) have order 2.

Proof. Since 1R ∈ S \ IR(r), the group S/IR(r) is nontrivial. If r = x + y + z, the map [s] 7→ [(x + y + z)s]: S/IR(x + y + z) → S4/S is an injection. If r = x, the map s 7→ xs : S → R has image in S1 ∩ S2. Hence it induces an injection S/IR(x) → S1/S ∩ S2/S, and so [S : IR(x)] ≤ 2. Apply symmetry and the previous lemma.

Lemma 7.8. We have [S : IR] = [S : IL]=2.

Proof. We know IR(r) ⊂ IR(x + y + z), which forces S/IR(x + y + z) to be a quotient of S/IR(r). But both have order 2, so IR(r)= IR(x + y + z), and hence IR = IR(r). The case of IL is symmetric.

Lemma 7.9. We have IR ∩ IL =0.

Proof. Let t ∈ IR∩IL. Since z ∈ S2∩S3, and x ∈ S1∩S2, we get xtz ∈ S1∩S2 ∩S3 = S. Similarly ztx, xty, ytx, ytz, zty ∈ S. We have (x+y +z)tr and rt(x+y +z) lying in S for r ∈{x, y, z}, since these lie in S4 ∩ Si = S with 1 ≤ i ≤ 3. Thus xtx, yty, ztz ∈ S. So RtR ⊂ S which forces t = 0.

Corollary 7.10. If (R,S1,S2,S3,S4) is a good tuple with [S1 : S] = [S2 : S] = [S3 : S]=4 and [S4 : S]=2, then R has order 16 or 32.

Proof. We have 2 ≤|S| = [S : IR ∩ SL] ≤ [S : IR][S : IL] = 4 and [R : S] = 8.

7.3. Rings in S(4) of order 16, 32, 64, 128, and 256. Now we will proceed to complete the proof of Theorem 3.11. The next proposition shows that S(4) contains no rings of order 16 besides F4 × F4 and M2(F2), and that S(4) contains no rings of order 32. Note that noncommutative rings are coverable. Proposition 7.11. Let R be a ring with σ(R) > 3. ∼ a) If R is commutative, coverable, and of order 16 then R = F4 × F4 and σ(R)=4. ∼ b) If R is commutative, coverable, and of order 32 then R = F4×F4×F2 and σ(R)=4. In particular, R 6∈ S(4).

c) If R is noncommutative of order 16 then either R = M2(F2) and σ(R)=4, or R is the unique ring of order 16 with σ(R)=5.

d) If R is noncommutative of order 32 then either R = F2 × M2(F2) and σ(R)=4, or R is one of three rings with σ(R)=5. In particular, R 6∈ S(4).

Proof. a) Suppose first that R is local. If R has residue field F16 then R = F16 and is not coverable. If R had residue field F8 then |R| would be a power of 8. By Proposition 5.1, if R has residue field F2 then σ(R) ∈ {3, ∞}. Last, suppose R has ∼ 2 residue field F4. If R has characteristic 2, then R = F4[x]/(x ), which is not coverable ONRINGSASUNIONSOFFOURSUBRINGS 15

4 4 since it is generated by b = a + x for any a ∈ F4 \ F2: we have b = a and b − b = x. If R has charactersitic 4, then σ(R)= σ(R/2R) so R has the same covering number as a local ring of smaller order with residue field F4. The only way this can happen is R/2R = F4, so R is not coverable. Now suppose that R is a product of smaller local rings. If there are more than two factors, then R = T × F2 × F2 for some ring T of order 4, so R surjects onto F2 × F2 and σ(R) = 3. Thus R = R1 × R2, with R1 and R2 local. Since quotients of local rings are local, either R1 and R2 have the same residue field or they have no common quotients. In the latter case, Lemmas 3.9, 4.1, and 4.3 imply that σ(R1 × R2) = min(σ(R1), σ(R2)) ∈{3, ∞}. If R1 and R2 both have residue field F2 then R surjects onto F2 × F2 and σ(R) = 3. If both have residue field F4 then R = F4 × F4, and we showed in example 3.3 that R is coverable with σ(R) = 4. b) Suppose R is local. If R has residue field F32 then R = F32 and is not coverable. Since 32 is not a power of 16, 8, or 4, R does not have residue field F16, F8 or F4. By Proposition 5.1, if R has residue field F2 then either R is not coverable or σ(R) = 3.

