RELATIVE COMMUTATOR THEORY IN SEMI-ABELIAN CATEGORIES TOMASEVERAERTANDTIMVANDERLINDEN Abstract. Basingourselvesontheconceptofdoublecentralextensionfrom categorical Galois theory, we study a notion of commutator which is defined relativetoaBirkhoffsubcategoryB ofasemi-abeliancategoryA. Thiscom- mutator characterises Janelidze and Kelly’s B-central extensions; when the subcategory B is determined by the abelian objects in A, it coincides with Huq’scommutator;andwhenthecategoryAisavarietyofΩ-groups,itcoin- cideswiththerelativecommutatorintroducedbythefirstauthor. 1. Introduction The aim of this article is to fill in the question mark in the diagram (cid:111)(cid:111)(cid:111)(cid:111)(cid:111)(cid:111)(cid:111) ? (cid:79)(cid:79)(cid:79)(cid:79)(cid:79)(cid:79)(cid:79)(cid:79)(cid:79) Janelidze & Kelly Huq Everaert (cid:111)(cid:111)(cid:111)(cid:111)(cid:111) (cid:79)(cid:79)(cid:79)(cid:79)(cid:79) Froehlich Higgins whichrelatesseveralnon-equivalentconceptsofcommuting normal subobjects,here named after the authors who introduced them. This diagram is meant to be read in the following manner. ThebottomtrianglerestrictsitselftotheorieswhichmakesenseforvarietiesofΩ- groups,whilethetoptriangleextendsthosetheoriestoacategoricalcontext. Inthe lefthandsidecolumnwehavetheorieswhichareone-dimensional andrelative; the theories in the right hand side column, however, are two-dimensional and absolute, while the ones in the middle column are two-dimensional and relative. So we are looking for a categorical commutator theory which is both relative and two- dimensional. Let us explain in more detail what this means for us. 1.1. Thebottomtriangle. RecallthatavarietyofΩ-groups[26]isavarietyin thesenseofuniversalalgebrawhichispointed(i.e.,ithasexactlyoneconstant)and hasamongstitsoperationsandidentitiesthoseofthevarietyofgroups. Apartfrom groups, the examples include the varieties of abelian groups, of non-unital rings, of commutative algebras, of modules and of Lie algebras, and also the categories of Date:7thOctober2010. 2010 Mathematics Subject Classification. 18E10,17D10,20J. Keywordsandphrases. CategoricalGaloistheory,semi-abeliancategory,commutator,double centralextension. The first author was supported by Fonds voor Wetenschappelijk Onderzoek (FWO-Vlaande- ren). The second author was supported by Centro de Matema´tica da Universidade de Coimbra, byFunda¸ca˜oparaaCiˆenciaeaTecnologia(grantnumberSFRH/BPD/38797/2007)andbyFonds delaRechercheScientifique–FNRS. 1 2 TOMASEVERAERTANDT.VANDERLINDEN crossed modules and of precrossed modules are known (essentially from [34]) to be equivalent to varieties of Ω-groups. In this context there are two classical approaches to commutator theory. On the one hand, there is the Higgins commutator of normal subobjects [26] which has as particular cases the ordinary commutators of groups, rings, etc. It is two- dimensional in the sense that any two normal subobjects (i.e., ideals or kernels) N and M of an object A in a variety of Ω-groups A have a commutator [N,M]Ω, namely, the normal subobject of the join M ∨N = M ·N of M and N generated by the set {w(mn)w(n)−1w(m)−1 |w is a term, m∈M and n∈N}. Call an object A of A abelian when it can be endowed with the structure of an internal abelian group (necessarily in a unique way). The subcategory of A determined by the abelian objects is denoted by AbA. It is well known (and easily verified) that when A is a variety of Ω-groups, an algebra A is in AbA precisely whentheproductmapA×A→A(sendingapairofelements(a,a(cid:48))toitsproduct aa(cid:48)) is a homomorphism in the variety. From this it follows immediately that the Higgins commutator characterises the abelian objects: A is abelian if and only if [A,A]Ω =0. On the other hand there is the relative notion of central extension due to Fr¨ohlich [23] (see also Lue [35] and Furtado-Coelho [24]). This notion of central extension corresponds to a one-dimensional commutator. Here one starts from a variety of Ω-groups A together with a chosen subvariety B of A. The subvariety B is completely determined by a set of identities of terms