Henkin/branching quantifier: say what?

421 Views Asked by At

This is a substantial revision of the original question, to cut out material that turned out to be irrelevant. Most of the comments and answers and comments-to-answers relate to the O.P. so are no longer applicable.

The question was triggered by a bald sub-heading in this informal write-up

A Problem With First Order Logic

The author doesn't say what specific example he was working with, but gives a simpler case

$$ S(x,y,a,b) \text{ could be } (a^2 = x) \wedge (b^2 = y) $$

He wants to express $\forall x, y ... [S(x, y, a, b)]$ in which $a$ is dependent purely on $x$, $b$ dependent purely on $y$; and those (in)dependencies is what he wants to express in the $...$.

That simpler case using roots is expressible in FOL, so is too over-simplified to illustrate anything.

He goes on to introduce 'Henkin' aka 'Branching' quantifiers. So his general case could be expressed as

$$ \pmatrix{\forall x \exists a\\ \forall y \exists b}S(x, y, a, b) $$

Although this appears to involve quantifying only over individuals, its semantics are in fact given in (a restricted form of) Second-Order. So the sub-heading's 'Problem with First Order Logic' is that it is not Second Order. Nothing to see here; move right along.

Some relevant resources: wikipedia branching quantifier; Stanford Encyclopedia Dependence logic; a critique of so-called 'Hintikka sentences' with their claim of non-expressibility in FOL. I found math.stackexchange had very little, hence this question. Google has lots of hits but mostly about Henkin's contributions to logic in general.

Then the nub of my question (before I realised a Henkin quantifier is merely Second-Order in sheep's clothing) was: I just don't get what the examples in those write-ups are trying to say that can't be said in FOL.

tl;dr: some Second-Order formulas (including some using Henkin quantification) can be equivalently expressed in FOL; some can't. Again nothing to see here. The literature (including in the Stanford Encyclopedia) mostly shows examples that in fact can be expressed in FOL. Again perhaps over-simplifying for the sake of explanation.

@Noah's answer below brings forward an example entailing an infinitary model; it is substantially more complex than the natural language examples in the literature.

A note for non-maths readers: I came at this from Dependency Theory in the database Relational Model, which is entirely equivalently expressible in FOL. That's a different sense of "Dependenc/y".

3

There are 3 best solutions below

24
On

EDIT: Based on the comments below, let me explain in detail the key point:

There is no first-order sentence $\theta$ which is true in exactly the infinite structures.

Proof: Suppose $\theta$ is a first-order sentence true in exactly the infinite structures. We'll consider its negation, $\neg\theta$; by our assumption on $\theta$, $\neg\theta$ is true in exactly the finite structures.

So what? Well, we use the compactness theorem. This is in my opinion the most important basic result in logic, and is definitely the first thing you want to master to understand first-order logic and its extensions. The compactness theorem says that if $\Gamma$ is a set of first-order sentences, and every finite subset of $\Gamma$ has a model, then $\Gamma$ itself has a model.

For $n\in\mathbb{N}$, let $\chi_n$ be the sentence $$\exists x_1\exists x_2...\exists x_n(\bigwedge_{i<j\le n} x_i\not=x_j).$$ Each $\chi_n$ is a first-order sentence which is true in exactly those structures with at least $n$ elements. Now let's think about the set of sentences $$\Gamma=\{\chi_n:n\in\mathbb{N}\}\cup\{\neg\theta\}.$$

Any finite subset $\Gamma_0\subseteq\Gamma$ is satisfiable: if we let $k\in\mathbb{N}$ be large enough so that no $\chi_m$ is in $\Gamma_0$ for $m\ge k$ (note that such a $k$ exists since $\Gamma_0$ is finite!), we see that any structure of size at least $k$ satisfies $\Gamma_0$. But $\Gamma$ itself has no model: any model of $\{\chi_n: n\in\mathbb{N}\}$ must be infinite, but no infinite structure is a model of $\neg\theta$. So we have our contradiction. $\quad\Box$

Note that we can write this as a direct proof instead, showing:

If $\eta$ is a first-order sentence with arbitrarily large finite models, then $\eta$ has an infinite model.

(Above, take $\neg\theta$ to be $\eta$.)

The two crucial facts we've used about first-order logic are:

  • The compactness theorem, and

  • the fact that first-order logic has negation.

It turns out, surprisingly, that the latter is the issue with dependence(-style) logic(s)! But that's something we should look at only after we understand the basic argument above in detail.


The bulk of this question I find difficult to follow; let me address the main question:

I just don't get what the examples in those write-ups are trying to say that can't be said in FOL.

I think it's easier to just do an explicit example: a sentence $\varphi$ using a Henkin quantifier, in the empty language, which is true exactly if the ambient domain is infinite. Namely, $$\varphi\equiv\exists zQ(x, y; a, b)[a\not=z\wedge (x\not=y\iff a\not=b)].$$ (Here I'm writing "$Q(x, y; a, b)$" to denote "for all $x, y$ there are $a, b$ depending only on $x, y$ respectively such that.")