t Now let R be a product of local rings, R = Ri, with 2 ≤ t ≤ 5. If t ≥ 3, then either i=1 Q ∼ two (or more) factors have residue field F2, so σ(R) = 3, or else R = F4 × F4 × F2, which has covering number 4 by Lemma 3.9. This just leaves the case R = R1 × R2 with R1 and R2 local. Suppose |R1| = 16, so R2 = F2. If R1 has residue field F2, R surjects onto F2 × F2 and σ(R) = 3. Otherwise R1 = F16 or R1 has residue field F4. We already showed that none of these are coverable, so by Lemma 3.9, R1 × F2 is not coverable. Finally, suppose |R1| = 8 and |R2| = 4. If they both have residue field F2 then R surjects onto F2 × F2 and σ(R) = 3. Otherwise they must have different residue fields, hence no common quotients, and by Lemmas 3.9, 4.1, and 4.3 we have σ(R1 × R2) = min(σ(R1), σ(R2)) ∈{3, ∞}. c) We consider R in cases according to the size of its radical J = J(R). Since R has a unit, |J|∈{0, 2, 4, 8}. If J = 0 then R is a semisimple ring, so by order considerations and noncommutativity R = M2(F2). We showed that σ(M2(F2)) = 4 in example 3.4. 3 Suppose |J| = 2. If R/J = F2 , then R surjects onto F2 × F2, giving σ(R) = 3. We cannot have R/J = F8 or R/J = F2 × F4; see [Derr94]. Suppose |J| = 4. If R/J = F2 × F2, then σ(R) = 3. The only other possibility is R/J = F4. By [Derr94] or [Cor I], there is a unique such R, isomorphic to the ring of matrices of the form a b 0 a2 where a, b ∈ F4. The proof that σ(R) = 5 is completely analogous to the proof in Proposition 6.4 that the corresponding ring of order 81 had covering number 10, so we omit the details. Finally, if |J| = 8 then σ(R) = 3 by Proposition 5.3. d) Suppose R is local. Since R =6 F32, the residue field of R is F2. Then σ(R) = 3 by Proposition 5.3. 16 JON COHEN

If R is decomposable, then at least one of its summands is noncommutative. Unital rings of order less than 8 are commutative. The only noncommutative ring of order 8 is T2(F2), which has covering number 3, and so σ(R) = 3 if T2(F2) is a summand ′ of R. The remaining case is R = F2 × R for some noncommutative ring of order 16. ′ ′ ′ If σ(R ) = 3 then σ(R)=3. If σ(R ) = 4 then R = M2(F2) by part b). By part b) again, the only remaining possibility is that R′ is the unique noncommutative ring ′ of order 16 and covering number 5. Since R does not surject onto F2, Lemma 3.9 ′ ′ implies σ(F2 × R )= σ(R ) = 5. Finally, suppose that R is indecomposable and nonlocal. From the classiﬁcation in [Cor I], there are exactly 2 nonlocal, indecomposable rings of order 32 that do not surject onto F2 × F2; they are opposite rings of one another, so they have the same covering number. Concretely, we may realize one of them as the ring of matrices a b 0 c

where a ∈ F2, b, c ∈ F4. The radical J of R consists of elements with a = c = 0. If x ∈ R is a matrix with a = 1, arbitrary b, and c 6∈ F2, and 0 =6 y ∈ J, then x and y generate R. Thus no proper subring of index 2 can contain such an element 3 x, since any subring of index 2 intersects J nontrivially. Since x = 1R, the subring 2 Sx generated by x consists of elements of the form i11R + i2x + i3x , where ij ∈ F2. It is easy to check that these are 8 distinct elements in R, so [R : Sx] = 4, and Sx is therefore a maximal subring generated by a single element. Thus Sx must be 2 included in any cover of R. Since no element of Sx besides x and x have a = 1 ′ 2 and c 6∈ F2, we see that Sx = Sx′ if and only if x ∈ {x, x }. There are 8 elements with a = 1 and c 6∈ F2, so there are 4 distinct subrings of the form Sx. The four Sx cannot cover R, being of index 4 and having nonempty intersection, so σ(R) > 4. Let T be the index-2 subring consisting of matrices with c ∈ F2. One sees that 2 2 ′ 2 Sx \ (T ∪ Sx′ ) = {x, x , 1R + x, 1R + x } for any x =6 x, x . Since |R \ T | = 16 and there are four elements in each of the four Sx that lie in no other of our maximal proper subrings, we have R = T ∪ ( Sx) and σ(R) = 5. S From Proposition 7.11, Corollary 7.5, and Corollary 7.10, we now know that if (R,S1,...,S4) is a good tuple and R is not F4 ×F4 or M2(F2), then R is noncommutative, |R|∈{64, 128, 256},[R : S1] = [R : S2] = 2, [R : S3] = [R : S4] = 4, [R : S] = 8, and 1R ∈ S. We also saw that S = Si ∩ Sj if i =6 j except that [S1 ∩ S2 : S] = 2. In particular, Si ∩ Sj ∩ Sk = S for any three distinct indices i, j, k. For the remainder of this section, R will denote a hypothetical ring in S(4) satisfying these conditions. We will show that no such ring exists, thereby completing the proof of Theorem 3.11. Proposition 7.12. R is indecomposable.

