2.1 Algebraic effects and the free monad

We encode algebraic effects in Agda by representing computations as an abstract syntax tree given by the free monad over an effect signature. Such effect signatures are traditionally (Swierstra, Reference Swierstra2008; Awodey, Reference Awodey2010; Kammar et al., Reference Kammar, Lindley and Oury2013; Wu et al., Reference Wu, Schrijvers and Hinze2014; Kiselyov & Ishii, Reference Kiselyov and Ishii2015) given by a functor; i.e., a type of kind together with a (lawful) mapping function.Footnote ¹¹ In our Agda implementation, effect signature functors are defined by giving a container (Abbott et al., Reference Abbott, Altenkirch and Ghani2003, Reference Abbott, Altenkirch and Ghani2005). Each container corresponds to a value of type that is both strictly positive Footnote ¹² and universe consistent Footnote ¹³ (Martin-Löf, Reference Martin-Löf1984), meaning they are a constructive approximation of endofunctors on . Effect signatures are given by a (dependent) record type: ^{Footnote 14,Footnote 15}

Here, is the set of operations, and defines the return type for each operation in the set . The extension of an effect signature, $[\![\_]\!]$ , reflects its input of type as a value of type :Footnote ¹⁶

The extension of an effect $\Delta$ into is indeed a functor, as witnessed by the following function:Footnote ¹⁷

As discussed in the introduction, computations may use multiple different effects. Effect signatures are closed under co-products:Footnote ¹⁸ ,Footnote ¹⁹

We compute the co-product of two effect signatures by taking the disjoint sum of their operations and combining the return type mappings pointwise. We use co-products to encode effect rows. For example, the effect corresponds to the row union denoted as $\Delta_{1},\Delta_{2}$ in the introduction.

The syntax of computations with effects $\Delta$ is given by the free monad over $\Delta$ . We encode the free monad as follows:

Here, is a computation with no side-effects, whereas is an operation whose syntax is given by the functor . By applying this functor to , we encode an operation whose continuation may contain more effectful operations.Footnote ²⁰ To see in what sense, let us consider an example.

Example. The data type on the left below defines an operation for outputting a string. On the right is its corresponding effect signature.

The effect signature on the right says that returns a unit value ( is the unit type). Using this, we can write a simple hello world corresponding to the ${hello}$ program from Section 1:

Section 2.1 shows how to make this program more readable by using monadic notation.

The program above makes use of just a single effect. Say we want to use another effect, , with a single operation, , which represents throwing an exception (therefore having the empty type as its return type):

Programs that use multiple effects, such as and , are unnecessarily verbose. For example, consider the following program which prints two strings before throwing an exception:Footnote ²¹

To reduce syntactic overhead, we use row insertions and smart constructors (Swierstra, Reference Swierstra2008).

2.2 Row insertions and smart constructors

A smart constructor constructs an effectful computation comprising a single operation. The type of this computation is polymorphic in what other effects the computation has. For example, the type of a smart constructor for the effect is

Here, the type declares the row insertion witness as an instance argument of . Instance arguments in Agda are conceptually similar to type class constraints in Haskell: when we call , Agda will attempt to automatically find a witness of the right type, and implicitly pass this as an argument.Footnote ²² Thus, calling will automatically inject the effect into some larger effect row $\Delta$ .

We define the order on effect rows in terms of a different which witnesses that any operation of $\Delta$ is isomorphic to either an operation of $\Delta_{1}$ or an operation of $\Delta_{2}$ :Footnote ²³ ,Footnote ²⁴

Using this, the order is defined as follows:

It is straightforward to show that is a preorder; i.e., that it is a reflexive and transitive relation.

We can also define the following function, which uses a $\Delta_{1}$ $\Delta_{2}$ witness to coerce an operation of effect type $\Delta_{1}$ into an operation of some larger effect type $\Delta_{2}$ .Footnote ²⁵

Furthermore, we can freely coerce the operations of a computation from one effect row to a different effect row:Footnote ²⁶ ,Footnote ²⁷

Using this infrastructure, we can now implement a generic function which lets us define smart constructors for operations such as the operation discussed in the previous subsection.

2.3 Fold and monadic bind for

Since is a monad, we can sequence computations using monadic bind, which is naturally defined in terms of the fold over .

Besides the input computation to be folded (last parameter), the fold is parameterized by a function (first parameter) which folds a computation, and an algebra $\Delta$ A (second parameter) which folds an computation. We call the latter an algebra because it corresponds to an F-algebra (Arbib & Manes, Reference Arbib and Manes1975; Pierce, Reference Pierce1991) over the signature functor of , denoted $F_{\Delta}$ . That is, a tuple $(A, \alpha)$ where A is an object called the carrier of the algebra, and $\alpha$ a morphism $F_{\Delta}(A) \to A$ . Using , monadic bind for the free monad is defined as follows:

Intuitively, concatenates g to all the leaves in the computation m.

Example. The following defines a smart constructor for :

Using this and the definition of above, we can use do-notation in Agda to make the program from Section 2.1 more readable:

This illustrates how we use the free monad to write effectful programs against an interface given by an effect signature. Next, we define effect handlers.

2.4 Effect handlers

An effect handler implements the interface given by an effect signature, interpreting the syntactic operations associated with an effect. Like monadic bind, effect handlers can be defined as a fold over the free monad. The following type of parameterized handlers (Leijen, Reference Leijen2017, §2.2) defines how to fold, respectively, and computations:Footnote ²⁸

A handler of type is parameterized in the sense that it turns a computation of type into a parameterized computation of type . The following function does so by folding using , and a function:Footnote ²⁹

Comparing with the syntax, we used to explain algebraic effects and handlers in the introduction, the field corresponds to the $\mathbf{return}{}$ case of the handlers from the introduction, and corresponds to the cases that define how operations are handled. The parameterized handler type corresponds to the type ${A}\mathop{!}{\Delta,\Delta^\prime} \Rightarrow P \to {B}\mathop{!}{\Delta^\prime}$ , and h m corresponds to $\textbf{with}~{h}~\textbf{handle}~{m}$ .

Using this type of handler, the ${hOut}$ handler from the introduction can be defined as follows:

The handler ${hOut}$ in Section 1.1 did not bind any parameters. However, since we are encoding it as a parameterized handler, now binds a unit-typed parameter. Besides this difference, the handler is the same as in Section 1.1. We can use the handler to run computations. To this end, we introduce a effect with no associated operations which we will use to indicate where an effect row ends:

Using these, we can run a simple hello world program:Footnote ³⁰

An example of parameterized (as opposed to unparameterized) handlers is the state effect. Figure 1 declares and illustrates how to handle such an effect with operations for reading ( ) and changing ( ) the state of a memory cell holding a natural number.

