Intuitionistic Higher-Order Logic and Equality


I have posted plenty of motivation for considering equality and higher-order logic elsewhere on this blog by now. Let me finally get to the intuitionistic version of higher-order logic, since I talked about the classical version previously and then also talked at length about "Intuitionistic Logic and Natural Deduction" without really getting closer to the matter. Intuitionistic higher-order logic is what let's you talk about an elementary topos, and using equality to characterize the logical connectives gives an insightful perspective on those commutative diagrams for people who are familiar with $\lambda$ calculus. \( \newcommand{\qeq}[0]{\color{blue} {\mathrm{eq}}} \newcommand{\qSigma}[0]{\color{blue} \Sigma} \newcommand{\qPi}[0]{\color{blue} \Pi} \newcommand{\qor}[0]{\color{blue} {\mathrm{or}}} \newcommand{\tprop}[0]{\omicron} \newcommand{\qtrue}[0]{{\color{blue} {\mathrm{true}}}} \newcommand{\qand}[0]{\color{blue} {\mathrm{and}}} \newcommand{\qimplies}[0]{\color{blue} {\mathrm{implies}}} \newcommand{\qPi}[0]{\color{blue} \Pi} \newcommand{\qfalse}[0]{{\color{blue} {\mathrm{false}}}} \newcommand{\qnot}[0]{{\color{blue} {\mathrm{not}}}} \)

I am going to refer to material in "Higher order logic and equality" often. Like in that post, there is actually no original material here; I am just gathering definitions that is probably well-known to some grad students and experts (yet nobody calls it standard since neither higher-order logic nor intuitionistic logic are known very well outside certain circles who deal with formal logic and proof assistants and the like). This is my feeble attempt at an exposition.

The Logic of a Topos

We take a $\lambda$ calculus with surjective pairing and a type $\omicron$ which we take as a substitute for "truth values" (topos theory texts use $\Omega$ but we want easy comparison with the classical version linked above, and we spell out $\qPi \lambda x. t$ without setting up an abbreviation $\forall x. t$ and we also spell out $\qSigma \lambda x. t$ without setting up an abbreviation $\exists x. t$). We want to set this up in an intuitionistic manner, which means doing away with the idea that $\omicron$ stands for two truth values, and not using classical logical equivalences to define $\qor$ and $\qSigma$. The terms would then correspond to a particular cartesian-closed category, the free topos.

We again want to define everything in terms of equality $\qeq_\alpha: \alpha \rightarrow \alpha \rightarrow \omicron$. What is this equality doing? In our langugage, it talks about equality of terms, and in categorical logic, terms correspond to arrows and thus equality should talk about equality of arrows. (There is some detail hidden here as our language will also let us talk about "global elements" $1 \stackrel{c}{\longrightarrow} C$ which are written as constant terms $c: C$). We take the following elements of our logical language: $$ \begin{align*} \qtrue:\ & \omicron\\ \qand:\ & \omicron \rightarrow \omicron \rightarrow \omicron\\ \qimplies:\ & \omicron \rightarrow \omicron \rightarrow \omicron\\ \qPi_\alpha:\ & (\alpha \rightarrow \omicron) \rightarrow \omicron\\ \end{align*} $$

... and define them using the only primitive $\qeq$: $$ \begin{align*} \qtrue &:= \qeq_{\tprop \rightarrow \tprop \rightarrow \tprop} \qeq_\tprop \qeq_\tprop\\ \qand &:= \lambda x:\tprop. \lambda y:\tprop. \\ &\quad \quad \qeq_{\tprop \rightarrow \tprop \rightarrow \tprop} \\ &\quad \quad \quad (\lambda z: \tprop \rightarrow \tprop \rightarrow \tprop. z~\qtrue~\qtrue) \\ &\quad \quad \quad (\lambda z: \tprop \rightarrow \tprop \rightarrow \tprop. z~x~y) \\ \qimplies &:= \lambda x:\tprop. \lambda y:\tprop. \qeq~x~(\qand~x~y)\\ \qPi_\alpha &:= \qeq_{(\alpha \rightarrow \tprop) \rightarrow (\alpha \rightarrow \tprop) \rightarrow \tprop}~(\lambda x : \alpha. \qtrue) \end{align*} $$

These definitions are the same as for classical logic. An intro to topos theory will typically describe $\qand$ and $\qimplies$ using commutative diagrams, but here we are defining the language first. Moreover, the language has only one primitive $\qeq$, though that primitive is instantiated at different types.


