Lawvere’s Frobenius reciprocity, the modular connections of Grandis and Dilworth’s abstract principal ideals Amartya Goswami, Zurab Janelidze, Graham Manuell Abstract The purpose of this short note is to fill a gap in the literature: Frobenius reciprocity in the theory of doctrines is closely related to modular connections in projective homological algebra and the notion of a principal element in abstract commutative ideal theory. These concepts are based on particular properties of Galois connections which play an important role also in the abstract study of group-like structures from the perspective of categorical/universal algebra; such role stems from a classical and basic result in group theory: the lattice isomorphism theorem. keywords Abstract ideal theory, adjunction, connection, Frobenius reciprocity, Grandis exact category, lattice, lattice isomorphism theorem, modular connection, modular lattice, noetherian form, principal element, principal mapping, protomodular category, quantale, range-closed mapping, residuated map 1991 Mathematics Subject Classification 06A06, 13A15, 03G30, 20J99, 18G50, 06C05, 18E13, 08B05 1. Introduction This paper concerns three fundamental concepts: Frobenius reciprocity from categorical logic, modular connections from non-abelian homological algebra and principal elements from abstract ideal theory. • Frobenius reciprocity is a condition relating images, preimages and intersection that makes existential quantification interact well with logical conjunction [ 15 ] . • Modular connections arise in an attempt to capture the homological behaviour of modules in lattice theoretical terms (see [ 11 , 12 ] and the references there). • Principal elements in a multiplicative lattice approximate the behaviour of principal ideals of a commutative ring (with identity) in the integral commutative quantale of all ideals [ 8 ] . We will explain how each of these concepts are instances of the same idea, and also point out links with closely related concepts in residuation theory [ 3 ] and various interlinked abstract approaches in the study of group-like structures [ 16 , 10 , 4 , 20 ] . 2. Connections Given posets P and Q , by a connection we mean a relation R\colon P\to Q such that a\leqslant b\;R\;c\leqslant d\quad\Rightarrow\quad a\;R\;d, for all a , b\in P and c , d\in Q . Note that a relation R\colon P\to Q is a connection if and only if the opposite relation R^{\mathsf{op}} is connection between dual posets, R^{\mathsf{op}}\colon Q^{\mathsf{op}}\to P^{\mathsf{op}} . Remark 2.1 . What we call connections in this paper are often referred to as “weakening relations” in, especially, computer science literature (see e.g., [ 14 ] , and the references there). When posets are viewed as categories, connections are \mathbf{2} -valued profunctors. We say that a connection R is a left adjoint connection when there is a map f_{R}\colon P\to Q (called a left adjoint ) such that f_{R}(x)\leqslant y\quad\Leftrightarrow\quad x\;R\;y, for all x\in P and y\in Q . Symmetrically, we say that a connection is a right adjoint connection when there is a map f^{R}\colon Q\to P (called a right adjoint ) such that x\;R\;y\quad\Leftrightarrow\quad x\leqslant f^{R}(y). Note that the concepts of left and right adjoint maps are dual to each other: left/right adjoint for R is the same as a right/left adjoint for R^{\mathsf{op}} . Piecing the conditions defining left/right adjoint maps together, we get: f_{R}(x)\leqslant y\quad\Leftrightarrow\quad x\;R\;y\quad\Leftrightarrow\quad x\leqslant f^{R}(y). We call a connection R admitting both a left adjoint and a right adjoint, an adjoint connection . Of course, every adjoint connection R gives rise to a Galois connection (f_{R},f^{R}) , since the middle expression in the pair of equivalences above can be dropped: f_{R}(x)\leqslant y\quad\Leftrightarrow\quad x\leqslant f^{R}(y). But it can just as well be inserted back: define R by these equivalent formulas. It is obvious that the relation reconstructed from (f_{R},f^{R}) will be the R we started with. The following result then gives that there is a one-to-one correspondence between adjoint connections and Galois connections. The left adjoints in