Fig 1.

A state effect (upper), its handler ( below), and a simple test ( , also below) which uses (the elided) smart constructors for and .

2.5 The modularity problem with higher-order effects, revisited

Section 1.2 described the modularity problem with higher-order effects, using a higher-order operation that interacts with output as an example. In this section, we revisit the problem, framing it in terms of the definitions introduced in the previous section. To this end, we use a different effect whose interface is summarized by the record below. The record asserts that the computation type M : has at least a higher-order operation and a first-order operation :

The idea is that throws an exception, and m $_{1}$ m $_{2}$ handles any exception thrown during evaluation of m $_{1}$ by running m $_{2}$ instead. The problem is that we cannot give a modular definition of operations such as using algebraic effects and handlers alone. As discussed in Section 1.2, the crux of the problem is that algebraic effects and handlers provide limited support for higher-order operations. However, as also discussed in Section 1.2, we can encode in terms of more primitive effects and handlers, such as the following handler for the effect:

The handler modifies the return type of the computation by decorating it with a . If no exception is thrown, wraps the yielded value in a constructor. If an exception is thrown, the handler never invokes the continuation k and aborts the computation by returning instead. We can elaborate into an inline application of . To do so, we make use of effect masking which lets us “weaken” the type of a computation by inserting extra effects in an effect row:

Using this, the following elaboration defines a semantics for the operation:Footnote ³¹ ,Footnote ³²

If m $_{1}$ does not throw an exception, we return the produced value. If it does, m $_{2}$ is run.

As observed by Wu et al. (Reference Wu, Schrijvers and Hinze2014), programs that use elaborations such as are less modular than programs that only use plain algebraic operations. In particular, the effect row of computations no longer represents the interface of operations that we use to write programs, since the elaboration is not represented in the effect type at all. So we have to rely on different machinery if we want to refactor, optimize, or change the semantics of without having to change programs that use it.

In the next subsection, we describe how to define effectful operations such as modularly using scoped effects and handlers and discuss how this is not possible for, e.g., operations representing $\lambda$ -abstraction.

2.6 Scoped effects and handlers

This subsection gives an overview of scoped effects and handlers. While the rest of the paper can be read and understood without a deep understanding of scoped effects and handlers, we include this overview to facilitate comparison with the alternative solution that we introduce in Section 3.

Scoped effects extend the expressiveness of algebraic effects to support a class of higher-order operations that Wu et al. (Reference Wu, Schrijvers and Hinze2014), Piróg et al. (Reference Piróg, Schrijvers, Wu and Jaskelioff2018), Yang et al. (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022) call scoped operations. We illustrate how scoped effects work, using a freer monad encoding of the endofunctor algebra approach of Yang et al. (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022). The work of Yang et al. (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022) does not include examples of modular handlers, but the original paper on scoped effects and handlers by Wu et al. (Reference Wu, Schrijvers and Hinze2014) does. We describe an adaptation of the modular handling techniques due to Wu et al. (Reference Wu, Schrijvers and Hinze2014) to the endofunctor algebra approach of Yang et al. (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022).

2.6.1 Scoped programs

Scoped effects extend the free monad data type with an additional row for scoped operations. The and constructors of below correspond to the and constructors of the free monad, whereas is new:

Here, the constructor represents a higher-order operation with sub-scopes; i.e., computations that themselves return computations:

This type represents scoped computations in the sense that outer computations can be handled independently of inner ones, as we illustrate in sec:scoped-effect-handlers. One way to think of inner computations is as continuations (or join-points) of sub-scopes.

Using , the catch operation can be defined as a scoped operation:

The effect signature indicates that has two scopes since has two inhabitants. Following Yang et al. (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022), scoped operations are handled using a structure-preserving fold over :

The first argument represents the case where we are folding a node; the second and third correspond to, respectively, and .

2.6.2 Scoped effect handlers

The following defines a type of parameterized scoped effect handlers:

A handler of type handles operations of $\Delta$ and $\gamma$ simultaneously and turns a computation into a parameterized computation of type . The and cases are similar to the and cases from sec:effect-handlers. The crucial addition that adds support for higher-order operations is the case.

The field is given by an case. This case takes as input a scoped operation whose outer and inner computation have been folded into a parameterized computation of type and returns as output an interpretation of that operation as a computation of type . The function is used for modularly weaving (Wu et al., Reference Wu, Schrijvers and Hinze2014) side effects of handlers through sub-scopes of yet-unhandled operations.

2.6.3 Weaving

To see why is needed, it is instructional to look at how the fields in the record type above are used to fold over :

The second to last line above shows how is used. Because eagerly folds the current handler over scopes (sc), there is a mismatch between the type that the continuation expects (B) and the type that the scoped computation returns (G B). The function fixes this mismatch for the particular return type modification G : of a parameterized scoped effect handler.

The scoped effect handler for exception catching is thus:

The field for the operation first runs m $_{1}$ . If no exception is thrown, the value produced by m $_{1}$ is forwarded to k. Otherwise, m $_{2}$ is run and its value is forwarded to k, or its exception is propagated. The field of says that, if an unhandled exception is thrown during evaluation of a scope, the continuation is discarded and the exception is propagated; and if no exception is thrown the continuation proceeds normally.

2.6.4 Discussion and limitations

As observed by Berg et al. (2021), some higher-order effects do not correspond to scoped operations. In particular, the record shown below is not a scoped operation:

The field represents an operation that constructs a $\lambda$ value. The field represents an operation that will apply the function value in the first parameter position to the argument computation in the second parameter position. The operation has a computation as its second parameter so that it remains compatible with different evaluation strategies.

To see why the operations summarized by the record above are not scoped operations, let us revisit the constructor of :

As summarized earlier in this subsection, lets us represent higher-order operations (specifically, scoped operations), whereas does not (only algebraic operations). Just like we defined the computational parameters as scopes (given by the outer in the type of ), we might try to define the body of a lambda as a scope in a similar way. However, whereas the operation always passes control to its continuation (the inner ), the effect is supposed to package the body of the lambda into a value and pass this value to the continuation (the inner computation). Because the inner computation is nested within the outer computation, the only way to gain access to the inner computation (the continuation) is by first running the outer computation (the body of the lambda). This does not give us the right semantics.

It is possible to elaborate the operations into more primitive effects and handlers, but as discussed in Sections 1.2 and 2.5, such elaborations are not modular. In the next section, we show how to make such elaborations modular.

