Recall that a geometric theory is given by implications between
positive statements: these are first order statements that do not
contain implications, negations or universal quantifiers. For any
such theory T there is a universal T-model U
T
living in a topos S(T) dubbed the
classifying
topos of T. For example, the theory of rings has for its classifying
topos the category of Sets-valued functors on the category of finitely
presented rings and homomorphisms between them, the universal ring
(Ring, rather) being the forgetful functor to Sets.
The category of T-models, and maps of T-models in a topos E is
equivalent to the category Top(E,S(T)) of geometric morphisms
E → S(T) and natural transformations between their inverse image
functors. A geometric morphism f : E → S(T) corresponds to a
T-model f
*
(U
T
) in E. The theory of local rings is obtained from that of rings by adjoining
the two extra axioms given above. The classifying topos S(locring)
has a geometric morphism to S(ring)
e: S(locring) → S(ring)
classifying the universal local ring U[
locring
]. If f : E → S(ring) classifies a Ring A in E then the pullback
of f along e yields the map
Spec(A) → S(locring)
which classifies the canonical sheaf of local rings A
~
on Spec(A). Pulling the e back along f yields a geometric
morphism Spec(A) → E, which is part of the homomorphism of
Rings A → A
~
which can be seen as the universal localization of A through which all
homomorphisms of Rings from A to a local Ring factor. In other words U
locring
is the universal localization of U
ring
and S(locring) is Spec of the universal ring. The fact that A
~
is not necessarily living in the same topos as A is what drives
the idea of extending rings to Rings. It is my belief that most of the
topos constructions that have appeared in algebraic geometry can be
more simply described in terms of appropriate geometric theories.
This is a game which can clearly be played in a great
number of fields. Anyone for rigged toposes?