Connecting nominal and LN semantics

Our final goal is to show that the abstract nominal machine implements the same semantics as the LN substitution-based small step relation. We'll do this by proving that any time the abstract machine takes a step, we can decode the machine configuration before and after the step to LN expressions, and those expressions are either identical or related by the LN step relation.

                    machine_step
             (h,t,s)  ------>  (h',t',s')            nominal terms
                |                  |
                | decode           | decode
                v                  v
                e       - - ->     e'           locally nameless terms
                      step e e'
                         or
                        e = e'

This result is the simulate_step lemma near the end of this file.

Translating nominal terms to LN terms

We decode named terms to LN terms through the use of the close_exp_wrt_exp function. This function replaces all occurrences of a given atom by a new bound variable. This function is defined in the Stlc.Lemmas module and can be automatically generated by LNgen.

We also define a translation from machine configurations to LN terms. In this case we must substitute all definitions in the heap through the terms and create all of the applications in the stack.

The full decode function puts all of these parts together.

Here's an example translation from the machine step demo.

Connecting free variable functions.

Here is the first result about our decoding: that the two free variable functions agree. In this part of the file, we will take advantage of automation provided by the Lemmas module. In particular, this module defines a database of rewriting hints (called lngen) that can be used to automatically rewrite LN terms into simpler form.

As we prove new lemmas, we can also extend the hint database with new rewritings.

The Metatheory library contains two powerful tactics for simplifying goals:

• default_steps: repeat a bunch of simplifying steps, such as simplifying the goal, inverting simple hypotheses, etc.
• default_simp: above plus case analysis for booleans and other sums

Below, we modify the behavior of these tactics by updating the following two definitions, so that the lngen hint databases will be available.

We also add a few more rewriting lemmas to the hint database to automate our proofs.

Decoded terms are locally closed

Next, we show that our decoding of nominal terms and configurations produces locally closed LN terms. Again, these proofs are highly automatable.

Exercise: apply_stack_lc

State and prove a lemma called apply_stack_lc and add it to the lngen hint database so that the proof for decode_lc goes through.

Properties of apply_heap

Since the heap is just an iterated substitution, it inherits properties from subst_exp. Below, alist induction is a tactic from the metatheory library for induction over association lists (such as the heap).

Exercise: apply_heap_open

This function is the apply_heap analogue to subst_exp_open_exp_wrt_exp, commuting the heap-based substitution with the open_exp_wrt_exp operation.

This last lemma "unsimpl"s the apply_heap function.

Stacks as evaluation contexts

Here is a quick lemma that uses the properties defined above. It shows that the stack behaves like an *evaluation context*, lifting a small-step reduction to a larger term.

Connecting "freshening"

The abstract machine uses the nominal swap operation to make sure that the variables in abstractions are "fresh". To be able prove that this machine implements the step relation for LN terms, we need to connect this operation to a LN version of "freshening". The swap_spec lemma below states that connection. If a variable y is "fresh", then substituting with it (in the LN representation) is the same as swapping with it (in the nominal representation).

Lemma swap_spec : forall  n w y,
    y `notin` fv_exp (nom_to_exp n) ->
    w <> y ->
    [w ~> var_f y] (nom_to_exp n) =
    nom_to_exp (swap w y n).

Exercise close_exp_wrt_exp_freshen

The swap_spec proof depends on the following auxiliary lemma about the close operation --- that we can equivalently rename the atom that we are closing with.

One difficulty of swap_spec is that we need to use the induction not on direct subterms, but on those that have had a swapping applied to them. Swapping preserves the size of a nominal term, so we can prove this result by induction on m, a bound on the size of the term. The difficulty in this proof comes from the n_abs case. In this case, the term is is of the form n_abs x t for some binding variable x. We don't know much about x; it could be w, or y or some other variable. The first and last case are straightforward, but the middle case causes difficulty: even though x is not free in n_abs x t, it is free in t. Therefore, our induction hypothesis doesn't apply. To solve this problem, we need to generate a completely fresh variable z for the binder, use the lemma above to replace y and w with it, and then use the induction hypothesis.

Connection for alpha-equivalence

Challenge Exercise: aeq_nom_to_exp

Show that alpha-equivalence for the nominal representation is definitional equality for LN terms. This result is not necessary for the simulation lemmas below, but it is another example of the a proof that takes advantage of the swap_spec and close_exp_wrt_exp_freshen lemmas above. The second proof is much more challenging than the first and requires lemmas from Lemmas.v.

Scoped configurations

Not all abstract machine steps simulate small-steps of the substitution-based STLC. For example, if the domain of the heap is not unique (i.e. it has multiple definitions for the same variable) then its evaluation may not produce expected results. Similarly, if the avoid set that we pass to each machine step does not include initial the free variables of the term, then we could capture them in strange ways. Therefore, we will restrict our correctness lemmas so that they only apply to well-scoped configurations; as defined below.

Scoped heaps

Well-scoped heaps behave "telescopically". Each binding (x,e) added to the heap is for a unique name x and the free variables of e are bound in the remainder of the heap. The scoping relation is parameterized by D, an "ambient scope". This will let us reason about the execution of the abstract machine for terms with free variables.

Recommended (Challenge) Exercise apply_heap_get

We can use get to look up expressions in the heap. However, to know that we have the right result we need to know that the heap is well-scoped, i.e. that later bindings do not affect earlier ones. State a lemma about the scoping of expressions that appear in the heap and use it to finish the apply_heap_get lemma below.

Scoped stacks

We also care about the free variables that can appear in stacks. (We will use this in the definition of well-scoped configurations below.)

Stacks that are well-scoped can discard irrelevant bindings from the heap.

Simulation

A scoped configuration is one where all free variables in terms are either bound in the heap or come from D

Exercise simulate_step

After stepping through the simulation result below, finish the missing case.

Exercise simulate_done

Show that if the machine says Done then the LN term is a value.

Challenge exercise simulate_error

Show that if the machine produces an error, the small step relation is stuck. This is a challenge exercise because you will need to figure out at least one nontrivial auxiliary lemma.

Challenge exercise machine_is_scoped

Show that if the abstract machine is scoped, then any resulting configurations are also scoped.