We present preliminary work on a framework for constructing identities that takes as its point of departure Barwise's channel theory, founded on category theory, and allied work on situation semantics. The framework takes actions relating to identity as fundamental. Actions and events are situations, although situations can include larger context. Identity-related situations are classified under (identity) types, which can be fused. Identity-related activity is partitioned according to the 'form of life' involved into (identity) environments, similar to Barwise's classifications. An identity-related action is often based on another identity-related action as a resource, and all these actions may reference events in an individual's life. The result is a layered picture of identity-related situations. Identities retain references to the situations from which they are constructed, so provenance, context, and narrative detail are retained. The category-theoretical constructs of Barwise's channel theory, however, must be relaxed when identity in multiple environments is considered.
This paper presents preliminary work on a framework for constructing identities that takes its point of departure work by Barwise and his associates on the “flow of information” and situations. Barwise’s channel theory is an application of an area of abstract mathematics known as category theory, which addresses the form of mathematical theories and interrelates their structures. Channel theory gives an account of how x being P carries the information that y is Q, where x and y are “tokens” and P and Q are “types,” x and P being in one classification system, y and Q being in another. Barwise and his associates also developed situation semantics, which accounts for the meaning of an expression as the relation between situations, viz., between an utterance (situation) and a described situation. Channel theory and situation semantics were two parts of a coherent program, and situation theory was developed as mathematical support for situation semantics.
The framework sketched here takes actions relating to identity, both judgments and assertions, as fundamental. Actions are events (with an agent), and events are basically situations, although situations can include larger regions of space and time and generally accommodate context. Identity-related situations are classified under (identity) types, which can be fused. Identity-related activity is partitioned according to the ‘form of life’ involved into environments, similar to Barwise’s classifications. An identity-related action is often based on another identityrelated action as a resource, and all these actions may reference events in an individual’s life. The result is a layered picture of identity-related situations, and this layering was part of the motivation for situation semantics.
Identities in this framework retain references to the situations from which they are constructed, so provenance, context, and narrative detail are retained. Multiple identities are easily accommodated yet fusion results in constraints that converge to real-life individuals behind the perspectives that arise in the environment, somewhat as a genotype is behind phenotypes.
It turns out that the category-theoretical constructs of Barwise’s channel theory must be relaxed when identity in multiple environments is considered. We explain the reasons for this and sketch what needs to be done to recover the category-theoretic foundation.
The next two sections provide brief introductions to category theory and channel theory. Section 4 introduces the technical notion of a situation. Section 5 introduces the notion of identity as a type in a classification (system), that is, an environment, which involves a coherent set of practices and standards used to identify and characterize individuals. Different environments provide different perspectives on a given person, but note that the same idrelevant situation (or id-situation) may be classified under several types since it may involve several individuals. Section 5 concludes with a sketch of three environments. Section 6 reconciles identification-relevant actions (idactions) that are first person assertions with third-person identifications and characterizations in terms of the notion of theory of mind from developmental psychology. Section 7 introduces resource situations as meta-representations in an id-situation, and it introduces referenced events as realworld facts that are the background content of the idsituations. The next section analyzes examples where identities (as types in environments) are fused. Finally, Section 9 considers how this framework can be expressed in category-theoretic terms similar to channel theory.
A category consists of a class of objects and a class of morphisms (or arrows or maps) between the objects. Each morphism f has a unique source object and target object b; we write : → . The composition of : → and : → is written as ∘ and is required to be associative: if in addition ℎ: → , then ℎ ∘ ( ∘ ) = (ℎ ∘ ) ∘ . It is also required that, for every object x, there exists a morphism 1 : → (the identity morphism for x) such that, for every morphism : → , we have 1 ∘ = = ∘ 1 . It follows from these properties that there is exactly one identity morphism for every object. A functor from one category to another is a structure-preserving mapping, preserving the identity and composition of morphisms. More exactly, if and are categories, then a functor F from to is a mapping that associates with each object ∈ an object, ( ) ∈ and, with each morphism : → ∈ , a morphism ( ): ( ) → ( ) ∈ such that ( ) = for every object x ∈ C and ( ∘ ) = ( ) ∘ ( ) for all morphisms f: x → y and g: y → z .