3.1 Generalizing Free to support higher-order operations

As summarized in Section 2.1, $\Delta$ A is the type of abstract syntax trees representing computations over the effect signature $\Delta$ . Our objective is to arrive at a more general type of abstract syntax trees representing computations involving (possibly) higher-order operations. To realize this objective, let us consider how to syntactically represent this variant of the ${censor}$ operation (Section 1.2), where M is the type of abstract syntax trees whose type we wish to define:

We call the second parameter of this operation a computation parameter. Using , computation parameters can only be encoded as continuations. But the computation parameter of is not a continuation, since

The crux of the issue is how signature functors are defined. Since this is an endofunctor on , the only suitable option in the constructor is to apply the functor to the type of continuations:

A more flexible approach would be to allow signature functors to build computation trees with an arbitrary return type, including the return type of the continuation. A higher-order signature functor of some higher-order signature H, written would fit that bill. Using such a signature functor, we could define the case as follows:

Here, is the type of the free monad using higher-order signature functors instead. In the rest of this subsection, we consider how to define higher-order signature functors H, their higher-order functor extensions , and the type of trees.

Recall how we defined plain algebraic effects in terms of containers:

Here, is the type of operations, and defines the return type of each operation. In order to allow operations to have sub-computations, we generalize this type to allow each operation to be associated with a number of sub-computations, where each sub-computation can have a different return type. The following record provides this generalization:

The set of operations is still given by a type field ( ), and each operation still has a return type ( ). associates each operation with a type that indicates how many sub-computations (or forks) an operation has, and indicates the return type of each such fork. For example, say we want to encode an operation op with two sub-computations with different return types, and whose return type is of a unit type. That is, using M as our type of computations:

The following signature declares a higher-order effect signature with a single operation with return type , and with two forks (we use to encode this fact), and where each fork has, respectively, and as return types.

The extension of higher-order effect signatures implements the intuition explained above:

Let us unpack this definition.

The extension of a higher-order signature functor is given by (1) the sum of operations of the signature, where each operation has (2) a continuation (of type M X) that expects to be passed a value of the operation’s return type, and (3) a set of forks where each fork is (4) a computation that returns the expected type for each fork.

Using the higher-order signature functor and its extension above, our generalized free monad becomes:

This type of trees can be used to define higher-order operations with an arbitrary number of computation parameters, with arbitrary return types. Using this type, and using a co-product for higher-order effect signatures ( ) which is analogous to the co-product for algebraic effect signatures in Section 2.2, Figure 2 represents the syntax of the operation.

Fig 2.

A higher-order censor effect and operation, with a single computation parameter (declared with in the effect signature top right) with return type (declared with top right).

Just like trees can be sequenced using monadic bind. Unlike for , the monadic bind of is not expressible in terms of the standard fold over trees. The difference between and is that is a regular data type, whereas is a nested datatype (Bird & Paterson, Reference Bird and Paterson1999). The fold of a nested data type is limited to describe natural transformations. As Bird & Paterson (Reference Bird and Paterson1999) show, this limitation can be overcome by using a generalized fold, but for the purpose of this paper, it suffices to define monadic bind as a recursive function:

The bind behaves similarly to the bind for ; i.e., concatenates g to all the leaves in the continuations (but not computation parameters) of m.

In Section 3.4, we show how to modularly elaborate higher-order operations into more primitive algebraic effects and handlers (i.e., computations over ), by folding modular elaboration algebras (hefty algebras) over trees. First, we show (in Section 3.2) how trees support programming against an interface of both algebraic and higher-order operations. We also address (in Section 3.3) the question of how to encode effect signatures for higher-order operations whose computation parameters have polymorphic return types, such as the highlighted below:

3.2 Programs with algebraic and higher-order effects

Any algebraic effect signature can be lifted to a higher-order effect signature with no fork (i.e., no computation parameters):

Using this effect signature and using higher-order effect row insertion witnesses analogous to the ones we defined and used in Section 2.2, the following smart constructor lets us represent any algebraic operation as a computation:

Using this notion of lifting, trees can be used to program against interfaces of both higher-order and plain algebraic effects.

3.3 Higher-order operations with polymorphic return types

Let us consider how to define as a higher-order effect. Ideally, we would define an operation that is parameterized by a return type of the branches of a particular catch operation, as shown on the left, such that we can define the higher-order effect signature on the right:Footnote ³³

The field on the right says that has two sub-computations (since has two constructors), and that each computation parameter has some return type A. However, the signature on the right above is not well defined!

The problem is that, because has a constructor that quantifies over a type ( ), the type lives in . Consequently, it does not fit the definition of , whose operations live in . There are two potential solutions to this problem: (1) increase the universe level of to allow to live in or (2) use a universe of types (Martin-Löf, Reference Martin-Löf1984). Either solution is applicable here; we choose type universes.

A universe of types is a (dependent) pair of a syntax of types ( ) and a semantic function ( ) defining the meaning of the syntax by reflecting it into Agda’s :

Section 4.1 contains a concrete example usage this notion of type universe. Using type universes, we can parameterize the constructor on the left below by a syntactic type of some universe u and use the meaning of this type ( ) as the return type of the computation parameters in the effect signature on the right below:

While the universe of types encoding restricts the kind of type that catch can have as a return type, the effect signature is parametric in the universe. Thus the implementer of the effect signature (or interface) is free to choose a sufficiently expressive universe of types.

3.4 Hefty algebras

As shown in Section 2.5, the higher-order catch operation can be encoded as a non-modular elaboration:

We can make this elaboration modular by expressing it as an algebra over trees containing operations of the signature. To this end, we will use the following notion of hefty algebra ( ) and fold (or catamorphism Meijer et al., Reference Meijer, Fokkinga and Paterson1991, ) for :

Here, defines how to transform an node of type H A into a value of type F A, assuming we have already folded the computation parameters and continuation into F values. A nice property of algebras is that they are closed under higher-order effect signature sums:

By defining elaborations as hefty algebras (below) we can compose them using .

An H $\Delta$ elaborates higher-order operations of signature H into algebraic operations of signature $\Delta$ . Given an elaboration, we can generically transform any hefty tree into more primitive algebraic effects and handlers:

This elaboration matches the ${eCensor}$ elaboration discussed in sec:solving-the-modularity-problem.

The elaboration is essentially the same as its non-modular counterpart, except that it now uses the universe of types encoding discussed in Section 3.3, and that it now transforms syntactic representations of higher-order operations instead. Using this elaboration, we can, for example, run the following example program involving the state effect from Figure 1, the throw effect from Section 2.1, and the catch effect defined here:

The program first sets the state to 1; then to 2; and then throws an exception. The exception is caught, and control is immediately passed to the final operation in the program which gets the state. By also defining elaborations for and , we can elaborate and run the program:

The program in Example 2 uses a so-called global interpretation of state, where the operation in the “try block” of causes the state to be updated globally. In Section 4.2.2, we return to this example and show how we can modularly change the elaboration of the higher-order effect to yield a so-called transactional interpretation of state where the operation in the try block is rolled back when an exception is thrown.