adjoint connections are known as “residuated mappings” (see e.g., [ 3 ] ). Theorem 2.2 . Every connection R\colon P\to Q has at most one left adjoint map f_{R} . Moreover, if f_{R} exists, it is monotone. In fact, there is a one-to-one correspondence between left adjoint connections R\colon P\to Q and monotone maps f\colon P\to Q . Proof. Suppose f_{R} and f^{\prime}_{R} are both left adjoint maps for the same connection R . Then f_{R}(x)\leqslant f^{\prime}_{R}(x)\quad\Leftrightarrow\quad x\;R\;f^{\prime}_{R}(x)\quad\Leftrightarrow\quad f^{\prime}_{R}(x)\leqslant f^{\prime}_{R}(x), for every x\in P . We have a similar chain of equivalences with the roles of f^{\prime}_{R} and f_{R} swapped around, and so f^{\prime}_{R}=f_{R} . Next, we prove that f_{R} is monotone. Suppose x\leqslant y in P . We have y\;R\;f_{R}(y) since f_{R}(y)\leqslant f_{R}(y) . Then x\;R\;f_{R}(y) and so f_{R}(x)\leqslant f_{R}(y) . Clearly, the relation R can be recovered from f_{R} . It is also obvious that a monotone f is the same as f_{R} for R defined by x\;R\;y\quad\Leftrightarrow\quad f(x)\leqslant y. So all it remains to show is that when f is monotone, R is a connection. Suppose a\leqslant b\;R\;c\leqslant d . Then f(a)\leqslant f(b)\leqslant c\leqslant d and so a\;R\;d . ∎ Dually to the previous theorem, we have: Theorem 2.3 . Every connection R\colon P\to Q has at most one right adjoint map f^{R} . Moreover, if f^{R} exists, it is monotone. In fact, there is a one-to-one correspondence between right adjoint connections R\colon P\to Q and monotone maps f\colon Q\to P . Remark 2.4 . The concept of a Galois connection between posets, in its full generality, goes back at least to [ 19 ] . In categorical terms, they are adjunctions between posets and the above theorems are specializations of well-known facts about representable profunctors. 3. Modular connections When P and Q are bounded modular lattices, a connection R\colon P\to Q is modular in the sense of M. Grandis (see [ 12 ] and the references there) when it is an adjoint connection and the two dual laws below hold, where \bot denotes the bottom element of Q and \top denotes the top element of P . (RM0) f^{R}(f_{R}(x))=x\vee f^{R}(\bot) , (LM0) f_{R}(f^{R}(y))=y\wedge f_{R}(\top) . In the setting of modular lattices, the laws (RM0) and (LM0) are equivalent to (RF0) and (LF0) (which we state in Section 4 ). The laws above play a significant role in projective homological algebra [ 11 , 12 ] . They are also important beyond the setting of modular lattices. As shown in [ 13 ] , they characterize protomodularity [ 5 ] of regular categories (when applied to direct-inverse image adjunctions between subobject posets), which play an important role both in categorical algebra and in universal algebra, where they are closely related to BIT varieties in the sense of A. Ursini [ 20 ] (see [ 4 ] and the references there). The work [ 13 ] led to “noetherian forms” [ 10 ] , a self-dual framework for establishing homomorphism theorems for group-like structures. The laws above are among the axioms for this framework. These developments are closely related to the fact that in group theory, the laws (RM0) and (LM0) describe the Lattice Isomorphism Theorem, which, as shown in [ 16 ] , can be used to derive other isomorphism theorems. The significance of modular connections in the broader theory of Galois connections is recognized in [ 3 ] , where they are called “weakly regular” mappings. See Example 13.2 there for how modular connections relate to the property of modularity of a lattice (see also [ 17 ] ). The following result is a direct consequence of Proposition 1 in [ 13 ] , hence we omit the proof. The equivalence of (LM0) and (LM1) in this theorem can also be found in [ 18 ] ; in the slightly more restricted context of bounded lattices, this equivalence has been proved already in [ 3 ] . Theorem 3.1 . For any adjoint connection R\colon P\to Q between posets P and Q , the following conditions are equivalent: (LM1) The direct image of f_{R} is down-closed, i.e., whenever c\leqslant f_{R}(b) , we have c=f_{R}(a) for some a\in P . (LM2) Whenever c\leqslant d in Q , the element f_{R}(f^{R}(c)) is the meet of c and