Brit. J. Phil. Sci.
64 (2013), 61–68
Criteria of Identity: Strong
and Wrong
Hannes Leitgeb
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
ABSTRACT
We show that finitely axiomatized first-order theories that involve some criterion of
identity for entities of a category C can be reformulated as conjunctions of a non-triviality
statement and a criterion of identity for entities of category C again. From this, we draw
two conclusions: First, criteria of identity can be very strong deductively. Second,
although the criteria of identity that are constructed in the proof of the theorem are
not good ones intuitively, it is difficult to say what exactly is wrong with them once the
modern metaphysical view of identity criteria is presupposed.
1 Criteria of Identity: Strong . . .
2 . . . and Wrong
Let us call a sentence a criterion of identity for entities of a category C if and
only if it is of the following form1:
For all x, y of category C : x ¼ y iff ½x, y:
Criteria of identity still constitute an important topic in the philosophy and
the foundations of mathematics and science. In set theory, the Extensionality
criterion is fundamental for our understanding of sets; in Neo-Fregean foun-
dations of mathematics, second-order logic is augmented by abstraction prin-
ciples that are very close in form to criteria of identity in the sense above; and
in the philosophy of quantum physics, there is an ongoing discussion on which
particles are subject to which criterion of identity and whether for particles of
a certain kind there might be no criterion of identity at all.
Intuitively, a criterion of identity can be good or bad qua identity criterion
for entities of category C depending on the properties of its right-hand side,
1
These are so-called one-level criteria of identity in the terminology of Williamson ([1990]). We
will not deal with abstraction principles, or what Williamson calls two-level criteria of identity, in
this article. Note that we do not regard ’[x,y] as the criterion of identity, but rather the univer-
sally quantified sentence itself.
ß The Author 2012. Published by Oxford University Press on behalf of British Society for the Philosophy of Science. All rights reserved.
doi:10.1093/bjps/axr058 For Permissions, please email: [email protected]
Advance Access published on April 25, 2012
62 Hannes Leitgeb
but it is less clear how one ought to understand ‘good’ here. One slightly
old-fashioned view on good criteria of identity is that they would have to be
analytic(ally true) and hence deductively weak in the sense of not entailing any
synthetic claims. In fact, criteria of identity are sometimes assimilated to
(conditional) definitions and, just as with proper definitions, they would
have to be deductively weak in the sense of being non-creative, that is, not
entailing any ‘¼’-free statements about the objects of the category in question
that would not have been derivable anyway even before the introduction of
the criterion.
When philosophy took a more metaphysical turn again, however, good
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
criteria of identity rather came to be viewed as true metaphysical principles,
maybe with some other virtuous properties but none of them being analyticity
(cf. Williamson [1990]; Lowe [1998]). Consequently, there was no presumption
anymore that a good criterion of identity had to be logically weak in any sense,
and this view seems to be the dominant one these days. For the same reason,
the questions arise: How strong exactly can a criterion of identity be deduct-
ively, and how strong can a good criterion of identity be deductively?
1 Criteria of Identity: Strong . . .
First of all, it is easy to see that one criterion of identity for C may be deduct-
ively stronger than another criterion of identity for C without either of them
being false, let alone with one criterion being inconsistent with the other. For
instance, Extensionality for sets,
For all sets x, y : x ¼ y iff for all z ½z 2 x iff z 2 y
logically implies (and is in fact strictly stronger than) the following criterion of
identity for sets which is also a true set-theoretic principle, which is also de-
rivable in standard ZFC set theory, and which was investigated, e.g. by Hilbert
and Bernays ([1934]):
For all sets x, y : x ¼ y iff for all z ½ðz 2 x iff z 2 yÞ & ðx 2 z iff y 2 zÞ:
Indeed, as far as the relationship of logical entailment between the two
criteria of identity is concerned, this example is an instance of a more general
pattern: Let ’1[x,y], ’2[x,y] be first-order formulas that express reflexive rela-
tions of entities of category C by logic alone, that is, where ‘For all x of
category C: ’1[x,x]’ and ‘For all x of category C: ’2[x,x]’ are logical truths.2
Let the equality sign ‘¼’ be understood to be a logical sign in the standard
2
If a criterion of identity is true, then of course its right-hand side expresses a reflexive (and
symmetric and transitive) relation simply because ‘¼’ does, but it is not necessarily the case that
the right-hand side of a criterion of identity is reflexive already for logical reasons. However, we
will need this in the following as an additional premise.