3.5 Discussion and limitations

Which (higher-order) effects can we describe using hefty trees and algebras? Since the core mechanism of our approach is modular elaboration of higher-order operations into more primitive effects and handlers, it is clear that hefty trees and algebras are at least as expressive as standard algebraic effects. The crucial benefit of hefty algebras over algebraic effects is that higher-order operations can be declared and implemented modularly. In this sense, hefty algebras provide a modular abstraction layer over standard algebraic effects that, although it adds an extra layer of indirection by requiring both elaborations and handlers to give a semantics to hefty trees, is comparatively cheap and implemented using only standard techniques such as F-algebras. As we show in Section 5, hefty algebras also let us define higher-order effect theories, akin to algebraic effect theories.

Conceptually, we expect that hefty trees can capture any monadic higher-order effect whose signature is given by a higher-order functor on . Filinski (Reference Filinski1999) showed that any monadic effect can be represented using continuations, and given that we can encode the continuation monad using algebraic effects (Schrijvers et al., Reference Schrijvers, Piróg, Wu and Jaskelioff2019) in terms of the sub/jump operations (Section 4.2.2) by Thielecke (Reference Thielecke1997), Fiore & Staton (Reference Fiore and Staton2014), it is possible to elaborate any monadic effect into algebraic effects using hefty algebras. The current Agda implementation, though, is slightly more restrictive. The type of effect signatures, , approximates the set of higher-order functors by constructively enforcing that all occurrences of the computation type are strictly positive. Hence, there may be higher-order effects that are well-defined semantically, but which cannot be captured in the Agda encoding presented here.

Recent work by Berg & Schrijvers (2023) introduced a higher-order free monad that coincides with our type. Their work shows that hefty trees are, in fact, a free monad. Furthermore, they demonstrate that a range of existing effects frameworks from the literature can be viewed as instances of hefty trees.

When comparing hefty trees to scoped effects, we observe two important differences. The first difference is that the syntax of programs with higher-order effects is fundamentally more restrictive when using scoped effects. Specifically, as discussed at the end of Section 2.6.4, scoped effects impose a restriction on operations that their computation parameters must pass control directly to the continuation of the operation. Hefty trees, on the other hand, do not restrict the control flow of computation parameters, meaning that they can be used to define a broader class of operations. For instance, in Section 4.1, we define a higher-order effect for function abstraction, which is an example of an operation where control does not flow from the computation parameter to the continuation.

The second difference is that hefty algebras and scoped effects and handlers are modular in different ways. Scoped effects are modular because we can modularly define, compose, and handle scoped operations, by applying scoped effect handlers in sequence; i.e.:

(‡)

As discussed in Section 2.6.3, each handler application modularly “weaves” effects through sub-computations, using a dedicated function.applying different scoped effect handlers in different orders.

Hefty algebras, on the other hand, work by applying an elaboration algebra assembled from modular components in one go. The program resulting from elaboration can then be handled using standard algebraic effect handlers; i.e.:

(§)

The algebraic effect handlers $h_1,\ldots,h_k$ in (§) serve the same purpose as the scoped effect handlers $h_1',\ldots,h_n'$ in ( $\dagger$ ); namely, to provide a semantics of operations. The order of handling is significant for both algebraic effect handlers and for scoped effect handlers: applying the same handlers in different orders may give a different semantics.

In contrast, the order that elaborations ( $E_1,\ldots,E_m$ ) are composed in (§) does not matter. Hefty algebras merely mediate higher-order operations into “first-order” effect trees that then must be handled, using standard effect handlers. While scoped effects supports “weaving”, standard algebraic effect handlers do not. This might suggest that scoped effects and handlers are generally more expressive. However, many scoped effects and handlers can be emulated using algebraic effects and hanlders, by encoding scoped operations as algebraic operations whose continuations encode a kind of scoped syntax, inspired by Wu et al. (Reference Wu, Schrijvers and Hinze2014, §7-9).Footnote ³⁴ We illustrate how in Section 4.2.2.

4.1 $\lambda$ as a higher-order operation

As recently observed by van den Berg et al. (Reference van den Berg, Schrijvers, Poulsen and Wu2021), the (higher-order) operations for $\lambda$ abstraction and application are neither algebraic nor scoped effects. We demonstrate how hefty algebras allow us to modularly define and elaborate an interface of higher-order operations for $\lambda$ abstraction and application, inspired by Levy’s call-by-push-value (Levy, Reference Levy2006). The interface we will consider is parametric in a universe of types given by the following record:

The field represents a function type, whereas is the type of thunk values. Distinguishing thunks in this way allows us to assign either a call-by-value or call-by-name semantics to the interface for $\lambda$ abstraction, given by the higher-order effect signature in Figure 3 and summarized by the following smart constructors:

Fig 3.

Higher-order effect signature of $\lambda$ abstraction and application.

Here, is a higher-order operation with a function typed computation parameter and whose return type is a function value ( ). The operation accepts a thunk value as argument and yields a value of a matching type. The operation is also a higher-order operation: its first parameter is a function value type, whereas its second parameter is a computation parameter whose return type matches that of the function value parameter type.

The interface above defines a kind of higher-order abstract syntax (Pfenning & Elliott, Reference Pfenning and Elliott1988), which piggy-backs on Agda functions for name binding. However, unlike most Agda functions, the constructors above represent functions with side effects. The representation in principle supports functions with arbitrary side effects since it is parametric in what the higher-order effect signature H is. Furthermore, we can assign different operational interpretations to the operations in the interface without having to change the interface or programs written against the interface. To illustrate we give two different implementations of the interface: one that implements a call-by-value evaluation strategy, and one that implements call-by-name.

4.1.1 Call-by-value

We give a call-by-value interpretation of by generically elaborating to algebraic effect trees with any set of effects $\Delta$ . Our interpretation is parametric in proof witnesses that the following isomorphisms hold for value types ( is the type of isomorphisms from the Agda standard library):

The first isomorphism says that a function value type corresponds to a function which accepts a value of type t $_{1}$ and produces a computation whose return type matches that of the function type. The second says that thunk types coincide with value types. Using these isomorphisms, the following defines a call-by-value elaboration of functions:

The case passes the function body given by the sub-tree $\psi$ as a value to the continuation, where the function mediates the sub-tree of type to a value type , using the isomorphism . The case uses the function to mediate a value to a value, using the isomorphism . The case first eagerly evaluates the argument expression of the application (in the sub-tree $\psi$ ) to an argument value and then passes the resulting value to the function value of the application. The resulting value is passed to the continuation.

