Вы находитесь на странице: 1из 20

Interpolation in extensions of first-order logic

arXiv:1807.11848v1 [math.LO] 31 Jul 2018

G. Gherardi∗ P. Maffezioli† E. Orlandelli‡

Abstract
We provide a constructive proof of the interpolation theorem for exten-
sions of classical first order logic with a special type of geometric axioms,
called singular geometric axioms. As a corollary, we obtain a direct proof
of interpolation for first-order logic with identity.
Interpolation is a central result in first-order logic. It asserts that for any
theorem A → B there exists a formula C, called interpolant, such that A → C
and C → B are also theorems and C only contains non-logical symbols that are
contained in both A and B. The aim of this paper is to extend interpolation
beyond first-order logic. In particular, we show how to prove interpolation
in Gentzen’s sequent calculi with singular geometric rules, a special case of
geometric rules investigated in [5]. Interestingly, singular geometric rules include
the rules for first-order logic with identity (as well as those for the theory of strict
partial orders and other order theories). Thus, from interpolation for singular
geometric rules, we obtain a direct proof of interpolation for first-order logic
with identity, while the existing techniques are based on indirect maneuvers
such as the reduction to the case of pure first-order logic.
Traditionally, interpolation1 has been established for logics allowing a cut-
free systematization and it is often regarded as a consequence, or even an appli-
cation, of cut elimination. In Gentzen’s LK, for example, interpolation follows
from cut elimination and a lemma, reportedly due to Maehara, which shows
how to effectively find an interpolant for any sequent with a cut-free deriva-
tion (cf. [9, ch.1, §6] for a proof of interpolation from Maehara’s Lemma and
cut-elimination in LK). When it comes to extensions of LK, however, it is not
entirely obvious how to apply Maehara’s lemma, since such extensions are not
generally cut free. For example, it is a known fact that in the calculus LKe ,
an extension of LK for first-order logic with identity, cuts on identities s = t
∗ Dipartimento di Filosofia e Comunicazione, Università di Bologna,
guido.gherardi@unibo.it
† Institut für Philosophie II, Ruhr-Universität Bochum, paolo.maffezioli@ruhr-uni-

bochum.de
‡ Dipartimento di Filosofia e Comunicazione, Università di Bologna, euge-
nio.orlandelli@unibo.it
1 Here we shall only be interested in constructive proofs of interpolation and we will disre-

gard semantic or, more generally, non-constructive methods. Moreover, interpolation must be
understood here as short for Craig’s interpolation, named after William Craig who was first
to prove it in [1] as early as 1957.

1
are not eliminable (cf. Theorem 6 in [9, ch.1 §7], where these cuts are called
inessential). Fortunately, even without Maehara’s lemma, interpolation can still
be proved for first-order logic with identity. The drawback of the existing proof
is, however, that there is no direct procedure as to find an interpolant. In
particular, in [10, ch.4 §4.4] it is shown how to use a translation so as to re-
duce interpolation for first-order logic with identity to interpolation for pure
first-order logic. Interestingly, during the last twenty years the situation has
changed. In a series of works starting from [6], Negri and von Plato have shown
how to recover cut elimination for extensions of G3c, a calculus equivalent to LK
in which all structural rules (not just cut) are admissible. Of particular interest
for us are the extensions of G3c with geometric rules, for which Negri provided
a cut-elimination procedure in [5].
Building on these works, the primary aim of this paper is to answer the
question as to whether, and to what extent, it is possible to recover interpo-
lation as well. We first introduce a subclass of geometric rules, the class of
singular geometric rules, and we show that Maehara’s lemma holds when G3c
is extended with any of such rules (Lemma 6). The proof, though quite long, is
not conceptually more difficult than Maehara’s and can be thought of as being
a generalization of it. Then we consider some applications. First comes first-
order logic with identity. We show that all the rules of the cut-free calculus of
[6] for first-order logic with identity are singular geometric. Thus, we obtain as
a corollary of Lemma 6, that Maehara’s lemma holds for first-order logic with
identity. Unlike the standard proof, however, our does give a procedure to find
an interpolant, with no need to use any sort of identity-elimination translation.2
Second, as an application of Lemma 6 to mathematical theories, we show that
Maehara’s lemma holds for the theory of strict partial orders.

The language. The language L is a first-order language with individual


constants and no functional symbols. Terms (s, t, u, . . . ) are either variables
(x, y, z, . . . ) or individual constants (a, b, c . . . ). L contains also denumerably
many k-ary predicates P k , Qk , Rk , . . . for each k > 0. L may also contain
the identity. We agree that all predicates, except identity, are non-logical. It
is convenient to have two propositional constants ⊥ (falsity) and ⊤ (truth).
Formulas are built up from atoms P k (t1 , . . . , tk ), ⊥ and ⊤ using propositional
connectives ∧ (conjunction), ∨ (disjunction), → (implication), ∃ (existential
quantifier) and ∀ (universal quantifier) as usual. We use P, Q, R, . . . for arbi-
trary atoms, A, B, C, . . . for arbitrary formulas and Γ, ∆, Π, . . . for multisets
of formulas. The substitution of a variable x with a term t in a term s (in a
formula A, in a multiset Γ) will be indicated as s[ tx ] (A[ tx ] and Γ[ tx ], respec-
tively) and defined as usual. To indicate the simultaneous substitution of the
list of variables x1 , . . . , xn (abbreviated in x̄) with the list of terms t1 , . . . , tn
(abbreviated in t̄), we use [ t̄x̄ ] in place of [ xt11... tn
... xn ]. Later on we shall also need a
more general notion of substitution of terms (not just variables) for terms which
2 It should be noted, however, that the standard proof accommodates also n-ary functional

symbols, whereas ours does not, the only functional symbols being considered here are the
individual constants.

2
will be proved to preserve derivability (Lemma 5).
Let FV(A) be the set of free variables of a formula A and let Con(A) be the
set of its individual constants. We agree that the set of terms Ter(A) of A is
FV(A) ∪ Con(A). Moreover, if Rel(A) is the set of non-logical predicates of A
then we define the language
S L(A) of A as Ter(A) ∪ Rel(A) and the language
L(Γ) of a multiset Γ as L(A), for A ∈ Γ. Notice that = ∈ / L(A), for all A.

