For some time now I've been practicing what can only be described as the slowest software development methodology in the world: a three step waltz of Prototyping, Verification and Export.
Prototyping: I begin quite normally by writing some code in Haskell, testing out different ways of solving some problem that I'm interested in. In this step I'm mainly focused on achieving a clean design with reasonable performance, and wherever possible I use existing Haskell packages (but restricted to those that have been developed with the same software development methodology). I also write QuickCheck tests to check various properties of the functions in the design. At this point I'm essentially done with the traditional coding phase, though usually I spend what might be perceived as an excessive amount of time polishing to make the design and properties as simple and clear as possible.
Verification: Next I use the HOL Light interactive theorem prover to develop a logical theory that (i) defines the types and functions in my Haskell design, and (ii) proves the QuickCheck properties as theorems that follow from the definitions. If the Haskell package depends on other packages that follow the same software development methodology then their contents must also be formalized in theories, and the logical definitions can build on each other in the same way as the Haskell definitions. This step is the most labour-intensive, and often some degree of creativity is required to formulate properties of recursive functions that can be proved by induction. This is where the effort spent polishing the prototype repays itself, since any unnecessary complexity in the design will inevitably thread its way through the proof of many properties.
Export: The final step is to use the opentheory tool to export the logical theory as a Haskell package. I first package up the HOL Light theory in OpenTheory format, by rendering its proof in a low-level article format and adding some meta-data such as the package name and the list of OpenTheory packages it depends on. There is also some Haskell-specific meta-data, such as the mapping from OpenTheory symbol names to Haskell names, and selecting which theorems should be exported as QuickCheck properties. Once that is complete the opentheory tool can automatically export the OpenTheory package as a Haskell package (in a form ready to be uploaded to Hackage).
At this point you might reasonably ask: why develop a Haskell package, throw it away, and then develop a logical theory only to export the same Haskell package again? Excellent question, and the rest of this post is devoted to answering it, but let's begin with a depressing truth: it's not the proof. Despite the countless hours and myriad creative insights you poured into that final Q.E.D., no one wants to see or even hear about your proof. Formal verification proofs are write-only. Sorry about that.
Specification: The proofs might be irrelevant, but the proved theorems constitute a precise specification of the Haskell package, grounded all the way down to standard mathematical objects formalized in higher order logic. I view Haskell as a computational platform for higher order logic (rather than higher order logic as a verification platform for Haskell), a perspective that emphasizes the consistent set of logical theories underlying the Haskell package that defines its semantics. A practical consequence of this is that any differences in the behaviour of the Haskell package and its higher order logic semantics (for example, due to bugs in the export code) must be fixed by changing the Haskell computational platform to conform to its higher order logic semantics.
Let's take a look at an example specification for an extended GCD function
egcd defined in the Haskell package opentheory-divides:
egcd :: Natural -> Natural -> (Natural, (Natural, Natural))
The following theorem about
egcd is one of many proved in the corresponding logical theory natural-divides:
∀a b g s t. ¬(a = 0) ∧ egcd a b = (g,(s,t)) ⇒ t * b + g = s * a
This says that the three components of the
egcd result (g, s and t) will satisfy a particular equation with its arguments (a and b), so long as the first argument a is non-zero. Side-conditions like this don't fit naturally into the Haskell type system, and so the burden falls on programmers to discover them and code around them when necessary. This may be one reason why programmers prefer to write their own software components (with side-conditions tailored to the application) rather than reuse existing components (with unknown and/or inconvenient side-conditions). Using this software development methodology I shift the burden of discovering side-conditions to HOL Light, which will force me to prove that the first argument of
egcd is non-zero in any code that relies on the above property of the result. If the code just uses the property that the result g of
egcd is the greatest common divisor then HOL Light will understand that no side-condition proof is necessary, because another theorem in the logical theory proves that this is true for all arguments to
∀a b. fst (egcd a b) = gcd a b
Proof: It may be surprising to learn that verified Haskell packages can be useful as a building block to formalize mathematics in higher order logic. Once we have defined Haskell types and functions in the logic and proved that they satisfy their specification, we can build on this formal development to prove mathematical theorems that do not refer to any Haskell symbols. To continue the
egcd example, suppose we have two natural numbers a and b (b is greater than 1) that are coprime (i.e.,
gcd a b = 1). Then we can reduce modulo b both sides of the first property of
egcd to show the existence of a multiplicative inverse for a modulo b:
∀a b. 1 < b ∧ gcd a b = 1 ⇒ ∃s. (s * a) `mod` b = 1
This is a non-trivial higher order logic theorem which refers only to standard mathematical concepts, but its proof rests on the properties of the verified Haskell
Maintenance: Software maintenance is a pain, and it is here that this software development methodology really shines (and who knows, maybe eventually cause the tortoise to pass the hare?). Formally verified software offers a new capability: by examining the logical theory of a Haskell package, we can see exactly which properties of dependent packages it relies on. We can use this information to automatically check whether or not a new version of a dependent package will preserve functional correctness: if so then extend the acceptable version range; and if not then report an error to the maintainer. The opentheory tool uses this technique to automatically generate maximal acceptable version ranges for dependent packages whenever it exports Haskell packages, removing another cognitive burden from the programmer. Earlier I said that formal verification proofs were write-only, but really all I meant is that humans won't read them: as this version analysis shows machines are quite happy to read them and extract useful information.
Fun: Actually I retract the whole depressing truth: it is totally about the proof. Not for showing to people (humans still won't read formal verification proofs) but rather for the pure fun of writing them. Those myriad creative insights you made over the course of countless hours? For me formal verification is as immersive as a video game, and my high scores come with a new artefact to play with and build upon. Free toy inside!