of the form w(x)=1; the set of all corresponding terms w(x) is denoted by W ={w(x)|w(b)=1,∀B ∈B,∀b∈B}, B and an object A of A belongs to B if and only if w(a) = 1 for all w ∈ W and B all a∈A. An extension f: A→B in A is a regular epimorphism, i.e., a surjective ho- momorphism. Let K denote the kernel of f. The normal subobject [K,A]Ω of A B generated by the set {w(ka)w(a)−1 |w ∈W , k∈K and a∈A} B is called the relative commutator (with respect to B) of K and A. (Note that Fr¨ohlich uses the notation V for the relative commutator.) The extension f 1 is central (with respect to B) when [K,A]Ω is zero. It is easily seen that this B relative commutator characterises objects of B as follows: A belongs to B if and only if [A,A]Ω is zero. B In the absolute case when the subvariety B consists of all abelian objects in A, it was shown in [24] that the two commutators coincide, [K,A]Ω =[K,A]Ω. AbA (NoteherethatK∨A=A.) Themainadvantageoftherelativeapproachisthatone may consider many situations which are not covered by the Higgins commutator. For instance, the notion of central extension of precrossed modules relative to the subvariety of crossed modules is of this type. The main advantage of the Higgins commutator is that it is two-dimensional. So the Higgins commutator is two- dimensionalandabsolute,theFr¨ohlichcommutatorisone-dimensionalandrelative, and in the one-dimensional absolute case the two commutators coincide. What about the two-dimensional relative case? Inhisarticle[15]thefirstauthorofthepresentarticleaimsatansweringprecisely this question. He introduces a two-dimensional relative commutator for varieties RELATIVE COMMUTATOR THEORY IN SEMI-ABELIAN CATEGORIES 3 of Ω-groups which restricts to the Higgins commutator in the absolute case and whichcharacterisesFr¨ohlich’srelativecentralextensions. Givenanypairofnormal subobjects M and N of an object A of A, the commutator [M,N] is the normal B subobject of M ∨N generated by the set {w(mn)w(n)−1w(m)−1w(p)|w ∈W ,m∈M,n∈N,p∈M ∧N}. B The examples give an indication of how good his definition is. For instance, when consideringthevarietyofprecrossedmodulestogetherwiththesubvarietyofcrossed modules, the relative commutator obtained is the so-called Peiffer commutator, which is exactly what one would expect. 1.2. The left hand side column. Basing themselves on ideas from categorical Galois theory [28, 4], in the article [31] Janelidze and Kelly introduce a general notion of central extension, relative with respect to a Birkhoff subcategory B of a (Barr)exactcategoryA. Thisnotionofrelativecentralextensionisageneralisation of Fr¨ohlich’s definition. In what follows, we shall restrict ourselves to the case of semi-abelian cat- egories [32]: pointed, exact and protomodular with binary sums. So let A be a semi-abelian category and B a Birkhoff subcategory of A—full, reflective and closed under subobjects and regular quotients; a Birkhoff subcategory of a variety is nothing but a subvariety. Let I: A→B denote the reflector, and η: 1 ⇒I the A unit of the adjunction. Recall from [31] that the closure of B under subobjects and regular quotients is equivalent to the condition that the commutative square f (cid:44)(cid:50) A B ηA ηB (A) (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) IA IB If is a pushout of regular epimorphisms, for any regular epi f: A→B. An extension in A is a regular epimorphism. Such an extension f: A→B is called trivial (with respect to B) when the induced commutative square (A) is a pullback. f is central (with respect to B) when it is locally trivial in the sense that there exists a regular epimorphism p: E → B such that the pullback p∗(f): E × A → E of f along p is a trivial extension. Since, in the present B context, this implies that f∗(f) is trivial, we have that f is central if and only if it is normal: either one of the projections in the kernel pair (R[f],f ,f ) of f is 0 1 a trivial extension. It is explained in the article [31] why these central extensions reduce to Fr¨ohlich’s when the category A is a variety of Ω-groups. This notion of relative central extension