In category theory, a commutative diagram is a diagram of objects (as vertices) and morphisms (arrows between objects) such that all directed paths in the diagram with the same start and end points lead to the same result by composition.
The classic presentation of category theory is
The point of departure for channel theory is the notion of a
classification, which consists of a set of individuals
(“tokens”), a set of types, and a binary relation indicating
that a given token is of a given type. Barwise and Seligman
For classifications A and C, an infomorphism f from A to C is a pair of functions ( ∧, ∨), ∧: ( ) → ( ), and ∨: ( ) → ( ) satisfying, for all tokens ∈ ( ) and all types ∈ ( ),
∨( ) ⊨ iff ⊨ ∧( ) That is, the image of token c (under ∨) is classified to be of type α if and only if c is classified to be of the type that is of the image of α (under ∨). Intuitively, an infomorphism is a part-to-whole, A-to-C, informational relationship. Figure 1.a shows function pairs ( ∧, ∨) and ( ∧, ∨), each depicted in Figure 1.b as a single infomorphism with codomain C, while f has domain A and g domain B.
Information (not the “amount” of information, as per Shannon) is assumed to “flow” among the components of a system. Components may, but need not, be distant from one another in time and space, and they may be very different things. The system is “distributed” in this sense (not necessarily in the sense in which that term is used in computer science). For exemplar, the students, classrooms, scheduling system, and attendance records at a school together form a distributed system.
An information channel is a family of infomorphisms with a common codomain, called the core. Essentially, a channel consists of a set {A1,...,An} of classifications that represent the parts of the distributed system, a classification C (the core) that represents the system as a whole, and a set of infomorphisms {f1,...,fn} from each of the parts onto C. Tokens in C are the connections of the system: a given token c in C connects the tokens it is related to by means of {f1,...,fn}. Parts {A1,...,An} carry information about each other as long as they all are part of C. In category-theoretic terms, the core is the sum (or colimit) of the parts. Figure 2 shows the sense in which the core is a universal construction, making it unique.
A distributed system D is a collection of elements that carry information about each other. Formally, D consists of an indexed class ( ) of classifications together with a class ( ) of infomorphisms whose domains and codomains are all in ( ). An information channel K covers distributed system D if and only if ( ) are the classifications of the channel and, for every infomorphism in ( ), there are infomorphisms from both the domain and codomain of to the core of K such that the diagram formed by these three infomorphisms commutes. Basically, all classifications in D are “informational parts” of the core whose channel covers D.
Turning to regularities in a classification’s type, let A be a classification and Γ and ∆ be sets of types in A. A token a of A satisfies the “sequent” 〈Γ,∆〉 provided that, if a is of every type in Γ, then it is of some type in ∆. If every token of A satisfies 〈Γ,∆〉, then Γ is said to entail ∆ and 〈Γ,∆〉 is called a constraint supported by A. The set of all constraints supported by A is called the complete theory of A, denoted by Th(A). These constraints are system regularities, and it is by virtue of regularities among connections that information about some components of a distributed system carries information about other components. These regularities are relative to the analysis of the distributed system in terms of information channels. Barwise and Seligman’s summary statement of their analysis of information flow, restricted to the simple case of a systems with two components, a and b, is as follows.
Suppose that the token a is of type α. Then a’s being of type α carries the information that b is of type β, relative to channel C, if a and b are connected in C and if the translation α′ of α entails the translation β′ of β in the theory Th(C), where C is the core of . (Barwise and Seligman 1997, p.35)
Channel theory is neutral about what tokens are, for
example, objects (or object states) or system states. One
alternative, which is quite common and natural, is to take
tokens to be situations. Situation semantics, articulated in
Barwise and Perry’s Situations and Attitudes
According to Devlin, a situation is some part of the ongoing happenings in the world and allows situation theory to handle context. Situations include not only connected space-time regions but also spatially disconnected (e.g., telephone calls) and temporally disconnected situations (e.g., multi-day sporting events).
Situation theory introduces the notion of an infon as the basic item of information. We assume as given sets A of individuals, R of properties and relations, L of spatial locations, and T of temporal locals. The general form of an infon, then, is
〈〈 , 1, … , , , , 〉〉, where ∈ is an n-place relation, 1, … , ∈ are objects appropriate for the corresponding argument places of , ∈ , ∈ , and ∈ {0,1} is the polarity. A polarity of 1 indicates that the objects are thus related in l at t; a polarity of 0 indicates otherwise. Where s is a situation and σ an infon, it might be that s supports σ, that is, that σ holds in s.