The sequent calculus. The calculus G is a variant of LK for classical first-


order logic, originally introduced by Gentzen in [2]. In the literature, especially
in [10] and [7], G is commonly referred to as G3c but we shall deliberately use G
as it is more readable. Recall that a sequent is a pair of multisets Γ, ∆ indicated
as Γ ⇒ ∆. G consists of the following initial sequents and logical rules (where
the variable y in R∀ and L∃ must not occur free in the conclusion of these rules):

Initial sequents

Γ ⇒ ∆, ⊤ ⊥, Γ ⇒ ∆
P, Γ ⇒ ∆, P

Logical rules

A, B, Γ ⇒ ∆ Γ ⇒ ∆, A Γ ⇒ ∆, B
L∧ R∧
A ∧ B, Γ ⇒ ∆ Γ ⇒ ∆, A ∧ B

A, Γ ⇒ ∆ B, Γ ⇒ ∆ Γ ⇒ ∆, A, B
L∨ R∨
A ∨ B, Γ ⇒ ∆ Γ ⇒ ∆, A ∨ B

Γ ⇒ ∆, A B, Γ ⇒ ∆ A, Γ ⇒ ∆, B
L→ R→
A → B, Γ ⇒ ∆ Γ ⇒ ∆, A → B

A[ tx ], ∀xA, Γ ⇒ ∆ Γ ⇒ ∆, A[ yx ]
L∀ R∀
∀xA, Γ ⇒ ∆ Γ ⇒ ∆, ∀xA

A[ yx ], Γ ⇒ ∆ Γ ⇒ ∆, ∃xA, A[ tx ]
L∃ R∃
∃xA, Γ ⇒ ∆ Γ ⇒ ∆, ∃xA

A derivation in G is a tree of sequents which grows according to the rules of


G and whose leaves are initial sequents. A derivation of a sequent is a derivation
concluding that sequent and a sequent is derivable when there is a derivation
of it. As usual, we consider only pure-variable derivations: bound and free
variables are kept distinct, and no two rule instances have the same variable as
eigenvariable, see [10, p. 38]. The height of a derivation is defined inductively as
follows: the derivation height of an initial sequent is 0, the derivation height of a
derivation of a conclusion of a one-premise logical rule is the derivation height of
its premise plus 1, and the derivation height of a derivation of a conclusion of a

3
two-premise logical rule is the maximum of the derivation heights of its premises
plus 1. A sequent is n-derivable if it is derivable and its derivation height is
at least n. A rule is admissible if the conclusion is derivable whenever the
premises are derivable; a rule is height-preserving admissible if the conclusion is
n-derivable whenever the premises are n-derivable. Derivations will be denoted
by D, D1 , D2 , . . . . We agree to use D ⊢ Γ ⇒ ∆ to indicate that D is a derivation
of Γ ⇒ ∆ and ⊢ Γ ⇒ ∆ to indicate that Γ ⇒ ∆ is derivable; finally ⊢n Γ ⇒ ∆
indicates that Γ ⇒ ∆ is n-derivable.
Although not primitive, G has also the following structural rules (weakening,
contraction and cut).

Structural rules

Γ⇒∆ W
Γ⇒∆ W
A, Γ ⇒ ∆ Γ ⇒ ∆, A

A, A, Γ ⇒ ∆ Γ ⇒ ∆, A, A
Ctr Ctr
A, Γ ⇒ ∆ Γ ⇒ ∆, A

Γ ⇒ ∆, A A, Π ⇒ Σ
Cut
Γ, Π ⇒ ∆, Σ

The reason for not taking any of them as primitive is that they are all admis-
sible in G. In fact, weakening and contraction are height-preserving admissible.
Moreover, substitution is height-preserving admissible and initial sequents with
arbitrary formulas as principal are derivable.
Theorem 1 (Troelstra & Schwichtenberg). In G the following hold:

1. if ⊢n Γ ⇒ ∆ and t is free for x in Γ, ∆ then ⊢n Γ[ tx ] ⇒ ∆[ tx ];


2. ⊢ A, Γ ⇒ ∆, A;
3. if ⊢n Γ ⇒ ∆ then ⊢n A, Γ ⇒ ∆;
4. if ⊢n Γ ⇒ ∆ then ⊢n Γ ⇒ ∆, A;
5. if ⊢n A, A, Γ ⇒ ∆ then ⊢n A, Γ ⇒ ∆;
6. if ⊢n Γ ⇒ ∆, A, A then ⊢n Γ ⇒ ∆, A;
7. if ⊢ Γ ⇒ ∆, A and ⊢ A, Π ⇒ Σ then ⊢ Γ, Π ⇒ ∆, Σ.

For a proof of Theorem 1, the interested reader is referred to [10] and [7].

From axioms to rules. Extensions of G are not, in general, cut free. This
means Theorem 1 does not hold in the presence of new initial sequents or rules.
For example, a natural way to extend G to cover first-order logic with identity
is to allow derivations in G to start with initial sequents of the form ⇒ s = s

4
and s = t, P [sx ] ⇒ P [tx ], corresponding to the reflexivity of identity and Leibniz’s
principle of indescernibility of identicals, respectively (we call these sequents S1
and S2 ). Notice that S2 is in fact a scheme which becomes s = t, s = s ⇒ t = s,
when P is x = s. From this, via cut on ⇒ s = s, one derives s = t ⇒ t = s,
namely the symmetry of identity. However, such a sequent has no derivation
without cut. Therefore, cut is not admissible in G + {S1 , S2 }, though it is
admissible in the underlying system G.
In [6] Negri and von Plato have shown how to recover cut elimination for
first-order logic with identity by transforming S1 and S2 into an equivalent pair
of rules of the form:

s = s, Γ ⇒ ∆ P [tx ], s = t, P [sx ], Γ ⇒ ∆
Ref Repl
Γ⇒∆ s = t, P [sx ], Γ ⇒ ∆
If one replaces S1 and S2 with the corresponding rules, it is easy to derive
s = t ⇒ t = s without any application of cut. More generally, cut elimination
holds in G + {Ref , Repl } (cf. Theorem 4.2 in [6] and [7, ch.6 §5]). This result
can be, and has been, extended in different directions. Here we are particularly
interested in the fact, established by [5], that cut elimination holds in extensions
of G with geometric rules (of which the rules of identity are special cases). The
result will be reviewed briefly in the next section, while for a more thorough
discussion on this topic the reader is referred to [5] or the monograph [8].

Geometric theories. A geometric axiom is a formula following the geo-


metric axiom scheme below:

∀x̄(P1 ∧ · · · ∧ Pn → ∃ȳ1 M1 ∨ · · · ∨ ∃ȳm Mm )


where each Mi is a conjunction of a list of atoms Qi1 , . . . , Qik and none of
the variables in any ȳi are free in the Pj s. We shall conveniently abbreviate
Qi1 , . . . , Qik in Qi . In a geometric axiom if m = 0 then the consequent of →
becomes ⊥, whereas if n = 0 the antecedent of → becomes ⊤. A geometric theory
is a theory containing only geometric axioms. In sequent calculus a geometric
theory can be formulated by adding on top of G a rule with m premises following
the geometric rule scheme below:
Q∗1 , P1 , . . . , Pn , Γ ⇒ ∆ · · · Q∗m , P1 , . . . , Pn , Γ ⇒ ∆
P1 , . . . , Pn , Γ ⇒ ∆
where each Q∗i is obtained from Qi by replacing every variable in ȳi with a
variable which does not occur free in the conclusion. Without loss of generality,
we assume that each ȳi consists of a single variable. A geometric rule is a rule
following the geometric rule scheme.
We depart from [5] in treating a geometric axiom in which m = 0 as an
inference rule with one premise. More precisely, we do not formulate a geometric
axiom of the form ∀x̄(P1 ∧ · · · ∧ Pn → ⊥) as a 0-ary rule

P1 , . . . Pn , Γ ⇒ ∆

5
Instead, ∀x̄(P1 ∧ · · · ∧ Pn → ⊥) will be analyzed into
⊥, P1 , . . . , Pn , Γ ⇒ ∆
P1 , . . . , Pn , Γ ⇒ ∆
The reason for this choice will be discussed at some length while proving
interpolation for the theory of strict partial orders. Importantly, cut elimination
for geometric rules, due to Negri in [5], is still in place. Let Rg be a finite set of
geometric rules. Then,
Theorem 2 (Negri). Theorem 1 holds in G + Rg .
For a proof of Theorem 2 one can consult the original article [5] by Negri or
the monograph [8, ch. 8].
At this point one may hope to prove interpolation in G + Rg . However, the
class of geometric rules seems to be too broad and some restriction is needed
to adapt Maehara’s original technique to geometric rules. We shall therefore
narrow down our attention to a subclass of Rg , called the class of singular
geometric rules.