OK, it's probably not clear what $\varphi$ means, so let's take it apart.

The semantics of branching quantifiers is best described, in my opinion, in terms of Skolem functions. According to this semantics we can express $\varphi$ in second-order logic as $$\hat{\psi}\equiv\exists F_1, F_2, z\forall x, y[(F_1(x)\not=z)\wedge (F_1(x)=F_2(y)\iff x=y)].$$ However, I claim we can do one better, and express it as $$\psi\equiv \exists z, F\forall x, y(F(x)\not=z\wedge (x\not=y\implies F(x)\not=F(y))).$$ (Note that here "$\iff$" isn't needed since we're only using one function symbol now - "$x=y\implies F(x)=F(y)$" is a tautology.) Clearly we have $\psi\implies\hat{\psi}$ (take $F_1=F_2=F$); conversely, the point is that in order for a pair of Skolem functions $F_1, F_2$ to witness that $\hat{\psi}$ is true we must have $F_1=F_2$ (if $F_1(a)\not=F_2(a)$ for $a$ in the domain, then $x=a, y=a$ would give a counterexample to $\psi$).

Now looking at $\psi$, we see that $\psi$ asserts the existence of an injection from the model to itself (namely, $F$) which is not surjective (it misses $z$). So $\psi$ is a sentence true in every infinite structure but false in every finite structure, and by the compactness theorem (and the argument above) this can't happen in a first-order way.


Let me try to give a quick "moral" of the story. The crucial point of difference between first-order and dependence sentences here is:

When Skolemizing a first-order sentence in prenex normal form, each existential quantifier corresponds to a function taking as inputs all variables in prior universal quantifiers.

E.g. the Skolemized version of the first-order sentence $$\forall x\forall y\exists z\forall u\exists v(\theta(x, y, z, u, v))$$ is $$\exists F\exists G\forall x\forall y\forall u(\theta(x, y, F(x, y), u, G(x, y, u))).$$

By contrast, in a dependence sentence we have more freedom with regard to how the Skolem functions are "entangled," and we've seen that in play above.


Now let me try to explain the claim that dependence logic only involves quantifying over individuals (a claim, incidentally, that I disagree with). What dependence logic does is argue that general expressions of the form $$\mbox{[second-order existential quantifiers](first-order formula)}$$ form a natural generalization of first-order sentences while staying in the realm of quantification over individuals, since all the new power we get this way really just comes down to expressing dependencies. (These are in fact essentially the sentences expressible in dependence logic.)

I haven't treated here the question of whether dependence logic is in fact more first-order or more second-order. Vaanaanen has written at length in defense of the former position; personally, I take the latter view. But you asked specifically for something expressible using dependencies which isn't first-order expressible, and I've done that above. So hopefully this is helpful.

3
On

This is a response to/critique of @Noah's answer and comprehensive explanation (thank you again). I offer an alternative formulation, which is why I'm posting as an answer.

tl;dr:

  • I can see Noah's formula $\psi$ is true only in infinite structures,
    and skolem function $\mathbf{F}$ must be injective but not surjective.
    I give my workings, to verify I'm on the right page.
  • In moving from the Henkin-quantified $\varphi$ to $\psi$, there seems to be a step omitted.
    I give grounds why that makes them not equivalent.
  • I give a formulation for $\varphi$ which is first-order.

    Edit: Per @DanielV's comment on Noah's answer, I've changed the formulas to use $\Longleftrightarrow$ rather than $\Longrightarrow$.

Only infinite structures:

For definiteness, take the domain to be the Naturals. The formula begins $\psi\equiv\exists z,\mathbf{F} ...$. We can take $z$ as some arbitrary $n \in \mathbb{N}$. For $\psi$ to be true, the skolem function $\mathbf{F}$ must never produce $n$, because then $a\neq z$ would fail. Then we could choose:

$$ \begin{align} & \mathbf{F}(x) = x + 1 & , x = n\\ & \mathbf{F}(x) = x & , x < n & \hspace{20pt}\text{-- no danger of producing n}\\ & \mathbf{F}(x) = x + 1 & , x > n \end{align} $$

Why produce $x + 1$ if already $x > n$? Remember $\mathbf{F}$ is blind as to whether it's applying to $x$ so producing $a$ or applying to $y$ so producing $b$. Consider the case in $P(z, x, y, \mathbf{F}(x), \mathbf{F}(y))$ of $x = n, y = n+1$. We have $x\neq y$ so for the implication to hold we must have $\mathbf{F}(x) \neq \mathbf{F}(y)$. Then consider $x = n+1, y = n+2$, and so on. So $x+1$ ripples all the way up the domain. We cannot choose to stop short of infinity.