A real situation (according to Devlin) is a part of reality (and considered a single entity) that supports an indefinite number of infons, while an abstract situation is a set of infons. Real situations are tokens for us. An abstract situation for us is a type. If abstract situation s is the set { 1, … , } of infons and we identify the classification as I, then, for real situation r, ⊨ if and only if r supports all infons in s.
We take an identity to be a classification in the sense of
channel theory, where types are abstract situations and
tokens are real situations. The similarity between
situations and events has been frequently noted
An (identity) environment is a coherent set of practices
and standards used to identify and to characterize
individuals. This notion of environment relates to
Wittgenstein’s observation that “to imagine a language
means to imagine a form of life”
Different environments provide different perspectives on a given person. The question of how these perspectives are fused into a single identity is deferred to Section 8. How one partitions identity-related activity into environments is somewhat arbitrary, and our example environments could easily be partitioned finer. Indeed, we mention fusing identities within each of these environments.
The same id-situation may be classified under several types, as when an action relates to the identities of several different individuals. One generally tries to identify idsituations that relate to as few distinct identities as possible although some events inherently involve multiple players.
Some id-actions amount to asserting an identity (which is a first-person perspective) while others identify or characterize an individual distinct from the identifying or characterizing agent. We handle both these perspectives within a given environment.
Our first example of an identity environment is that in which we establish online identities in a rather narrow sense. Here we can look at the FOAF properties that have Person (or, more broadly, Agent) as domain. This includes, for example, mbox (a person’s email address) and workplaceHomepage. As per FOAF (and the Semantic Web in general), resources are generally identified with URIs. A type is associated with a pattern that (in the first instance) is a URI. As a situation type, such a type is a context in which the URI is used (e.g., a given email address is used in the ‘To’ field of an email). Tokens are situations where someone engages in an online activity or reflects on such an activity. Such a situation is classified as of a given type if the activity involved makes use of the URI in question. The situation could involve an individual asserting their identity (as in the ‘From’ field of an email) or characterizing an individual (as in the ‘To’ field of an email or when someone analyzes the content of an email).
Another identity environment is that in which one establishes what we refer to as physiological identity. Types here relate to the physical features of a person that may figure in their identification or characterization, such things as fingerprints, iris patterns, photographs of faces, and recordings of voices. Thus, for example, we have a situation type in which a given fingerprint pattern is used for identification of an individual. A token here is a situation in which one uses physical evidence for identification or characterization. Perhaps the more common cases are where one identifies or characterizes an individual, but there are notable cases where one asserts one’s identity (e.g., “shows one’s face”). To ground the notion of a pattern and relate it to officially sanctioned evidence, we distinguish an instance of a pattern that is preserved by an authority (such as a law-enforcement agency) and accepted as a standard against which other instances of the pattern may be compared for authenticity. Examples of such standards include, for example, the images that the FBI keeps of convicted criminals’ fingerprints. The pattern instantiated by such a standard may also be instantiated, for example, on a doorknob (in, say, a crime scene investigation), but it is also instantiated in the texture of the thumb of an actual person. Note that a standard need not be pictorial. For example, an iris pattern may be stored as a multitude of properties and distributions in computer memory.
Our final example of an identity environment is where we establish what we call (in a narrow sense) social identities, concerned with social aspects of a person that may figure in an attempt to identify or characterize them. Paradigmatic examples come from historical research. A token here is a situation in which one uses social evidence for identifying or characterizing an individual. Types include such things as the observing a given name in a given military roster or a church record. A name is to be understood in terms of its denotation independent of language and variants within the same language. Thus, for example, ‘Charlemagne’, ‘Carolus Magnus’, and ‘Karl der Große’ all count as the same name. On the other hand, a phrase structured as a proper name (e.g., “John Smith”) often (if not usually) denotes different individuals in different circumstances and may not denote a real individual in any circumstance (e.g., “Lieutenant Kijé”). A definite description is a way of identifying an individual based on characteristics that are distinguishing in the context in which they are used. Extended discourse can provide rich characterization of an individual, who is denoted with various linguistic devices, including proper name, definite descriptions, and anaphoric pronouns. While examples of actions relevant to social identity that come readily to mind generally involve identifying or characterizing an individual, there are notable cases of asserting social identity (e.g., as per Shelley, “My name is Ozymandias, king of kings: Look on my works, ye Mighty, and despair!”).