Proof. Suppose that R is decomposable. If R = R1 × R2 where R1 and R2 have no common quotients. By Lemma 3.9, we then have σ(R) = min(σ(R1), σ(R2)). So ONRINGSASUNIONSOFFOURSUBRINGS 17

R 6∈ S(4) because R has a quotient with the same covering number, contradicting the deﬁnition of S(n). In any decomposition of R ∈ S(4) into a sum of indecomposables, there must then be pairwise common quotients.

If R = R1 × R2 where R1 is of order 2 or 4, then the minimal quotient of R1 is either F2 or F4, so R2 will have to surject onto F2 or F4 for there to be a common quotient. In the former case R surjects onto F2 × F2 and σ(R) = 3, while in the latter case R surjects onto F4 × F4 so R 6∈ S(4). Summands of R therefore have order at least 8.

Clearly R must have a noncommutative indecomposable summand. Since T2(F2) is the only noncommutative ring of order 8 and σ(T2(F2)) = 3, R has a noncommutative summand of order at least 16. Write R′ for the unique ring of order 16 and covering ′ number 5. Suppose R = T × R , where T is any ring of order |R|/16. Since F4 is ′ the only proper quotient of R , F4 must be a quotient of T . But then R surjects onto F4 × F4, so R 6∈ S(4). By Proposition 7.11 no noncommutative ring of order 16 other than R′ can be a summand of R. Thus R must have a noncommutative indecomposable summand of order at least 32.

These observations rule out |R| = 64 and |R| = 128. Let R = R1 × R2 where |R1| = 8, |R2| = 32, and R2 is noncommutative and indecomposable. If R1 and R2 have a common quotient of F2 or F4, then we can argue as before to deduce R 6∈ S(4). The only other apparent possibility would be a common quotient of F8. However, this would contradict R2 being noncommutative and indecomposable; see the classiﬁcation in [Cor I], for example.

Lemma 7.13. R contains no 2-sided ideals I with 0 =6 S ∩ I = S3 ∩ I = S4 ∩ I.

Proof. If I is such an ideal in R then Si ∩ I is an ideal in Si. Therefore S ∩ I is an ideal in both S3 and in S4, hence is an ideal in the ring generated by S3 and S4. Maximality of the Si then implies that S ∩ I is an ideal of R, contained in S. So σ(R)= σ(R/(S ∩ I)), contradicting R ∈ S(4). Proposition 7.14. R is not a local ring.

7 Proof. By Proposition 5.3, R cannot have residue ﬁeld F2. Thus |R|= 6 2 since R =6 F128. Let J = J(R) be the radical of R, and f : R → R/J the quotient map.

Case 1a: |R| = 64 and R/J = F8. The relevant subrings have orders |S1| = |S2| = 32, |S3| = |S4| = 16, and |S| = 8. Since 1R ∈ S \ J, we have |S ∩ J|∈{0, 2, 4}.

If S ∩ J = 0 then f|S : S → R/J is injective, hence surjective. If 0 =6 j ∈ J then 2 {sj}s∈S = J, since J = J/J is a 1-dimensional vector space over F8. The subring generated by S and j then contains J, and R = S + J. But this means j 6∈ Si for any i, contradicting the deﬁnition of a covering. If S ∩ J had order 2 then f(S) = S/S ∩ J would have order 4 and be a subring of R/J = F8. This is impossible. 18 JON COHEN

Suppose S ∩ J has order 4. The maps f|Si : Si → R/J contain S ∩ J in their kernels. So f(S3) and f(S4) have order at most 4, hence at most 2 since F8 contains no subring −1 of order 4. But f (F2) is then a proper subring of R containing the maximal subrings S3 and S4, a contradiction.