Singular geometric theories. A singular geometric axiom is a geometric


axiom in which (a) there is at at most one non-logical predicate and (b) if such
a non-logical predicate occurs in the consequent of → then it also occurs in the
antecedent. Moreover, for the sake of simplicity, we assume that no constant
occurs in a singular geometric axiom. In principle, we could also consider ax-
ioms dealing with some constant respecting the following additional restraint
analogous to (b) above: if c occurs in some Qi , then it occurs also in all Pj s;
but this wouldn’t add new particularly interesting cases. A singular geometric
theory is a theory containing only singular geometric axioms. In sequent calcu-
lus a singular geometric theory can be formulated by adding to G a geometric
rule satisfying the following condition.
|Rel(Q∗1 , . . . , Q∗m , P1 , . . . , Pn )| 6 1 & Rel(Q∗1 , . . . , Q∗m ) ⊆ Rel(P1 , . . . , Pn ) (⋆)

Singular geometric axioms are ubiquitous in mathematics and, more gener-


ally, in relational theories. Here, for example, is an incomplete list of singular
geometric axioms for a binary relation R.

6
R is irreflexive ∀x(xRx → ⊥)
R is transitive ∀x∀y∀z(xRy ∧ yRz → xRz)
R is intransitive ∀x∀y∀z(xRy ∧ yRz ∧ xRz) → ⊥
R is co-transitive ∀x∀y∀z(xRy → xRz ∨ zRy)
R is symmetric ∀x∀y(xRy → yRx)
R is asymmetric ∀x∀y(xRy ∧ yRx → ⊥)
R is anti-symmetric ∀x∀y(xRy ∧ yRx → x = y)
R is Euclidean ∀x∀y∀z(xRz ∧ yRz → xRy)
R is left-unique ∀x∀y∀z(xRz ∧ yRz → x = y)
R is right-unique ∀x∀y∀z(zRx ∧ zRy → x = y)
R is connected ∀x∀y∀z(xRy ∧ xRz → yRz ∨ zRy)
R is nilpotent ∀x∀y∀z(xRz ∧ zRy → ⊥)
R is a left ideal ∀x∀y∀z(xRy → xRz)
R is a right ideal ∀x∀y∀z(xRy → zRy)
R is rectangular ∀x∀y∀z∀v(xRz ∧ vRy → xRy)
R is dense ∀x∀y(xRy → ∃z(xRz ∧ zRy))
R is confluent ∀x∀y∀z(xRy ∧ xRz → ∃u(yRu ∧ zRu))
Singular geometric axioms play an important role in logic, too. Since identity
is a logical predicate, the axioms of identity are singular geometric.

= is reflexive ∀x(x = x)
= satisfies the indescernibility of identicals ∀x∀y(x = y ∧ P [xz ] → P [yz ])
Clearly the reflexivity of identity satisfies (⋆) as no non-logical predicate
occurs in it at all. But also the indescernibility of identicals satisfies (⋆), since if
the atom P contains a non-logical predicate then such a predicate occurs both
in the consequent and antecedent of →. It follows that first-order logic with
identity is axiomatized by singular geometric axioms. Also the theory of strict
partial orders resulting by adding on top of first-order logic with identity the
irreflexivity and transitivity of the relation < is a singular geometric theory.
Cut elimination for singular geometric rules clearly follows from cut elim-
ination for geometric rules. More precisely, let Rs be a finite set of singular
geometric rules, then:
Corollary 3. Theorem 1 holds in G + Rs .

Proof. Immediate from Theorem 2 and the fact that, by definition, all singular
geometric rules are geometric rules.

Interpolation. The standard proof of interpolation rests on a result due


to Maehara which appeared (in Japanese) in [3] and was later made available
to international readership by Takeuti in his [9]. While interpolation is a result
about logic, regardless the formal system (sequent calculus, natural deduction,
axiom system, etc), Maehara’s lemma is a “sequent-calculus version” of interpo-
lation. Thus, Maehara’s lemma comes close to be synonymous of interpolation
and we will use these expressions interchangeably. Although originally Maehara

7
proved his lemma for LK, it is easy to adapt the proof so that it holds also in G
(cf. [10, ch.4 §4.4] for a proof of Maehara’s lemma in G).
We recall from [10] some basic definitions. A partition of a sequent Γ ⇒ ∆
is an expression Γ1 ; Γ2 ⇒ ∆1 ; ∆2 , where Γ = Γ1 , Γ2 and ∆ = ∆1 , ∆2 . A
split-interpolant of a partition Γ1 ; Γ2 ⇒ ∆1 ; ∆2 is a formula C such that:

I ⊢ Γ1 ⇒ ∆1 , C
II ⊢ C, Γ2 ⇒ ∆2
III L(C) ⊆ L(Γ1 , ∆1 ) ∩ L(Γ2 , ∆2 )
C
We use Γ1 ; Γ2 = ⇒ ∆1 ; ∆2 to indicate that C is a split-interpolant for
Γ1 ; Γ2 ⇒ ∆1 ; ∆2 . Moreover, we say that a C satisfying conditions (I)
and (II) satisfies the derivability conditions for being a split-interpolant for the
partition Γ1 ; Γ2 ⇒ ∆1 ; ∆2 , whereas if C satisfies (III) we say that it satisfies
the language condition for being a split-interpolant for the same partition.
Lemma 4 (Maehara). In G every partition of a derivable sequent has a split-
interpolant.
The proof is by induction on the height h of the derivation. If h = 0 then
Γ ⇒ ∆ is an initial sequent and the proof is as in [10].3 If h > 0 one uses as
induction hypothesis (IH) the fact that any partition of the premises of a rule R
has a split-interpolant. For a detailed proof the reader is again referred to [10].
Now we extend Lemma 4 to extensions of G with singular geometric rules.
Preliminary to this is a result concerning a general notion of substitution. The
result states that an arbitrary term u (possibly a constant) can be replaced
by a term t without impairing derivability in G, as long as no rule involves u
essentially – i.e., we exclude geometric rules representing axioms where u occurs
and it is a constant (the lemma wouldn’t hold, e.g., for rules representing Peano’s
axioms for 0). Notice that this restraint is indeed satisfied by our singular
geometric rules because we have excluded axioms containing constants.
Lemma 5. If ⊢n Γ ⇒ ∆, t is free for u in Γ, ∆, and no geometric rule involves
u essentially, then ⊢n Γ[tu ] ⇒ ∆[tu ].
Proof. If u is a variable see Theorem 1.1. Otherwise, let u be an individual
constant. We can think of the derivation D of Γ ⇒ ∆ as
Γ′ ⇒ ∆′ [u ]
⇒ ∆′ [uz ] z
Γ′ [uz ]

where Γ′ ⇒ ∆′ is like Γ ⇒ ∆ save that it has a fresh variable z in place of u.