Note that physiological identity lacks the conventional or social nature of the other two. In some standard sense of “identity,” it is quite common for people to have multiple online identities, and FBI Associate Director Mark Felt was revealed to be the character Deep Throat only 31 years after Watergate. In contrast, perspectives that apparently have a life of their own do not arise with physiological evidence unless the person himself undergoes radical change.
We have maintained that, for a given identity environment,
there are id-actions that are assertions of identity and
others that identify or characterize an individual distinct
from the identifying agent. That the same ‘form of life’ is
involved in both aspects (when the identity environment is
held constant) is supported by research in developmental
psychology addressing what is known as theory of mind
(ToM)
A quite comprehensive coverage of ToM is presented by
Alvin Goldman in Simulating Minds
Regarding identity-relevant actions, what is known of
ToM suggests that, in identifying or characterizing an
individual distinct from the agent doing the identifying or
characterizing, the agent uses the ToM relevant to the
identity environment in question. According to TT, all
states of the mindreader are meta-representations, none of
which are attributed to the target, although the contents of
the final state are
The denotations of the meta-representations invoked by TT are essentially what Devlin identifies as resource situations. For example, the utterance "The dog I saw yesterday is back" carries information about the presence of a dog and does so by reference to a previous situation (in which the same dog was present). Regarding idactions, when one inspects an email, for example, and notes that it is from someone, the id-action is the inspection and the resource situation is the sending of the email. When an id-action asserts identity, no resource situation is generally needed, but when it identifies or characterizes an individual, it generally (but not always) references a resource situation. An example of the latter that does not involve a resource situation is taking a picture of an individual. Taking a picture, however, can be a resource situation for any number of id-situations because that id-action leaves an enduring object.
Another component in many id-situations is what we call the referenced event (or, sometimes, the referenced state), a real-world fact that is the background content of the id-situation. For example, the content of an email often makes reference to one or more events, fingerprints on the doorknob may relate to someone opening the door at a certain time, and analyzing a passage in a chronicle usually involves reading about events that happened in the society being chronicled. Frequently, a referenced event is what a resource situation is about. It may also be what an assertion of identity is about (as in, e.g., a confession). An id-action often involves a referenced event without a resource situation as when someone describes another's personality, background, or work. When an object is provided for reference, we take the referenced 'event' to be the event producing the object. For example, when a mug shot is attached to an email, the referenced event is the subject posing for the mug shot (which includes provenance and topic information missing in the bare picture).
For establishing identity, sometimes we focus on the idsituation itself, sometime on resource situation, and sometimes on the referenced event. Times and spatial locations are often details that bind together events to form a coherent picture, and these parameters are relevant to all three of these aspects. These details, however, are often lacking for some or all of these aspects. In such cases, sometimes we can impose a non-metric (and possibly only partial) temporal order on situations and events and determine relative locations. Spatial and temporal inclusion and overlap relations are also often indicated.
Often the more interesting issues that arise regarding identity relate to fusing identities from several identity environments. Suppose we do the following. (i) We observe an email from John.Doe@Acme.com that states that he was at Joe’s Bar at 11:00 PM on the eleventh (ii) We observe a picture of the inside of Joe’s Bar taken at 11:00 PM on the eleventh that contains a face that closely matches the face in a picture labeled ‘John Doe.’ Call the situation that involves the id-action in (i) o1 and the situation that involves the id-action in (ii) o2. Note that both involve a resource situation: in o1, it is the action of sending the email, and in o2 it is the action of taking and labeling John’s picture. The id-situation in (i) is of type, say, JohnDowOL, a type in the online environment, and we assert the identity proposition (1) o1 |= JohnDowOL The id-situation in (ii) is of type, say, JohnDowPhy, a type in the physiological environment, and we assert the identity proposition (2) o2 |= JohnDowPhy Apparently, there is a common referenced event, ref, of John Doe being at Joe’s Bar at 11:00 PM on the eleventh. If we judge that ref is indeed a common referenced event, we thereby fuse JohnDoeOL and JohnDoePhy, forming the super-type, say, JohnDoe, and o1 and o2 are then taken as evidence for ref by way of propositions (1) and (2).