Criteria of Identity 63
sense, such that it expresses the identity relation in every first-order model and
hence the usual first-order logical laws and rules of identity apply. Then it is
straightforward to show:
If
For all x, y of category C: if ’2[x,y] then ’1[x,y]
is a logical truth, then
For all x, y of category C: x ¼ y iff ’1[x,y]
logically implies, i.e. is deductively at least as strong as,
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
For all x, y of category C: x ¼ y iff ’2[x,y].
(The reflexivity property of ’2[x,y] for C-entities by logic alone is needed
for the derivation of the left-to-right direction of the latter criterion.) In
short: given the assumptions mentioned before, the logically weaker the
right-hand side of a criterion of identity, the logically stronger the criterion
is itself.3
Now assume that we are given a finitely axiomatized theory F of which
some criterion of identity CI for entities of category C is an axiom. Is there
anything like the strongest criterion of identity for entities of category C that is
still entailed by F and which one gets by ‘maximally’ weakening the right-hand
side of CI? The following Proposition supplies a positive answer under fairly
general conditions:
Proposition: Let F be a consistent, finitely-axiomatized, first-order
theory, such that F logically implies that there are at least two entities
of category C, and where F consists of the following (non-logical)
axioms:
. a conjunction 1 & . . . & n,
. a criterion of identity
ðCIÞ For all x, y of category C: x ¼ y iff ’½x, y,
where ’[x,y] expresses a reflexive relation for entities of category C by
logic alone.
Then there exists a consistent finitely axiomatized theory F* which has
as its (non-logical) axioms
. a ‘non-triviality’ statement
(NT) There are at least two entities of category C,
3
Note that the converse does not hold. For instance, ‘For all x, y of category C: x ¼ y iff [x ¼ y or
there are at least two objects of category C]’ logically implies ‘For all x, y of category C: x ¼ y iff
[A or non-A]’, since both are satisfied precisely in such models in which there is at least one
object to which the predicate C applies. However ‘A or non-A’ does not logically imply the open
formula ‘x ¼ y or there are at least two objects of category C’.
64 Hannes Leitgeb
. a criterion of identity
ðCI Þ For all x, y of category C: x ¼ y iff ’ ½x, y,
where ’*[x,y] expresses a reflexive relation for entities of category C by
logic alone,
such that:
. F logically implies F*, F* logically implies F,
. For all x, y of category C: if ’[x,y] then ’*[x,y]
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
is a logical truth, and
. CI* is the logically strongest criterion of identity for entities of category C
(unique up to logical equivalence) among those which are logically
implied by F and the right-hand sides of which express reflexive relations
for C-entities by logic alone.
In words: Given a theory F as described, it is possible to reformulate it in
terms of a logically equivalent theory F* which consists just of two
(non-logical) axioms: a non-triviality thesis NT, which requires the existence
of at least two entities of category C, and an identity criterion CI* for the
entities in that category. NT expresses that, even by the lights of the theory
itself, there are some non-trivial questions concerning the identity or differ-
ence of entities of category C that may be settled by an identity criterion. CI* is
such a criterion, and it arises from the original criterion of identity CI in the
way considered before, that is, by weakening the right-hand side of CI. Since
NT is itself very deductively weak, almost all of the content of F must have
been shifted to the new criterion CI*. Indeed, by the Proposition above, CI* is
at least as deductively strong as any other criterion of identity that it is entailed
by the given theory F and whose right-hand side expresses a reflexive relation
of C-entities on purely logical grounds.