Case 1b: |R| = 64 and R/J = F4. The sizes of the subrings Si and S are as before. Since 1R ∈ S \ J, and S ∩ J is the kernel of the map f|S : S → R/J = F4, we have |S ∩ J|∈{2, 4}.

3 If S ∩ J has order 2 then f|S : S → R/J = F4 is onto. Since J = 0 we have

2 |S ∩ J| = |(S ∩ J i)/(S ∩ J i+1)|. Yi=1 i i+1 i i+1 Observe that (S ∩ J )/(S ∩ J ) is an F4-subspace of J /J since it is stable under i i+1 multiplication by S/(S ∩ J)= F4. So |(S ∩ J )/(S ∩ J )| is a power of 4, but their product equals |S ∩ J| = 2.

−1 If S ∩ J has order 4, then f(S)= S/(S ∩ J)= F2. The preimage f (F2) is a subring of index 2, so it cannot contain the maximal subrings S3 or S4, which have index 4. Therefore f(S3)= f(S4)= R/J, and it follows that S3 ∩ J and S4 ∩ J have order 4,

being the kernels of f|Si : Si → R/J = F4. But then the inclusion S ∩ J ⊂ Si ∩ J and the assumption that S ∩ J has order 4 means that S ∩ J = Si ∩ J for i =3, 4. Now apply Lemma 7.13

Case 2a: |R| = 256 and R/J = F16. We have |S1| = |S2| = 128, |S3| = |S4| = 64, and |S| = 32. Note that J, which has order 16, is not contained in S because S contains no 2-sided ideals of R. Thus |S ∩ J|∈{2, 4, 8}.

2 If S ∩J has order 2 then f|S : S → R/J is onto. Noting that J = 0, we see that S ∩J is a vector space over S/(S ∩ J)= f(S)= F16. But 16 does not divide |S ∩ J| = 2.

If S ∩J has order 4 then f(S)= S/(S ∩J) has order 8 and is a subring of R/J = F16. This is impossible.

Finally, suppose S ∩ J has order 8. The maps f|Si : Si → R/J contain S ∩ J in their kernels. In particular, f(S3) and f(S4) have order at most |S3/(S ∩ J)| = |S4/(S ∩ J)| = 8. Since F16 contains no subring of order 8, both S3 and S4 are −1 contained in f (F4), since F4 is the unique subring of F16 of order 4. This contradicts the maximality of S3 and S4.

Case 2b: |R| = 256 and R/J = F4. The sizes of the subrings Si and S are as before. Since f(S)= S/(S ∩ J) injects into R/J and 1R ∈ S \ J, we have |S ∩ J|∈{8, 16}.

4 If S ∩ J has order 8, then f|S : S → R/J is onto. Since J = 0 we may write

3 |S ∩ J| = |(S ∩ J i)/(S ∩ J i+1)|. Yi=1 ONRINGSASUNIONSOFFOURSUBRINGS 19

i i+1 i i+1 Observe that (S ∩ J )/(S ∩ J ) is an F4-subspace of J /J since it is stable under i i+1 multiplication by S/(S ∩ J) = f(S) = F4. So |(S ∩ J )/(S ∩ J )| is a power of 4, but their product equals |S ∩ J| = 8.

−1 If S ∩ J has order 16, then f(S) = F2. Since f (F2) is an index-2 subring of R, it cannot contain the index-4 maximal subrings S3 or S4. So f(S3) = f(S4) = R/J.

Thus S3 ∩ J and S4 ∩ J have order 16, as kernels of the maps f|Si : Si → R/J = F4. But then the inclusions S ∩ J ⊂ Si ∩ J and the assumption |S ∩ J| = 16 imply that S ∩ J = Si ∩ J for i =3, 4. Now apply Lemma 7.13.

