I’m confused by this. Are you saying that in Agda typechecking is
exponential in the number of files? Or exponential in the number of
nested abstractions? Or something else? Do you have a toy example
demonstrating this behavior?
No toy example, so far, but I think such can be provided.
I have a real-world example of the DoCon-A library for algebra:
Index of /pub/local/Mechveliani/docon-A/2.02
This is a small part of the intended general purpose library for algebra
(for algebra methods, it is very small, but comparing to the Agda
practice, it is large).
It is written in install.txt
"for the -M15G key (15 Gb heap) installation takes about 50 minutes
on a
3 GHz personal computer.
"
It looks like the type checker has exponential cost in the depth of the
tree of the
introduced parametric module instances.
There is a hierarchy of “classes” (classical abstract structures):
Magma, (commutative)Semigroup, (Commutative)Monoid, (Commutative)Group,
… ,
(Commutative)Ring, Field, IntegralDomain, EuclideanDomain, GCDDomain,
LeftModuleOverARing …
– this tree depth may grow up to, may be, about 25.
And there are domain constructors: integer, vector, fraction,
polynomial, residueRing …
And these constructors are provided with instances of some of the above
abstract structures.
These instances include implementation for their needed operations, with
proofs.
The type checker deals with a hierarchy of such instances. And it
performs evaluation
(normalization …) with very large terms representing these instances.
For example, the Integer domain has may be 20 instances in it, and this
large term is
substituted many times on other terms, because almost every domain uses
some features of
the Integer domain. Anyway there appear internally very large terms that
repeat many
times, and their embracing terms need to normalize.
Further, the domain
Vector (EuclideanRingResidue f (Polynomial (Fraction Integer)))
(D)
is supplied with five instances of Magma, five instances of Semigroup,
five instances of CommutativeSemigroup, five instances of Monoid,
five instances of CommutativeMonoid, and also many other instances.
And the class operations in these instances (and their proofs) are
implemented
each in its individual way.
The number of different instances of the classical operations grows
exponentially
in the constructor depth for the domains like (D).
I do not expect that in mathematical textbooks appear domain constructs
as (D)
of the level greater than 10.
But Agda has practical difficulties with the level of about 4.
Because each construct like (D) is further substituted to different
parametric modules.
Because the method M1 uses one item from (D), so it is implemented in
the environment of
a parametric program module to which (D) is substituted for a parameter,
the method M2 uses another item from (D), and so on. And large subterms
get internally copied.
In a mathematical textbook, all these substitutions are mentioned or
presumed, and are
understood by the reader. So the informal “rigorous” proofs fit, say,
200 pages
(~ 100 Kb of memory).
But when a type checker tries to understand these constructions, it
creates many copies of large subterms and spends the cost in normalizing
them.
And formal proofs take 15 Gb memory to check.
First, this copy eagerness can probably be reduced (probably, this is
not easy to implement).
Second, something can be needed to arrange in the style of the
application programs.
There is a paper of Coq about this style, I recall A.Mahboubi is among
the authors.
So, there is a fundamental restriction – which hopefully can be handled
by introducing a certain
programming style (I never looked into this).
And also there is probably something to fix in the type checker in Agda.
Regards,