sanjoy / pyano

A formalization of first-order logic and Peano's axioms in Python

Pyano

Pyano (pronounced "piano") is a formalization of first-order logic and Peano's axioms in Python.

It consists of a few things:

1. A representation for proofs in first-order logic.
2. An automated proof verifier.
3. Python helpers to generate proofs.

Pyano is a toy project and not suitable for production use. Having said that, I'd love to hear from you if you found it useful in some way.

Getting started

Pyano uses pytest to run its unit tests and black to keep the Python code formatted. I recommend installing these into a local virtual environment in the following way:

python3 -m venv venv && \
	source venv/bin/activate && \
	pip install --upgrade pip && \
	pip install -r requirements.txt

Running pytest in the pyano directory will execute all of Pyano's unit tests.

If you learn by doing and enjoy throwing yourself in the deep end, try extending prove_one_less_than_or_eq_two in theorems.py (which proves 1 is less than or equal to 2) to prove that 1 is less than or equal to 3. See theorems_test.py on how to formally verify the proof.

Code structure

• formula.py contains the data structures to represent first-order formulae and functions to manipulate and query them. formula_helpers.py has some ergonomic helpers that seemed natural to split out.
• axioms.py contains the list of axioms allowed in Pyano.
• proof_checker.py contains assert_proof_is_valid which checks whether a formal proof is valid.
• proof_builder.py defines ProofBuilder which is a (stateful) builder class that makes writing proofs easier.
• theorems.py proves a couple of theorems. The full proofs for these theorems are checked-in at pyano/proved_theorems, and running theorems.py regenerates these.

What is a correct proof?

A formally correct proof in first-order logic is a sequence of formulae where each statement is either an axiom (and thus assumed to be true) or follows from two previous axioms through modus ponens. There are equivalent formulations, but this is the definition used in Pyano.

So the following is a valid proof that shows 0=0. The text in square brackets are comments which aren't part of the formal proof.

1. forall x. x = x. [This is an axiom.]
2. forall x. x = x => 0=0 [This is an axiom.]
3. 0=0 [Follows from modus ponens in 1 and 2.]

As a counterexample, the following proof is incorrect:

1. forall x. x = x. [This is an axiom.]
2. 0=0 [Invalid step.]

How are proofs verified?

The description above immediately suggests a straightforward algorithm to verify proofs:

for i in range(0, len(proof)):
  if is_axiom(proof[i]):
    continue

  modus_ponens = False
  for j in range(0, i):
    for k in range(0, i):
      if (proof[j] is P=>Q and
          proof[k] is P and
          proof[i] is Q)
        modus_ponens = True

  if not modus_ponens:
    fail_verification()

proof_checker.py implements a slightly more optimized version of this algorithm.

Axioms

The axioms encode much of the firepower of this proof system.

I picked this set from Antonio Montalban's excellent Math 125A -- Mathematical Logic course.

ForAll Axioms

These two are hopefully intuitively clear:

• For any k the following is an axiom forall x. P(x) => P(k).
• P => forall x. P(x), assuming x is not a free variable in P.

Finally, we have an axiom that lets us split forall formulae: (forall x. P(x) => Q(x)) => forall x. P(x) => forall x. Q(x).

I understood it intuitively as: if everyone who loves chocolate also loves ice-cream, and everyone loves chocolate then everyone loves ice-cream

Note that the converse, (forall x. P(x) => forall x. Q(x)) => (forall x. P(x) => Q(x)) cannot be an axiom since it is false -- consider P(x) = x == 2 and Q(x) = x == 3.

Sentential Tautology

Any And/Not/Implies expression that is true for all possible values of its "leaves" is an axiom.

So Not(Not(forall x. P(x))) => forall x. P(x) is an axiom since it is true irrespective of whether forall x. P(x) is true or not.

It may seem like this axiom says "true things are true" but that's not correct. We can easily (well, in exponential time :) ) decide if an expression is an axiom by this rule or not so it's well founded.

Reflexivity

This says things are equal to themselves: forall x. x = x.

Substitution

This says you can substitute expressions with another expression equal to it.

Formally it says x = y => F => G where you get G by replacing some (not necessarily all) instance of x in F with y.

Natural number axioms

These are the standard Peano axioms stating the various properties of natural numbers.

• forallx(Not(Eq(Zero(), Succ(x))))
• forallxy(Implies(Eq(Succ(x), Succ(y)), Eq(x, y)))
• forallx(Eq(Add(x, Zero()), x))
• forallxy(Eq(Add(x, Succ(y)), Succ(Add(x, y))))
• forallx(Eq(Mul(x, Zero()), Zero()))
• forallxy(Eq(Mul(x, Succ(y)), Add(Mul(x, y), x)))

Induction Axiom

This says that for any predicate P(x) the following is an axiom: P(0) & (forall k. P(k) => P(S(k))) => forall x. P(x).

It may seem tempting to have a more direct representation for the induction axiom, like forall P. (P(0) & (forall k. P(k) => P(S(k))) => forall x. P(x)). However, quantifiers in first-order logic only range over natural numbers so this won't work.

pyano's Issues

docs: "ponens" is misspelled as "pones" in `README.md`

sed -i "s/pones/ponens/g" README.md