By the previous two propositions, R is indecomposable and nonlocal. Following the ideas in [Cor I], there exist orthogonal idempotents e1, e2,...,em ∈ R such that m ei = 1, Ri := eiRei is a full matrix ring over a local ring, and Mij := eiRej iP=1 is an (Ri, Rj)-bimodule contained in the radical J of R. Let M = Mij. Then i=6 j m L J = J(R1) ⊕···⊕ J(Rm) ⊕ M. In particular, if R/J = Ri/J(Ri) has covering iL=1 number 3 or 4, then R 6∈ S(4). Since |R| ≤ 256, the only way some Ri could be nonlocal is if Ri = M2(T ), where T is a local ring of order 2 or 4. If T surjects onto F2 then Ri/J(Ri)= M2(F2) which has covering number 4, and R 6∈ S(4). Otherwise, Ri = R = M2(F4), as this has order 256. But σ(M2(F4)) > 4 by [Wer18]. Thus the subrings Ri are all local rings. We now study some constraints on M and R, and then proceed to eliminate each possible value of |R|.

m Lemma 7.15. Write R = Ri ⊕ M as above. The following hold. iL=1 a) |Mij| ≥ 4 for all 1 ≤ i =6 j ≤ m. In particular, |M| ≥ 4. b) 8 ≤ [R : M] ≤ 64. c) If [R : M]=8 then m =2, {R1, R2} = {F2, F4}, and |R| 6∈ {64, 256}. 2 d) If [R : M]=16 then m =2. If |R|∈{64, 256} then {R1, R2} = {F2[x]/(x ), F4}. e) If [R : M]=32 or 64, then m =2.

Proof. a) Since R is indecomposable and nonlocal, we have m ≥ 2 and M =6 0. If some Mij has order 2 then the homomorphisms Ri → EndZ(Mij) = F2 and Rj → EndZ(Mij)= F2 coming from the bimodule structure would force Ri and Rj to have residue field F2, from which we get a surjection R → R/J → F2 × F2, and σ(R) = 3. b) If [R : M] = 4 then m = 2 and R1 = R2 = F2, so we get a surjection R → R/J → F2 × F2, and σ(R) = 3. Thus [R : M] ≥ 8. The inequality [R : M] ≤ 64 is clear since |M| ≥ 4 and |R| ≤ 256. 20 JON COHEN c) Suppose [R : M]=8. If m = 3 then R1 = R2 = R3 = F2, which forces σ(R) = 3. If m = 2 then up to indexing, R1 = F2 and R2 is a local ring of order 4. If R2 has residue field F2 then once again σ(R) = 3, so R2 = F4. Thus |M| is a power of 4, which together with [R : M] = 8 rules out |R|∈{64, 256}. d) Suppose [R : M] = 16. Then m = 2 since m ≥ 3 would imply at least two Ri being F2. If, say, R1 = F2 then R2 is a local ring of order 8 that does not surject onto F2, so R2 = F8. This forces |M| to be a power of 8, which rules out |R|∈{64, 256}. If neither R1 nor R2 is F2, then both have order 4, and since R does not surject onto 2 F2 × F2 or F4 × F4, up to indexing we have R1 = F4 and R2 ∈{Z4, F2[x]/(x )}. Since 2 R has characteristic 2, R2 = F2[x]/(x ). e) Suppose [R : M] = 32. Note this cannot happen if |R| = 64. If m ≥ 4 then at least 2 of the Ri equal F2, and σ(R)=3. If m = 3, then at least one Ri = F2. If more than one is F2 then again σ(R) = 3. Otherwise, either the other two local rings Ri both equal F4, or at least one of them surjects onto F2. In either case σ(R/J) ∈{3, 4} and R 6∈ S(4). So m = 2. Finally, suppose [R : M] = 64, which is only possible if |R| = 256. It is easy to see that if m ≥ 4 then either two Ri have residue field F2 or equal F4, and both of these cases force R 6∈ S(4). If m = 3, and no two of R1, R2, and R3 have common residue field F2 or F4, then R1 = F2, R2 = F4, and R3 = F8, up to indexing. But |M| =4 so M is not vector space over F8.

In particular, we see that m = 2 in all cases. The following lemma will rule out several cases below. Note that MijMji ⊂ J(Ri), so in general M is not a 2-sided ideal of R, but it is if M 2 = 0.

Lemma 7.16. The following cannot occur: R = Ri ⊕Rj ⊕M, with Ri and Rj having 2 residue ﬁelds F2 and F4, respectively, and M = J(Rj)M = MJ(Rj )=0.

Proof. Let f : R → R/J = F2 × F4 be the quotient map. Suppose T is a subring of R with [R : T ] = 2 and 1R ∈ T . Since [f(R): f(T )] ≤ [R : T ] = 2, and f(1R)=(1, 1) ∈ f(T ), either f(T )= F2 × F2 or f(T )= F2 × F4.