Using the elaboration above, we can evaluate programs such as the following which uses both the higher-order lambda effect, the algebraic state effect, and assumes that our universe has a number type :

The program first sets the state to . Then it constructs a function that binds a variable x, dereferences the variable twice, and adds the two resulting values together. Finally, the application in the second-to-last line applies the function with an argument expression which increments the state by and returns the resulting value. Running the program produces since the state increment expression is eagerly evaluated before the function is applied.

4.1.2 Call-by-name

The key difference between the call-by-value and the call-by-name interpretation of our $\lambda$ operations is that we now assume that thunks are computations. That is, we assume that the following isomorphisms hold for value types:

Using these isomorphisms, the following defines a call-by-name elaboration of functions:

The case for is the same as the call-by-value elaboration. The case for now needs to force the thunk by running the computation and passing its result to k. The case for passes the argument sub-tree ( $\psi$ ) as an argument to the function f, runs the computation resulting from doing so, and then passes its result to k. Running the example program from above now produces as result, since the state increment expression in the argument of is thunked and run twice during the evaluation of the called function.

4.2 Optionally transactional exception catching

A feature of scoped effect handlers (Wu et al., Reference Wu, Schrijvers and Hinze2014; Piróg et al., Reference Piróg, Schrijvers, Wu and Jaskelioff2018; Yang et al., Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022) is that changing the order of handlers makes it possible to obtain different semantics of effect interaction. A classical example of effect interaction is the interaction between state and exception catching that we briefly considered at the end of sec:hefty-algebras in connection with this program:

The state and exception catching effect can interact to give either of these two semantics:

1. Global interpretation of state, where the program returns since the operation in the “try” block causes the state to be updated globally.

2. Transactional interpretation of state, where the program returns since the state changes of the operation are rolled back when the “try” block throws an exception.

With monad transformers (Cenciarelli & Moggi, Reference Cenciarelli and Moggi1993; Liang et al., Reference Liang, Hudak and Jones1995), we can recover either of these semantics by permuting the order of monad transformers. With scoped effect handlers, we can also recover either by permuting the order of handlers. However, the elaboration in Section 3.4 always gives us a global interpretation of state. In this section, we demonstrate how we can recover a transactional interpretation of state by using a different elaboration of the operation into an algebraically effectful program with the operation and the off-the-shelf sub/jump control effects due to Thielecke (Reference Thielecke1997), Fiore & Staton (Reference Fiore and Staton2014). The different elaboration is modular in the sense that we do not have to change the interface of the catch operation nor any programs written against the interface.

4.2.1 Sub/Jump

We recall how to define two operations, sub and jump, due to Thielecke (Reference Thielecke1997), Fiore & Staton (Reference Fiore and Staton2014). We define these operations as algebraic effects following Schrijvers et al. (Reference Schrijvers, Piróg, Wu and Jaskelioff2019). The algebraic effect signature of Ref is given in Figure 4 and is summarized by the following smart constructors:

Fig 4.

Effect signature of the sub/jump effect.

An operation f g gives a computation f access to the continuation g via a reference value Ref t which represents a continuation expecting a value of type . The operation invokes such continuations.

The operations and their handler (abbreviated to ) satisfy the following laws:

The last law asserts that and are algebraic operations, since their computational sub-terms behave as continuations. Thus, we encode and its handler as an algebraic effect.

4.2.2 Optionally transactional exception catching

By using the and operations in our elaboration of , we get a semantics of exception catching whose interaction with state depends on the order that the state effect and sub/jump effect is handled.

The elaboration uses to capture the continuation of a higher-order operation. If no exception is raised, then control flows to the continuation k without invoking the continuation of . Otherwise, we jump to the continuation of , which runs ${m_{2}}$ before passing control to k. Capturing the continuation in this way interacts with state because the continuation of may have been pre-applied to a state handler that only knows about the “old” state. This happens when we handle the state effect before the sub/jump effect: in this case, we get the transactional interpretation of state, so running gives . Otherwise, if we run the sub/jump handler before the state handler, we get the global interpretation of state and the result .

The sub/jump elaboration above is more involved than the scoped effect handler that we considered in Section 2.6. However, the complicated encoding does not pollute the higher-order effect interface and only turns up if we strictly want or need effect interaction.

4.3 Logic programming

Following Schrijvers et al. (Reference Schrijvers, Wu, Desouter and Demoen2014), Wu et al. (Reference Wu, Schrijvers and Hinze2014), Yang et al. (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022), we can define a nondeterministic choice operation ( ) as an algebraic effect, to provide support for expressing the kind of non-deterministic search for solutions that is common in logic programming. We can also define a operation that indicates that the search in the current branch was unsuccessful. The effect signature for is given in Figure 5. The following smart constructors are the lifted higher-order counterparts to the and operations:

Fig 5.

Effect signature of the choice effect.

A useful operator for cutting non-deterministic search short when a solution is found is the operator. The operator, whose higher-order effect signature is given in Figure 6, is not an algebraic effect, but a scoped (and thus higher-order) effect.

Fig 6.

Higher-order effect signature of the once effect.

We can define the meaning of the operator as the following elaboration:

The elaboration runs the branch ( $\psi$ ) of under the handler to compute a list of all solutions of $\psi$ . It then tries to choose the first solution and pass that to the continuation k. If the branch has no solutions, we fail. Under a strict evaluation order, the elaboration computes all possible solutions which is doing more work than needed. Agda 2.6.2.2 does not have a specified evaluation strategy, but does compile to Haskell which is lazy. In Haskell, the solutions would be lazily computed, such that the operator cuts search short as intended.

4.4 Concurrency

Finally, we consider how to define higher-order operations for concurrency, inspired by Yang et al.’s (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022) resumption monad (Schmidt, Reference Schmidt1986; Claessen, Reference Claessen1999; Piróg & Gibbons, Reference Piróg and Gibbons2014) defined using scoped effects. We summarize our encoding and compare it with the resumption monad. The goal is to define the two operations, whose higher-order effect signature is given in Figure 7, and summarized by these smart constructors:

The operation ${m_{1}}~{m_{2}}$ spawns two threads that run concurrently, and returns the value produced by ${m}_{1}$ after both have finished. The operation m represents a block to be executed atomically; i.e., no other threads run before the block finishes executing.

