Fixed Points, Induction, and Coinduction in Order Theory, Set Theory, Type Theory, Category Theory, and Logic: A Concise Summary

12/25/2018 ∙ by Moez A. AbdelGawad, et al. ∙ 0

In this note we present the formulation of the induction and co-induction principles and related notions, such as fixed points, using the language and conventions of each of order theory, set theory, (first-order) logic, category theory, and the theory of types in (object-oriented and functional) programming languages, for the purpose of examining some of the similarities and dissimilarities between these six mathematical subdisciplines. As a side-benefit that is of relevance to programming language researchers in particular, our comparison demonstrates one of the fundamental differences between structural typing (dominant in FP) and nominal typing (dominant in mainstream OOP).

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Fixed points and concepts related to them, such as induction and coinduction, have applications in numerous scientific fields. These mathematical concepts have applications also in social sciences, and even in common culture, usually under disguise.

These precise formal mathematical concepts are in fact quite common in everyday life to the extent that in many activities or situations (or “systems”) a mapping and a set of its fixed points can be easily discovered when some little mental effort is exercised. In particular, in any situation or system where there is some informal notion of repetition or iteration or continual “come-and-go” or “give-and-take” together with some informal notion of reaching a stable condition (e.g., equilibrium in chemical reactions or “common ground” between spouses) it is usually the case that an underlying mapping, i.e., a mathematical function/map, together with a set of fixed points of that mapping, can be easily revealed. This mapping is sometimes called a ‘state transition function’, and a fixed point of the mapping is what is usually called a ‘stable condition’ or a ‘steady state’ (of the “system”).

These concepts also show up thinly-disguised in visual art and music [29]. However, applications of these mathematical concepts in scientific fields, in particular, are plenty and more explicit, including ones in economics and econometrics [38], in mathematical physics [40], in computer science, and in many other areas of mathematics itself. Their applications in computer science, in particular, include ones in programming language semantics—which we touch upon in this paper—as well as in relational database theory (e.g., recursive or iterated joins) and in concurrency theory.111For more details on the use of induction, coinduction, and fixed points in computer science, and for concrete examples of how they are used, the reader is invited to check relevant literature on the topic, e.g., [27, 42, 13, 46, 34].

Fixed points, induction and coinduction have been formulated and studied in various subfields of mathematics, usually using different vocabulary in each field. Fields for studying these concepts formally include order theory, set theory, programming languages (PL) type theory (i.e., the study of data types in computer programming languages), (first-order) logic, and category theory. In this paper—which started as a brief self-note—we compare the various formulations of these concepts by presenting a summary of how these concepts, and many related ones—such as pre-fixed points, post-fixed points, inductive sets/types, coinductive sets/types, algebras, and coalgebras—are defined in each of these mathematical subfields.

Of particular interest to programming languages researchers, we also give attention to the fundamental conceptual difference between structural type theory, where type equality and type inclusion are defined based on type structures but not on type names, and nominal type theory, where type equality and type inclusion are defined based on the names of types in addition to their structures. Given their difference in how each particularly defines the inclusion relation between types (i.e., the subtyping relation), the conceptual difference between structural typing and nominal typing expresses itself, prominently, when fixed points and related concepts are formulated in each of structural type theory and nominal type theory.

As such this paper is structured as follows. In 2 (2 Order Theory) we start by presenting how these important mathematical concepts are formulated in a natural and simple manner in order theory, then we present their standard formulation in set theory in 3 (3 Set Theory). In 3 we also illustrate the concepts, for pedagogic purposes, using examples from number theory, set theory and real analysis.  Since PL type theory builds on set theory, we next present the formulation of these concepts in the theory of types of functional programming languages (which is largely structurally-typed) in 4.1 (4.1 Inductive and Coinductive Functional Data Types), then we follow that by presenting their formulation in the type theory of object-oriented programming languages (which is largely nominally-typed) in 4.2 (4.2 Object-Oriented Type Theory). Building on intuitions gained from the formulation of the concepts we presented in 3, we then suggest in 5 (5 First-Order Logic) a formulation of these concepts in first-order logic. Then we present the most general formulation of these concepts, in the context of category theory, in 6 (6 Category Theory).

In 7 (7 Comparison Summary) we summarize the paper by presenting tables that collect the formulations given in the previous sections. Based on the tabular comparison in 7 and the discussion in 4 (4 Programming Languages Theory), we also briefly discuss in 8 (8 Structural Type Theory versus Nominal Type Theory) a consequence of the fundamental difference between structural typing and nominal typing that we discussed in 4. We conclude in 9 (9 A Fundamental and More Abstract Treatment) by hinting at the possibility of a more abstract, and more unified formulation of the concepts that we discussed in this paper, using concepts from category theory, namely monads and comonads.

It should be noted that in each particular subfield we discuss in this paper we try to use what seems (to us) to be the most natural terminology in that field. Hence, in agreement with standard mathematical practice, the same mathematical concepts may have significantly different names when presented in different fields in this paper. It should be noted also that mathematical-but-non-computer-science readers—presumably interested mainly in comparing “pure” mathematical subdisciplines, i.e., in comparing formulations in order theory, set theory, first-order logic, and category theory, but not in PL type theory—may safely skip, at least in a first reading, the lengthier 4 and the short 8. (Those readers may like to also ignore the rightmost column of Table 1 and the leftmost column of Table 2 in 7.) This skipped material is mainly of interest to PL theorists only.