−1 If f(T ) = F2 × F2 then T ⊂ f (F2 × F2) ( R, and maximality of T implies −1 T = f (F2 × F2). In particular, T ⊃ J ⊃ M.

Suppose f(T ) = F2 × F4. Let t ∈ T such that f(t) = (0,c) ∈ F2 × F4 where c 6∈ F2. Since [Rj : T ∩ Rj] ≤ 2 and [Rj : J(Rj)] > 2, we have T ∩ Rj 6⊂ J(Rj). × Therefore T intersects the unit group Rj = Rj \ J(Rj). Finiteness then implies T ∋ ej, the idempotent of R giving the unit in Rj = ejRej. Then ejtej ∈ T ∩ Rj and f(ejtej) = (0,c). It follows that T ∩ M is an F4-subspace of M, since T contains a set of representatives for Rj/J(Rj) = F4. But [M : T ∩ M] ≤ 2, so we must have T ⊃ M. ONRINGSASUNIONSOFFOURSUBRINGS 21

Now, 1R ∈ S1 ∩ S2 and [R : S1] = [R : S2] = 2, so the argument just given shows that S1 ∩ S2 ⊃ M. Since [S1 ∩ S2 : S] = 2, we have [M : S ∩ M] ≤ 2. We cannot have S ⊃ M, because the assumption M 2 = 0 means M is a 2-sided ideal of R. Since S = S1 ∩ S2 ∩ S3, we have S3 6⊃ M and 2 ≤ [M : S3 ∩ M] ≤ [M : S ∩ M] ≤ 2. The inclusion S ∩M ⊂ S3 ∩M now forces S ∩M = S3 ∩M, and similarly S ∩M = S4 ∩M. Now apply Lemma 7.13. Proposition 7.17. |R|= 6 64.

Proof. Suppose R has order 64. From Lemma 7.15 and its proof, [R : M] = 16, 2 m = 2, {R1, R2} = {F4, F2[x]/(x )} and M = M12 has order 4. Because the ei are 2 2 orthogonal idempotents, M = M12 = 0. Now apply Lemma 7.16. Proposition 7.18. |R|= 6 128.

Proof. Suppose R has order 128. By Lemma 7.15 and its proof, there are three possible values for [R : M], all of which have m = 2.

Case 1: [R : M] = 8, so |M| = 16. Lemma 7.15 shows {R1, R2} = {F2, F4}. In 2 particular, J = M is a 2-dimensional vector space over F4. Then M = M12M21 ⊕ M21M12 ⊂ J(R1) ⊕ J(R2) = 0. Now apply Lemma 7.16.

Case 2: [R : M] = 16, so |M| = 8. If R1 =6 F2 and R2 =6 F2, then {R1, R2} = 2 {F4, F2[x]/(x )}. But M has order 8 so is not an F4-vector space. So up to indexing, R1 = F2 and R2 = F8. Thus M = J is a 1-dimensional F8-vector space. As R2 = F8 has no index-2 subrings and [R2 : S1 ∩ R2] ≤ 2, we have S1 ⊃ R2. Since |S1||J| > |R| there exists 0 =6 j ∈ S1 ∩ J. Then R2 and j generate the 1-dimensional F8-vector space J, so S1 ⊃ J. Therefore S1/J ⊂ R/J = F2 × F8 has order 8 and contains (1, 1). This is impossible.

Case 3: [R : M] = 32, so |M| = 4. By Lemma 7.15 either M = M12 or M = M21. In particular, M 2 = 0.

If R1 = F2, then |R2| = 16. Since |M| < 16, R2 =6 F16, and R2 does not have residue ﬁeld F2 since that would give a surjection R → F2 × F2. So R2 has residue ﬁeld F4. Suppose M = M12. By Nakayama’s lemma, M/MJ(R2) is a nonzero (right) module over R2/J(R2)= F4, so has order at least 4. Thus MJ(R2) = 0. Similarly, if R2M =6 0 then J(R2)M = 0. Now apply Lemma 7.16. The case M = M21 is similar.

If R1 =6 F2, we can assume R1 has order 4 and R2 has order 8. Since |M| < 8, we conclude that R2 has residue ﬁeld F2, which forces R1 = F4. Since MJ(R2) is an F4-subspace of M and MJ(R2) =6 M by Nakayama’s Lemma, we have MJ(R2) = 0. Similarly J(R2)M = 0, so apply Lemma 7.16. Proposition 7.19. |R|= 6 256.