f_{R}(f^{R}(d)) , i.e., f_{R}(f^{R}(c))=c\wedge f_{R}(f^{R}(d)) . (LM3) Whenever some meet c\wedge d exists in Q , we have: f_{R}(f^{R}(c\wedge d))=c\wedge f_{R}(f^{R}(d)) . Whenever P has a top element, these conditions are further equivalent to (LM0). Dually, we have: Theorem 3.2 . For any adjoint connection R\colon P\to Q between posets P and Q , the following conditions are equivalent: (RM1) The direct image of f^{R} is up-closed, i.e., whenever b\geqslant f^{R}(c) , we have b=f^{R}(d) for some a\in P . (RM2) Whenever b\geqslant a in P , the element f^{R}(f_{R}(b)) is the join of b and f^{R}(f_{R}(a)) , i.e., f^{R}(f_{R}(b))=b\vee f^{R}(f_{R}(a)) . (RM3) Whenever some join a\vee b exists in P , we have: f^{R}(f_{R}(a\vee b))=a\vee f^{R}(f_{R}(b)) . Whenever Q has a bottom element, these conditions are further equivalent to (RM0). 4. Principal elements and Frobenius reciprocity Consider now a commutative quantale, i.e., a complete lattice L equipped with an associative and commutative binary operation L\times L\rightarrow L,\quad(a,b)\mapsto a\cdot b=ab such that \cdot preserves joins in each argument. By well-known properties of Galois connections, for each fixed e\in L , the map b\mapsto be is a left adjoint in an adjoint connection, where the values of the right adjoint are written as follows: a\mapsto a:e. In the context of such lattice, in abstract commutative ideal theory [ 8 ] (see also [ 2 ] and the references there), an element e is said to be principal when the following dual laws hold: (i) c\wedge de=((c:e)\wedge d)e , (ii) a\vee(b:e)=(ae\vee b):e . These properties are identical to the laws (LF0) and (RF0), respectively, of a general adjoint connection: (RF0) a\vee f^{R}(c)=f^{R}(f_{R}(a)\vee c) . (LF0) c\wedge f_{R}(b)=f_{R}(f^{R}(c)\wedge b) . The law (LF0) was called Frobenius reciprocity in [ 15 ] . It plays a significant role in the theory of doctrines – a functorial approach to mathematical logic. In a slightly more abstract form, it is one of the axioms (“modular law”) for the concept of an allegory [ 9 ] , which is an abstraction of the category of relations in a regular category. The laws (RF0) and (LF0) are easily seen to be stable under composition. Note that we can get (RM0) from (RF0) by setting c=\bot . Dually, by setting b=\top in (LF0) we obtain (LM0). The laws (RM0) and (LM0) also feature in abstract commutative ideal theory. They define “weak principal elements”. Lemma 1(a) from [ 1 ] is a special case of a general property of adjoint connections: the equivalence of (LM0) and (LM1). Lemma 1(d) from [ 1 ] generalizes similarly to arbitrary adjoint connections. For semilattices, the equivalence of (RM0) and (RM5) below has been noted in [ 18 ] , whose very goal is to abstract the theory of principal elements to mappings (connections satisfying (LF0) and (RF0) are called “principal mappings” in [ 18 ] ). For complete lattices, equivalence of (RM0) and (RM4) is noted in [ 17 ] . Theorem 4.1 . Let R be a connection from a poset P to a poset Q , where P has binary joins and Q has a bottom element. Then (RM0) is equivalent to each of the the following laws: (RM4) f_{R}(a)=f_{R}(b) implies a\vee f^{R}(\bot)=b\vee f^{R}(\bot) . (RM5) f_{R}(a)\leqslant f_{R}(b) implies a\leqslant b\vee f^{R}(\bot) . Proof. (RM0) clearly implies (RM4). Suppose (RM4) holds and f_{R}(a)\leqslant f_{R}(b) . Then f_{R}(b)=f_{R}(a)\vee f_{R}(b)=f_{R}(a\vee b) (since in a Galois connection, the left adjoint map preserves joins). By (RM4), we then have b\vee f^{R}(\bot)=a\vee b\vee f^{R}(\bot), i.e., a\leqslant b\vee f^{R}(\bot) . Thus, (RM4) implies (RM5). Next, we show that (RM5) implies (RM0). Suppose (RM5) holds. Since f_{R}(f^{R}(f_{R}(b)))=f_{R}(b) (by a general property of Galois connections), we have f^{R}(f_{R}(b))\leqslant b\vee f^{R}(\bot). The reverse inequality holds for any adjoint connection. This proves (RM0). ∎ Dually, we have: Theorem 4.2 . Let R be a connection from a poset P to a poset Q , where P has a top element and Q has binary meets. Then (LM0) is equivalent to each of the following laws: (LM4) f^{R}(c)=f^{R}(d) implies c\wedge f_{R}(\top)=d\wedge f_{R}(\top) . (LM5) f^{R}(d)\geqslant f^{R}(c) implies d\geqslant c\wedge