We elaborate by interleaving the sub-trees of its computations. To this end, we use a dedicated function that interleaves the operations in two trees and yields as output the value of the left input tree (the first computation parameter):

Here, the effect is the sub/jump effect that we also used in Section 4.2.2. The function ensures atomic execution by only interleaving code that is not wrapped in a operation. We elaborate into as follows, where the and functions use the row insertion witness w $_a$ to move the effect to the front of the row and back again:

The elaboration uses as a delimiter for blocks that should not be interleaved, such that the function only interleaves code that does not reside in atomic blocks. At the end of an block, we to the (possibly interleaved) computation context, k. By using to explicitly delimit blocks that should not be interleaved, we have encoded what Wu et al. (Reference Wu, Schrijvers and Hinze2014, § 7) call scoped syntax.

Example. Below is an example program that spawns two threads that use the effect. The first thread prints , and ; the second prints and .

Since the effect is elaborated to interleave the effects of the two threads, the printed output appears in interleaved order:

The following program spawns an additional thread with an block

Inspecting the output, we see that the additional thread indeed computes atomically:

The example above is inspired by the resumption monad, and in particular by the scoped effects definition of concurrency due to Yang et al. (Reference Yang, Paviotti, Wu, van den Berg and Schrijvers2022). Yang et al. do not (explicitly) consider how to define the concurrency operations in a modular style. Instead, they give a direct semantics that translates to the resumption monad which we can encode as follows in Agda (assuming resumptions are given by the free monad):

We could elaborate into this type using a hefty algebra but that would be incompatible with our other elaborations which use the free monad. For that reason, we emulate the resumption monad using the free monad instead of using the type directly.

5.1 Effect theories and implementation correctness

Let us consider the state effect as an example, which comprises the and operations. With the state effect, we typically associate a set of equations (or laws) that specify how its implementations ought to behave. One such law is the get-get law, which captures the intuition that the state returned by two subsequent operations does not change if we do not use the operation in between:

We can define equational laws for higher-order effects in a similar fashion. For example, the following catch-return law for the operation of the effect, stating that catching exceptions in a computation that only returns a value does nothing.

Correctness of an implementation of an algebraic effect with respect to a given theory is defined by comparing the implementations of programs that are equal under that theory. That is, if we can show that two programs are equal using the equations of a theory for its effects, handling the effects should produce equal results. For instance, a way to implement the state effect is by mapping programs to functions of the form ${S}~\to~{S}\times{A}$ . Such an implementation would be correct if programs that are equal with respect to a theory of the state effect are mapped to functions that give the same value and output state for every input state.

For higher-order effects, correctness is defined in a similar manner. However, since we define higher-order effects by elaborating them into algebraic effects, correctness of elaborations with respect to a higher-order effect theory is defined by comparing the elaborated programs. Crucially, the elaborated programs do not have to be syntactically equal, but rather we should be able to prove them equal using a theory of the algebraic effects used to implement a higher-order effect.

Effect theories are known to be closed under the co-product of effects, by combining the equations into a new theory that contains all equations for both effects (Hyland et al., Reference Hyland, Plotkin and Power2006). Similarly, theories of higher-order effects are closed under sums of higher-order effect signatures. In Section 5.9, we show that composing two elaborations preserves their correctness, with respect to the sum of their respective theories.

5.2 Theories of algebraic effects

Theories of effects are collections of equations, so we start defining the type of equations in Agda. At its core, an equation for an effect $\Delta$ is given by a pair of effect trees of type , that define the left- and right-hand side of the equation. However, looking at the get-get law above, we see that this equation contains a term metavariable; i.e., ${k}$ . Furthermore, when considering the type of ${k}$ , which is , we see that it refers to a type metavariable; i.e., ${A}$ . Generally speaking, an equation may refer to any number of term metavariables, which, in turn, may depend on any number of type metavariables. Moreover, the type of the value returned by the left-hand side and right-hand side of an equation may depend on these type metavariables as well, as is the case for the get-get law above. This motivates the following definition of equations in Agda.

An equation consists of five components. The field defines the number of type metavariables used in the equation. Then, the fields and , respectively, define the term metavariables ( ) and return type ( ) of the equation.

Example. To illustrate how the record captures equational laws of effects, we consider how to define the get-get as a value of type .

The fields and define the left- and right-hand sides of the equation. Both sides only use a single term metavariable, representing a continuation of type . The field declares this term meta-variable. For equations with more than $n>1$ metavariables, we would define as an n-ary product instead.

5.3 Modal necessity

The current definition of equations is too weak, in the sense that it does not apply in many situations where it should. The issue is that it fixes the set of effects that can be used in the left- and right-hand side. To illustrate why this is problematic, consider the following equality:

(5.1)

We might expect to be able to prove this equality using the get-get law, but using the embedding of the law defined above—i.e., —this is not possible. The reason for this is that we cannot pick an appropriate instantiation for the term metavariable k: it ranges over values of type , inhibiting all references to effectful operation that are not part of the state effect, such as .

Given an equation for the effect $\Delta$ , the solution to this problem is to view $\Delta$ as a lower bound on the effects that might occur in the left-hand and right-hand side of the equation, rather than an exact specification. Effectively, this means that we close over all posible contexts of effects in which the equation can occur. This pattern of closing over all possible extensions of a type index is well-known (Allais et al., Reference Allais, Atkey, Chapman, McBride and McKinna2021; van der Rest et al., Reference van der Rest, Poulsen, Rouvoet, Visser and Mosses2022), and corresponds to a shallow embedding of the Kripke semantics of the necessity modality from modal logic. We can define it in Agda as follows.Footnote ³⁵

Intuitively, the modality transforms, for any effect-indexed type (P : ), an exact specification of the set of effects to a lower bound on the set of effects. For equations, the difference between terms of type and amounts to the former defining an equation relating programs that have exactly effects $\Delta$ , while the latter defines an equation relating programs that have at least the effects $\Delta$ but potentially more. The modality is a comonad: the counit ( below) witnesses that we can always transform a lower bound on effects to an exact specification, by instantiating the extension witness with a proof of reflexivity.

We can now redefine the get-get law such that it applies to all programs that have the effect, but potentially other effects too.

The above definition of the get-get law now lets us prove the equality in Equation (5.1); the term metavariable k ranges ranges over all continuations that return a tree of type , for all ${\Delta^\prime}$ such that . This way, we can instantiate ${\Delta^\prime}$ with an effect signature that subsumes both the and the , which in turn allows us to instantiate k with .

5.4 Effect theories

Equations for an effect $\Delta$ can be combined into a theory for $\Delta$ . A theory for the effect $\Delta$ is simply a collection of equations, transformed using the modality to ensure that term metavariables can range over programs that include more effects than just $\Delta$ .

