Consider a model \(\mathbb{M}\) of some amount of arithmetic, say induction for bounded formulas and the totality of the exponential function. This could be a structure in the basic language of ordered rings \(\{+,\cdot, 0,1, \le\}\) or in \(\{+,\cdot, 0,1, \le\} \cup \{A_1,\ldots,A_k\}\) for some finite number of unary predicates \(A_i\) that can be treated as “oracles”. The latter case has become important in recent decades due e.g. to its relevance for reverse mathematics: we may want to expand \(\mathbb{M}\) by adding more oracle predicates to the language and eventually declare all the \(A_i\) to be bona fide objects of a “second-order” sort, turning \(\mathbb{M}\) into a model of some fragment of second-order arithmetic. This is a very natural technique in the study of the first-order strength of second-order arithmetic theories.

If \(\mathbb{M}\) does not satisfy the full induction scheme in its language – in the language of ordered rings, that would be Peano Arithmetic – then exactly one of two things must hold. Either there is \(n \ge 0\) such that \(\mathbb{M}\) satisfies \(\Sigma^0_n\)-induction, \(\mathrm{I}\Sigma^0_n\), but not the \(\Sigma^0_{n+1}\)-bounding (or collection) scheme, \(\mathrm{B}\Sigma^0_{n+1}\), i.e. \(\forall x \! \le \! a \exists y \psi(x,y) \to \exists b \forall x \! \le \! a \exists y \! \le \! b \psi(x,y),\) for \(\Sigma^0_{n+1}\) formulas \(\psi\); or there is \(n \ge 1\) such that \(\mathbb{M}\) satisfies \(\mathrm{B}\Sigma^0_n\) but not \(\mathrm{I}\Sigma^0_{n}\). Sometimes, one refers to models of the first kind as I-models and to those of the second kind as B-models.

A typical example of an I-model is a nonstandard pointwise \(\Sigma^0_n\)-definable one, i.e. one in which every singleton subset can be defined without using parameters. Not every I-model is elementarily equivalent to a {pointwise \(\Sigma^0_n\)-definable} one, but the example is canonical at least in the following sense: every countable model of \(\Sigma^0_n\)-induction can be expanded (by adding a new oracle predicate) to a pointwise \(\Sigma^0_n\)-definable model. (This is implicit in a theorem proved in 2015 by Towsner.)

B-models, on the other hand, are not particularly well-understood and have a number of surprising properties. For example, it has been known since the 1990’s that each countable B-model has continuum many automorphisms. Moreover, no B-model admits a definable injection from the universe into a bounded initial segment, regardless of the complexity of the definition. Neither of these things is true in a pointwise definable model. More recently, some results have emerged showing that, in contrast to the case of I-models, the possibilities of expanding a B-model so that the expanded structure remains a B-model are quite restricted.

In my talk, I will discuss some old and new results on the specific features of B-models, as well as some open problems. The emphasis will be on countable models and on features that have relevance to proof-theoretic issues such as provability and conservativity.