Focal Rewriting in Sequent Calculus

• Focal rewriting is a rewriting system where classical cut elimination is restricted by focusing, turning a non-confluent process into a confluent one.
• It employs the Curien–Herbelin syntactic kit to reveal a duality between terms, contexts, and commands, supporting dual evaluation strategies like call-by-name and call-by-value.
• The framework introduces finitary syntax with patterns and counterpatterns that quotient away negative connective decomposition order, streamlining proof construction.

Searching arXiv for the cited paper and closely related work on focal rewriting, focusing, and rewriting frameworks. Focal rewriting denotes cut elimination as a rewriting system restricted by the focusing discipline, so that only reductions compatible with the phase structure of proof search are allowed. In the formulation developed in “The duality of computation under focus” (Curien et al., 2010), it is the operational face of classical sequent calculus under polarity, expressed through the Curien–Herbelin syntactic kit of terms, contexts, and commands. Its central claim is that the focused restriction turns general classical cut elimination—which is non confluent—into a simple confluent rewriting system; its strongest formulation replaces trees of focal decompositions by a finitary syntax of patterns and counterpatterns that quotients away the order of decomposition of negative connectives (Curien et al., 2010).

1. Duality of computation and the focused viewpoint

The underlying framework is the Curien–Herbelin syntactic kit, a sequent-calculus-inspired $\lambda$-calculus with an explicit duality between terms (expressions), contexts (stacks, continuations), and commands. In core $\lambda\mu\tilde\mu$ style, commands are of the form $c ::= \langle M \mid e\rangle$ (written in the paper as ${M}{e}$), expressions include variables, abstractions, and $\mu\alpha.c$, and contexts include continuation variables, stacks, and $\tilde\mu x.c$. The typing discipline uses three judgments: $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$ Here $\Gamma$ and $\Delta$ are multisets of formulas, and $(\Gamma \vdash \Delta)$ is a classical sequent. Terms and contexts are dual; $\lambda\mu\tilde\mu$0 captures a continuation $\lambda\mu\tilde\mu$1 and gives a term, while $\lambda\mu\tilde\mu$2 captures a variable $\lambda\mu\tilde\mu$3 and gives a context (Curien et al., 2010).

This duality is also the basis for the paper’s treatment of evaluation strategy. Call-by-name and call-by-value appear as a dual pair of focal evaluation strategies: in CBV, contexts consume values, while in CBN, contexts consume expressions in a stack-like way. The paper identifies these as two dual focused sequent calculi, $\lambda\mu\tilde\mu$4 and $\lambda\mu\tilde\mu$5, built on the same syntactic kit. A plausible implication is that focal rewriting is not merely a normalization discipline for proofs; it is also a typed operational semantics for classical computation in which continuations and variables are treated symmetrically.

The term focus is inherited from Andreoli’s focalisation discipline. In the classical setting of the paper, proof search alternates between a left phase (negative phase), where one repeatedly applies reversible rules on the left and may stop at any time, and a right phase (positive phase), where one chooses one formula on the right to focus on and decomposes it hereditarily using irreversible rules until reaching either an axiom or a right negation, which returns the derivation to a left phase (Curien et al., 2010).

2. Abstract machines, sequent calculus, and the $\lambda\mu\tilde\mu$6 kit

A distinctive feature of the theory is that it is obtained by reading sequent-calculus structure directly out of abstract machines. For call-by-name, a Krivine machine configuration has the form $\lambda\mu\tilde\mu$7, where $\lambda\mu\tilde\mu$8 is a $\lambda\mu\tilde\mu$9-term and $c ::= \langle M \mid e\rangle$0 is a stack or continuation. The operational rules

$c ::= \langle M \mid e\rangle$1

already have the shape of sequent-calculus cut. Typing is again stratified into expressions, contexts, and commands, and the command-typing rule

$c ::= \langle M \mid e\rangle$2

is explicitly identified as cut (Curien et al., 2010).

For control, Felleisen’s $c ::= \langle M \mid e\rangle$3 is internalized by Parigot’s $c ::= \langle M \mid e\rangle$4-binder: $c ::= \langle M \mid e\rangle$5 Typing becomes fully classical,

$c ::= \langle M \mid e\rangle$6

with activation

$c ::= \langle M \mid e\rangle$7

and the reduction rule

$c ::= \langle M \mid e\rangle$8

The call-by-value side introduces the dual binder $c ::= \langle M \mid e\rangle$9 through

${M}{e}$0

with typing

${M}{e}$1

and reduction

${M}{e}$2

Combining ${M}{e}$3 and ${M}{e}$4 yields the Curien–Herbelin ${M}{e}$5-calculus: ${M}{e}$6 equipped with a fourth judgment for values. The resulting CBV-flavored reductions are

${M}{e}$7

These equations are the direct computational substrate from which focal rewriting is later extracted (Curien et al., 2010).