Note that this is always feasible for purely logical derivations, and it is feasible
3 Notice, however, the the proof given in [10] contains a misprint and the split-interpolant

for the partition of the initial sequent Γ1 , P ; Γ2 ⇒ ∆1 , P ; ∆2 (their notation adjusted to


ours) is ⊥ (and not ⊥ → ⊥ as stated in [10, p.117])

8
for derivations involving geometric rules as long as these rules do not involve
essentially the constant u. We transform D into

Γ′ ⇒ ∆′
[t ]
⇒ ∆′ [tz ] z
Γ′ [tz ]

where t is free for z since we assumed it is free for u in Γ ⇒ ∆. We have


thus found a derivation of Γ[tu ] ⇒ ∆[tu ] of same height of the derivation D of
Γ ⇒ ∆.
As above, let Rs be a finite set of singular geometric rules. In fact, in the
proof of Lemma 6 we shall only consider singular geometric rules where each
Q∗i is a single atom Q∗i . More precisely, we consider singular geometric rules of
the form
Q∗1 , P1 , . . . , Pn , Γ ⇒ ∆ · · · Q∗m , P1 , . . . , Pn , Γ ⇒ ∆
P1 , . . . , Pn , Γ ⇒ ∆
This allows some notational simplification and will significantly improve the
readability of the proof. It does not impair the generality of the result.
Lemma 6. In G+Rs every partition of a derivable sequent has a split-interpolant.
Proof. The proof extends that of Lemma 4. Let R be a rule in Rs and let
P1 , . . . , Pn , Γ ⇒ ∆ be its conclusion. We have four exhaustive cases, corre-
sponding to the possible partitions of the conclusion:

1. P1 , . . . , Pn , Γ1 ; Γ2 ⇒ ∆1 ; ∆2
2. Γ1 ; P1 , . . . , Pn , Γ2 ⇒ ∆1 ; ∆2
3. Pi1 , . . . , Pij , Γ1 ; Pij+1 , . . . , Pin , Γ2 ⇒ ∆1 ; ∆2
4. Γ1 ; Γ2 ⇒ ∆1 ; ∆2

where the fourth one is obtained as a limit case when n = 0.

Case 1. By IH every partition of each premise of R has a split-interpolant.


In particular, for all k, 1 6 k 6 m, there is a Ck such that

Ik ⊢ Q∗k , P1 , . . . , Pn , Γ1 ⇒ ∆1 , Ck
IIk ⊢ Ck , Γ2 ⇒ ∆2
IIIk L(Ck ) ⊆ L(Q∗k , P1 , . . . , Pn , Γ1 , ∆1 ) ∩ L(Γ2 , ∆2 )

Consider the following derivation D1 , where the topmost sequents are all
derivable by (Ik ). Notice that the double inference line stands for simultaneous
applications of logical or structural rules. Let Π be P1 , . . . , Pn .

9
Q∗1 , Π, Γ1 ⇒ ∆1 , C1 Q∗m , Π, Γ1 ⇒ ∆1 , Cm
W W
Q∗1 , Π, Γ1 ⇒ ∆1 , C1 , . . . , Cm · · · Q∗m , Π, Γ1 ⇒ ∆1 , C1 , . . . , Cm
R
Π, Γ1 ⇒ ∆1 , C1 , . . . , Cm
R∨
Π, Γ1 ⇒ ∆1 , C1 ∨ · · · ∨ Cm

Notice that by condition (IIIk ) every variable occurring free in each Ck must
occurs also in L(Γ2 , ∆2 ). It follows that the variable conditions on R, if any,
are satisfied and the application of R is legitimate. Then consider a derivation
D2 , the topmost sequents of which are all derivable by (IIk ):4
C1 , Γ2 ⇒ ∆2 · · · Cm , Γ2 ⇒ ∆2
L∨
C1 ∨ · · · ∨ Cm , Γ2 ⇒ ∆2
Consider all terms t1 , . . . , tl such that t1 , . . . , tl ∈ Ter(C1 ∨ · · · ∨ Cm ) and
t1 , . . . , t l ∈
/ Ter(P1 , . . . , Pn , Γ1 , ∆1 ) ∩ Ter(Γ2 , ∆2 ) and let t̄ denote such vector
of terms. By (IIIk ) it follows immediately that t 6∈ Ter(P1 , . . . , Pn , Γ1 , ∆1 ).
Moreover, let z̄ be a vector of variables z1 , . . . , zl not occurring free in the
conclusions of D1 and D2 . Now we apply the substitution [ z̄t̄ ] to the conclusion
of D1 so as to obtain by Lemma 5 that:

⊢ P1 , . . . , Pn , Γ1 ⇒ ∆1 , (C1 ∨ · · · ∨ Cm )[ z̄t̄ ]
Notice that the substitution leaves P1 , . . . , Pn , Γ1 , ∆1 unchanged, since none
of the terms in t̄ is in Ter(P1 , . . . , Pn , Γ1 , ∆1 ). Now, by l applications of R∀
(which is applicable since the variables in z̄ are fresh), we obtain

(I) ⊢ P1 , . . . , Pn , Γ1 ⇒ ∆1 , ∀z̄((C1 ∨ · · · ∨ Cm )[ z̄t̄ ])


Similarly, we apply the same substitution [ z̄t̄ ] to the conclusion of D2 and
conclude by Lemma 5 that:

⊢ (C1 ∨ · · · ∨ Cm )[ z̄t̄ ], Γ2 [ z̄t̄ ] ⇒ ∆2 [ z̄t̄ ]


And from here, by l applications of L∀, we conclude

⊢ ∀z̄((C1 ∨ · · · ∨ Cm )[ z̄t̄ ]), Γ2 [ z̄t̄ ] ⇒ ∆2 [ z̄t̄ ]


Now, the inverse substitution [ t̄z̄ ] on the latter gives

(II) ⊢ ∀z̄((C1 ∨ · · · ∨ Cm )[ z̄t̄ ]), Γ2 ⇒ ∆2


Thus, (I) and (II) guarantee that the formula ∀z̄((C1 ∨ · · · ∨ Cm )[ z̄t̄ ]) sat-
isfies the derivability condition for being a split-interpolant for the partition
under consideration. It remains to be shown that it also satisfies the language
condition:
4 To shorten proofs, we represent m consecutive applications of rules L∨ and R∧ as if they

were applicable in parallel.