Now suppose the eleventh was a Saturday, and suppose we have the following. (i′) We observe an email from John.Doe@Acme.com that states that he was at Joe’s Bar at 11:00 PM on the fourth (ii′) We observe a picture of the inside of Joe’s Bar taken at 11:00 PM on the eighteenth that contains a face that closely matches the face in a picture labeled ‘John Doe.’ Call the situation that involves the id-action in (i′) o1a and the situation that involves the id-action in (ii) o2a. Both involve resource situations similar to those in o1 and o2, respectively. The referenced event for o1a, call it ref1, is John being in the bar at 11:00 PM on the fourth, and the referenced event for o2a, call it ref2, is John being there at 11:00 on the eighteenth. We assert the following. (1′) o1a |= JohnDowOL (2′) o2a |= JohnDowPhy Id-situations (i′) and (ii′) may be further evidence, by way of propositions (1′) and (2′), for fusing JohnDowOL and JohnDowPhy. Fusing the identity types produces an additional correspondence between the referenced event ref, ref1, and ref2 in that they involve the same identity type. These referenced events (and hence, indirectly, idsituations (i), (ii), (i′), and (ii′)) may provide evidence regarding John Doe’s habit of visiting Joe’s Bar Saturday evenings. A large number of diverse id-situations relating to a single, fused identity may allow us to build up not just a history of an individual but even a full-blown identity.
As another example, suppose we do the following. (iii) We determine that the DNA from skeletal remains found beneath a car park in Leicester match the DNA from the blood on a garment known to have been worn by Richard III. (iv) We read in a chronicle that Richard III was hastily buried at a certain location near medieval Leicester. Call the id-situation that involves the id-action in (iii) o3 and the id-situation that involves the id-action in (iv) o4. Note that o4, but not o3, involves a resource situation, namely, the situation in which the chronicler wrote about Richard III’s burial. The id-situation in (iii) is of type, say, RichardIIIPhy, while that in (iv) is of type, say, RichardIIISoc, and we assert (3) o3 |= RichardIIIPhy (4) o4 |= RichardIIISoc Relevant to what is of interest here, o3 relates to a referenced event, call it ref3, to do with the demise of the last monarch of the House of York in a certain time interval and his remains being interred at a given location. Id-situation o4 relates to a referenced event, call it ref4, to do with the same monarch being buried near a certain city on a particular day in 1485. If we fuse RichardIIIPhy and RichardIIISoc, giving the super-type, say, RichardIII, we thereby relate ref3 and ref4, in effect inferring a new, more specific event, say, ref34, that (unlike ref3) involves a specific day and (unlike ref4) involves a specific location. Where e is an event, let prop(e) be a statement in some sufficiently well-defined language whose sense is e. Then we have prop(ref34) ⇒ prop(ref3) ∧ prop(ref4), where ⇒ and ∧ are the usual truth-functional material implication and conjunction connectives. Also note that the referenced event in (iv) helps us rule out such things as a reburial.
For a systematic account of the interrelation and fusion of identities, we would like something like Barwise and Seligman’s information channels. The “parts” would be the identity environments, and an infomorphism f from classification A to classification B would at least include the type function ∧: ( ) ⎯⎯ ( ). The token function, however, is problematic, as we shall see. The core C of such an “identity channel” would compose the types of the various identity environments (not necessarily as a sum), and it would support the following. Suppose that id-situation o0 in environment E0 is of identity type T0 and that id-situation o1 in environment E1 is of identity type T1. If the translation T0′ of T0 entails the translation T1′ of T1 in the theory Th(C), then o0 being of type T0 carries the information that o1 is type T1.
The problem is that not every token (i.e., id-situation) in one “part” (i.e., environment) can be related to some idsituation in some fixed other environment, nor is there a guarantee that an id-situation in one environment relates to at most one in the other environment For example, perhaps I inspect an email from Ed@Acme.com to Al@Acme.com. There is no reason to expect any id-actions relating to Ed’s or Al’s physical aspects; on the other hand, there might be several such id-actions. Resource situations and referenced events do not help here since they might simply be missing from the record for a given environment or duplicated with trivial changes.