2 Order Theory

Order theory, which includes lattice theory as a subfield, is the branch of mathematics where the concepts fixed point, pre-fixed point, post-fixed point, least fixed point, greatest fixed point, induction, and coinduction were first formulated in some generality and the relations between them were proven [51]. Order theory seems to be the simplest and most natural setting in which these concepts can be defined and studied. We thus start this paper by presenting the order-theoretic formulation of these concepts in this section.

Formulation

Let (‘is less than or equal to’) be an ordering relation—also called a partial order—over a set and let be an endofunction over (also called a self-map over , i.e., a function whose domain and codomain are the same set, thus mapping a set into itself222We are focused on unary functions in this paper because we are interested in discussing fixed points and closely-related concepts, to which multi-arity makes little difference. Note that a binary function can be transformed into an equivalent unary one via “currying” (also known in logic as exportation). By iterating it, currying can be applied to multi-ary functions, i.e.

, ones with (finite) arity greater than two. Currying does preserve monotonicity/variance, and currying seems to be applicable to all fields of interest to us in this paper since, in each field, objects of that field—

i.e., posets, power sets, types, etc.—and “morphisms/arrows” between these objects seem to form a closed monoidal category.).

Given a point , we call point —the image of point under —the ‘-image333In this paper, nonstandard names (suggested by the author) are single-quoted like this ‘—’ when first introduced. of .

A point is called a pre-fixed point of if its -image is less than or equal to it, i.e., if

(A pre-fixed point of is sometimes also called an -closed point, ‘-lower bounded point’, or ‘-large point’.) The greatest element of , if it exists (in ), is usually denoted by , and it is a pre-fixed point of for all endofunctions . In fact , when it exists, is the greatest pre-fixed point of for all .

A point is called a post-fixed point of if it is less than or equal to its -image, i.e., if

(A post-fixed point of is sometimes also called an -consistent point, -‘(upper)’ bounded point, or ‘-small point’.) The least element of , if it exists (in ), is usually denoted by , and it is a post-fixed point of for all endofunctions . In fact , when it exists, is the least post-fixed point of for all .

A point is called a fixed point (or ‘fixed element’) of if it is equal to its -image, i.e., if

As such, a fixed point of is simultaneously a pre-fixed point of and a post-fixed point of .

Now, if is a complete lattice over (i.e., if is an ordering relation where meets and joins of all subsets of are guaranteed to exist in ) and if, in addition, is a monotonic endofunction over , i.e., if

then is called a generating function (or generator) and a point , called the least pre-fixed point of , exists in , and—as was proven by Tarski [51] is also the least fixed point (or, lfp) of , and a point , called the greatest post-fixed point of , exists in , and is also the greatest fixed point (or, gfp) of .444See Table 1 for the definitions of and in order theory. Further, for any element we have:

  • (induction)co if , then ,
    which, in words, means that if is a pre-fixed point of (i.e., if the -image of is less than or equal to it), then point is less than or equal to , and

  • (coinduction) if , then ,
    which, in words, means that if is a post-fixed point of (i.e., if is less than or equal to its -image), then point is less than or equal to point .

References

See [19, 44].

3 Set Theory

In set theory, the set of subsets of any set—i.e., the powerset of the set—is always a complete lattice under the inclusion () ordering. As such, Tarski’s result in lattice theory (see 2) was first formulated and proved in the specific context of powersets [33], which present a simple and very basic example of a context in which the order-theoretic formulation of induction/coinduction and related concepts can be applied and demonstrated. In fact, given the unique foundational significance of set theory in mathematics—unparalleled except, arguably, by the foundational significance of category theory—the set-theoretic formulation of the induction and coinduction principles are the standard formulations of these principles. The set-theoretic formulation also forms the basis for very similar formulations of these concepts in (structural) type theory (see 4.1) and in first-order logic (see 5).

Formulation

Let (‘is a subset of’) denote the inclusion ordering of set theory and let (‘is a member of’) denote the membership relation. Further, let be the partially-ordered set of all subsets of some fixed set , under the inclusion ordering (as such , where is the powerset operation, and is always a complete lattice) and let be an endofunction over .

A set (equivalently, ) is called an -closed set if its -image is a subset of it, i.e., if

(An -closed subset is sometimes also called an -lower bounded set or -large set.) Set , the largest set in , is an -closed set for all endofunctions —in fact is the largest -closed set for all .

A set is called an -consistent set if it is a subset of its -image, i.e., if

(An -consistent subset is sometimes also called an -(upper) bounded set555From which comes the name -bounded polymorphism in functional programming. (See 4.1.) , -correct set, or -small set.) The empty set, , the smallest set in , is an -consistent set for all endofunctions —in fact is the smallest -consistent set for all .

A set is called a fixed point (or ‘fixed set’) of if it is equal to its -image, i.e., if

As such, a fixed point of is simultaneously an -closed set and an -consistent set.

Now, given that is a complete lattice, if, in addition, is a monotonic endofunction, i.e., if

then is called a sets-generating function (or generator) and an inductively-defined subset , the smallest -closed set, exists in , and is also the smallest fixed point of , and a coinductively-defined subset , the largest -consistent set, exists in , and is also the largest fixed point of .666See Table 1 for the definitions of and in set theory. Further, for any set we have:

  • (induction)co
    (i.e., ),
    which, in words, means that if (we can prove that) set is an -closed set, then (by induction, we get that) set is a subset of , and

  • (coinduction)
    (i.e., ),
    which, in words, means that if (we can prove that) set is an -consistent set, then (by coinduction, we get that) is a subset of set .