Proof. Suppose R has order 256. From Lemma 7.15 and its proof, there are three possible values for [R : M], all of which have m = 2. 22 JON COHEN

2 Case 1: [R : M] = 16, so |M| = 16. By Lemma 7.15, R1 = F4 and R2 = F2[x]/(x ), up to indexing.

2 2 Suppose M =6 0. Then M = M12M21 ⊕M21M12 = J(R2) has order 2. It follows that any subring of R containing M also contains J = J(R2) ⊕ M. The proof of Lemma 7.16 shows that if T is an index-2 subring of R and T ∋ 1R then T ⊃ M. We note 2 that this did not assume M = 0. Therefore S1 ⊃ J and S2 ⊃ J. But then the image of S1 and S2 in R/J = F2 × F4 is F2 × F2, which means S1 ∪ S2 is contained in the 2 preimage of F2 × F2, a contradiction. Therefore M = 0 and Lemma 7.16 applies.

Case 2: [R : M] = 32, so |M| = 8. By Lemma 7.15, M has only one summand Mij of order 8, else M would have a summand of order 2. In particular, M 2 = 0.

If R1 = F2 then R2 is a local ring of order 16 and with residue ﬁeld not equal to F2. Since |M| < 16, the residue ﬁeld of R2 must be F4. We have |J(R2)| = 4 and 2 J(R2) = 0. Assume M = M12; the case M = M21 is similar and corresponds to the opposite ring. By Nakayama’s Lemma, M/MJ(R2) is a nonzero right module over R2/J(R2) = F4. Since M has order 8, this prevents MJ(R2) from being zero. So 2 MJ(R2) = MJ(R2)/MJ(R2) is also a nonzero right module over R2/J(R2). But then both M/MJ(R2) and MJ(R2) have orders divisible by 4, and with product 8.

If R1 =6 F2 then, up to indexing, R1 has order 4 and R2 order 8. Since |M| is not a power of 4, we cannot have R1 = F4. So R1 must have residue field F2, and since 2 R has characteristic 2, this forces R1 = F2[x]/(x ). The local ring R2 cannot also have residue field F2, so R2 = F8. We claim that there is exactly one subring of R of index 2 that contains 1R. This will contradict the existence of S1 and S2. Towards that end, let T be an index-2 subring of R containing 1R. Then [R2 : T ∩ R2] ≤ 2, which forces T ⊃ R2 since R2 = F8 has no subrings of index 2. Similarly, there exists 0 =6 m ∈ T ∩ M since [M : T ∩ M] ≤ 2. But M is a 1-dimensional vector space over R2 = F8 ⊂ T , so T ⊃ M. Since 1R = e1 + e2 ∈ T and e2 ∈ T , we have T ⊃ F2 ⊕R2 ⊕M. The right side being an additive subset of R of index 2, we conclude T = F2 ⊕ R2 ⊕ M. Case 3: [R : M] = 64, so |M| = 4. By Lemma 7.15, M has only one summand, 2 so M = 0. If R1 = F2, then R2 is a local ring of order 32. Such a ring either has residue field F2, in which case σ(R) = 3, or else is equal to F32, which is impossible since |M| < 32. If R1 and R2 have order 8, then either they both surject onto F2 and σ(R) = 3, or else at least one of them is F8. But this is again impossible since |M| < 8. Thus R1 has order 4 and R2 order 16. The residue field of R2 cannot be F16 because |M| =4. If R1 and R2 both have the same residue field F2 or F4, then R 6∈ S(4). So one has residue field F2 and the other F4. Since R has characteristic 2, 2 the two choices for R1 are F4 and F2[x]/(x ).

If R1 = F4 and R2 has order 16 with residue ﬁeld F2, Lemma 7.16 applies.

2 Suppose R1 = F2[x]/(x ) and R2 has order 16 and residue ﬁeld F4. If M = M12, then M/MJ(R2) is nonzero by Nakayama’s Lemma and is a module for R2/J(R2) = F4. ONRINGSASUNIONSOFFOURSUBRINGS 23

This and |M| = 4 forces MJ(R2) = 0. The case M = M21 is similar. Now apply Lemma 7.16.

This completes the proof of part b) of Theorem 3.11.