10
(III) L(∀z̄((C1 ∨ · · · ∨ Cm )[ z̄t̄ ])) ⊆ L(P1 , . . . , Pn , Γ1 , ∆1 ) ∩ L(Γ2 , ∆2 )
Suppose that P h is a predicate in Rel(C1 ∨ · · · ∨ Cm ). We need to show that
(III.1) P h ∈ Rel(P1 , . . . , Pn , Γ1 , ∆1 ) ∩ Rel(Γ2 , ∆2 ) .
From the hypothesis that P h ∈ Rel(C1 ∨ · · · ∨ Cm ) it follows that that P h ∈
Rel(Ck ), for some k such that 1 6 k 6 m. Let l be such k. If P h ∈
Rel(Cl ) then by (IIIk ) we conclude both (i) P h ∈ Rel(Q∗l , P1 , . . . , Pn , Γ1 , ∆1 )
and (ii) P h ∈ Rel(Γ2 , ∆2 ). From (i) we know that P h ∈ Rel(Q∗l ) or P h ∈
Rel(P1 , . . . , Pn , Γ1 , ∆1 ). We show that (III.1) follows from either. Suppose
P h ∈ Rel(P1 , . . . , Pn , Γ1 , ∆1 ). Then (III.1) follows from it and (ii). Suppose
P h ∈ Rel(Q∗l ) ⊆ Rel(Q∗1 , . . . , Q∗m ). Since R is a singular geometric rule, we know
by (⋆) that Rel(Q∗1 , . . . , Q∗m ) ⊆ Rel(P1 , . . . , Pn ). Hence P h ∈ Rel(P1 , . . . , Pn ),
from which (iii) P h ∈ Rel(P1 , . . . , Pn , Γ1 , ∆1 ) follows as well. Finally from (ii)
and (iii) we obtain (III.1).
Suppose that s is a term in Ter(∀z̄((C1 ∨ · · · ∨ Cm )[ z̄t̄ ])). We need to show
that
(III.2) s ∈ Ter(P1 , . . . , Pn , Γ1 , ∆1 ) ∩ Ter(Γ2 , ∆2 ) .
Given that we have quantified away all terms ti ∈ t̄ that would have been
problematic for (III.2), this condition obviously holds.
Finally, (III.1) and (III.2) jointly give (III).
From (I), (II) and (III) we conclude that the split-interpolant for the parti-
tion under consideration is ∀z̄((C1 ∨ · · · ∨ Cm )[ z̄t̄ ]) (recall that Π is P1 , . . . , Pn ).
In brief, we have shown that:

C C
Q∗1 , Π, Γ1 ; Γ2 =⇒
1
∆1 ; ∆2 ··· Q∗m , Π, Γ1 ; Γ2 ==⇒
m
∆1 ; ∆2
∀z̄((C1 ∨···∨Cm )[ z̄ ])
Π, Γ1 ; Γ2 ============t̄=⇒ ∆1 ; ∆2

Case 2. This case is similar to the previous one. To build the interpolant
we use conjunction and existential quantifier in place of disjunction and univer-
sal quantifier, respectively. It is easy to check that a split-interpolant for the
partition under consideration is ∃z̄((C1 ∧ · · · ∧ Cm )[ z̄t̄ ]). As above, we agree that
Π is P1 , . . . , Pn .

C C
Γ1 ; Q∗1 , Π, Γ2 =⇒
1
∆1 ; ∆2 ··· Γ1 ; Q∗m , Π, Γ2 ==⇒
m
∆1 ; ∆2
∃z̄((C1 ∧···∧Cm )[ z̄ ])
Γ1 ; Π, Γ2 ============t̄=⇒ ∆1 ; ∆2

Case 3. Consider the two sets Rel(Pij+1 , . . . , Pin ) and Rel(Pi1 , . . . , Pij ). The
proof is organized in three exhaustive sub-cases, according to whether one of
the two sets is empty or none:

11
3.1 Rel(Pij+1 , . . . , Pin ) = ∅
3.2 Rel(Pi1 , . . . , Pij ) = ∅
3.3 Rel(Pi1 , . . . , Pij ) 6= ∅ and Rel(Pij+1 , . . . , Pin ) 6= ∅

Sub-case 3.1. By the inductive hypothesis (IH) every partition of each


premise of R has a split-interpolant. In particular, for all k, 1 6 k 6 m,
there is a Ck such that

Ik ⊢ Q∗k , Pi1 , . . . , Pij , Pij+1 , . . . , Pin , Γ1 ⇒ ∆1 , Ck


IIk ⊢ Ck , Γ2 ⇒ ∆2
IIIk L(Ck ) ⊆ L(Q∗k , Pi1 , . . . , Pij , Pij+1 , . . . , Pin , Γ1 , ∆1 ) ∩ L(Γ2 , ∆2 )

Let Π1 and Π2 be Pi1 , . . . , Pij and Pij+1 , . . . , Pin , respectively. Consider the
following derivation D1 , where the topmost sequents are derivable by (Ik ):

Q∗1 , Π1 , Π2 , Γ1 ⇒ ∆1 , C1 Q∗m , Π1 , Π2 , Γ1 ⇒ ∆1 , Cm
W W
Q∗1 , Π1 , Π2 , Γ1 ⇒ ∆1 , C1 , . . . , Cm · · · Q∗m , Π1 , Π2 , Γ1 ⇒ ∆1 , C1 , . . . , Cm
R
Π 1 , Π 2 , Γ1 ⇒ ∆1 , C 1 , . . . , C m
R∨
Π 1 , Π 2 , Γ1 ⇒ ∆1 , C 1 ∨ · · · ∨ C m
L∧
Π1 , Pij+1 ∧ · · · ∧ Pin , Γ1 ⇒ ∆1 , C1 ∨ · · · ∨ Cm
R→
Π1 , Γ1 ⇒ ∆1 , Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm

Consider another derivation D2 , where the left-topmost sequents are initial


sequents and the right-topmost sequents are derivable by (IIk ):

C 1 , Γ2 ⇒ ∆ 2 ··· C m , Γ2 ⇒ ∆ 2
L∨
Π2 , Γ2 ⇒ ∆2 , Pij+1 ··· Π2 , Γ2 ⇒ ∆2 , Pin C 1 ∨ · · · ∨ C m , Γ2 ⇒ ∆ 2
R∧ W
Π2 , Γ2 ⇒ ∆2 , Pij+1 ∧ · · · ∧ Pin C 1 ∨ · · · ∨ C m , Π 2 , Γ2 ⇒ ∆2
L→
Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm , Π2 , Γ2 ⇒ ∆2

Let t1 , . . . , tl be all terms such that t1 , . . . , tl ∈ Ter(Pij+1 ∧ · · · ∧ Pin → C1 ∨


· · · ∨ Cm ) and t1 , . . . , tl ∈ / Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ Ter(Pij+1 , . . . , Pin , Γ2 , ∆2 ).
We use t to denote t1 , . . . , tl . We show that t1 , . . . , tl ∈ / Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ).
Let ti be an arbitrary term in t. Since ti ∈ Ter(Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨
Cm ), it follows ti ∈ Ter(Pij+1 , . . . , Pin ) or ti ∈ Ter(C1 ∨ · · · ∨ Cm ). In the first
case, we also have ti ∈ Ter(Pij+1 , . . . , Pin , Γ2 , ∆2 ). Therefore, we conclude that
/ Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ) since by hypothesis we have we know that
t1 , . . . , t l ∈
ti ∈ / Ter(Pi1 , . . . , Pij , Γ1 , ∆1 )∩Ter(Pij+1 , . . . , Pin , Γ2 , ∆2 ). In the second case, by
(IIIk ), ti ∈ Ter(Γ2 , ∆2 ), and consequently also ti ∈ Ter(Pij+1 , . . . , Pin , Γ2 , ∆2 ).
Hence ti ∈ / Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ) from the latter and the hypothesis ti ∈ /
Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ Ter(Pij+1 , . . . , Pin , Γ2 , ∆2 ).