3. LK as a rewrite system and the role of explicit substitutions

The next step is a term language for classical sequent calculus ${M}{e}$8 with conjunction, disjunction, and negation. Sequents occur in three forms: ${M}{e}$9 The syntax is deliberately dissymmetric in a way that anticipates polarity. On the right, negation introduction is $\mu\alpha.c$0, conjunction introduction is $\mu\alpha.c$1, and disjunction introduction uses two separate rules $\mu\alpha.c$2 or $\mu\alpha.c$3. On the left, conjunction and disjunction are reversible through $\mu\alpha.c$4 and $\mu\alpha.c$5. The crucial design choice is that right disjunction is an irreversible choice, while left conjunction is reversible; this is exactly the later split between positive and negative behavior (Curien et al., 2010).

To make cut elimination rewrite-theoretic, the paper refines the syntax into an “atomic” form with explicit substitutions: $\mu\alpha.c$6

$\mu\alpha.c$7

$\mu\alpha.c$8

An explicit substitution is a list

$\mu\alpha.c$9

This allows multi-cut to be represented uniformly, including cases where the active cut formula is not the one immediately visible in a $\tilde\mu x.c$0 redex (Curien et al., 2010).

Cut elimination is then split into three classes of rules. The control rules initiate substitution: $\tilde\mu x.c$1 The logical rules eliminate principal cuts, for example

$\tilde\mu x.c$2

$\tilde\mu x.c$3

and the two disjunction branches. The commutative rules are given not as separate permutative conversions for each connective, but as propagation of explicit substitutions: $\tilde\mu x.c$4 together with variable lookup and pushing substitutions under binders. The paper summarizes this by the slogan:

“Commutative cut elimination rules are explicit substitution propagation rules.”

This slogan is exact, not metaphorical: the commutative cut rules of proof theory are identified with the uniform propagation behavior of $\tilde\mu x.c$5 (Curien et al., 2010).

At the unrestricted $\tilde\mu x.c$6 level, rewriting is non confluent. The paper explicitly contrasts this with Lafont’s critical pair

$\tilde\mu x.c$7

which witnesses the “wild non-confluence” of general classical cut elimination. Focal rewriting arises by forbidding precisely these unfocused symmetric situations (Curien et al., 2010).

4. Focused restriction, phase structure, and confluence

The focused system $\tilde\mu x.c$8 and its term language $\tilde\mu x.c$9 impose polarity directly on syntax and typing. Positive formulas are generated by

$$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$0

while negative formulas are their De Morgan duals. The calculus adds a stoup judgment for values,

$$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$1

and distinguishes values from general expressions: $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$2 Typing enforces that right introduction of $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$3, $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$4, and $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$5 always builds a value; left introduction is performed by the $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$6-binders; and values may be coerced to expressions via $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$7. The paper annotates phase transitions explicitly by $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$8 for positive-phase steps and $$\Gamma \vdash M : A \mid \Delta,\qquad \Gamma \mid e : A \vdash \Delta,\qquad c : (\Gamma \vdash \Delta).$$9 for negative-phase steps (Curien et al., 2010).

The focal rewrite system retains the same three classes of rules—control, logical, and commutation—but only for focal redexes. The key examples are

$\Gamma$0

$\Gamma$1

$\Gamma$2

plus the two disjunction branches and the same generic substitution propagation machinery. What changes is the shape of redexes: the critical pair $\Gamma$3 has only one reduction path, because the focusing discipline blocks the symmetric alternative (Curien et al., 2010).

The paper proves that the system is an orthogonal higher-order rewrite system in the sense of Nipkow: the rules are left-linear and there are no critical pairs. Hence it is confluent. It also proves weak normalization, subject reduction, a characterization of normal forms as contractions such as $\Gamma$4 and $\Gamma$5, and completeness: if $\Gamma$6 is provable in $\Gamma$7, then it is provable in $\Gamma$8. Together with cut elimination for $\Gamma$9, this yields the focalisation theorem that every sequent provable in $\Delta$0 has a cut-free focalised proof (Curien et al., 2010).

System	Restriction	Rewriting behavior
$\Delta$1 term language	none	non confluent
$\Delta$2 / $\Delta$3	values and phase discipline	confluent
$\Delta$4	patterns/counterpatterns	quotients negative-order bureaucracy

A common misconception is that focusing is only a proof-search heuristic. In this formulation, it also supplies a canonical operational semantics for classical calculi with control: the focused restriction is what removes Lafont’s critical pair and makes rewriting orthogonal (Curien et al., 2010).

5. Strong focalisation: patterns, counterpatterns, and synthetic connectives

The final strengthening of focal rewriting is the passage to a pattern–counterpattern game and the synthetic calculus $\Delta$5. The first step restricts left contexts to a single $\Delta$6, where counterpatterns are generated by

$\Delta$7

A counterpattern encodes the whole left decomposition of a positive formula in one object. For example, for $\Delta$8 the counterpattern

$\Delta$9

captures the conjunction/disjunction tree shape uniformly (Curien et al., 2010).