We are forced to accept not functions between sets of tokens but rather relations between sets of tokens: for a given identity environment (i.e., classification) A, where C denotes the core, the relation ∨: ( ) ⟷ ( ) in general relates some elements in tok(C) to multiple elements in tok(A) and relates some to no element in tok(A) (when the id-situations are “missing”). This requires us to relax the sum construction. The “core” is no longer unique, that is, the construction is no longer universal. We have, then, a cocone, but not a universal cocone (which is a sum). The token relations have to be constructed ad hoc for each environment. Some indication of how this is done was given in the above discussion, where, for various cases, we referred to various combinations of the idsituations, resource situations, and referenced events. The type function can remain a total function ∧: ( ) → ( ) and must be such that the type of an id-situation in a given environment is reflected in its type in the core. We call such a function-relation pair ( ∧, ∨) an id-map, denoted by undecorated : → . Instead of an information channel, we might refer to an identity channel.
Each identity proposition c |= γ in the core C can be given a confidence indicating that some identity proposition in one identity environment is evidence for some identity proposition in another environment. Suppose we have environments A and B with id-maps f: A → C and g: B → C. Suppose also that c |= γ in the core has a certain confidence p ∈ [0, 1]. Let = ∨( ), = ∨( ), and ∧( ) = ∧( ) = . Then a |= α carries the information that b |= β with confidence p (and vice versa). (Fig. 3 shows a variation with domains Ai, Aj, and Ak, associated identity propositions ai |= αi, aj |= αj, and ak |= αk, respectively, and id-maps fi, fj, and fk, respectively.) Confidence in a core proposition is less than certainty since tokens corresponding to the connection token are “missing” in some environments. Note, however, that some environments may contribute more tokens, which would generally increase confidence.
Fig. 3. An identity channel with corresponding identity propositions for the “parts” and the “whole” (core). The confidence in the core proposition is the confidence that a proposition in a part carries the information expressed by any of the corresponding propositions in the other parts.
As mentioned, the partition of identity activity into environments is somewhat arbitrary. In particular, there is a general-to-specific hierarchy of environments. Fusion of identity types, therefore, is often a multistage process as one moves up this hierarchy on multiple branches. Indeed, in our sketches of three environments, we noted fusions within the environments. Since what are parts in Fig. 3 could themselves be cores in subsumed identity channels, we would typically associate a confidence less than 1.0 with the identity propositions associated with these parts.
Fusing types generally reduces confidence in identity propositions since fusion is an inductive process. On the other hand, fusion generally increases confidence in the identification of the players in the referenced events and asserted identities. For example, fusing the identity types relating to the email fom and picture of John Doe in Section 8 results in an identity proposition with confidence less than 1.0 but increases the confidence that we have put our finger on a particular person. The theory associated with the core of an identity channel can be thought of as merged constraints that pick out a noumenal individual (a sort of genotype) behind the id-situations that arise (collectively forming various phenotypes) from its interactions with the various environments. As we fuse more and more of the identity types, these constraints progressively narrow the field. Such increase in one aspect accompanied by a decrease in a mirror aspect is characteristic of Galois connections in category theory.
10. Conclusion and Future Work
We presented preliminary work on a framework for constructing identities that takes as its point of departure Barwise’s channel theory, founded on category theory, and allied work on situation semantics and situation theory. The framework takes actions relating to identity, both judgments and assertions, as fundamental. Actions are events (with an agent), and events are basically situations, although situations can include larger regions of space and time and generally accommodate context. Identity-related situations are classified under (identity) types, which can be fused. Identity-related activity is partitioned according to the ‘form of life’ involved into (identity) environments, similar to Barwise’s classifications. An identity-related action is often based on another identity-related action as a resource, and all these actions may reference events in an individual’s life. The result is a layered picture of identityrelated situations, which is easily accommodated in situation theory. The category-theoretical constructs of Barwise’s channel theory, however, must be relaxed when identity in multiple environments is considered. We explained the reasons for this and sketched what needs to be done to recover the category-theoretic foundation.
Identities in this framework retain references to the
situations from which they are constructed, so provenance,
context, and narrative detail are retained. The role of the
“spectator,” as well as those of the “subject” and “sample”
(data or evidence), are acknowledged
We note that how individuals construct their identities in
online settings has become an active area of research