phase semantics

Girard’s *phase semantics* is a way of building star-autonomous posets with exponential modalities, hence models of classical linear logic.

Let $M$ be a commutative monoid, and $\bot\subseteq M$ a specified subset. ($M$ equipped with $\bot$ is sometimes called a *phase space*.) Then $P M$, the powerset of $M$, is a commutative quantale, with tensor product

$X\otimes Y = \{ m n \mid m\in X \wedge n \in Y \}$

and internal-hom

$X\multimap Y = \{ m \mid \forall n \in X, m n \in Y \}.$

Indeed, if we regard $M$ as a discrete poset, hence as a category enriched over $\mathbf{2} = \{0\le 1\}$, then $P M$ is its $\mathbf{2}$-enriched presheaf category, and this is its Day convolution monoidal structure.

Now $\bot$ is an object of $P M$, hence induces a contravariant self-adjunction $(-\multimap\bot) \dashv (-\multimap \bot)$ of $P M$, which is idempotent since $P M$ is a poset. We define a **fact** to be a fixed point of this adjunction, i.e. a subset $X\in P M$ of the form $Y\multimap \bot$, or equivalently such that $X = (X\multimap\bot)\multimap\bot$.

The poset of facts is star-autonomous.

It is closed under $\multimap$, since $(X\multimap (Y\multimap\bot)) = (X\otimes Y\multimap \bot)$. And it is reflective, with reflector $(-\multimap\bot)\multimap\bot$. Thus, it is closed symmetric monoidal with tensor product $((X\otimes Y)\multimap\bot)\multimap\bot$. Since it also contains $\bot$, as $\bot = (I\multimap \bot)$, it is therefore star-autonomous by construction.

The poset of facts is also, of course, a complete lattice, since it a reflective sub-poset of the complete lattice $P M$. In addition, it admits exponential modalities $!$ and $?$. There are different ways to obtain these, but perhaps the simplest (see here) is to take

$!X = ((X \cap Idem \cap 1)\multimap\bot)\multimap\bot$

where $Idem= \{ m \mid m m = m \}$ is the set of idempotents in $M$, and $1 = (\{1\}\multimap\bot)\multimap\bot$ is the unit object of the monoidal category of facts.

Phase semantics is complete for provability in linear logic, i.e. if a sequent of linear logic is valid in the phase semantics for all choices of $M$ and $\bot$, then it is provable. This follows from a construction of a particular $M$ and $\bot$ out of the syntax: let $M$ be the free commutative monoid on the formulas of linear logic subject to the relation that formulas of the form $?A$ are idempotent, and let $\bot$ be the set of $\Gamma\in M$ that are provable when regarded as one-sided sequents $\vdash\Gamma$, and interpret a formula $A$ by the set of $\Gamma\in M$ such that $\vdash \Gamma,A$ is provable. See Girard for details.

When phrased categorically as above, there is an obvious generalization to the case when $M$ is a commutative monoidal *poset*, with $P M$ replaced by its $\mathbf{2}$-enriched presheaf category, i.e. the poset $D M$ of downsets in $M$. We can also generalize to other enrichments, although in that case the idempotence of the adjunction $(-\multimap\bot) \dashv (-\multimap \bot)$ is no longer automatic but has to be assumed.

If $M$ is assumed only to be a promonoidal $\mathbf{2}$-enriched category, i.e. equipped with a suitable relation $M\times M ⇸ M$, then it is a ternary frame. This can be used to construct models of more general substructural logics.

- Jean-Yves Girard,
*Linear logic*, Theoretical Computer Science 50:1, 1987. (pdf)

- Jean-Yves Girard,
*Linear logic, its syntax and semantics*(pdf)