A Brief on Blame Calculus

1. What is Blame Calculus and what is it for?

Blame calculus is an extension of a mixed-typed lambda calculus (i.e. containing both typed terms and untyped terms working on the catch-all type Dyn).

The purpose is to create a language where:

• Both untyped terms and typed terms co-exist, and typed terms are well-typed.
• There exists a decidable well-form check, and a semi-decidable well-typedness check.
• All well-typed terms (both typed and untyped) have a single type.

Additions to mixed-typed calculus

It features the following extensions to the type system:

• Refinement types: $\text{Nat} = {x : \text{Int} \mid x \ge 0}$, with runtime casting.
• A fully dynamic type Dyn, with compile-time upcasting and runtime downcasting.

It features the following extensions to the expression list:

• Casting operation: $\left<S\Leftarrow T\right>^{p} s$ is a cast from $s: T$ to $S$ (when $S \sim T$ - $S$ is compatible to $T$) , either
• succeeding returning $s: S$, or
• failing, returning $\Uparrow p$. We call this blaming on $p$.
• An internal transformation expression: $t \triangleright^p s_B$ where $t: \text{Bool}$, either
• succeeding with $s: {B \mid t}$ when $t = \text{True}$, or
• failing with $\Uparrow p$ otherwise.

A blame on $p$ can either be

• positive, where the cast fails due to the inner expression having an incompatible type with the destination type.

• negative, where the cast fails due to the surrounding context failing to comply with the destination type.

• This is possible only from the delayed casting behavior of Blame Calculus:

$$(\left<S \rightarrow T \Leftarrow S' \rightarrow T'\right>^p f)(v : S) \longrightarrow \left<T \Leftarrow T'\right>^p\left(f \left(\left<S' \Leftarrow S\right>^{\overline{p}}v\right)\right)$$

(notice the flip from $p$ to $\overline p$ on the cast of $v$)

Subtyping in Blame Calculus

Instead of the standard subtyping relation, which is undecidable in Blame Calculus (due to refinement types), it opts to an equally powerful system called positive (safe casts without positive blames) (denoted as $S <:^+ T$), negative ($S <:^- T$) and naive ($S <:_n T$) (which means $S <:^+ T$ and $T <:^- S$).

Unlike standard subtyping which is contravariant on the argument and covariant on the return type, naive subtyping is covariant on both sides.

Well-formednes

Two additional syntax extensions are made in order to mix typed and untyped expressions:

• $\lfloor M \rfloor$ casts the typed expression $M$ into an untyped expression. This is well-formed if and only if $M : \text{Dyn}$ (so a cast is most likely to happen somewhere).
• $\lceil M \rceil$ casts the untyped expression $M$ into a typed expression returning $\text{Dyn}$. This is well-formed if and only if
• all free variables in it has type $\text{Dyn}$, and
• all of its subterms have type $\text{Dyn}$.

2. Remarks

The central result of the paper, the Blame Theorem, states that:

Let $t$ be a well-typed term with a subterm $\left<T \Leftarrow S\right>^p s$ containing the only occurrences of $p$ in $t$.

Then:

• If $S <:^+ T$ then $t \not\longrightarrow^* \Uparrow p$.
• If $S <:^- T$ then $t \not\longrightarrow^* \Uparrow \overline p$.
• If $S <: T$ then $t$ will not blame on either $p$ nor $\overline p$.

In other words, from the decidable positive/negative subtyping relations, we can safely deduce which casts will not create a positive or negative blame.

This solidifies the intuition of upcasting/downcasting in the simple OOP sense, and allows us to extend this blame calculus with more structures, as long as the subtyping relations are still decidable.

3. Semantics

Please refer to the paper.