Professional Documents
Culture Documents
Jigpal/jzt 038
Jigpal/jzt 038
classical logic
DAVID MAKINSON∗ , Department of Philosophy, Logic and Scientific Method,
London School of Economics, Houghton St, London WC2A 2AE, UK
Abstract
1 Introduction
Sandqvist [3, 4] devised a semantics for first-order logic, in which connectives are evaluated on sets
of ‘basic rules’ rather than on assignments of truth-values, and showed that despite its inferential
character it yields full classical logic. It thus provides the first known inferential semantics for
classical logic, upsetting the widely held presumption that such a style of semantics must lead to a
weaker, typically intuitionistic, logic.
As recognized by Sandqvist and emphasized in a proof-theoretic discussion by de Campos Sanz,
Piecha and Schroeder-Heister [1], the system is not very robust in its choice of primitive connectives.
If, like Sandqvist, we take → and ⊥ as primitive for the propositional part, with ¬, ∨, ∧, treated as
abbreviations in a familiar manner, then the semantics yields all of classical logic; but if we treat ∨
as primitive with a classical or Kripke-style evaluation rule, then it becomes sub-classical.
Despite this sensitivity, the construction remains interesting and, moreover, rather puzzling. Why
does it yield classical logic despite its inferentialist character and its similarity to Kripke seman-
tics, which yields only intuitionistic logic? In this note we seek a better understanding of the
situation.
Sandqvist’s original proof is not very helpful in this regard. It proceeds by checking the validity
of the axioms of minimal logic under his semantics, plus the axiom of double negation elimination,
drawing on proof-theoretic results from Prawitz [2] to verify the latter. The desired result then
follows, since those axioms are known to be classically complete.
We give a direct proof with a ‘maxiconsistent basis’ argument that is analogous to, but not
quite the same as, the familiar Lindenbaum/Henkin maxiconsistent set-of-formulas construction.
This gives a clearer picture of the inner mechanisms, and allows easy identification of the places
where proof would snag under various small changes in the evaluation procedure. It also per-
mits one to formulate some variants and extensions of the Sandqvist semantics without loss of
classicality.
The proof is given initially in the propositional context as that allows the clearest view of what is
going on, without distractions. It is then extended routinely to the first-order context.
∗
E-mail: david.makinson@gmail.com
Vol. 22 No. 1, © The Author 2013. Published by Oxford University Press. All rights reserved.
For Permissions, please email: journals.permissions@oup.com
doi:10.1093/jigpal/jzt038 Advance Access published 7 November 2013
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 147 147–154
148 Inferential semantics for classical logic
(a) Thus for any basis B, the function B(·) is a closure operation (i.e. satisfying inclusion, monotony,
idempotence) taking each set K of sentence letters to a set B(K) of sentence letters.
(b) Clearly, for every basis B the operation B(·) is compact in two distinct senses. It is sentence-
compact in the sense that whenever p ∈ B(K) then p ∈ B(Kf ) for some finite set Kf ⊆ K of
sentence letters. It is also rule-compact, in the sense that whenever p ∈ B(K) then p ∈ Bf (K) for
some finite basis Bf ⊆ B.
(c) Sentence-compactness will not be important for us, for we will never have to consider values of
B(K) for K = ∅. But rule-compactness will be essential for our proof of completeness, which
would snag (at Lemma 3.3) if one allowed it to fail (e.g. by allowing rules with countably many
premises).
(a) Some readers may not be happy with the term ‘semantics’ for this construction, feeling that
it—and perhaps any inferential semantics—is too syntactic in nature to deserve that title. They
may prefer to use the neutral term ‘evaluation system’; none of our formal results depend on
the choice of terminology.
(b) This way of evaluating the falsum has a long history in inferentialist circles, but is quite different
from the one customary in classical logic (or even in Kripke models for intuitionistic logic)
where ⊥ is counted false under every classical valuation (resp. is satisfied in no state in any
Kripke model). In the present context, the analogue of that would be: B ⊥ for all bases B,
and for future reference we will call that the classical-style evaluation of ⊥. We will see that it
makes a difference which evaluation rule is used.
(c) If one wishes to evaluate sequents as well, one may follow Sandqvist in putting B (A ⇒ ψ),
where A is a set of formulae and ψ is a formula, iff for all C ⊇ B, if C ϕ for all ϕ ∈ A then
C ψ. However, we will not need to consider sequents for our results.
A formula ϕ is called valid in the Sandqvist semantics iff B ϕ for every basis B. Equivalently (once
one proves Lemma 3.1a below), iff ∅ ϕ.
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 148 147–154
Inferential semantics for classical logic 149
(a) Given Sandqvist’s way of evaluating the falsum, it is important that the language has infinitely
many sentence letters. If there were only n letters p1 ,...,pn then the semantics would val-
idate the non-tautology p1 → (... → (pn → ⊥)...), thus destroying its soundness wrt classical
logic.
(b) At first glance, this feature might seem to prevent the semantics being used in any decision
procedure. But one could get around it by requiring, without loss of classical soundness, that the
language is ‘potentially infinite’ in the sense that it always contains at least one more sentence
letter than the formula whose status one wishes to decide (cf. the proof of soundness in section 6
below).
(a) From his inferentialist perspective, Sandqvist sees this as a soundness theorem, stating the
soundness of the classical semantics wrt his inferential one. With classical logic as our ‘home
base’, we see it as a completeness theorem, reporting the completeness of Sandqvist’s seman-
tics wrt the classical one. This is a terminological contrast engendered by radically different
philosophical vantage points, but in no way affects the formal results.