Induction Instances

An instance of the set-theoretic induction principle presented above is the standard mathematical induction principle. In this well-known instance induction is setup as follows: is the “successor” function (of Peano)777Namely, , where, e.g., ., (the smallest fixed point of the successor function ) is the set of natural numbers ,888In fact defines the set as the set of all finite non-negative whole numbers. Its existence (as the least infinite inductive set) is an axiom of set theory. and is any inductive property/set of numbers.

  • An example of an inductive set is the set of natural numbers defined as . Set is an inductive set since i.e., —can be proven inductively by proving that is -closed, which is sometimes, equivalently, stated as proving that preserves (property) or that is preserved by . The -closedness (or ‘-preservation’) of can be proven by proving that (i.e., that ). Using the definition of the successor function , this means proving that , or, in other words, proving that (i.e., that the base case, , holds) and that from one can conclude that (i.e., that the inductive case, , holds). This last form is the form of proof-by-induction presented in most standard discrete mathematics textbooks.

Another instance of the induction principle is lexicographic induction, defined on lexicographically linearly-ordered (i.e.

, “dictionary ordered”) pairs of elements 

[42, 45]. In 4.1 we will see a type-theoretic formulation of the induction principle that is closely related to the set-theoretic one above. The type-theoretic formulation is the basis for yet a third instance of the induction principle—called structural induction—that is extensively used in programming semantics and automated theorem proving (ATP), including reasoning about and proving properties of (functional) software.

Even though coinduction is the dual of induction, and thus apparently very similar to it, practical uses of coinduction are relatively obscure compared to those of induction. For some applications of coinduction see, e.g.[14, 46, 34].

Illustrations

In this section we strengthen our understanding of pre-fixed points, post-fixed points, and fixed points by illustrating these concepts in set theory, and also by presenting a pair of examples (and exercises) from analysis—that are thus likely to be familiar to many readers—that exemplify these concepts.

In Set Theory

Figure 1 visually illustrates the main concepts we discuss in this paper in the context of set theory, by particularly presenting the pre-/post-/fixed points of some abstract generator over the powerset (a complete lattice) for some abstract set .

Figure 1: Illustrating the pre-fixed points, post-fixedpoints, and fixed points of a generator (i.e., monotonic function) where .

Figure 1 illustrates that subsets , and of are among the -closed/-large/-pre-fixed subsets of (i.e., the upper diamond, approximately), while subsets , and are among its -consistent/-small/-post-fixed subsets (i.e., the lower diamond, approximately). Of the sets illustrated in the diagram, only sets and are -fixed subsets of , i.e., belong to the intersection of -pre-fixed and -post-fixed subsets (i.e., the inner diamond, exactly). It should be noted that some other subsets of (i.e., points/elements of ) may neither be among the pre-fixed subsets of nor be among the post-fixed subsets of (and thus are not among the fixed subsets of too). Such subsets are not illustrated in Figure 1. (If drawn, such subsets would lie, mostly, outside all three diamonds illustrated in Figure 1.)

Figure 1 also illustrates that is true for any set and any generator . However, depending on the particular , it may (or may not) be the case that (e.g., if is a constant function or if happens to have only one fixed point).999In the context of category theory, the symbol is usually interpreted as denoting an equivalence or isomorphism relation between objects, rather than denoting the equality relation. See 6.

With little alterations (such as changing the labels of its objects, the direction of its arrows, and the symbols on its arrows) the diagram in Figure 1 can be used to illustrate these same concepts in the context of order theory, type theory, first-order logic or category theory. (Exercise: Do that.)

In Analysis

Most readers will definitely have met (and have been intrigued by?) fixed points in their high-school days, particularly during their study of analysis (the subfield of mathematics concerned with the study of real numbers and with mathematical concepts related to them). To further motivate the imagination and intuitions of readers, we invite them to consider the function defined as

and graphed in

Figure 2: Graph of function .

Figure 2. In particular, we invite the reader to decide: (1) which points of the totally-ordered real line (the x-axis) are pre-fixed points of , (2) which points/real numbers are post-fixed points of and, (3) most easily, which points are fixed points of .101010To decide these sets of points using Figure 2, readers should particularly take care: (1) not to confuse the fixed points of (i.e., crossings of the graph of with the graph of the identity function, which are solutions of the equation ) with the zeroes of (i.e., crossings with the -axis, which are solutions of the equation ), and (2) not to confuse pre-fixed points of for its post-fixed points and vice versa.
Since is not a monotonic function (i.e., is not a generator) we compensate by giving some visual hints in Figure 2 so as to make the job of readers easier. The observant reader will recall that diagrams such as that in Figure 2 are often used (e.g., in high-school/college math textbooks), with “spirals” that are zeroing in on fixed points overlaid on the diagrams, to explain iterative methods—such as the renowned Newton-Raphson method—that are used in numerical analysis to compute numerical derivatives of functions in analysis. These methods can be easily seen to be seeking to compute fixed points of some related functions. As such, these numerical analysis methods for computing numerical derivatives can also be explained in terms of pre-fixed points and post-fixed points.

(Exercise: Do the same for the monotonic function defined as

and depicted in

Figure 3: Graph of monotonic function .

Figure 3, and then relate your findings to Figure 1. Now also relate your findings regarding in Figure 2—or regarding monotonic/monotonically-increasing and antimonotic/monotonically-decreasing sections of —to Figure 1. Based on your relating of Figure 2 and Figure 3 to Figure 1, what do you conclude?).