12
Let now z̄ be all variables z1 , . . . , zl not occurring in the conclusions of D1
and D2 . As above, we apply the substitution [ z̄t̄ ] to the conclusion of D1 in
order to obtain by Lemma 5 that

⊢ Pi1 , . . . , Pij , Γ1 ⇒ ∆1 , (Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm )[ z̄t̄ ]


Once again, the substitution does not affect Pi1 , . . . , Pij , Γ1 , ∆1 since, as we
/ Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ). By l applications of R∀ to
have just seen, t1 , . . . , tl ∈
the latter we obtain

(I) ⊢ Pi1 , . . . , Pij , Γ1 ⇒ ∆1 , ∀z̄((Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm )[ z̄t̄ ])

Similarly, the same substitution [ z̄t̄ ] on the conclusion of D2 gives by Lemma


5:

⊢ (Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm )[ z̄t̄ ], (Pij+1 , . . . , Pin )[ z̄t̄ ], Γ2 [ z̄t̄ ] ⇒ ∆2 [ z̄t̄ ]

From this by l applications of L∀ we conclude

⊢ ∀z̄((Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm )[ z̄t̄ ]), (Pij+1 , . . . , Pin )[ z̄t̄ ], Γ2 [ z̄t̄ ] ⇒ ∆2 [ z̄t̄ ]

Finally, the inverse substitution [ t̄z̄ ] applied to the latter gives

(II) ⊢ ∀z̄((Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm )[ z̄t̄ ]), Pij+1 , . . . , Pin , Γ2 ⇒ ∆2

Let A be ∀z̄((Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm )[ z̄t̄ ]). We now check that


A is suitable as split-interpolant. By (I) and (II), A satisfies the derivability
conditions for being a split-interpolant for the partition under consideration.
We now show that it also satisfies

(III) L(A) ⊆ L(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ L(Pij+1 , . . . , Pin , Γ2 , ∆2 )

Suppose that P h is a predicate in Rel(Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm ). We


show that

(III.1) P h ∈ Rel(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ Rel(Pij+1 , . . . , Pin , Γ2 , ∆2 )

Notice first that from Rel(Pij+1 , . . . , Pin ) = ∅ it follows that P h ∈ Rel(C1 ∨ · · · ∨


Cm ). From this, (IIIk ) and Rel(Pij+1 , . . . , Pin ) = ∅ again, we conclude that both
(i) P h ∈ Rel(Q∗k , Pi1 , . . . , Pij , Γ1 , ∆1 ) and (ii) P h ∈ Rel(Γ2 , ∆2 ) hold. Since
R satisfies (⋆), we know that Rel(Q∗1 , . . . , Q∗m ) ⊆ Rel(P1 , . . . , Pn ) from which,
together with (i), it follows that (iii) P h ∈ Rel(Pi1 , . . . , Pij , Γ1 , ∆1 ). Now, (ii)
and (iii) imply (III.1).

13
Suppose s is a term in Ter(A). We need to show that from this it follows
that

(III.2) s ∈ Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ Ter(Pij+1 , . . . , Pin , Γ2 , ∆2 )

Given that we have quantified away all terms ti ∈ t that would have been
problematic for (III.2), this condition obviously holds.
We conclude that:

C C
Q∗1 , Π1 , Π2 , Γ1 ; Γ2 =⇒
1
∆1 ; ∆2 ··· Q∗m , Π1 , Π2 , Γ1 ; Γ2 ==⇒
m
∆1 ; ∆2
∀z̄((Pij+1 ∧···∧Pin →C1 ∨···∨Cm )[ z̄t̄ ])
Π1 , Γ1 ; Π2 , Γ2 =======================⇒ ∆1 ; ∆2

Recall that here Π1 and Π2 stand for Pi1 , . . . , Pij and Pij+1 , . . . , Pin , respec-
tively.

Sub-case 3.2. The proof differs substantially with respect to sub-case 3.1.
only as far as the derivability conditions are concerned. Thus, we give a detailed
analysis of these and leave to the reader the task to check that also the language
condition is satisfied. By the inductive hypothesis (IH) every partition of each
premise of R has a split-interpolant. In particular, for all k, 1 6 k 6 m, there
is a Ck such that:

Ik ⊢ Γ1 ⇒ ∆1 , Ck
IIk ⊢ Ck , Q∗k , Pi1 , . . . , Pij , Pij+1 , . . . , Pin , Γ2 ⇒ ∆2
IIIk L(Ck ) ⊆ L(Γ1 , ∆1 ) ∩ L(Q∗k , Pi1 , . . . , Pij , Pij+1 , . . . , Pin , Γ2 , ∆2 )

As in the previous sub-case, let Π1 and Π2 be Pi1 , . . . , Pij and Pij+1 , . . . , Pin ,
respectively. Let D1 be the following derivation, where the topmost sequents
are all derivable by (IIk ):

C1 , Q∗1 , Π1 , Π2 , Γ2 ⇒ ∆2 Cm , Q∗m , Π1 , Π2 , Γ2 ⇒ ∆2
W W
C1 , . . . , Cm , Q∗1 , Π1 , Π2 , Γ2 ⇒ ∆2 · · · C1 , . . . , Cm , Q∗m , Π1 , Π2 , Γ2 ⇒ ∆2
R
C 1 , . . . , C m , Π 1 , Π 2 , Γ2 ⇒ ∆ 2
L∧
C1 ∧ · · · ∧ Cm ∧ Pi1 ∧ · · · ∧ Pij , Π2 , Γ2 ⇒ ∆2

Consider now another derivation D2 where the left topmost sequents are
derivable by (Ik ) and the right ones are initial sequents:

Γ1 ⇒ ∆1 , C 1 ··· Γ1 ⇒ ∆1 , C m
R∧
Γ1 ⇒ ∆1 , C 1 ∧ · · · ∧ C m Π1 , Γ1 ⇒ ∆1 , Pi1 ... Π1 , Γ1 ⇒ ∆1 , Pij
W R∧
Π 1 , Γ1 ⇒ ∆1 , C 1 ∧ · · · ∧ C m Π1 , Γ1 ⇒ ∆1 , Pi1 ∧ · · · ∧ Pij
R∧
Π1 , Γ1 ⇒ ∆1 , C1 ∧ · · · ∧ Cm ∧ Pi1 ∧ · · · ∧ Pij

14
Let t̄ be are all terms t1 , . . . , tl such that t1 , . . . , tl ∈ Ter(C1 ∧ · · · ∧ Cm ∧ Pi1 ∧
/ Ter(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ Ter(Pij+1 , . . . , Pin , Γ2 , ∆2 ).
· · · ∧ Pij ) and t1 , . . . , tl ∈
Moreover let z̄ be variables z1 , . . . , zl not occurring free in the conclusions of D1
and of D2 . We apply the same reasoning of the previous sub-case to obtain

(I) ⊢ Pi1 , . . . , Pij , Γ1 ⇒ ∆1 , ∃z̄((C1 ∧ · · · ∧ Cm ∧ Pi1 ∧ · · · ∧ Pij )[ z̄t̄ ])

As above we also obtain

(II) ⊢ ∃z̄((C1 ∧ · · · ∧ Cm ∧ Pi1 ∧ · · · ∧ Pij )[ z̄t̄ ]), Pij+1 , . . . , Pin , Γ2 ⇒ ∆2

Let A be ∃z̄((C1 ∧ · · · ∧ Cm ∧ Pi1 ∧ · · · ∧ Pij )[ z̄t̄ ]). Then we have:

(III) L(A) ⊆ L(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ L(Pij+1 , . . . , Pin , Γ2 , ∆2 )

We conclude that ∃z̄((C1 ∧ · · · ∧ Cm ∧ Pi1 ∧ · · · ∧ Pij )[ z̄t̄ ]) is a split-interpolant


for the partition under consideration.