induces a one-dimensional relative commutator as follows [20, 19]. Let [−] : A→A denote the radical induced B by B: the functor which maps an object A of A to the object [A] defined through B the short exact sequence 0 (cid:44)(cid:50)[A]B µA (cid:44)(cid:50)A ηA (cid:44)(cid:50)IA (cid:44)(cid:50)0, and a morphism a: A(cid:48) →A to its (co)restriction [a] : [A(cid:48)] →[A] . Let again B B B f: A→B be an extension and let K be its kernel. By protomodularity, f is B- central if and only if for the kernel pair (R[f],f ,f ) of f, the (co)restrictions 0 1 [f ] ,[f ] : [R[f]] →[A] 0 B 1 B B B ofthetwoprojectionsareisomorphisms(see[11]). Hencethekernel[K,A] of[f ] B 0 B measures how far f is from being central: f is B-central if and only if [K,A] is B zero. 4 TOMASEVERAERTANDT.VANDERLINDEN The object [K,A] may be considered as a normal subobject of A via the com- B posite µ ◦[f ] ◦ker[f ] : [K,A] →A; A 1 B 0 B B the induced extension A/[K,A] →B is the B-centralisation of f. We interpret B [K,A] as a commutator of K with A, relative to the Birkhoff subcategory B of A. B When A is a variety of Ω-groups, [K,A] coincides with the relative commutator B [K,A]Ω, because they induce the same central extensions. And as in the varietal B case,anobjectAofAbelongstoBifandonlyif[A,A] =0,becausetheextension B A→0 is a split epimorphism, and therefore central if and only if it is trivial [31]. 1.3. The right hand side column. In his article [27], Huq introduces a catego- rical notion of commutator of coterminal morphisms which makes sense in quite diverse algebraic settings. Using “old-style” axioms, he formulates his results for those categories we would nowadays call semi-abelian [32]. Recast in more modern terminologybyBourn,hisdefinitiontakesthefollowingshape[9]. Inasemi-abelian category, consider two coterminal morphisms, m: M →A and n: N →A, and the resulting square of solid arrows M(cid:63) (cid:104)1M,(cid:122)(cid:4)(cid:127)0(cid:127)(cid:105)(cid:127)(cid:127)(cid:127)(cid:127)(cid:127) (cid:12)(cid:18) (cid:63)(cid:63)(cid:63)(cid:63)m(cid:63)(cid:63)(cid:26)(cid:36) (cid:44)(cid:50) (cid:108)(cid:114) M(cid:104)0×,1N(cid:90)(cid:100)(cid:63)N(cid:63)(cid:105)(cid:63)(cid:63)(cid:63)(cid:63)(cid:63)Q(cid:76)(cid:82) (cid:127)(cid:127)(cid:127)q(cid:127)(cid:127)n(cid:127)(cid:127)(cid:58)(cid:68)A. N The colimit of this square consists of an object Q together with four morphisms with codomain Q as indicated in the diagram. The morphism q turns out to be a normal epimorphism; its kernel is denoted [m,n]H: [M,N]H →A and called the Huq commutator of m and n. It is convenient for us to restrict its use to the situation when M and N are normal subobjects of A, i.e., m and n arekernels. Thecommutator[M,N]H becomestheordinarycommutatorofnormal subgroups M and N in the case of groups, the ideal generated by MN +NM in the case of rings, the Lie bracket in the case of Lie algebras, and so on. More generally,whencomputedinthejoinM∨N,weknowfrom[27]thatinanyvariety of Ω-groups the Huq commutator [M,N]H coincides with the Higgins commutator [M,N]Ω. Just as the Higgins commutator, the Huq commutator characterises the Birkhoff subcategory AbA of A of abelian objects in A. This is a consequence of thefactthat,inasemi-abeliancategoryA,anobjectAadmitsatmostoneinternal abeliangroupstructure,andsuchastructureisentirelydeterminedbyamorphism m: A×A→A which satisfies m◦(cid:104)1 ,0(cid:105)=1 =m◦(cid:104)0,1 (cid:105) [27, 8]. A A A 1.4. The question mark. By now it is clear, we hope, that the purpose of the present article is to introduce a categorical version of the relative commutator for varieties of Ω-groups, in such a way that (1) it characterises the B-central extensions of A, (2) it coincides with the Huq commutator when B is AbA. In [21] the present authors already introduced a relative concept of commuting normal subobjects, based on categorical Galois theory and valid in the context of semi-abelian categories. This notion was shown to be compatible with the relative commutator for varieties of Ω-groups. What we still have to do now is · explain