f_{R}(\top) . The following is a generalization of Lemma 1(b,e) from [ 1 ] to arbitrary adjoint connections. Connections between lattices that satisfy (LF0) and (RF0) are called “strongly range-closed” mappings in [ 3 ] , which contains the result about the equivalence of (LF1) and (LF0) for bounded lattices, and for lattices: in the form of an exercise with a hint; the proof presented below makes use of this hint too (this can also be found in [ 18 ] ). Below we use the notation a^{\downarrow}=\{b\mid b\leqslant a\} and a^{\uparrow}=\{b\mid b\geqslant a\} for an element a in a given poset. Theorem 4.3 . Let R be a left adjoint connection between posets, P\to Q . The following conditions are equivalent: (LF1) If c\leqslant f_{R}(b) for c\in Q and b\in P , then c=f_{R}(a) for some a\in P such that a\leqslant b . (LF2) (LM1) holds for the restriction a^{\downarrow}\to f_{R}(a)^{\downarrow} of f_{R} , for each a\in P . When R is an adjoint connection and P and Q have binary meets, the conditions above are further equivalent to (LF0). Proof. The equivalence of (LF1) and (LF2) is trivial. Now assume R is an adjoint connection and that P and Q have binary meets. Suppose (LF1) holds. The inequality f_{R}(f^{R}(c)\wedge b)\leqslant c\wedge f_{R}(b) holds for any adjoint connection. Since c\wedge f_{R}(b)\leqslant f_{R}(b) , we get that c\wedge f_{R}(b)=f_{R}(a) for some a\leqslant b . Then f_{R}(a)\leqslant c and so a\leqslant f^{R}(c) . This gives c\wedge f_{R}(b)=f_{R}(a)\leqslant f_{R}(f^{R}(c)\wedge b), as desired. Conversely, the implication (LF0) \Rightarrow (LF1) is immediate. ∎ Dually, we have: Theorem 4.4 . Let R be a right adjoint connection between posets, P\to Q . The following conditions are equivalent: (LF1) If b\geqslant f^{R}(c) for b\in P and c\in Q , then b=f^{R}(d) for some d\in Q such that d\geqslant c . (LF2) (RM1) holds for the restriction c^{\uparrow}\to f^{R}(c)^{\uparrow} of f^{R} , for each c\in P . When R is an adjoint connection and P and Q have binary joins, the conditions above are further equivalent to (RF0). Characterizations of adjoint connections between complete lattices satisfying the conditions studied in this paper, motivated by viewing connections as morphisms between complete lattices, can be found [ 17 ] . This approach relates closely to the work of Grandis (see [ 11 , 12 ] and the references there), where adjoint connections between various lattices are studied as morphisms in suitable categories. See also [ 18 , 3 ] (cited in [ 17 ] ) for related results. It is shown in [ 18 ] that for an adjoint connection R\colon P\to Q between bounded lattices, if P is modular, then (LM0) \Leftrightarrow (LF0) and if Q is modular, then (RM0) \Leftrightarrow (RF0). This refines the observation made in [ 11 ] that: when P and Q are both modular lattices, the conjunction of (LM0) and (RM0) is equivalent to the conjunction of (LF0) and (RF0). It follows from this observation that modular connections in Grandis exact categories [ 11 ] satisfy Frobenius reciprocity and its dual. The same is not true for noetherian forms, as noted in [ 7 ] , where the link between [ 11 ] and [ 15 ] was first pointed out. Remark 4.5 . As noted in [ 13 ] , in categorical terms, (LM0) states that the counit of the adjunction is a cartesian natural transformation. As also noted there, (RM0) is equivalent to the monad corresponding to the adjunction being nullary in the sense of [ 6 ] . As far as we know, apart from a minor remark in [ 13 ] , the extension of the characterizations of (LM0) and (LF0) contained in this paper to adjunctions between general categories has not been pursued. 5. Conclusion The present paper brings together two streams of literature that explore conditions on connections considered in this paper. These two streams are represented by [ 7 ] and [ 17 ] , and appropriate references there. As these conditions, in one or another form, have arisen independently in a number of unrelated works [ 8 , 15 , 20 , 16 , 4 , 11 ] , it is quite likely that there are also other streams of literature that we have missed and that rediscover and/or exploit the same conditions. Compiling a complete library of all such sources could be worthwhile.