C C
Γ1 ; Q∗1 , Π1 , Π2 , Γ2 =⇒
1
∆1 ; ∆2 ... Γ1 ; Q∗m , Π1 , Π2 , Γ2 ==⇒
m
∆1 ; ∆2
∃z̄((C1 ∧···∧Cm ∧Pi1 ∧···∧Pij )[ z̄t̄ ])
Π1 , Γ1 ; Π2 , Γ2 =====================⇒ ∆1 ; ∆2

Sub-case 3.3. The proof differs substantially with respect to sub-case 3.1.
only as far as the language condition is concerned. Thus, we give a detailed
analysis of this. For the derivability condition, on the other hand, the procedure
can be reduced indifferently to either sub-case 3.1 or sub-case 3.2. For instance,
as in sub-case 3.1, the formula ∀z̄((Pij+1 ∧· · · ∧Pin → C1 ∨· · ·∨Cm )[ z̄t̄ ]) satisfies
the derivability conditions. Let A be such a formula. We need to show that

(III) L(A) ⊆ L(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ L(Pij+1 , . . . , Pin , Γ2 , ∆2 )

The proof for the term part is the same as in sub-case 3.1. We only need to
show the relational part of (III). Let P h be a predicate in Rel(Pij+1 ∧ · · · ∧ Pin →
C1 ∨ · · · ∨ Cm ). We show that

(III.1) P h ∈ Rel(Pi1 , . . . , Pij , Γ1 , ∆1 ) ∩ Rel(Pij+1 , . . . , Pin , Γ2 , ∆2 )

First notice that from the fact that Rel(Pi1 , . . . , Pij ) and Rel(Pij+1 , . . . , Pin )
are both not empty and from the condition (⋆) we have (i) P h ∈ Rel(Pi1 , . . . , Pij )
if and only if P h ∈ Rel(Pij+1 , . . . , Pin ).
Now, if P h ∈ Rel(Pij+1 ∧ · · · ∧ Pin → C1 ∨ · · · ∨ Cm ) then either (ii) P h ∈
Rel(Pij+1 , . . . , Pin ) or (iii) P h ∈ Rel(C1 , . . . , Cm ). On the one hand, if (ii) is

15
the case then clearly also P h ∈ Rel(Pij+1 , . . . , Pin , Γ2 , ∆2 ) and, by (i), P h ∈
Rel(Pi1 , . . . , Pij , Γ1 , ∆1 ). Therefore (III.1) holds. On the other hand, if (iii) is
the case, by (IIIk ) and (⋆) we conclude (iv) P h ∈ Rel(Pi1 , . . . , Pin , Γ1 , ∆1 ) and
(v) P h ∈ Rel(Γ2 , ∆2 ). From (v) we have (vi) P h ∈ Rel(Pij+1 , . . . , Pin , Γ2 , ∆2 ).
Moreover, for (iv) we have two cases: either P h ∈ Rel(Pi1 , . . . , Pij , Γ1 , ∆1 ) or
P h ∈ Rel(Pij+1 , . . . , Pin ). If the former is the case then (III.1) follows from it
and (vi). If the latter is the case then by (i) we also know P h ∈ Rel(Pi1 , . . . , Pij ),
hence (vii) P h ∈ Rel(Pi1 , . . . , Pij , Γ1 , ∆1 ) as well. Therefore, (III.1) follows from
(vii) and (vi).
We conclude:

C C
Q∗1 , Π1 , Π2 , Γ1 ; Γ2 =⇒
1
∆1 ; ∆2 ··· Q∗m , Π1 , Π2 , Γ1 ; Γ2 ==⇒
m
∆1 ; ∆2
∀z̄((Pij+1 ∧···∧Pin →C1 ∨···∨Cm )[ z̄t̄ ])
Π1 , Γ1 ; Π2 , Γ2 =======================⇒ ∆1 ; ∆2

Case 4. This can be seen as a limit case of the previous ones. Notice that
since there are no P1 , . . . , Pn , the set Rel(Q∗1 , . . . Q∗m ) is empty. Then, by a
simplification of the proof for case 3.1, we conclude:

C C
Q∗1 , Γ1 ; Γ2 =⇒
1
∆1 ; ∆2 ··· Q∗m , Γ1 ; Γ2 ==⇒
m
∆1 ; ∆2
∀z̄((C1 ∨···∨Cm )[ z̄ ])
Γ1 ; Γ2 ============t̄=⇒ ∆1 ; ∆2

First-order logic with identity. Let G= be G + {Ref , Repl }:

s = s, Γ ⇒ ∆ P [tx ], s = t, P [sx ], Γ ⇒ ∆
Ref Repl
Γ⇒∆ s = t, P [sx ], Γ ⇒ ∆
In [6] it is shown that G= is deductively equivalent to any given axiomatiza-
tion of first-order logic with identity and is cut-free. As we have noted earlier,
the rules Ref and Repl satisfy the condition (⋆) and therefore are examples of
singular geometric rules. It follows from our Lemma 6 that Maehara’s lemma
holds in G= :
Corollary 7. In G= every derivable sequent has a split-interpolant.
We determine the split-interpolants as applications of the procedures given
in the proof of Lemma 6. Rule Ref is treated as an instance of case 4:

C
s = s, Γ1 ; Γ2 =
⇒ ∆1 ; ∆2
∀z(C[z ])
s
Γ1 ; Γ2 =====⇒ ∆1 ; ∆2
For Repl there are four possible partitions of the conclusion:

16
1. s = t, P [sx ], Γ1 ; Γ2 ⇒ ∆1 ; ∆2
2. Γ1 ; s = t, P [sx ], Γ2 ⇒ ∆1 ; ∆2
3. P [sx ], Γ1 ; s = t, Γ2 ⇒ ∆1 ; ∆2
4. s = t, Γ1 ; P [sx ], Γ2 ⇒ ∆1 ; ∆2

We must then consider four sub-cases. As in case 1 of Lemma 6, the inter-


polant for the first partition is as follows:

C
P [tx ], s = t, P [sx ], Γ1 ; Γ2 =
⇒ ∆1 ; ∆2
C
s = t, P [sx ], Γ1 ; Γ2 =
⇒ ∆1 ; ∆2

The interpolant for the second partition reasoning as in case 2 of Lemma 6:

C
Γ1 ; P [tx ], s = t, P [sx ], Γ2 =
⇒ ∆1 ; ∆2
C
Γ1 ; s = t, P [sx ], Γ2 =
⇒ ∆1 ; ∆2

The interpolant for the third partition is found as in subcase 3.1 of Lemma
6, depending on whether t ∈ Ter(P [sx ], Γ1 , ∆1 ) or not:

C C
P [tx ], s = t, P [sx ], Γ1 ; Γ2 =
⇒ ∆1 ; ∆2 P [tx ], s = t, P [sx ], Γ1 ; Γ2 =
⇒ ∆1 ; ∆2
s=t→C ∀x(s=x→C)
P [sx ], Γ1 ; s = t, Γ2 ====⇒ ∆1 ; ∆2 P [sx ], Γ1 ; s = t, Γ2 =======⇒ ∆1 ; ∆2

Lastly, the interpolant for the fourth partition is found as in subcase 3.2 of
Lemma 6, depending on whether t ∈ Ter(P [sx ], Γ2 , ∆2 ) or not:

C C
Γ1 ; P [tx ], s = t, P [sx ], Γ2 =
⇒ ∆1 ; ∆2 Γ1 ; P [tx ], s = t, P [sx ], Γ2 =
⇒ ∆1 ; ∆2
s=t∧C ∃x(s=x∧C)
s = t, Γ1 ; P [sx ], Γ2 ====⇒ ∆1 ; ∆2 s = t, Γ1 ; P [sx ], Γ2 =======⇒ ∆1 ; ∆2

Notice that, when P is an identity, we can freely choose between the im-
plicative and the conjunctive interpolant.

