Existence of well-ordering of epsilon_0 in weak theories

In a discussion on a youtube video on the hydra game I jokingly mentioned how everyone was assuming that $\varepsilon_0$ was well-ordered. This lead to a bit of disagreement (in a nice way!) about the existence of the well-ordering of $\varepsilon_0$, assuming that $\varepsilon_0$ exists in whatever weak theory one was working in. The person with whom I was discussing seemed to think it obvious that if $\varepsilon_0$ existed, then it was an ordinal, hence well-ordered. I am very skeptical, to say the least. To clear things up, I said I would ask here:

Is it possible, in some suitably weak theory, to define (by a code, or outright) the object $\varepsilon_0$ whose elements (or 'elements') are the usual countable ordinals given by Cantor normal form using only smaller ordinals, but the well-ordering on $\varepsilon_0$ is not available in the theory? What is a good reference for the answer, in either case?

To me this seems like it should be true, but I can't pin down what the required weak theory is. Some form of second-order arithmetic seems reasonable to try, but I don't want one that implies the consistency of (first-order) PA, since then clearly $\varepsilon_0$ is well-ordered.

• I'm not sure I understand the question. The order on $\epsilon_0$ as such can be constructed in very weak fragments of PA, but those theories cannot prove it is a well order (i.e., satisfies appropriate forms of induction). – Emil Jeřábek Mar 6 '17 at 9:19
• @Emil this is exactly the sort of answer I'm after. Details would be good, to settle the argument conclusively. I don't even care about the ordering, merely the existence of $\varepsilon_0$, but the actual definition is perhaps important for my purposes. – David Roberts Mar 6 '17 at 9:39
• @EmilJeřábek or, perhaps you mean the order type of $\varepsilon_0$ can be defined in a weak fragment of PA, so we can consider $\varepsilon_0$ to be 'defined' in this way... – David Roberts Mar 6 '17 at 10:50
• The comment from @EmilJeřábek answers the question, but it might be worthwhile to amplify it to say "but those theories and even PA itself cannot prove it is a well order." – Andreas Blass Mar 6 '17 at 16:05

The existence of $\epsilon_0$ and its order is not a problem, its well-foundedness is.

Cantor normal forms (recursively expanded) of ordinals below $\epsilon_0$ can be written as strings over a finite alphabet, and in that form can be manipulated in weak fragments of arithmetic, say in $I\Delta_0+\mathrm{EXP}$. (Even much weaker theories would suffice, such as a theory of polynomial-time functions.)

That is, $I\Delta_0+\mathrm{EXP}$ can define $\epsilon_0$ (as a “definable class” of Cantor normal forms) in a natural way by a low-complexity ($\Delta_0(\exp)$) formula, and it can also define the order $\alpha<\beta$ on its elements, and basic operations like $\alpha+\beta$, $\alpha\cdot\beta$, and $\alpha^\beta$. The theory can prove elementary properties of these operations such as associativity, and in particular, it can prove that $<$ is a linear order. When interpreted in the standard model of arithmetic $\mathbb N$, these definitions give a structure isomorphic to the actual $(\epsilon_0,<,+,\cdot,x^y)$ from the outside world.

In a “second-order” theory of arithmetic like $\mathrm{RCA}_0^*$, we can even turn the above definition of $\epsilon_0$ into an actual object of the theory, rather than just a definable class.

However, what these relatively weak theories cannot prove is that $<$ is a well order, i.e., that it is well founded. More precisely, well foundedness is a second-order property, and thus can be properly stated only in second-order theories: $\forall X\,[\forall\alpha\in\epsilon_0\,((\forall\beta<\alpha\,\beta\in X)\to\alpha\in X)\to\forall\alpha\in\epsilon_0\,\alpha\in X]$. In first-order theories of arithmetic, it is approximated by a transfinite induction schema; $I\Delta_0+\mathrm{EXP}$, or even $\mathrm{PA}$, cannot prove the $\epsilon_0$-induction schema even for formulas of low complexity ($\Delta_0$).

• +1. And for the OP, although (an order of type) $\epsilon_0$ is definable as a class in PA in an appropriate sense, we can improve this - working in ACA$_0$, a conservative extension of PA, there actually is an object which is a linear order of type $\epsilon_0$ (again, in an appropriate sense). The "appropriate sense" is that the class/object needs to be interpreted in the standard model. – Noah Schweber Mar 6 '17 at 16:41
• @NoahSchweber thanks, that extra detail helps. – David Roberts Mar 6 '17 at 22:51
• Emil, while the weak theories and even PA do not prove that the $\epsilon_0$ order on representations is a well order, could you tell me whether they nevertheless prove of every particular computable sequence (or other complexity-limited sequence), that it is a not a descending sequence in the order? In other words, in the models where $\epsilon_0$ is not well-ordered, what is the complexity of the simplest counterexample to well-foundedness? I guess this is closely related to the failure of induction that you mention in your final remark. – Joel David Hamkins Mar 7 '17 at 0:03
• @JoelDavidHamkins I don't think that's the case. I suspect that via the equivalence WF($\epsilon_0$) iff Con(PA), any model in which $\epsilon_0$ is illfounded will have a computable (with standard index, even!) descending chain, corresonding to a putative proof of inconsistency in PA (although that proof will be nonstandard, a standard index can say "look for the lex. least proof of $0=1$," and I think that will be the only parameter needed). But I'm not sure. – Noah Schweber Mar 7 '17 at 1:18
• Yes, there may be a computable descending chain. – Emil Jeřábek Mar 7 '17 at 10:00