An effect theory consists of an that defines the number of equations in the theory, and a function that maps arities to equations. We can think of effect theories as defining a specification for how implementations of an effect ought to behave. Although implementations may vary, depending for example on whether they are tailored to readability or efficiency, they should at least respect the equations of the theory of the effect they implement. We will make precise what it means for an implementation to respect an equation in sec:handler-correctness.

Effect theories are closed under several composition operations that allow us to combine the equations of different theories into single theory. The most basic way of combining effect theories is by summing their arities.

This way of combining effects is somewhat limiting, as it imposes that the theories we are combining are theories for the exact same effect. It is more likely, however, that we would want to combine theories for different effects. This requires that we can weaken effect theories with respect to the relation.

Categorically speaking, the observation that for a given effect-indexed type P we can transform a value of type $P\ {\Delta_{1}}$ to a value of type $P\ {\Delta_{2}}$ if we know that is equivalent to saying that P is a functor from the category of containers and container morphisms to the category of sets. From this perspective, the existence of weakening for free , as witnessed by the operation discussed in Section 3, implies that it too is a such a functor.

With weakening for theories at our disposal, we can combine effect theories for different effects into a theory of the coproduct of their respective effects. This requires us to first define appropriate witnesses relating coproducts to effect inclusion.

It is now straightforward to show that effect theories are closed under the coproduct of effect signatures, by summing the weakened theories.

While this operation is in principle sufficient for our purposes, it forces a specific order on the effects of the combined theories. We can further generalize the operation above to allow for the effects of the combined theory to appear in any order. This requires the following instances.

We show that effect theories are closed under coproducts up to reordering by, again, summing the weakened theories.

Since equations are defined by storing the syntax trees that define their left-hand and right-hand side, and effect trees are weakenable, we would expect equations to be weakenable too. Indeed, we can define the following function witnessing weakenability of equations.

This begs the question: why would we opt to use weakenability of the modality (or, bother with the modality at all) to show that theories are weakenable, rather than using directly? Although the latter approach would indeed allow us to define the composition operations for effect theories defined above, the possible ways in which we can instantiate term metavariables remains too restrictive. That is, we would still not be able to prove the equality in Equation (5.1), despite the fact that we can weaken the get-get law so that it applies to programs that use the effect as well. Instantiations of the term metavariable k will be limited to weakened effect trees, precluding any instantiation that use operations of effects other than , such as .

Finally, we define the following predicate to witness that an equation is part of a theory.

We construct a proof that an equation ${eq}$ is part of a theory ${T}$ by providing an arity together with a proof that ${T}$ maps to ${eq}$ for that arity.

5.5 Syntactic equivalence of effectful programs

Propositional equality of effectful programs is too strict, as it precludes us from proving equalities that rely on a semantic understanding of the effects involved, such as the equality in Equation (5.1). The solution is to define an inductive relation that captures syntactic equivalence modulo some effect theory. We base our definition of syntactic equality of effectful programs on the relation defining equivalent computations by Yang &Wu (Reference Yang and Wu2021), Definition 3.1, adapting their definition where necessary to account for the use of modal necessity in the definition of .

A value of type witnesses that programs ${m_{1}}$ and ${m_{2}}$ are equal modulo the equations of theory ${T}$ . The first three constructors ensure that it is an equivalence relation.

Then, we add the following congruence rule, which establishes that we can prove equality of two programs starting with the same operation by proving that the continuations yield equal programs for every possible value.

The final constructor allows to prove equality of programs by reifying equations of an effect theory.

Since the equations of a theory are wrapped in the modality, we cannot refer to its components directly, but we must first provide a suitable extension witness.

With the constructor, we can prove equivalence between the left-hand and right-hand side of an equation, sequenced with an arbitrary continuation ${k}$ . For convenience, we define the following lemma that allows us to apply an equation where the sides of the equation do not have a continuation.

The definition of follows readily from the right-identity law for monads, i.e., , which allows us to instantiate with .

To construct proofs of equality, it is convenient to use the following set of combinators to write proof terms in an equational style. They are completely analogous to the combinators commonly used to construct proofs of Agda’s propositional equality, for example, as found in PLFA (Wadler et al., Reference Wadler, Kokke and Siek2020).

We now have all the necessary tools to prove syntactic equality of programs modulo a theory of their effect. To illustrate, we consider how to prove the equation in Equation (5.1). First, we define a theory for the effect containing the law. While this is not the only law typically associated with , for this example it is enough to only have the law.

Now to prove the equality in Equation (5.1) is simply a matter of invoking the law.

5.6 Handler correctness

A handler is correct with respect to a given theory if handling syntactically equal programs yields equal results. Since handlers are defined as algebras over effect signatures, we start by defining what it means for an algebra of an effect $\Delta$ to respect an equation of the same effect, adapting Definition 2.1 from the exposition of Yang & Wu (Reference Yang and Wu2021).

An algebra ${alg}$ respects an equation ${eq}$ if folding with that algebra produces propositionally equal results for the left- and right-hand side of the equation, for all possible instantiations of its type and term metavariables, and continuations k.

A handler ${H}$ is correct with respect to a given theory ${T}$ if its algebra respects all equations of ${T}$ (Yang & Wu, Reference Yang and Wu2021, Definition 4.3).

We can now show that the handler for the effect defined in Figure 1 is correct with respect to . The proof follows immediately by reflexivity.

5.7 Theories of higher-order effects

For the most part, equations and theories for higher-order effects are defined in the same way as for first-order effects and support many of the same operations. Indeed, the definition of equations ranging over higher-order effects is exactly the same as its first-order counterpart, the most major difference being that the left-hand and right-hand side are now defined as Hefty trees. To ensure compatibility with the use of type universes to avoid size-issues, we must also allow type metavariables to range over the types in a universe in addition to . For this reason, the set of type metavariables is no longer described by a natural number, but rather by a list of kinds, which stores for each type metavariable whether it ranges over a types in a universe, or an Agda .

A carries unapplied substitutions for a given set of type metavariables and is defined by induction over a list of kinds.Footnote ³⁶

Equations of higher-order effects are then defined as follows.

This definition of equations suffers the same problem when it comes to term metavariables, which here too can only range over programs that exhibit the exact effect that the equation is defined for. Again, we address the issue using an embedding of modal necessity to close over all possible extensions of this effect. The definition is analogous to the one in sec:modal-necessity, but this time we use higher-order effect subtyping as the modal accessibility relation:

To illustrate: we can define the catch-return law from the introduction of this section as a value of type a follows. Since the operation relies on a type universe to avoid size issues, the sole type metavariable of this equation must range over the types in this universe as well.

Theories of higher-order effects bundle extensible equations. The setup is the same as for theories of first-order effects.

