15-317 Constructive Logic
Lecture 24: A Taste of Linear Logic

Linear logic is the logic of precious resources: hypotheses may neither be duplicated nor discarded in the course of a proof. At the end of the day, every resource must be used exactly once. Earlier, we saw that the notion of hypothesis-as-resource may be fruitfully employed in logic programming to encode certain kinds of state transitions. We began class today by reviewing the program that decomposes a list and then rebuilds a permutation.

list([X|Xs]) ⊸ elem(X) ⊗ list(Xs).
list([]) ⊸ perm([]).
elem(X) ⊗ perm(Xs) ⊸ perm([X|Xs]).
In the notation of linear logic, we see that the encoding uses two connectives:
  • linear implication A ⊸ B, pronounced "A lolli B", which represents the ability to consume a resource A and replace it with a resource B.
  • simultaneous conjunction A ⊗ B, pronounced "A tensor B", which represents the simultaneous possession of a resource A and a resource B.
We reviewed the intuitionistic linear sequent calculus right and left rules for both connectives, showing how their interpretation matches our intuition: we must sometimes partition our resources in order to achieve our goals, and the init rule tells us we cannot discard any resources.

We also recalled a modification of the above program to make it return a permutation of a sub-list. To accomplish this, we explicitly permit ourselves to discard elements using the empty resource 1.

elem(X) ⊸ 1.
The left and right rules for 1 are nullary versions of those for tensor.

Then we saw another linear logic connective, A & B, pronounced "A with B", and its unit, , "top". The right rule is strikingly similar to that for A ⊗ B, except that it does not split the resources! This is accounted for in the left rules, though, which only give you one of A or B, not both. (Connectives whose rules split the context, such as ⊗ and ⊸ above, are called multiplicative. Connectives whose rules do not split the context, like &, are called additive.) This "alternative conjunction" or "internal choice" represents the ability to have either resource, but not simultaneously both. As a characteristic example, we considered the vending machine which, given a dollar, offers you your choice of Coke or Sprite:

$1 ⊸ coke & sprite.
We also considered a vending machine offering a limited time promotion, where each Coke comes with a pack of gum. We modelled the customer's indifference about this gum by using the consumptive unit ⊤ to eat up the undesired resource in the proof of:
($1 ⊸ coke ⊗ gum) ===> ($1 ⊸ coke ⊗ ⊤)

Given a notion of "internal choice", it is natural to wonder if there is a corresponding "external choice". In linear logic, there is: the connective A ⊕ B, pronounced "A plus B", representing either an A resource or a B resource, but you don't know which. The rules are very similar to the rules for ordinary disjunction from intuitionistic logic. This external choice can be used to model a Las Vegas gambling machine that sometimes gives you an all-expenses paid vacation to the Grand Caymans and somtimes just gives you a Coke:

$1 ⊸ coke ⊕ vacation.

Finally, we saw the unrestricted modality, !A, pronounced "bang A" or "of course A", which represents as many A resources as you'd like, possibly none at all. To interpret this connective, we introduced a second context to hold unrestricted hypotheses. In addition to the appropriate right and left rules, we posit a "copy" rule which lets a proof proceed by copying an unrestricted hypothesis to the linear context. In fact, we have implicitly made use of this context already: the rules of a logic program may be applied over and over again as long as their preconditions match the current state, and a rule may be used not at all in the execution of a program. Thus, the rules themselves are represented as unrestricted hypotheses.

To tie it all together, we worked through an extended example modelling a delicious brunch menu using the tasty and versatile connectives of linear logic. If only restaurateurs knew of such things!

  • Reading: lp:12-linear.pdf (Notes from an earlier course on Logic Programming)
  • Brunch menu example: menu.pdf
  • Key concepts:
    • Linear hypotheses as precious resources
    • Unrestricted hypotheses for unlimited use
    • Internal vs. external choice
    • Multiplicative vs. additive resource management
    • Modelling monetary transactions using linear logic
  • Previous lecture: Imperative Logic Programming
  • Next lecture: ??

[ Home | Schedule | Assignments | Handouts | Software ]

Frank Pfenning