how this induces a two-dimensional commutator; RELATIVE COMMUTATOR THEORY IN SEMI-ABELIAN CATEGORIES 5 · prove that this commutator satisfies (1) and (2) above; · explore the commutator’s basic properties. One may ask whether it is worth the effort at all to leave the context of Ω-groups and study a relative commutator from a categorical perspective. We claim that the categorical approach not only provides us with a conceptual explanation of the definitions (in terms of Galois theory) but also with interesting new examples. For instance,inthecaseofloopsvs.groupsconsideredin[21],thecommutatorbecomes an associator, and it effectively measures how well two normal subloops of a loop associate with each other. 1.5. Definition of the commutator. Let us now briefly sketch how the relative commutator [−,−] is defined. Let A again be a semi-abelian category and B a B Birkhoff subcategory of A. M and N will be normal subobjects of an object A of A. R and R are the equivalence relations on the join M ∨N (taken in the M N lattice of normal subobjects of A) corresponding to M and N, and R (cid:3)R r1 (cid:44)(cid:50)(cid:44)(cid:50)R M N N r0 p0 (cid:12)(cid:18) (cid:12)(cid:18)p1 (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) R (cid:44)(cid:50)M ∨N M is the largest double equivalence relation on R and R : the object R (cid:3)R M N M N “consists of” all quadruples (x,y,z,t) ∈ M ∨ N where (x,z), (y,t) ∈ R and M (x,y), (z,t)∈R . N The commutator of M and N is the meet [M,N] =K[[p ] ]∧K[[r ] ] B 0 B 0 B of the kernels of the morphisms [p ] and [r ] in the following diagram, obtained 0 B 0 B by applying the functor [−] to the diagram above. B [R (cid:3)R ] [r1]B(cid:44)(cid:50)(cid:44)(cid:50)[R ] M N B N B [r0]B (B) [p0]B (cid:12)(cid:18) (cid:12)(cid:18)[p1]B (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) [R ] (cid:44)(cid:50)[M ∨N] M B B It may be considered as a normal subobject of M ∨N. 1.6. Interpretation in terms of double central extensions. We have to ex- plain why [M,N] is defined the way it is. The reason comes from categorical B Galois theory, in particular the theory of higher central extensions. Just like the concept of central extension which is defined with respect to the adjunction I (cid:44)(cid:50) A(cid:108)(cid:114) ⊥ B, (C) ⊃ one may consider double central extensions which are defined with respect to the reflection of extensions to central extensions—the adjunction I1 (cid:44)(cid:50) ExtA(cid:108)(cid:114) ⊥ CExtBA (D) ⊃ where ExtA is the category of extensions and commutative squares between them, and CExt A its full subcategory determined by those extensions which are central. B The reflector I takes an extension f: A→B with kernel K and maps it to the 1 central extension I f: A/[K,A] →B. 1 B 6 TOMASEVERAERTANDT.VANDERLINDEN This may be repeated ad infinitum, so that notions of n-fold central extension are obtained, but for the present purposes the second step is sufficient. Double central extensions, first introduced by Janelidze for groups [29], are an important tool in semi-abelian(co)homology[19,30,39],andturnouttobepreciselywhatisneeded to understand how the relative commutator works. We refer the reader to the articles [19, 16] for more details on higher central extensions. As we explain below, the commutator [M,N] is zero if and only if any (hence, B all)ofthefourcommutativesquaresinthediagram(B)isapullback. Galoistheory shows that this condition is equivalent to the square M ∨N qM (cid:44)(cid:50) M∨N M qN (cid:12)(cid:18) (cid:12)(cid:18) (E) (cid:44)(cid:50) M∨N 0 N being a double central extension. (Here q denotes the cokernel of the normal M monomorphismM →M∨N.) Whenthishappens,wesaythatM andN commute (with respect to B). Accordingly, given any two normal subobjects M and N of an object A, the commutator [M,N] is the smallest normal subobject J of M ∨N such that M/J B andN/J commute; itisthenormalsubobjectwhichmustbedividedoutofM∨N to turn the double extension (E) into a double central extension. 