prakol16 / lean_complexity_theory_polytime_trees

Complexity Theory in Lean

This repository defines and proves lemmas about the class of polynomial time functions. It also introduces a tactic in polytime_tac.lean which tries to automatically prove that functions run in polynomial time.

Implementation

The basic implementation is based of the structure of to_partrec.code in mathlib and the start of BoltonBailey's PR.

However, one key difference is that rather than using list ℕ as the main data structure, we use ptree ("plain tree"), which is a binary tree with no data attached to the leaves (defined in plain_tree.lean):

inductive ptree
| nil
| node (left : ptree) (right : ptree)

The main reason for this is that otherwise, we would need a pairing function $\mathbb{N} \times \mathbb{N} \rightarrow \mathbb{N}$ with logarithmic overhead (I worked on this for a bit here, but I eventually got fed up with it, and ptree is much easier to use).

Structure

There are a few files containing essentially dead code that need to be cleaned up. The files of interest are:

• code.lean -- Defines the code model
• plain_tree.lean -- Defines ptree, a plain binary tree with no data
• time_bound.lean -- Defines what it means for a code c to run in time $T(n)$
• polytime.lean -- Very short file that specializes the lemmas from time_bound to the case of polynomial bounds.
• polycodable_init.lean -- Defines polycodable and polytime_fun. In general, functions on other types $f : \alpha \rightarrow \beta$ are said to run in polynomial time if they have an encoding to ptree (many encodings e.g. unit, bool, pair, list, are defined in plain_tree.lean), and there is a code which runs in polynomial time and equals the function under this encoding. The encoding being a polycodable is the stronger property that encode (decode x) runs in polynomial time .This file sets up some basic definitions that the tactic depends on.
• polytime_tac.lean -- A tactic, polyfun, to automatically discharge goals of the form polytime_fun f.
• polycodable.lean -- Proves many more encodings are polycodable using polyfun
• polysize.lean -- Defines predicates polysize_fun, polysize_fun_safe, and polysize_fun_uniform which are each some variant of the fact that some function's growth is bounded by a polynomial in the input (possibly uniformly with respect to another input).
• polyfix.lean -- Proves that repeatedly evaluating a polynomial time function $f : \sigma \rightarrow \sigma$ polynomially many times such that the state $\sigma$ doesn't grow to fast results in a polynomial time function.
• lists.lean -- Shows that many functions about lists run in polynomial time.
• nat.lean -- Shows that addition and multiplication of numbers encoded as bits run in polynomial time.

Example of the tactic

def list_product {α β} (l₁ : list α) (l₂ : list β) : list (α × β) :=
(l₁.map (λ x, l₂.map (λ y, (x, y)))).join

example {α β : Type*} [polycodable α] [polycodable β] :
  polytime_fun₂ (λ (l₁ : list α) (l₂ : list β), list_product l₁ l₂) :=
by { dunfold list_product, polyfun, }

TODO

• Add a tactic that automatically makes progress on polysize goals (a bit trickier than polytime)
• Show polynomial time equivalence between Turing machines and the code model
• Define NP, NP completeness, and show SAT is NP complete
• Show other problems like graph coloring, max cut, clique, etc. are NP complete.