$\mathbf{F}$ must be not surjective because it must never produce $n$; otherwise $a\neq z$ would fail. $\mathbf{F}$ must be injective because for each distinct input it must produce a unique result; otherwise the implication $(x\neq y \Longleftrightarrow F(x)\neq\mathbf{F}(y))$ would fail.

Step from Henkin-quantified to Second-Order skolem function:

For $\varphi\equiv\exists z, \mathbf{Q}(x,y;a,b)[ ... ]$ Noah writes

(Here I'm writing $\mathbf{Q}(x, y; a, b)$ to denote "forall $x,y$ there are $a,b$ depending only on $x,y$ respectively such that.")

If this were FOL, we'd say we could represent that dependence by two skolems: one for $a$, one for $b$.

In more natural language, $\varphi$ states the existence of a function from the domain to itself [**] which is injective but not surjective.

I agree with the "injective but not surjective", see above. I do not agree with the "a function" [singular]. We have two skolems, I expect two functions. I think $\varphi$ should be:

$$ \varphi'\equiv\exists z,\mathbf{F}', \mathbf{G}'\forall x,y[\mathbf{F}'(x)\neq z\wedge(x\neq y \Longleftrightarrow \mathbf{F}'(x)\neq\mathbf{G}'(y))]. $$

I can see of course that $\varphi'$ will only be true providing $\mathbf{F}', \mathbf{G}'$ are the same function. I've no idea how to prove that -- especially in an infinite model.

Note this has $\exists \mathbf{G}'$ quantified inside $\exists z,\mathbf{F}'$. So $\mathbf{G}'$ is dependent on $z$ (fair enough) but also dependent on $\mathbf{F}'$. Of course $\mathbf{G}'$ is a function therefore whose result is produced purely from $y$. The choice of function for $\mathbf{G}'$ is dependent on the choice of function for $\mathbf{F}'$. (They must be the same.) We seem to be back at the problem we started with. (Edit: No, the natural language 'dependent on' isn't the technical Dependence Logic sense. The $\exists$ can be permuted without losing logical equivalence; so they're not dependent.)

Noah does consider the way two skolems would be treated in FOL, as two functions. This example is carefully constructed so that we can use the same function $\mathbf{F}$ to produce both dependent values. That won't be the case in general.

I'm pointing this out because in all the examples I'm looking at of 'Hintikka sentences' Section 1, (reference from wikipedia on 'Branching quantifier'), I see exactly this sort of entanglement of choices to instantiate the dependent variables. Perhaps they're all poorly chosen examples. I'm increasingly thinking the Lipton 'less formal write-up' is poorly chosen.

EDIT: Furthermore the semantics Henkin gives for his own quantifier introduce two skolem functions -- that's the same paper, top of p 368.

([**] BTW neither do I agree with "from the domain to itself". We could have $z, a, b$ in a distinct domain from $x, y$. Then $x, y$'s domain could be finite; $z, a, b$'s would need to be larger by at least one. How does that affect the justification?)

First-Order formulation:

$$\begin{align} \exists z & [ \forall x \exists a \forall y \exists b & [a\neq z\wedge(x\neq y\Longleftrightarrow a\neq b)\\ & & \wedge \forall x', y', a', b' & [a'\neq z\wedge(x'\neq y'\Longleftrightarrow a'\neq b')\\ & & & \Longrightarrow (a\neq z\wedge(x\neq y'\Longleftrightarrow a\neq b') \wedge a'\neq z\wedge(x'\neq y\Longleftrightarrow a'\neq b)] ]\\ & \wedge \forall y \exists b \forall x \exists a & [a\neq z\wedge(x\neq y\Longleftrightarrow a\neq b)\\ & & \wedge \forall x', y', a', b' & [a'\neq z\wedge(x'\neq y'\Longleftrightarrow a'\neq b')\\ & & & \Longrightarrow (a'\neq z\wedge(x\neq y'\Longleftrightarrow a\neq b') \wedge a'\neq z\wedge(x'\neq y\Longleftrightarrow a'\neq b)] ]\\ & ]. \end{align} $$

The two branches mixing up the $∀x ∃a ∀y ∃b$ vs $∀y ∃b ∀x ∃a$ are per the strong form formula (5) in the 'Hintikka sentences' paper above. They ensure we get (a version of the assertion with) both $a$ dependent purely on $x$ and $b$ dependent purely on $y$.

The embedded $∀x', y', a', b'$ within each branch says: whereever you get $x', y', a', b'$ such that $a'\neq z\wedge(x'\neq y'\Longleftrightarrow a'\neq b')$ holds, you must be able to permute the dependencies $x ↦ a, y \mapsto b$.

Then the only solutions for which the overall formula holds is when $b$ is chosen independently of $x$, $a$ is chosen independently of $y$.