Now we look at $\qfalse$, which is also defined the same as in classical case: $$ \begin{align*} \qfalse &: \omicron\\ \qfalse &:= \qeq_{\tprop \rightarrow \tprop} (\lambda x:\tprop. \qtrue) (\lambda x:\tprop. x)\\ \end{align*} $$

In a topos theory text, one encounters $\forall v:\omicron. v$, which is shorthand for $\qPi\lambda v:\omicron. v$ and actually expands to the very same definition above. The definition that mentions $\qeq$ explicitly has the advantage that it explains what is going on: it morally says "if there is anything other than $\qtrue$ in $\omicron$, then $\lambda x:\omicron. \qtrue$ and $\mathsf{id}_\omicron$ are not the same." So we have a value $\qfalse$, but it need to be the only thing distinct from $\qtrue$ in $\omicron$.

Not, Or and Exists

On to intuitionistic negation $\qnot$. In intuitionistic logic, we would define negation like this: $$ \begin{align*} \qnot &: \omicron \rightarrow \omicron\\ \qnot &:= \lambda x: \tprop. \qimplies(x, \qfalse) \end{align*} $$

If we consult the classical version $\qnot := \qeq_{\tprop \rightarrow \tprop} \qfalse$, we again find that this is not that different.

Finally, let's look at $\qor$ and $\qSigma$. We define these using $\qPi$: $$ \begin{align*} \qor &: \omicron \rightarrow \omicron \rightarrow \omicron\\ \qor &:= \lambda x: \omicron. \lambda y: \omicron. \qPi \lambda r:\omicron. \\ & \quad \quad \qimplies(\qand(\qimplies(x, r), \qimplies(y,r)), r) \\ \\ \qSigma_\alpha &: (\alpha \rightarrow \omicron) \rightarrow \omicron \\ \qSigma_\alpha &:= \lambda x: \alpha \rightarrow \omicron. \qPi \lambda r:\omicron. \\ &\quad \quad (\qPi \lambda z:\alpha. \qimplies(\qimplies(x z, r), vr) \end{align*} $$

We can get an intuitive understanding of these definitions: for $\qor$, if we have a way of obtaining an element $r$ (result) from $x$ and one of obtaining the same from $y$, then also from $(\qor~x~y)$. Similarly for $\qSigma$, if we have a way of obtaining any result $r$ from the "instantiated" predicate $x z$, then also from $\qSigma x$.

The definition of $\qor$ in terms of $\qand$ and $\qPi$ is not very different from natural deduction rule for "or elimination": $$ \cfrac{\begin{array}[cc]{} \\ {} \\ A \vee B \end{array} \quad \begin{array}[cc]{} [A] \\ ~~\vdots \\ \phantom{[}C\phantom{]} \end{array}\quad \begin{array}[cc]{} [B] \\ ~~\vdots \\ \phantom{[}C\phantom{]} \end{array} } {C}~\color{blue}{(\vee{\mathrm{\small ELIM}})} $$

Checking the the natural deduction rule for existential elimination and comparing it with the encoding above is left as an exercise.

The end of the beginning

Higher order logic seems to deserve more attention. We have set up a logical system based on $\lambda$ calculus without making use of the Curry Howard correspondence. While the previous discussion focused on classical higher order logic as introduced by Church in 1940, in this post we looked at intuitionistic higher order logic. The interest in the latter is at least very heavily influenced by topos theory.

It is slightly confusing that when discussing elementary toposes, one talks of first-order logic when in reality the logic is inherently higher-order (because a topos is cartesian closed). I guess the charitable perspective is what people mean is "you can do everything that you would want to do in formal set theory using first-order logic."

The Lambek and Scott book "Introduction to higher order categorical logic" dwells a bit on the fact that it is always possible to get a formulation of the internal language of a topos using equality. Andrews $Q_0$ develops simple type theory / higher order logic using a single equality operator $Q$. Various theorem proving systems have been based on this (TPS and HOL).

One could say that the way higher-order logic admits the definition of logical connectives in terms of other is an "encoding" of natural deduction rules. Could one use this idea to build a logical framework, a representation of derivations? Indeed this is what is underlying Isabelle/Pure framework and $\lambda$Prolog.