The following predicate establishes that an equation is part of a theory. We prove this fact by providing an arity whose corresponding equation is equal to ${eq}$ .

Weakenability of theories of higher-order effects then follows from weakenability of its equations.

Theories of higher-order effects can be combined using the following sum operation. The resulting theory contains all equations of both argument theories.

Theories of higher-order effects are closed under sums of higher-order effect theories as well. This operation is defined by appropriately weakening the respective theories, for which we need the following lemmas witnessing that higher-order effect signatures can be injected in a sum of signatures.

The operation that combines theories under signature sums is then defined like so.

5.8 Equivalence of programs with higher-order effects

We define the following inductive relation to capture equivalence of programs with higher-order effects modulo the equations of a given theory.

To ensure that it is indeed an equivalence relation, we include constructors for reflexivity, symmetry, and transitivity.

Furthermore, we include the following congruence rule that equates two program trees that have the same operation at the root, if their continuations are equivalent for all inputs.

Finally, we include a constructor that equates two programs using an equation of the theory.

We can define the same reasoning combinators as in Section 5.5 to construct proofs of equivalence for programs with higher-order effects.

To illustrate, we can prove that the programs are equal under a theory for the afCatch effect that contains the catch-return law.

The equivalence proof above makes, again, essential use of modal necessity. That is, by closing over all possible extensions of the effe, the term metavariable in the catch-return law to range over programs that have higher-order effects other than , which is needed to apply the law if the second branch of the operation contains the operation.

5.9 Correctness of elaborations

As the first step toward defining correctness of elaborations, we must specify what it means for an algebra over a higher-order effect signature ${H}$ to respect an equation. The definition is broadly similar to its counterpart for first-order effects in Section 5.6, with the crucial difference that the definition of “being equation respecting” for algebras over higher-order effect signatures is parameterized over a binary relation between first-order effect trees. In practice, this binary relation will be instantiated with the inductive equivalence relation defined in Section 5.5; propositional equality would be too restrictive, since that does not allow us prove equivalence of programs using equations of the first-order effect(s) that we elaborate into.

Since elaborations are composed in parallel, the use of necessity in the definition of equations has additional consequences for the definiton of elaboration correctness. That is, correctness of an elaboration is defined with respect to a theory whose equations have left-hand and right-hand sides that may contain term metavariables that range over programs with more higher-order effects than those the elaboration is defined for. Therefore, to state correctness, we must also close over all possible ways these additional effects are elaborated. For this, we define the following binary relation on extensible elaborations.Footnote ³⁷

A proof of the form ${e_{1}}~\mathbin{\sqsubseteq}~{e_{2}}$ witnesses that the elaboration ${e_{1}}$ is included in ${e_{2}}$ . Informally, this means that ${e_{2}}$ may elaborate a bigger set of higher-order effects, for which it may need to refer to a bigger set of first-order effects, but for those higher-order effects that both ${e_{1}}$ and ${e_{2}}$ know how to elaborate, they should agree on how those effects are elaborated.

We then define correctness of elaborations as follows.

Which is to say that an elaboration is correct with respect to a theory of the higher-order effects it elaborates ( ${Th}$ ) and a theory of the first-order effects it elaborates into ( ${T}$ ), if all possible extensions of said elaboration respect all equations of the higher-order theory, modulo the equations of the first-order theory.

Crucially, correctness of elaborations is preserved under composition of elaborations. Figure 8 shows the type of the corresponding correctness theorem in Agda; for the full details of the proof we refer to the Agda formalization accompanying this paper (van der Rest & Poulsen, Reference van der Rest and Poulsen2024). We remark that correctness of a composed elaboration is defined with respect to the composition of the theories of the first-order effects that the respective elaborations use. Constructing a handler that is correct with respect to this composed first-order effect theory is a separate concern; a solution based on fusion is detailed in the work by Yang & Wu (Reference Yang and Wu2021).

Fig 7.

Higher-order effect signature of the concur effect.

Fig 8.

The central correctness theorem, which establishes that correctness of elaborations is preserved under composition.

5.10 Proving correctness of elaborations

To illustrate how the reasoning infrastructure build up in this section can be applied to verify correctness of elaborations, we show how to verify the catch-return law for the elaboration defined in Section 3.4. First, we define the following syntax for invoking a known elaboration.

When opening the module ${Elab}$ , we can use the syntax for elaborating m with some known elaboration, which helps to simplify and improve readability of equational proofs for higher-order effects.

Now, to prove that is correct with respect to a higher-order theory for the effect containing the catch-return law, we must produce a proof that the programs are equal (in the sense of the inductive equivalence relation defined in Section 5.5) with respect to some first-order theory for the effect. In this instance, we do not need any equations from this underlying theory to prove the equality, but sometimes it is necessary to invoke equations of the underlying first-order effects to prove correctness of an elaboration.

In the Agda formalization accompanying this paper (van der Rest & Poulsen, Reference van der Rest and Poulsen2024), we verify correctness of elaborations for the higher-order operations that are part of the 3MT library by Delaware et al. (Reference Delaware, Keuchel and Schrijvers2013). Table 1 shows an overview of first-order and higher-order effects included in the development, and the laws which we prove about their handlers respectively elaborations.

Table 1

Overview of effects, their operations, and verified laws in the Agda code

5.11 Discussion

In the introduction, we discussed the desired degrees of modularity that hefty algebras and their reasoning infrastructure should support. That is, they should support the modular composition of syntax, semantics, equational theories, and proofs.

Composability of the syntax and semantics of higher order effects follows readily from the fact that we define higher-order effect signatures and elaborations as higher-order functors and their algebras respectively (Section 3.1), which are closed under coproducts.

For proofs, we demonstrate in Section 5.3 that equational proofs for programs with higher-order effects are similarly modular. For instance, the proof of the law can be reused to reason about larger programs even if they involve additional (higher-order) effects. Crucially, correctness proofs of elaborations also compose: combining two correct elaborations automatically yields a proof of correctness for the composite elaboration (Section 5.9). This demonstrates that hefty trees enjoy the same, if not more, modularity properties than algebraic effects, which require sophisticated reasoning about fusion to compose proofs of handler correctness (Yang & Wu, Reference Yang and Wu2021).

Finally, we remark that elaborations and their correctness proofs support a more fine-grained composition than coproducts. Building on techniques developped by van der Rest et al. (Reference van der Rest, Poulsen, Rouvoet, Visser and Mosses2022), composition operators consume a separation witness, which act as a specification of which equations to identify across theories. This way, we avoid indiscriminate summation of effect theories. For more details we refer to the artifact accompanying this paper (van der Rest & Poulsen, Reference van der Rest and Poulsen2024).