The second step introduces patterns and synthetic values. Elementary values are

$(\Gamma \vdash \Delta)$0

patterns are

$(\Gamma \vdash \Delta)$1

and a value is written as

$(\Gamma \vdash \Delta)$2

Thus a value is a pattern skeleton decorated with basic leaves. Patterns and counterpatterns are linear in the precise sense stated in the paper: each variable or $(\Gamma \vdash \Delta)$3 appears at most once in a pattern, while in a counterpattern a variable may appear twice in a sum $(\Gamma \vdash \Delta)$4 but only in compatible ways (Curien et al., 2010).

Interaction is governed by the orthogonality relation $(\Gamma \vdash \Delta)$5, defined inductively by matching variables, negations, products, and sums. Contexts become records over patterns,

$(\Gamma \vdash \Delta)$6

and the focal logical rule $(\Gamma \vdash \Delta)$7 performs an entire focal phase in one step: given a value $(\Gamma \vdash \Delta)$8 with pattern $(\Gamma \vdash \Delta)$9 orthogonal to $\lambda\mu\tilde\mu$00, the reduction selects the branch $\lambda\mu\tilde\mu$01 and substitutes the leaves of the pattern into it. The paper formulates this as: pattern–counterpattern matching picks the appropriate branch $\lambda\mu\tilde\mu$02 and substitutes the leaves of the pattern into it (Curien et al., 2010).

The significance of the synthetic system is proof-theoretic quotienting. Because patterns abstract away the order of decomposition of negative connectives, different focalised proof trees that differ only by that order map to the same pattern/counterpattern derivation. The paper’s example is

$\lambda\mu\tilde\mu$03

where decomposing the left tensor first or the right tensor first yields different trees in $\lambda\mu\tilde\mu$04, but the same pattern

$\lambda\mu\tilde\mu$05

and the same counterpattern in $\lambda\mu\tilde\mu$06. This is why the paper describes the result as a fully focalised finitary syntax and as a quotient on focalised proofs (Curien et al., 2010).

6. Broader rewriting contexts and later interpretations

The sense of focal rewriting just described is specific to focused sequent calculus and classical computation. A broader literature suggests related, but distinct, uses of “focus” as locality, priority, or controlled overlap.

In network rewriting over PROPs, rewriting steps are inherently localized to a chosen subexpression or embedding. “Network Rewriting I: The Foundation” develops a subexpression concept with both algebraic and graph-theoretic characterisations, based on symmetric join, homeomorphisms, and embeddings, and treats ambiguities as interactions between such local foci (Hellström, 2012). This suggests a graph-theoretic analogue of focal rewriting in which the focus is an embedded subnetwork rather than a focal proof phase.

In higher-dimensional rewriting, “Rewriting in Gray categories with applications to coherence” studies critical branchings in precategories and Gray presentations. It shows that a finite rewriting system in precategories admits a finite number of critical pairs, extends Squier’s theorem to this setting, and proves coherence from convergence for structures such as monoids, adjunctions, and Frobenius monoids (Forest et al., 2021). Here the relevant notion of focus is the minimal local branching at which two 3-cells overlap.

In finite-state string rewriting, “Compiling Rewrite Rules to Finite-State Transducers with the Worsening Trick” presents what the details explicitly describe as a finite-state account of “focal rewriting”: generate all legal rewrite candidates, then filter candidates that are worse than another candidate for the same input. Obligatory, leftmost, longest, shortest, weighted, and parallel behaviors are all expressed by regular worsening relations over candidate rewrites (Hulden et al., 8 Jun 2026). The family resemblance to focal rewriting in logic lies in restricting the operational space to undominated candidates under a discipline of preference.

Other adjacent usages are more metaphorical. “RewriteLM” states that “focal rewriting” is not a term used in the paper, but the model is designed for targeted, controllable text rewriting under natural-language instructions such as formality, conciseness, elaboration, tone, or lexical substitution (Shu et al., 2023). Likewise, “Rewriting in Free Hypergraph Categories” does not use the phrase, but its DPOI account of rewriting modulo Frobenius structure yields interface-based local rewriting in which structural bureaucracy is absorbed into the hypergraph representation (Zanasi, 2017).

Across these settings, the strongest common thread is controlled locality: in focused proof theory, locality is phase-structured cut elimination; in network, hypergraph, Gray-category, and finite-state settings, it is a restriction on where or how rewrites are allowed to compete. The specific technical meaning of focal rewriting, however, remains the one fixed by the focused classical calculus of Curien–Herbelin syntax, explicit substitutions, and confluent cut elimination under focus (Curien et al., 2010).