The main goal of presenting these two examples from analysis is to let the readers strengthen their understanding of pre-/post-/fixed points and dispel any misunderstandings they may have about them, by allowing them to connect examples of fixed points that they are likely familiar with (i.e.

, in analysis) to the more general but probably less familiar concepts of fixed points found in order theory, set theory, and other branches of mathematics. Comparing the illustration of fixed points and related concepts in set theory (

e.g., as in Figure 1) with illustrations of them using examples from analysis (e.g., as in Figure 2 and Figure 3) makes evident the fact that a function over a partial order (such as ) is not as simple to imagine or draw as doing so for a function over a total order (such as ) is. However, illustrations and examples using functions over total orders can be too specific and thus misleading, since they may hide the more generality, the wider applicability, and the greater mathematical beauty of the concepts depicted in the illustrations.

References

See [42, 12].

4 Programming Languages Theory

Given the ‘types as sets’ view of types in programming languages, in this section we build on the set-theoretic presentation in 3 to present the induction and coinduction principles using the jargon of programming languages type theory. The presentation allows us to demonstrate and discuss the influence structural and nominal typing have on the theory of type systems of functional programming languages (which mostly use structural typing) and object-oriented programming languages (which mostly use nominal typing).

4.1 Inductive and Coinductive Functional Data Types

Formulation

Building on the concepts developed in 3, let be the set of structural types in functional programming.111111By construction/definition, the poset of structural types under the inclusion/structural subtyping ordering relation is always a complete lattice. This point is discussed in more detail below.

Let (‘is a subset/subtype of’) denote the structural subtyping/inclusion relation between structural data types, and let (‘has type/is a member of/has structural property’) denote the structural typing relation between structural data values and structural data types.

Now, if is a polynomial (with powers) datatype constructor121212That is, is one of the , , or data type constructors (i.e., the summation/disjoint union/variant constructor, the product/record/labeled product constructor, or the continuous function/exponential/power constructor, respectively) or is a composition of these constructors (and their compositions). By their definitions in domain theory [48, 43, 28, 31, 10, 24, 17], these structural datatype constructors, and their compositions, are monotonic (also called covariant) datatype constructors (except for the first type argument of , for which is an anti-monotonic/contravariant constructor, but that otherwise “behaves nicely” [37])., i.e., if

then an inductively-defined type/set , the smallest -closed set, exists in , and is also the smallest fixed point of , and a coinductively-defined type/set , the largest -consistent set, exists in , and is also the largest fixed point of .131313See Table 1 for the definitions of and .

Further, for any type (where , as a structural type, expresses a structural property of data values) we have:

  • (structural induction, and recursion)co
    (i.e., ),
    which, in words, means that if the (structural) property is preserved by (i.e., if is -closed), then all data values of the inductive type have property (i.e., ).
    Furthermore, borrowing terminology from category theory (see 6), a recursive function that maps data values of the inductive type to data values of type (i.e., having structural property ) is the unique catamorphism (also called a fold) from to (where is viewed as an initial -algebra and as an -algebra), and

  • (structural coinduction, and corecursion)
    (i.e.,
    which, in words, means that if the (structural) property is reflected by (i.e., if is -consistent), then all data values that have property are data values of the coinductive type (i.e., ).
    Furthermore, borrowing terminology from category theory, a corecursive function that maps data values of type (i.e., having structural property ) to data values of the coinductive type is the unique anamorphism from to (where is viewed as an -coalgebra and as a final -coalgebra).

Notes

  • To guarantee the existence of and in for all type constructors , and hence to guarantee the ability to reason easily—i.e., inductively and coinductively—about functional programs, the domain of types in functional programming is deliberately constructed to be a complete lattice under the inclusion ordering. This is achieved by limiting the type constructors used in constructing and over to structural type constructors only (i.e., to the constructors , , and their compositions, in addition to basic types such as Unit, Bool, Top, Nat and Int).

    • For example, the inductive type of lists of integers in functional programming is defined structurally (i.e., using , , and structural induction) as

      which defines the type as (isomorphic/equivalent to) the summation of type Unit (which provides the value unit as an encoding for the empty list) to the product of type Int with type itself.

    • In fact the three basic types Bool, Nat and Int can also be defined structurally. For example, in a functional program we may structurally define type Bool using the definition (for false and true), structurally define type Nat using the definition (for 0 and the successor of a natural number), and, out of other equally-valid choices, structurally define type Int using the definition (for negative integers, zero, and positive integers).

References

See [42, 23, 13, 54].

4.2 Object-Oriented Type Theory

The accurate and precise understanding of the generic subtyping relation in mainstream OOP languages such as Java, C#, C++, Kotlin and Scala, and the proper mathematical modeling of the OO subtyping relation in these languages, is one of our main research interests. Due to the existence of features such as wildcard types, type erasure, and bounded generic classes (where classes141414The notion of class in this paper includes that of an abstract class, of an interface, and of an enum in Java [25]. It also includes similar “type-constructing” constructs in other nominally-typed OO languages, such as traits in Scala [39]. And a generic class is a class that takes a type parameter (An example is the generic interface List in Java—that models lists/sequences of items—whose type parameter specifies the type of items in a list).

play the role of type constructors), the mathematical modeling of the generic subtyping relation in mainstream OOP languages is a hard problem that, in spite of much effort, seems to still have not been resolved, at least not completely nor satisfactorily, up to the present moment 

[52, 26, 5, 3, 8, 6].