It would seem worthwhile to write a proper book that takes a logical perspective on the $\lambda$ calculus, developing higher order logic as language, category theory as semantics and arriving at elementary toposes as models. It would probably feel very elementary and pedestrian in the academic world, but I do think it would be useful. It feels to me that there is so much "background knowledge" to gather the relevant theory on logic and computation that most practitioners will not have the energy (and this says something about the potential for adopting of formal methods). Until such a book arrives (I'm not sure if I would have the energy), I hope my out-of-order blog posts provide some help on the reader's journey. There are certainly a lot of of connections between logic, language, types and semantics that are left to be explored, besides Curry Howard and dependent types.


Programmers love Higher-Order Logic

My previous take on higher-order logic was classical, channeling Church's simple theory of types and Andrews $Q_0$. In that last post, I was hinting at motivation. So I want to expand a bit now how this stuff can be useful for building software at large. I am posting some ramblings here.

Equality is Fundamental

There are two different "levels" at which a programmer deals with some kind of reasoning that involves equality:

  • while writing code: need to be able to tell when two objects are the same.
  • when designing or arguing that code is correct: reasoning whether behavior like output values conform to a (however implicit) specification

In both cases, one needs to be able to identify whether objects (or behaviors or input-output pairs) are "the same" as expected ones. But the latter activity, reasoning about program correctness, happens in some form of logic. Much that has been written and is being written on formal logic is about mathematics and mathematical convention. This is true whether we talk about semantics of programming languages or program specification.

But for programmers, there are benefits to knowing formal reasoning even if one is not interested in the foundations of math, axiomatizations of set theory etc. It is no secret that in mathematics as in programming, you run into situations where there are multiple different ways to represent some data, and one needs to chose or translate between those representations. While programmers resign to writing "proto-to-proto" translation over and over (which often are not as trivial as it seems), in mathematics it is customary to leave out the awkward details, introduce "abuse of notiation" and quotienting to keep arguments informal, but concise.

The heart and soul of much mathematics consists of the fact that the “same” object can be presented to us in different ways. [1]

[1] Barry Mazur. When is one thing equal to some other thing?

Equality in First Order Logic

We take a leap towards "formal systems", where objects are represented using (strings of) symbols, along with strict rules to form statements and derive new true statements from existing ones.

A typical definition of a first-order logic language contains constant and function symbols, predicate symbols, logical operators.

There is a "design choice" here whether equality is considered as built-in, magically denoting exactly the equality in the underlying sets, or whether equality is a predicate with certain special properties. Wikipedia" talks about "first-order logic with equality" (when built-in) and first-order logic without equality (when it is not).

When equality does not come built-in, one needs to add axioms. The equality axioms are special, in the sense that they have to be preserved by all the function and relation symbols. This looks like:

  • $\mathsf{eq}$ is reflexitve, symmetric, transitive
  • if $\mathsf{eq}(x_1, y_1), \ldots, \mathsf{eq}(x_n, y_n)$, then $\mathsf{eq}(f(x_1,\ldots,x_n),f(y_1,\ldots,y_n)$ for all function symbols $f$ of arity $n$
  • if $\mathsf{eq}(x_1, y_1), \ldots, \mathsf{eq}(x_n, y_n)$, then $R(x_1,\ldots,x_n) \Leftrightarrow R(y_1,\ldots,y_n)$ for all relation symbols $R$ of arity $n$

With a "defined" equality, is it is well possible to interpret logical statements involving $\mathsf{eq}$ in such a way that two distinct elements are regarded as equal according to $\mathsf{eq}$.

Coding our own equality

When we are talking about practical software development, there is really no hope of defining a platonic ideals; yet when techniques of reasoning, informaticians are in various ways confined to results and methods that formalizing mathematics.

Take this example: We'd like our terms to represent free monoid on $X$, so we'd like these terms to satisfy the following equations (or identities): $$s \cdot 1 = s$$ $$t = 1 \cdot t$$ $$(s \cdot t) \cdot u = s \cdot (t \cdot u)$$

Consider definitions in some fictional language to express the "interface" of a monoid: $$ \begin{array}{ll} \mathsf{interface}~ M[X] \\ \quad \mathsf{unit}(): M[X]\\ \quad \mathsf{emb}(x: X): M[X]\\ \quad \mathsf{dot}(s: M[X], t: M[X]): M[X]\\ \quad \mathsf{eq}(s: M[X], t: M[X]): \mathsf{bool} \end{array}$$