1.7. Structure of the text. In the following sections we shall explain why the commutator has the properties (1) and (2) mentioned above. With this purpose in mind, thetextisstructuredasfollows. InSection2weprovidethenecessaryback- groundforunderstandingthedefinitionofthecommutator: semi-abeliancategories, normalsubobjects,doubleextensionsanddoublecentralextensions. Itsbasictech- nical properties and the proof of (1) are given in Section 3. In Section 4 we prove (2): thecommutator[−,−] coincideswiththeHuqcommutatorincaseB isAbA. B Finally, Section 5 brings up some further remarks and unanswered questions. 2. Preliminaries We recall some basic definitions and results which we shall need in the following sections. 2.1. Semi-abelian categories. Acategoryisregularwhenitisfinitelycomplete withcoequalisersofkernelpairsandwithpullback-stableregularepimorphisms[1]. In a regular category, any morphism f may be factored as a regular epimorphism followed by a monomorphism (called the image of f), and this image factor- isation is unique up to isomorphism. Given a monomorphism m: M →A and a regular epimorphism f: A → B, the direct image f(m): fM →B of m along f is the image of the composite f ◦m. When a category is pointed and regular, protomodularity can be defined via the following property [5, 7]: given any commutative diagram kerf(cid:44)(cid:50)(cid:48) f(cid:48) (cid:44)(cid:50) K[f(cid:48)] A(cid:48) B(cid:48) k (cid:12)(cid:18) a(cid:12)(cid:18) (cid:12)(cid:18)b (F) (cid:44)(cid:50) (cid:44)(cid:50) K[f] A B kerf f such that f and f(cid:48) are regular epimorphisms, k is an isomorphism if and only if the right hand square b◦f(cid:48) = f ◦a is a pullback. (Here, we use the notation kerf: K[f]→A for the kernel of f.) A homological category is pointed, regular RELATIVE COMMUTATOR THEORY IN SEMI-ABELIAN CATEGORIES 7 and protomodular [3]. In such a category, a regular epimorphism is always the cokernel of its kernel, and there is the following notion of short exact sequence. A short exact sequence is any sequence k (cid:44)(cid:50) f (cid:44)(cid:50) K A B with k =kerf and f a regular epimorphism. We denote this situation by (cid:44)(cid:50) k (cid:44)(cid:50) f (cid:44)(cid:50) (cid:44)(cid:50) 0 K A B 0. The following property holds. Lemma 2.2. [7] Consider a morphism of short exact sequences such as (F) above. The left hand side square kerf ◦k = a◦kerf(cid:48) is a pullback if and only if b is a mono. (cid:3) A (Barr) exact category is regular and such that every internal equivalence relationisakernelpair[1]. Ahomologicalcategoryisexactifandonlyifthedirect image of a normal monomorphism along a regular epimorphism is again a normal monomorphism. A semi-abelian category is homological and exact with binary coproducts [32]. A regular pushout square is a commutative square c (cid:44)(cid:50) X C d g (G) (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) D Z f suchthatallitsmapsandthecomparisonmap(cid:104)d,c(cid:105): X →D× C tothepullback Z of f with g are regular epimorphisms. In a semi-abelian category, every pushout of a regular epimorphism along a regular epimorphism is a regular pushout [14], and the following dual to Lemma 2.2 holds: Lemma 2.3. [11] Given a morphism of short exact sequences such as (F) above with a and b regular epi, the right hand side square f ◦a = b◦f(cid:48) is a (regular) pushout if and only if k is a regular epimorphism. (cid:3) 2.4. Normal subobjects. A normal subobject N of an object A of a semi- abeliancategoryisasubobjectrepresentedbyanormalmonomorphismn: N →A. LetM andN betwonormalsubobjectsofAwithrepresentingnormalmonomorph- isms m and n. Taking into account Lemma 2.2 and the stability of normal mono- morphismsunderregularimages,wemayalwaysformthe3×3diagraminFigure1 (in which all rows and columns are short exact sequences). The meet M ∧N and the join M ∨N of the subobjects M and N are taken in the lattice of normal subobjects of A. We see that M ∧N is computed as the pullback (i) and M ∨N is obtained through the pushout (ii), as the kernel of the composite morphism A→A/(M ∨N). Of course, M ∧N coincides with the meet M ∩N in the lattice of(all)subobjectsofA. OnecouldalsocomputethejoinofM andN