The majority of mainstream OO programming languages are class-based, and subtyping () is a fundamental relation in OO software development. In industrial-strength OOP, i.e., in statically-typed class-based OO programming languages such as Java, C#, C++, Kotlin and Scala, class names are used as type names, since class names—which objects carry at runtime—are assumed to be associated with behavioral class contracts by developers of OO software. Hence, the decision of equality between types in these languages takes type names in consideration—hence, nominal typing. In agreement with the nominality of typing in these OO languages, the fundamental subtyping relation in these languages is also a nominal relation. Accordingly, subtyping decisions in the type systems of these OO languages make use of the inherently-nominal inheritance declarations (i.e., that are explicitly declared between class names) in programs written using these languages.151515Type/contract inheritance that we discuss in this paper is the same thing as the inheritance of behavioral interfaces (APIs) from superclasses to their subclasses that mainstream OO software developers are familiar with. As is empirically familiar to OO developers, the subtyping relation in class-based OO programming languages is in one-to-one correspondence with API (and, thus, type/contract) inheritance from superclasses to their subclasses [53]. Formally, this correspondence is due to the nominality of the subtyping relation [1].

Formulation

Let (‘is a subtype of’) denote the nominal subtyping relation between nominal data types (i.e., class types), and let (‘has type’) denote the nominal typing relation between nominal data values (i.e., objects) and nominal data types.

Further, let be the set of nominal types in object-oriented programming, ordered by the nominal subtyping relation, and let be a type constructor over (e.g., a generic class).161616Unlike poset in 4.1 (of structural types under the structural subtyping relation), poset (of nominal types under the nominal subtyping relation) is not guaranteed to be a complete lattice.

A type is called an ‘-supertype’ if its -image is a subtype of it, i.e., if

and is said to be preserved by . (An -supertype is sometimes also called an -closed type, -lower bounded type, or -large type). The root or top of the subtyping hierarchy, if it exists (in ), is usually called Object or All, and it is an -supertype for all generic classes . In fact the top type, when it exists, is the greatest -supertype for all .

A type is called an ‘-subtype’ if it is a subtype of its -image, i.e., if

and is said to be reflected by . (An -subtype is sometimes also called an -consistent type, -(upper) bounded type171717From which comes the name -bounded generics in object-oriented programming., or -small type). The bottom of the subtyping hierarchy, if it exists (in ), is usually called Null or Nothing, and it is an -subtype for all generic classes . In fact the bottom type, when it exists, is the least -subtype for all .

A type is called a fixed point (or ‘fixed type’) of if it is equal to its -image, i.e., if

As such, a fixed point of is simultaneously an -supertype and an -subtype. (Such fixed types/points are rare in OOP practice).

Now, if is a covariant generic class (i.e., a types-generator)181818Generic classes in Java are in fact always monotonic/covariant, not over types but over interval types ordered by containment! (See [9]). In particular, for any generic class in Java we have

meaning that if interval type is a subinterval-of (or contained-in) interval type then the instantiation of generic class with (i.e., the parameterized type , usually written as ) is always a subtype-of the instantiation of class with (i.e., of the parameterized type , usually written as ). As such, a generic class in Java is not exactly an endofunction over types but rather an “indirect endofunction,” since it generates/constructs types not directly from types but from interval types that are derived from types (See [8, 9]), and the generic class, as a function, is monotonic/covariant with respect to the containment relation over these interval types (i.e., it “generates a subtype when provided with a subinterval”)., i.e., if

and if , the ‘least -supertype’ exists in , and is also the least fixed point of , and if , the ‘greatest -subtype’, exists in , and is also the greatest fixed point of ,191919See Table 2 for the definitions of and in the (rare) case when happens to be a complete lattice. then, for any type we have:

  • (induction)co
    (i.e., ),
    which, in words, means that if the contract (i.e., behavioral type) is preserved by (i.e., is an -supertype), then the inductive type is a subtype of , and

  • (coinduction)
    (i.e., ),
    which, in words, means that if the contract (i.e., behavioral type) is reflected by (i.e., is an -subtype), then is a subtype of the coinductive type .

Notes

  • As discussed earlier and in , in structural type theory type expressions express only structural properties of data values, i.e., how the data values of the type are structured and constructed. In nominal type theory type names are associated with formal or informal contracts, called behavioral contracts, which express behavioral properties of the data values (e.g., objects) in addition to their structural properties.

    • To demonstrate, in a pure structural type system a record type that has, say, one member (e.g., type plane fly() , type bird fly() and type insect fly() ) is semantically equivalent to any other type that has the same member (i.e., type plane is equivalent to type bird and to type insect)—in other words, in a pure structural type system these types are ‘interchangeable for all purposes’.

    • On the other hand, in a pure nominal type system any types that have the same structure but have different names (e.g., types plane, bird and insect) are considered distinct types that are not semantically equivalent, since their different names (e.g., ‘plane’ versus ‘bird’ versus ‘insect’) imply the possibility, even likelihood, that data values of each type maintain different behavioral contracts, and thus of the likelihood of different use considerations for the types and their data values.202020For another example, a float used for monetary values (e.g., in finanicial transactions) should normally not be confused with (i.e., equated to) a float used for measuring distances (e.g., in scientific applications). Declaring type money=float type distance=float does not help in a purely structural type system, however, since the types float, money, and distance are structurally equivalent. On the other hand, in a purely nominal type system the declarations of types money and distance do have the desired effect, since the non-equivalence of the types is implied by their different names.,212121Further, when (1) the functional components of data values are (mutually) recursive, which is typical for methods of objects in OOP [4], and when (2) data values (i.e., objects) are autognostic data values (i.e., have a notion of self/this, which is an essential feature of mainstream OOP [18])—which are two features of OOP that necessitate recursive types—then the semantic differences between nominal typing and structural typing become even more prominent, since type names and their associated contracts gain more relevance as expressions of the richer recursive behavior of the more complex data values. (For more details, see [2] and [42, 19.3].)