Here, $\mathsf{unit}$ stands for $1$, $\mathsf{emb}(x)$ is for embedding $x$ and $\mathsf{dot}$ is the combine operaion in $MX$. If our fictional language has algebraic data types, we can use those to build terms which represent the algebraic theory of monoids: $$ \begin{array}{ll} \mathsf{data}~ \mathsf{Mon}[X] \\ \quad \mathsf{Unit}(): \mathsf{Mon}[X]\\ \quad \mathsf{Emb}(x: X): \mathsf{Mon}[X]\\ \quad \mathsf{Dot}(s: \mathsf{Mon}[X], t: \mathsf{Mon}[X]): \mathsf{Mon}[X]\\ \quad \mathsf{eqMon}(s: \mathsf{Mon}[X], t: \mathsf{Mon}[X]): \mathsf{bool} \end{array}$$

It is now really obvious what these constructors stand for, but we note that while the equality $\mathsf{eqMon}$ returns true whenever two terms are equal when considered as elements of the free monoid, there is at least one more equality relation on terms that is different ("structural equality" or "syntactic equality", which will consider $\mathsf{dot}(\mathsf{dot}(s, t), u)$ as distinct from $\mathsf{dot}(s, \mathsf{dot}(t, u))$).

Mathematically speaking, we want to consider the quotient set $\mathsf{Mon}[X] / \mathsf{eqMon}$, and work with equivalence classes.

Formulas = Functions to Truth Values

After all this talk about equality in algebraic theories, we want to turn equality in the context of logic, specifically higher-order logic.

In classical logic, a proposition (logical statement) stands for a truth value, and there are exactly two of then (the two-value type "bool"). Boole was the first to associate propositional connectives with algebraic operations. In this perspective, the the truth values are part of the language, even though this was later abandoned as in Tarski-style semantics.

Programmers are quite used to having functions return $\mathsf{bool}$. In terms of mathematical logic, our custom-made equality "relation" $\mathsf{eqMon}$ above is in fact a function to the set of truth values. Could we not do away with relations and talk in terms of functions?

Yes, we can; using truth values and functions in a formal language of propositions is the essential summary of classical higher order logic (or Church's simple theory of types. There is a more compelling example: if you consider "set builder notation" $$ \{ x \in A \ |\ \phi(x) \} $$ then this describes a subset of $A$. We can choose to describe $\phi$ as defining a (unary) relation, but viewing it as a function to $\mathsf{bool}$ is no doubt more natural in a context like SELECT x FROM A WHERE $\phi$ where "relations" are given by their extensions (as tables) and computation in the form of expressions is available.

The point of (classical) higher-order logic is that using the $\lambda$ calculus and truth values, we have everything we need to define the meaning of all logical connectives, and the types prevent logical paradoxes of "naively-untyped set theory". Since the truth values are part of the language, it is possible to define the meaning purely through definitions that involve equality.

More on this and classical higher order logic in that last post on Higher Order Logic and Equality and the next one I'm going to write in a second, where I talk about the intuitionistic version.

I will just end with a note that there is no distinction between "predicate", "relation" and "set" is higher-order logic, and this is a good thing. An expression of type $\iota \rightarrow \omicron$ (where $\iota$ are the individuals and $\omicron$ the truth values) defines a (sub)set of individuals, a predicate on the set of individuals, a unary relation, all at the same time. $$\mathsf{even} := \lambda x:\iota. \Sigma \lambda y:\iota. \mathsf{eq}(x, \mathsf{times}(2, y))$$ $$\mathsf{greater} := \lambda x:\iota.\lambda y:\iota. \Sigma \lambda z:\iota. \mathsf{eq}(x, \mathsf{plus}(y,z))$$ $$\mathsf{lowerbound} := \lambda x:\iota.\lambda y:\iota \rightarrow \omicron. \Pi \lambda z:\iota. \mathsf{implies}(y(z), \mathsf{greater}(z, x))$$

The confusing bit for programmers may be that this thing is called simple type theory when in fact there is only one "sort" of individuals. This makes more sense when considering that it is no problem to come up with multiple "sorts" of individuals but that wouldn't give sets ($\iota \rightarrow \omicron$), sets of sets $(\iota \rightarrow \omicron) \rightarrow \omicron$, sets of sets of sets (and so on) of individuals. It's this distinction of hierarchies of sets/predicates/relations that earn Church's higher order logic the name "simple type theory": only terms with the right types are considered well-formed and meaningful.