as(ordinary) subobjects of A by taking the image M ∪N of the morphism (cid:104)m(cid:105): M +N →A. n It is known [2, 27] that both constructions yield the same result. We shall give an alternative proof of this fact below, but first we prove a weaker property. Letusfixsomenotation: wewritej forthenormalmonomorphismrepresenting M ∨N, and m(cid:48): M →M ∨N and n(cid:48): N →M ∨N for the induced factorisations. Since m(cid:48) and n(cid:48) are normal monomorphisms, we may also consider the join of M and N as normal subobjects of M ∨N. We denote it by M (cid:103)N and write j(cid:48): M (cid:103)N →M ∨N for the representing normal monomorphism. Lemma 2.5. The two joins M ∨N and M (cid:103)N coincide: j(cid:48) is an isomorphism. 8 TOMASEVERAERTANDT.VANDERLINDEN 0 0 0 (cid:12)(cid:18) (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) 0 M ∧N N N 0 M∧N (cid:12)(cid:18) (i) (cid:12)(cid:18)n (cid:12)(cid:18) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) 0 M A A 0 m M (ii) (cid:12)(cid:18) (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) 0 M A A 0 M∧N N M∨N (cid:12)(cid:18) (cid:12)(cid:18) (cid:12)(cid:18) 0 0 0 Figure 1. The 3×3 diagram induced by M, N normal in A Proof. First of all note that the commutative square (cid:44)(cid:50) M ∨N M∨N M j (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) A A M is a pullback by protomodularity, so that the right hand vertical morphism is a monomorphism because, in a protomodular category, pullbacks reflect monos [5]. (One could, alternatively, use Lemma 2.2 to prove that this morphism is a mono- morphism.) Now, the normal monomorphisms m(cid:48) and n(cid:48) induce a 3×3 diagram similar to Figure 1, and j induces a morphism between the two 3×3 diagrams, of which we consider only the last row: (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) 0 N M∨N M∨N 0 M∧N M M(cid:103)N (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) (cid:44)(cid:50) 0 N A A 0 M∧N M M∨N We have just explained why the middle vertical morphism is a monomorphism. Hence,usingthesameargumentsasabove,wefindthatalsotherighthandvertical morphisms is a mono. Since the composite M ∨N →(M ∨N)/(M (cid:103)N)→A/(M ∨N) iszero,wefindthat(M∨N)/(M(cid:103)N)=0,i.e.,thefactorisationj(cid:48) isanisomorph- ism. (cid:3) Now, taking this lemma into account, when A = M ∨N in the 3×3 diagram above,theobjectA/(M∨N)iszero,andweregaintheNoetherisomorphisms[3] N M ∨N M M ∨N ∼= and ∼= . (H) M ∧N M M ∧N N We are ready to prove the identity M ∨N =M ∪N. Notation 2.6. Given a normal subobject J of M ∨N, the induced quotient of M ∨N is denoted q : M ∨N →(M ∨N)/J; J we write R for the kernel pair of q . J J RELATIVE COMMUTATOR THEORY IN SEMI-ABELIAN CATEGORIES 9 Proposition 2.7. [2, 27] If M and N are normal in A, then their join as normal subobjectsM∨N coincideswiththeirjoinassubobjectsM∪N. Hencethemorphism (cid:104)cokern,cokerm(cid:105): A→(A/N)×(A/M) is a regular epimorphism if and only if such is the morphism (cid:104)m(cid:105): M +N →A. n Proof. If J is a subobject of M ∨N containing M and N, then by Lemma 2.2 it induces a factorisation of the first of the isomorphisms (H) as a morphism N/(M ∧N)→J/M followed by a monomorphism j: J/M →(M ∨N)/M. This j is also a split epimorphism; hence it is an isomorphism, and J is equal to M ∨N. Now M ∪ N is a subobject of M ∨ N, because the composite of the sum (cid:104)m(cid:105): M +N →AwiththequotientA→A/(M ∨N)iszero. SinceM∪N contains n M and N, the two joins coincide. As to the latter statement, since in a semi-abelian category every pushout is a regular pushout, the first condition holds if and only if A/(M ∨N) is zero; by definition, the second one holds when A is M ∪N. (cid:3) A Givenamonomorphismm: M →A,thenormal closureM ofM inAalways exists,andiscomputedasthekernelofthecokernelofm. Itisthesmallestnormal subobject of A that contains M. 2.8. Double (central) extensions. A double extension is a regular pushout square (G). For instance, given any two normal subobjects M and N of an object AofA,theinducedpushoutsquare(E)isadoubleextension. Recallfrom[19]that pullbacks of