That is, if that is a genuine requirement in this example. (I'm suspecting not.)

3
On

something couldn't be expressed in FOL, but still only involved quantifying over individuals, not predicates (which would make it Second-Order).

(That's the opening of my question.)

The semantics of the Henkin quantifier are given in wikipedia 'Branching quantifier' Definition. But I didn't want to rely entirely on that. It turns out the original Henkin 1961 paper (behind a pay wall) explored a family of quantifiers. This is the most basic. The semantics in wikipedia agree with this critique cited by wp and this survey.

The simplest Henkin quantifier $Q_{H}$ is

$$ (Q_{H}x_{1},x_{2},y_{1},y_{2})\phi (x_{1},x_{2},y_{1},y_{2})\equiv {\begin{pmatrix}\forall x_{1}\exists y_{1}\\ \forall x_{2}\exists y_{2}\end{pmatrix}}\phi (x_{1},x_{2},y_{1},y_{2}). $$

It (in fact every formula with a Henkin prefix, not just the simplest one) is equivalent to its second-order Skolemization, i.e.

$$ \exists f\exists g\forall x_{1}\forall x_{2}\phi (x_{1},x_{2},f(x_{1}),g(x_{2})). $$

Then Henkin has introduced some notation mentioning only individuals, that he's defining as equivalent to a Second-Order formula. That's fine: the claim is there's a Second-Order formula that (in general) can't be expressed in FOL. No need to question that.

Of course any FOL formula with $\exists$ quantification of individuals might be written using Second-Order Skolemisation functions. Again nothing to see here. Whether the Lipton 'less formal write-up' case can be expressed in FOL is a side-issue: Lipton has perhaps chosen a poor example which doesn't need the expressive power of a Henkin quantifier.

Noah's answer gives an informal description of the Henkin quantifier:

(Here I'm writing "$Q(x,y;a,b)$" to denote "for all $x,y$ there are $a,b$ depending only on $x,y$ respectively such that.")

Does Henkin's Skolemisation $\exists f \exists g$ achieve the effect of '$a$ depending only on $x$; $b$ depending only on $y$'? (To paraphrase slightly.)

Certainly $f(x)$ depends only on $x$ and $g(y)$ depends only on $y$. But putting $\exists f \exists g$ has merely introduced a different dependency: $f, g$ must be chosen together.

EDIT: What is a dependence, anyway?

In that last paragraph, I used "dependency" in a natural language sense. But "(in)dependent" seems to have a technical sense here: related-but-different to the natural language sense or the database Dependency Theory sense. To explain by example:

  • A $\forall$ quantified variable is always dependent only on the domain; therefore 'independent' of any other variable.
  • Two adjacent $\exists$ quantifications are not 'dependent', because permuting them yields a logically equivalent formula. That is: $$ \exists x \exists y \phi \equiv \exists y \exists x \phi $$

    Then this is what the Henkin Skolemisation is relying on to claim it has $f, g$ 'independent' -- in this technical sense.

  • Adjacent $\forall x \exists y$ cannot in general be permuted and retain logical equivalence. $$ \forall x \exists y \phi \not\equiv \exists y \forall x \phi $$

    The rules of this language game have already decided the $\forall x$ isn't 'dependent', so it must be that permuting $\exists y$ to its right makes $y$ 'dependent'. Informally $\forall x \exists y$ means we might need to choose a different $y$ for each $x$ in order to make $\phi$ true for that $x$. Whereas $\exists y \forall x$ asserts we need only find a single $y$ such that $\phi$ holds for each $x$.

OK. The test is logical equivalence of formulas. Then in general $$ \forall x_{1} \exists y_{1} \forall x_{2} \exists y_{2} \phi \not\equiv \forall x_{2} \exists y_{2} \forall x_{1} \exists y_{1} \phi $$

And we have to say the rightmost $\exists y_{n}$ is 'dependent' on all the $\forall x_{n}$ to its left.

But, but, but! If the test is logical equivalence (and only that) we can easily arrange to say we require the two formulations to be equivalent:

$$\begin{align} & \forall x_{1} \exists y_{1} \forall x_{2} \exists y_{2} \phi\\ \wedge & \forall x_{2} \exists y_{2} \forall x_{1} \exists y_{1} \phi\\ \wedge & (\forall x_{1} \exists y_{1} \forall x_{2} \exists y_{2} \phi \equiv \forall x_{2} \exists y_{2} \forall x_{1} \exists y_{1} \phi) \end{align}$$

I just expressed in purely FOL that each $y_{n}$ is dependent only on its corresponding $x_{n}$. This formula is not in Prenex Normal Form; and converting to it would indeed smash all the 'dependencies'. We already know that PNF preserves logical equivalence under classical logic, but not necessarily under other logics. Then Dependence Logic is another example.