(b) As noted by Sandqvist, the completeness theorem fails if ∨ is allowed as primitive with the
classical valuation B ϕ ∨ψ iff B ϕ or B ψ, or the tighter valuation B ϕ ∨ψ iff for all
C ⊇ B either C ϕ or C ψ, or the even more demanding one B ϕ ∨ψ iff either for all C ⊇ B,
C ϕ or for all C ⊇ B, C ψ. Specifically, under each of those valuation procedures, excluded
middle p∨(p → ⊥) would fail at the empty basis, i.e. ∅ p∨(p → ⊥).
(c) However, the theorem continues to hold if ∨ is taken as primitive but given an evaluation
procedure that mimics the definition of ϕ ∨ψ as (ϕ → ⊥) → ψ, such as the following: B ϕ ∨ψ
iff for all C ⊇ B, either (i) C ψ or (ii) D ϕ while p ∈
/ D(∅) for some D ⊇ C and sentence
letter p. Negation may likewise be taken as primitive, with an evaluation that similarly mimics
the definition of ¬ϕ as ϕ → ⊥: B ¬ϕ iff for all C ⊇ B, if C ϕ then C p for every sentence
letter p. However, it seems unlikely that these rather complex recursion clauses can be further
simplified, or rendered more intuitive.
(d) Finally, we note that the theorem fails if ⊥ is given the classical-style valuation B ⊥ for all
bases B while retaining Sandqvist’s evaluation of →. Specifically, double negation elimination
then fails at the empty basis, i.e. ∅ ((p → ⊥) → ⊥) → p. Verification: for every basis B there
is a basis C ⊇ B with C p; we need only take C to be the set of all basic rules. Thus, under
the classical-style valuation of ⊥ combined with Sandqvist’s evaluation of →, there is no
basis B with B p → ⊥. Hence, vacuously, ∅ (p → ⊥) → ⊥. Since also ∅ p we have ∅
((p → ⊥) → ⊥) → p.
3 Maxiconsistent bases
We do not hesitate to use full classical logic in our own reasoning, notably excluded middle for the
recursive definition of Bi+1 in the proof of Lemma 3.3. The first two lemmas are routine; the third
may be thought of as containing the ‘heart’ of the argument; the fourth reveals how the choice of
primitive connectives is critical.
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 149 147–154
150 Inferential semantics for classical logic
LEMMA 3.1
For bases B,C and formulae ϕ,ψ: (a) If B ϕ and B ⊆ C then C ϕ; (b) If B ϕ and B ϕ → ψ then
B ψ; (c) If B ⊥ then B ϕ for all ϕ.
PROOF. Part (a) is the first lemma of Sandqvist [4] and is proven by an easy induction on ϕ. As a
corollary, it tells us that the validity of a formula ϕ in the semantics may equivalently be defined
as ∅ ϕ. Part (b) is immediate using the evaluation rule for →. For part (c), suppose B ⊥. Then
by Sandqvist’s valuation of ⊥ we have B p for every sentence letter p, and an easy induction on ϕ
gives us B ϕ.
Remarks on Lemma 3.1
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 150 147–154
Inferential semantics for classical logic 151
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 151 147–154
152 Inferential semantics for classical logic
PROOF. Suppose ϕ is not a tautology; we want to show that B ϕ for some basis B. Take a classical
valuation v with v(ϕ) = 0. Since ϕ contains only finitely many sentence letters and the language has
a countable supply of them (or, failing that, at least one more letter than those in ϕ, cf. the remarks
in section 2 on the supply of letters) we may assume wlog that v(q) = 0 for some letter q, which
we now fix. Put A = {∅ ⇒ p : v(p) = 1}∪{p ⇒ q : v(p) = 0}. Clearly q ∈ / {p : v(p) = 1} = A(∅), so A q,
so by Lemma 3.3 there is a maxiconsistent B ⊇ A with B q. From the latter and the definition of A
we have B(∅) = {p : v(p) = 1} = A(∅), so for every letter p, B p iff v(p) = 1. By Lemma 3.5, since
B is maxiconsistent it is classically well-behaved wrt →, ⊥, and so we may conclude that B ϕ iff
v(ϕ) = 1 and thus B ϕ.
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 152 147–154
Inferential semantics for classical logic 153
LEMMA 6.1
For any basis B in the extended sense, and any formula ϕ, B ϕ iff B∗ ϕ.
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 153 147–154
154 Inferential semantics for classical logic
PROOF. If B ϕ then B* ϕ by Lemma 6.1, so by the completeness theorem in the original context
ϕ is not a tautology.
7 Philosophical perspectives
How far does Sandqvist’s semantics meet the conditions that inferentialists in the tradition of Prawitz
and Dummett would require for it to serve as a ‘justification’ of its outcome logic? This is a slippery
question to answer, for those requirements are not entirely clear. Indeed, it could be said that
Sandqvist’s completeness theorem reveals a situation that was quite unforeseen in that tradition, and
should force inferentialists to be more explicit about their acceptability criteria.
Acknowledgements
Thanks to Lloyd Humberstone, Tor Sandqvist, Peter Schroeder-Heister and two anonymous referees
for valuable comments and suggestions.
References
[1] W. de Campos Sanz, T. Piecha and P. Schroeder-Heister. Constructive semantics, admissibility
of rules, and the validity of Peirce’s law. Logic Journal of the IGPL (to appear).
[2] D. Prawitz. Natural Deduction: A Proof-Theoretical Study. Almqvist & Wiksell, 1965,
reprinted 2006 by Dover Publications, Mineola, NY.
[3] T. Sandqvist. An Inferentialist Interpretation of Classical Logic. PhD Thesis, Uppsala Uppsala
Prints and Preprints in Philosophy, 2005.
[4] T. Sandqvist. Classical logic without bivalence. Analysis, 69, 211–218, 2009.
[17:18 21/12/2013 jzt038.tex] Paper Size: a4 paper Job: JIGPAL Page: 154 147–154