double extensions exist in ExtA and are degree-wise pullbacks in A. Moreover,doubleextensionsarepullback-stable. Thecategoryofdoubleextensions in A and commutative cubes between them is denoted Ext2A. Double central extensions are defined with respect to the adjunction (D) in the same way as central extensions are defined with respect to the adjunction (C). More precisely, a double extension (G), considered as a map (c,f): d→g in the categoryExtA,istrivialwhenthelefthandcommutativesquarebelow,inducedby theunitoftheadjunction(D),isapullbackinExtA;thismeansthattherighthand commutative square, in which the vertical morphisms are the canonical quotient maps, is a pullback in A. (c,f) (cid:44)(cid:50) X c (cid:44)(cid:50)C d g (cid:12)(cid:18) (cid:12)(cid:18) (cid:12)(cid:18) (cid:44)(cid:50) (cid:12)(cid:18) X (cid:44)(cid:50) C I1d I1g [K[d],X]B [K[g],C]B The square (G) is a double central extension (with respect to B) when its pullback along some double extension is a trivial double extension. It is a double normal extension (with respect to B) when the first projection of its kernel pair R[c] c0 (cid:44)(cid:50)X r (cid:12)(cid:18) (cid:12)(cid:18)d (cid:44)(cid:50) R[f] D f0 is a trivial double extension. (Alternatively, one could use the square of second projections.) By protomodularity, this amounts to the (one-dimensional, relative) 10 TOMASEVERAERTANDT.VANDERLINDEN commutators [K[r],R[c]] and [K[d],X] being isomorphic. Similar to the one- B B dimensionalcase,doublecentralextensionsanddoublenormalextensionscoincide. 2.9. Higher extensions. Inwhatfollowsweshallalsoneedthree-foldextensions, so let us recall the definition of n-fold extension for arbitrary n. Given n ≥ 0, denote by ArrnA the category of n-dimensional arrows in A. (Zero-dimensional arrows—as well as zero-dimensional extensions—are just objects of A.) A (one- fold) extension is a regular epimorphism in A. For n ≥ 1, an (n + 1)-fold extension is a commutative square (G) in Arrn−1A (an arrow in ArrnA) such that allitsmapsandthecomparisonmap(cid:104)d,c(cid:105): X →D× C tothepullbackoff with Z g are n-fold extensions. Thus for n=2 we regain the notion of double extension. Basic results on higher-dimensional extensions and central extensions may be foundin[19]and[16]. Letusjustrecallherethat,foranyn≥0,asplitepimorphism of n-fold extensions is always an (n+1)-fold extension, and it is an (n+1)-fold central extension if and only if it is a trivial (n+1)-fold extension. Higher-dimensional central extensions are important in homology where they appear in the higher Hopf formulae, and in cohomology where (in the absolute case, and in low dimensions) they are classified by the cohomology groups [25, 39]. 3. Definition and basic properties Inthissectionwerecallthecategoricaldefinitionoftherelativecommutatorfrom the introduction and we explore its basic properties: compatibility with the cent- ral extensions introduced by Janelidze and Kelly (Proposition 3.2), basic stability properties (Theorem 3.9) and the case of Ω-groups (Proposition 3.10). Inwhatfollows, Awillbeasemi-abeliancategoryandB aBirkhoffsubcategory of A. Definition 3.1. Let M and N be normal subobjects of an object A of A. We say that M and N commute (with respect to B) when the double extension M ∨N qM (cid:44)(cid:50) M∨N M qN (cid:12)(cid:18) (cid:12)(cid:18) (I) (cid:44)(cid:50) M∨N 0 N is central (with respect to B). Is is immediately clear that this notion of commuting subobjects characterises the B-central extensions of A and the objects of B: Proposition 3.2. An extension f: A→B in A is B-central if and only if the object A and the kernel K of f commute. An object A of A lies in B if and only if A commutes with itself. Proof. The first result holds because the double extension A qA (cid:44)(cid:50)0 f=qK (cid:12)(cid:18) (cid:44)(cid:50) B 0, being a split epimorphism of extensions, is central if and only if it is trivial, which happens precisely when f is a central extension. The second result follows from the first, since A is in B if and only if the split epimorphism A→0 is a B-central extension. (cid:3)