8. Further directions

This paper established a strategy to classify unital rings of covering number n that requires a finite amount of computation. It is possible that S(n) is still finite if non- unital rings are permitted. This is known if n = 3; see [Maro12]. Perhaps the more general case requires no new ideas from those presented here. However, the actual computation of S(n) would likely require different techniques.

Propositions 5.1 and 5.3 applied only to local rings with residue field Fp. Having a similar criterion for local rings with arbitrary residue field would help significantly in classifying which commutative rings are coverable, and in any attempt to compute S(n) for larger n. For example, S(5) contains both a noncommutative and commutative local ring with residue field F4. It is unknown whether S(n) is nonempty for all n ≥ 3. Equivalently, it is not known if every n ≥ 3 is the covering number of some ring. If true, this would be in stark contrast to the corresponding situation for groups: it is shown in [Wer18] that σ(M2(F3)) = 7, but Theorem 3.6. of [Tom97] states that 7 is not the covering number of any group. As a more refined question, one may ask if, given n ≥ 3 such that S(n) is nonempty, and given a prime p < n, there is always a ring in S(n) of characteristic p. If R ∈ S(n) for some n ≥ 3, then there exists what we have called a good tuple

(R,S1,...,Sn) with R = i Si. It would be interesting to determine how unique this tuple is for given R; for SR ∈ S(3) and R ∈ S(4) there are exactly σ(R) maximal subrings. One can also ask if there are any common features of such covers. For n example, if R =6 F2 × F2 and R ∈ S(n), then must we always have 1R ∈ Si? iT=1 It is clear that if R is coverable and all its proper quotients are not, then R ∈ S(σ(R)). Based on the known examples, one can speculate whether the converse holds: if R ∈ S(σ(R)), does one always have σ(R/I) = ∞ for every proper ideal I? Such a theorem would, if true, imply that a noncommutative ring in any S(n) can only have commutative quotients. Following the example of groups, one might consider an expanded notion of coverings in which cosets of subrings are considered. Alternatively, one might restrict the notion of coverings to permit only special classes of subrings. Does doing so aﬀect the existence or basic properties of the corresponding set S(n)? 24 JON COHEN

References

[Ant82] Antipkin, V. G., & Elizarov, V. P. (1982). Rings of order p3. Siberian Mathematical Journal, 23(4), 457-464. [Bhar09] Bhargava, Mira. Groups as unions of proper subgroups. Amer. Math. Monthly 116 (2009), no. 5, 413–422. [Cohn94] Cohn, J. H. E. On n-sum groups. Math. Scand. 75 (1994), no. 1, 44–58. [Cor I] Corbas, B.; Williams, G. D. Rings of order p5. I. Nonlocal rings. J. Algebra 231 (2000), no. 2, 677–690 [Cor II] Corbas, B.; Williams, G. D. Rings of order p5. II. Local rings. J. Algebra 231 (2000), no. 2, 691–704. [Derr94] Derr, J. B.; Orr, G. F.; Peck, Paul S. Noncommutative rings of order p4. J. Pure Appl. Algebra 97 (1994), no. 2, 109116. [Eld68] Eldridge, K. E. Orders for finite noncommutative rings with unity. Amer. Math. Monthly 75 (1968), 512–514. [Lew67] Lewin, Jacques. Subrings of finite index in finitely generated rings. J. Algebra 5 (1967), 84–88. [Maro12] Lucchini, Andrea; Maróti, Attila. Rings as the unions of proper subrings. Algebr. Repre- sent. Theory 15 (2012), no. 6, 1035–1047. [Sco26] Scorza, Gaetano. ”I gruppi che possono pensarsi come somma di tre loro sottogruppi.” Boll. Un. Mat. Ital 5.216-218 (1926): 1. [Tom97] Tomkinson, M. J. Groups as the union of proper subgroups. Math. Scand. 81 (1997), no. 2, 191–198. [Wer15] Werner, Nicholas J. Covering numbers of finite rings. Amer. Math. Monthly 122 (2015), no. 6, 552–566. [Wer18] Peruginelli, G.; Werner, N. J. Maximal subrings and covering numbers of finite semisimple rings. Comm. Algebra 46 (2018), no. 11, 4724–4738. [Wer19] Cai, Merrick; Werner, Nicholas J. Covering numbers of upper triangular matrix rings over finite fields. Involve 12 (2019), no. 6, 1005–1013.