  • In industrial-strength OO programming languages (such as Java, C#, C++, Kotlin and Scala) where types are nominal types rather than structural ones and, accordingly, where subtyping is a nominal relation, rarely is poset a lattice under the subtyping relation , let alone a complete lattice. Further, many type constructors (i.e., generic classes) in these languages are not covariant. As such, and rarely exist in .222222Annihilating the possibility of reasoning inductively or coinductively about nominal OO types. Still, the notion of a pre-fixed point (or of an -algebra) of a generic class and the notion of a post-fixed point (or of an -coalgebra) of , under the names -supertype and -subtype respectively, do have relevance in OO type theory, e.g., when discussing -bounded generics [6].232323In fact, owing to our research interests (hinted at in the beginning of 4.2), inquiries in [6] have been a main initial motivation for writing this note/paper. In particular, we have noted that if is a generic class in a Java program then a role similar to the role played by the coinductive type is played by the wildcard type F<?>, since, by the subtyping rules of Java (discussed in [6], and illustrated vividly in earlier publications such as [8, 5]), every -subtype (i.e., every parameterized type constructed using —called an instantiation of —and every subtype thereof) is a subtype of the type F<?>. On the other hand, in Java there is not a non-Null type (not even type F<Null>; see [6]) that plays a role similar to the role played above by the inductive type (i.e., a type that is a subtype of all -supertypes, which are all instantiations of and all supertypes thereof). This means that in Java greatest post-fixed points (i.e., greatest -subtypes) that are not greatest fixed points do exist, while non-bottom least pre-fixed points (i.e., least -supertypes) do not exist. Also, since is rarely a complete lattice, greatest fixed points, generally-speaking, do not exist in Java, neither do least fixed points. These same observations apply more-or-less to other nominally-typed OOP languages similar to Java, such as C#, C++, Kotlin and Scala. (See further discussion in Footnote 29 of 6.)

References

See [16, 15, 14, 11, 23, 32, 52, 6].

5 First-Order Logic

Via the axiom of comprehension in set theory, first-order logic—abbr. FOL, and also called predicate calculus—is strongly tied to set theory. As such, in correspondence with the set-theoretic concepts presented in 3, one should expect to find counterparts in (first-order) logic. Even though seemingly unpopular in mathematical logic literature, we try to explore these corresponding concepts in this section. The discussion of these concepts in logic is also a step that prepares for discussing these concepts in 6 in the more general setting of category theory.

Formulation

Let (‘implies’) denote the implication relation between predicates/logical statements of first-order logic242424Note that, as in earlier sections, we use the long implication symbol ‘’ to denote the implication relation in the metalogic (i.e., the logic used to reason about objects of interest to us, e.g., points, sets, types, or—as is the case in this section—statements of first-order logic). The reader in this section should be careful not to confuse the metalogical implication relation (denoted by the long implication symbol ) with the implication relation of first-order logic (used to reason in FOL, and denoted by the short implication symbol )., and let juxtaposition or (‘is satisfied by/applies to’) denote the satisfiability relation between predicates and objects/elements. Further, let be the set of statements (i.e., the well-formed formulas) of first-order logic ordered by implication ( is thus a complete lattice) and let be a logical operator over .252525Such as [and] and [or] (both of which are covariant/monotonic logical operators), [not] (which is contravariant/anti-monotonic), or compositions thereof.

A statement is called an ‘-weak statement’ if its -image implies it, i.e., if

Statement is an -weak statement for all endofunctors —in fact is the weakest -weak statement for all .

A statement is called an ‘-strong statement’ if it implies its -image, i.e., if

Statement is an -strong statement for all endofunctors —in fact is the strongest -strong statement for all .

A statement is called a fixed point (or ‘fixed statement’) of if it is equivalent to its -image, i.e., if

As such, a fixed point of is simultaneously an -weak statement and an -strong statement.

Now, if is a covariant logical operator (i.e., a statements-generator), i.e., if

then , the ‘strongest -weak statement’, exists in , and is also the ‘strongest fixed point’ of , and , the ‘weakest -strong statement’, exists in , and is also the ‘weakest fixed point’ of .262626See Table 2 for the definitions of and . Further, for any statement we have:

  • (induction)co if , then ,
    which, in words, means that if is an -weak statement, then implies , and

  • (coinduction) if , then ,
    which, in words, means that if is an -strong statement, then implies .

References

See [7](!), and [20, 49, 35, 47, 22].

6 Category Theory

Category theory seems to present the most general context in which the notions of induction, coinduction, fixed points, pre-fixed points (called algebras) and post-fixed points (called coalgebras) can be studied.