Strict partial orders. The theory of strict partial orders consists of first-
order logic with identity plus the irreflexivity and transitivity of <. Here the
theory of strict partial orders is formulated by adding on top of G= the following
rules:
⊥, s < s, Γ =
⇒∆ s < u, s < t, t < u, Γ =
⇒∆
Irref Trans
s < s, Γ =
⇒∆ s < t, t < u, Γ =
⇒∆

17
Notice that irreflexivity is traditionally formulated as a 0-ary inference rule
of the form [7]:
Irref ′
s < s, Γ =
⇒∆
Such a formulation, however, is problematic for interpolation. Consider one
of the two possible partitions of the conclusion, namely s < s, Γ1 ; Γ2 = ⇒
∆1 ; ∆2 (the case of the other partition is similar). A natural candidate for
being a split-interpolant of the former is s < s, since s < s, Γ1 = ⇒ ∆1 , s < s
and s < s, Γ2 = ⇒ ∆2 are both derivable. However, neither s nor < have to be
in L(Γ2 , ∆2 ), though they both are in L(s < s). Thus, the language condition
L(s < s) ⊆ L(s < s, Γ1 , ∆1 )∩L(Γ2 , ∆2 ) is not, in general, satisfied. To overcome
this obstacle we treat the irreflexivity rule as a one-premise rule and replace the
standard of irreflexivity with Irref.
Also Irref and Trans are singular geometric axioms. Let SPO be G= +
{Ref , Trans}. From Lemma 6 it follows that
Corollary 8. In SPO every derivable sequent has a split-interpolant.
As above, we show how to find the interpolants for Irref and Trans. We start
with Irref. There are two possible partition of the conclusion of its conclusion,
namely

• s < s, Γ1 ; Γ2 ⇒ ∆1 ; ∆2
• Γ1 ; s < s, Γ2 ⇒ ∆1 ; ∆2

As in case 1 and case 2 of Lemma 6, we find the split-interpolant for each


partition as follows:

C C
⊥, s < s, Γ1 ; Γ2 =
⇒ ∆1 ; ∆2 Γ1 ; ⊥, s < s, Γ2 =
⇒ ∆1 ; ∆2
C C
s < s, Γ1 ; Γ2 =
⇒ ∆1 ; ∆2 Γ1 ; s < s, Γ2 =
⇒ ∆1 ; ∆2

For the case of Trans there are four possible partitions of the conclusion:

1. s < t, t < u, Γ1 ; Γ2 ⇒ ∆1 ; ∆2
2. Γ1 ; s < t, t < u, Γ2 ⇒ ∆1 ; ∆2
3. s < t, Γ1 ; t < u, Γ2 ⇒ ∆1 ; ∆2
4. t < u, Γ1 ; s < t, Γ2 ⇒ ∆1 ; ∆2

For each we find the split-interpolant as follows by reasoning as in case 1,


case 2 and case 3 of Lemma 6 respectively. More specifically, as in case 1 of
Lemma 6 a split-interpolant for the first partition is:

18
C
s < u, s < t, t < u, Γ1 ; Γ2 =
⇒ ∆1 ; ∆2
C
s < t, t < u, Γ1 ; Γ2 =
⇒ ∆1 ; ∆2
The split-interpolant for the second partition is found as in case 2 of Lemma
6 as follows:

C
Γ1 ; s < u, s < t, t < u, Γ2 =
⇒ ∆1 ; ∆2
C
Γ1 ; s < t, t < u, Γ2 =
⇒ ∆1 ; ∆2
Finally, by applying subcase 3.3 of Lemma 6 we find the split-interpolants
for the remaining partitions as follows, assuming, respectively, u 6∈ Ter(s <
t, Γ1 , ∆1 ) and s 6∈ Ter(t < u, Γ1 , ∆1 ) :

C C
s < u, s < t, t < u, Γ1 ; Γ2 =
⇒ ∆1 ; ∆2 s < u, s < t, t < u, Γ1 ; Γ2 =
⇒ ∆1 ; ∆2
∀z(t<z→C)[z
u] ∀z(z<t→C)[z ]
s
s < t, Γ1 ; t < u, Γ2 ==========⇒ ∆1 ; ∆2 t < u, Γ1 ; s < t, Γ2 ==========⇒ ∆1 ; ∆2

Conclusions. In this paper we have shown (Lemma 6) how to extend Mae-


hara’s lemma to sequent calculi with singular geometric rules and provided a
number of interesting instances of singular geometric rules that are important in
mathematics, especially in order theories. As case studies, we have shown that
Lemma 6 covers directly first-order logic with identity and its extension with
the theory of strict partial orders. For future work, it might be interesting to in-
vestigate in what directions our results can be further developed. For example,
it remains to be shown an analogous of Lemma 6 for extensions of intuition-
istic first-order logic. These extensions include, amongst others, the theory of
apartness for which Negri in [4] has already established a cut-elimination pro-
cedure. Interestingly, the axioms of apartness (irreflexivity, symmetry and co-
transitivity) fall within the class of singular geometric axioms. Thus, it should
be an easy task to prove interpolation in the same manner we have presented
here.

References
[1] W. Craig. Three uses of the Herbrand-Gentzen theorem in relating model
theory and proof theory. The Journal of Symbolic Logic, 22(3):269–285,
1957.
[2] G. Gentzen. Investigation into logical deductions. In M. E. Sazbo, editor,
The collected papers of Gerhard Gentzen, chapter 3, pages 68–131. North-
Holland, 1969.

19
[3] S. Maehara. On the interpolation theorem of Craig. Suugaku, 12:235–237,
1960. (in Japanese).
[4] S. Negri. Sequent calculus proof theory of intuitionistic apartness and order
relations. Archive for Mathematical Logic, 38(8):521–547, 1999.
[5] S. Negri. Contraction-free sequent calculi for geometric theories with an
application to Barr’s theorem. Archive for Mathematical Logic, 42(4):389–
401, 2003.
[6] S. Negri and J. von Plato. Cut elimination in the presence of axioms. The
Bullettin of Symbolic Logic, 4(4):418–435, 1998.
[7] S. Negri and J. von Plato. Structural Proof Theory. Cambridge University
Press, 2001.
[8] S. Negri and J. von Plato. Proof Analysis: A Contribution to Hilbert’s Last
Problem. Cambridge University Press, 2011.
[9] G. Takeuti. Proof Theory, volume 81 of Studies in Logic and the Founda-
tions of Mathematics. North-Holland, 2nd edition, 1987.
[10] A.S. Troelstra and H. Schwichtenberg. Basic Proof Theory. Cambridge
University Press, 2nd edition, 2000.

20