Formulation

Let (‘is related to’/‘arrow’) denote that two objects in a category are related (i.e., denote the ‘is-related-to’ relation, or, more concisely, denote relatedness).272727For mostly historical reasons, an arrow relating two objects in a category is sometimes also called a morphism. We prefer using is-related-to (i.e., has some relationship with) or arrow instead, since these terms seem to be more easily and intuitively understood, and also because they seem to be more in agreement with the abstract and general nature of category theory. Further, let be the collection of objects of a category (i.e., the category and the collection of its objects are homonyms) and let be282828Note that, similar to the situation for symbols and that we met in 5, in this section the same exact symbol is used to denote two (strongly-related but slightly different) meanings: the first, that two objects in a category are related (which is the meaning specific to category theory), while the second is the functional type of a self-map/endofunction/endofunctor that acts on objects of interest to us (i.e., points, sets, types, etc.), which is the meaning for that we have been using all along since the beginning of this paper. an endofunctor over .

An object is called an -algebra if its -image is related to it, i.e., if

An object is called an -coalgebra if it is related to its -image, i.e., if

Now, if is a covariant endofunctor, i.e., if

and if an initial -algebra exists in and a final -coalgebra exists in ,292929While referring to 2, note that a category-theoretic initial -algebra and final -coalgebra are not the exact counterparts of an order-theoretic least fixed point and greatest fixed point but of a least pre-fixed point and greatest post-fixed point. This slight difference is significant, since, for example, a least pre-fixed point is not necessarily a least fixed point, unless the underlying ordering is a complete lattice and is covariant/monotonic (Exercise: Prove this?), and also a greatest post-fixed point is not necessarily a greatest fixed point, unless the underlying ordering is a complete lattice and is covariant/monotonic. The difference is demonstrated, for example, by the subtyping relation in generic nominally-typed OOP (see Footnote 23 in 4.2). Hence, strictly speaking, an initial -algebra does not deserve the symbol ‘’ we used as a name for it, nor does a final -coalgebra deserve the symbol ‘’ as its name, since the symbols ‘’ and ‘’ are standard names for well-known concepts (i.e., they are, strictly speaking, reserved for least and greatest fixed points, or exact counterparts of them, respectively, and hence should always imply that the concepts denoted by them are indeed exactly such concepts). then for any object we have:

  • (induction)co if , then ,
    which, in words, means that if is an -algebra, then is related to (via a unique “complex-to-simple” arrow called a catamorphism), and

  • (coinduction) if , then ,
    which, in words, means that if is an -coalgebra, then is related to (via a unique “simple-to-complex” arrow called an anamorphism).

Notes

  • Even though each of an initial algebra and a final coalgebra is simultaneously an algebra and a coalgebra, it should be noted that there is no explicit concept in category theory corresponding to the concept of a fixed point in order theory, due to the general lack of an equality relation in category theory and the use of the isomorphism relation instead.

    • If such a “fixedness” notion is defined in category theory, it would denote an object that is simultaneously an algebra and a coalgebra, i.e., for a functor , a “fixed object” of will be related to the object and vice versa. This usually means that and , if not the same object, are isomorphic objects. (That is in fact the case for any initial -algebra and for any final -coalgebra, which—given the uniqueness of arrows from an initial algebra and to a final coalgebra—are indeed isomorphic to their -images).

  • Note also that, unlike the case in order theory, the induction and coinduction principles can be expressed in category theory using a “point-free style” (as we do in this paper) but they can be also expressed using a “point-wise style”303030By giving a name to a specific arrow that relates two objects of a category, e.g., using notation such as to mean not only that objects and are related but also that they are related by a particular arrow named .. As such, regarding the possibility of expressing the two principles using either a point-wise style or a point-free style, category theory agrees more with set theory, type theory, and (first order) logic than it does with order theory.

  • Incidentally, categories are more similar to preorders (sometimes, but not invariably, also called quasiorders) than they are similar to partial-orders (i.e., posets). This is because a category, when viewed as an ordered set, is not necessarily anti-symmetric.

    • Categories are more general than preorders however, since a category can have multiple arrows between any pair of its objects, whereas any two elements/points of a preorder can only either have one “arrow” between the two points (denoted by rather than ) or not have one. (This possible multiplicity is what enables, and sometimes even necessitates, the use of arrow names, so as to distinguish between different arrows when multiple arrows relate the same pair of objects). As such, every preorder is a category, or, more precisely, every preorder has a category—appropriately called a category or a bool(ean)-category—corresponding to it. However, generally-speaking, there is not a unique preorder corresponding to each category.313131For more on the very strong relation between order theory and category theory, i.e., metaphorically on , see 9. Indeed the relation between the two fields is so precise that the metaphor can be described, formally, as an adjunction between the category of preorders and the category of small categories.

References

See [54, 41, 50, 21].

7 Comparison Summary

Order Theory Set Theory FP Type Theory
Domain Points of a Set =Subsets of a Set Structural Data Types
Relation Abstract Ordering Inclusion Inclusion
Operator Endofunction Endofunction Type Cons.
Generator Monotonic Monotonic Poly. Type Constr.
(, and comp.)
Pre-F.P. Pre-Fixed Point -Large Set -Closed Type
Post-F.P. Post-Fixed Point -Small Set -Bounded Type
If Complete Lattice Smallest -Large Set Smallest -Closed Type
L.F.P.
(Least Pre.) ( denotes ‘meet’) (Inductive Type)
If Complete Lattice Largest -Small Set Largest -Bounded Type
G.F.P.
(Gr. Post.) ( denotes ‘join’) (Coinductive Type)
Induction
Principle (i.e., (i.e.,
) )
Coinduction
Principle (i.e., (i.e.,
) )
Domain is Sometimes Always Always
Comp. Lat.
Operator is Sometimes Sometimes Always
Generator
Table 1: Comparison of concepts by subdiscipline (Part 1).
OOP Type Theory First-Order Logic Category Theory
Class Types Statements Objects
Subtyping Implication Arrow
Generic Class Logical Operator Endofunctor
Covariant Covariant Covariant
-Supertype -Weak -Algebra
-Subtype -Strong -Coalgebra
Least -Supertype Strongest -Weak Stmt
If Complete Lattice Initial -Algebra
( denotes ‘meet’) ( denotes conjunction)
Greatest -Subtype Weakest -Strong Stmt
If Complete Lattice Final -Coalgebra
( denotes ‘join’) ( denotes disjunction)
(i.e., (i.e., (via unique catamorphism
) ) from to )
(i.e., (i.e., (via unique anamorphism
) ) from to )
Rarely Always Sometimes
Sometimes Sometimes Sometimes
Table 2: Comparison of concepts by subdiscipline (Part 2).

Table 1 and Table 2 summarize the formulations of the induction/coinduction principles and concepts related to them that we presented in 2-6.

8 Structural Type Theory versus Nominal Type Theory

The discussion in 4.1, together with that in 3 and 5, demonstrates that FP type theory, with its structural types and structural subtyping rules being motivated by mathematical reasoning about programs (using induction or coinduction), is closer in its flavor to set theory (and first-order logic/predicate calculus), since structural type theory assumes and requires the existence of fixed points and in for all type constructors . (For a discussion of the importance of structural typing in FP see [30, 36] and [42, 19.3].)

On the other hand, the discussion in 4.2, together with that in 6 and 2, demonstrates that OOP type theory, with its nominal types and nominal subtyping being motivated by the association of nominal types with behavioral contracts, is closer in its flavor to category theory and order theory, since nominal type theory does not assume or require the existence of fixed points and in for all type constructors . (For a discussion of why nominal typing and nominal subtyping matter in OOP see [2] and [42, 19.3].)

As such, we conclude that the theory of data types of functional programming languages is more similar in its views and its flavor to the views and flavor of set theory and first-order logic, while the theory of data types of object-oriented programming languages is more similar in its views and its flavor to those of category theory and order theory. This conclusion adds further supporting evidence to our speculation (e.g., in [5, 3]) that category theory is more suited than set theory for the accurate understanding of mainstream object-oriented type systems.

9 A Fundamental and More Abstract Treatment

As we hinted to in 6, order theory and category theory are strongly related. In fact the connection between the two fields goes much, much further than we hinted at.

Closure and Kernel Operators

In order theory a closure operator over a poset is an idempotent extensive generator [19]. An extensive (or inflationary) endofunction over is an endofunction where

meaning that all points of are -small (i.e., are post-fixed points of ). An endofunction is idempotent iff

i.e., iff , meaning that applying twice does not transform or change an element of any more than applying to the element once does.

Also, in order theory a kernel (or interior) operator over is an idempotent intensive generator. An intensive (or deflationary) endofunction over is an endofunction where

meaning that all points of are -large (i.e., are pre-fixed points of ).323232It may be helpful here to check Figure 1 again. If in Figure 1 is a closure operator then the upper diamond minus the inner diamond “collapses” (becomes empty), and we have . That is because, since is extensive, we have , and thus —if it exists in —is always a fixed point of , in fact the gfp (greatest fixed point) of . Dually, if in Figure 1 is a kernel operator then the lower diamond minus the inner diamond “collapses”, and we have . That is because, since is intensive, we have , and thus —if it exists in —is always a fixed point of , in fact the lfp (least fixed point) of .

Monads and Comonads

In category theory, on the other hand, when a partially-ordered set is viewed as a category, then a monad on a poset turns out to be exactly a closure operator over  [50]. By the definition of monads, all objects of a category are coalgebras of a monad, which translates to all elements of a poset being post-fixed points of the corresponding closure operator on (which we noted above). As such, the algebras for this monad correspond to the pre-fixed points of the closure operator, and thus correspond exactly to its fixed points.

Similarly, a comonad (the dual of a monad) on turns out to be exactly a kernel operator over . As such, by a dual argument, the coalgebras for this comonad are the post-fixed points of the kernel operator, and thus also correspond exactly to its fixed points.

An Alternative Presentation

Based on these observations that further relate category theory and order theory, the whole technical discussion in this paper can be presented, more succinctly if also more abstractly, by rewriting it in terms of the very general language of monads/comonads, then specializing the abstract argument by applying it to each of the six categories we are interested in.333333Namely, (1) category OrdE of partially-ordered sets together with self-maps, (2) category PSetE of power sets (ordered by inclusion) together with endofunctions, (3) category TypS of structural types (ordered by inclusion/structural subtyping) together with structural type constructors, (4) category TypN of nominal types (ordered by nominal subtyping) together with generic classes, (5) category FOL of first-order logical statements (ordered by implication) together with logical operators, and (6) category CatE of (small) categories (ordered/related by arrows) together with endofunctors. Given that our goal, however, is to compare the concepts in their most natural and most concrete mathematical contexts, we refrain from presenting such a fundamental treatment here, keeping the possibility of making such a presentation in some future work, e.g., as a separate paper, or as an appendix to future versions of this paper.

References

See [50, 21].

Acknowledgments

The author would like to thank John Greiner (Rice University) and David Spivak (MIT) for their suggestions and their valuable feedback on earlier versions of this paper.

References