It follows that the deductive power of theories of a very general kind may be
captured more or less just in terms of criteria of identity. That is: identity
criteria can be really strong! This might sound like a pretty amazing result.
However, the construction of ’*[x,y] is actually completely trivial: simply take
the formula
½if 1 & ::: & n then ’½x, y:
That ’*[x,y] expresses a reflexive relation for C-entities by logic alone then
follows from the corresponding property of ’[x,y]. Clearly, F entails NT and
CI* (by our assumptions on F). Vice versa, NT and CI* together logically
imply F: Since ‘For all x, y of category C: if ’[x,y] then ’*[x,y]’ is obviously a
logical truth, CI* entails CI by our initial considerations on comparisons of
Criteria of Identity 65
criteria of identity in terms of logical strength, and 1 & . . . & n follows from
NT and CI* taken together. Finally, assume that there is a criterion of identity
ðCI Þ For all x, y of category C: x ¼ y iff ½x, y
that is also entailed by F, where [x,y] expresses a reflexive relation for
C-entities by logic alone, but where CI** is not logically implied by CI*.
Then there must be a first-order model in which CI* holds but CI** does
not; by reflexivity, in the domain of that model there must therefore be distinct
entities a and b to which C applies in the model but for which [a,b] also holds
in the model.4 By CI*, since a is distinct from b, ’*[a,b] must then be false in
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
the model, which means that 1 & . . . & n must hold in the model. Because
CI* is true in the model, CI is satisfied by the model too, which is why F as a
whole is true in the model. But then CI** would have to be true in the model as
well, since by supposition F entails CI**, which contradicts the falsity of CI**
in the model as had been assumed initially. Hence, a criterion CI** as above
does not exist. So CI* logically implies all criteria of identity that satisfy the
given constraints, and therefore CI* is the strongest criterion of identity (up to
logical equivalence) subject to these constraints.
2 . . . and Wrong
Results like these, in which a theory gets reconstructed in some equivalent yet
diverse manner, always trigger the following question: Is the result merely
interesting to those who are occupied with the logical or methodological
study of science or could it be the case that, even for those who intend to
put forward a theory such as F above, it might actually be advantageous to do
so in terms of its reformulation F*? For instance, while William Craig’s
famous method of replacing a recursively axiomatized theory by a deductively
empirically equivalent one that uses only observational vocabulary does make
a very nice general point about the (maybe) in-principle-redundancy of the-
oretical terms for matters of empirical content only, no one would actually
apply the method to a theory and then use the result in one’s scientific or
philosophical work. After all, the axiomatic structure of the resulting system
is guaranteed to be cumbersome, unorganized, and non-explanatory
(cf. English [1973], p. 457). On the other hand, Frank P. Ramsey’s method
of eliminating theoretical terms from a theory by replacing them by instances
of second-order quantification also yields, in the deductive sense again, an
empirically equivalent theory, but it does leave the deductive structure of the
4
What we mean by [a,b] can be made formally precise either by introducing two object language
names for the two entities a and b and by substituting them for the relevant variables in [x,y],
or by using a variable assignment to set the values of the relevant variables in [x,y] to a and b,
respectively.
66 Hannes Leitgeb
original theory more or less intact (see English [1973], p. 458). And indeed,
while not being without problems either, Ramsification has been suggested
by philosophers of different schools and traditions—among whom, some
empiricists and structural realists—to lead to philosophically superior
theory formulations.
Now back to F* from above: Is it more like the outcome of Craig’s method
or more like Ramsey’s in that respect? Theories such as F* are certainly not as
muddled as those generated by Craig’s procedure and, of course, the point of
F* is not just to give an empirically equivalent variant of F but actually a
logically equivalent one. This said, it should still be clear that no one would
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
ever conceive of CI* as a good or even acceptable5 criterion of identity, and
therefore no one would ever want to reformulate F as described before and
then use the result in one’s scientific or philosophical work. For instance, take a
finitely axiomatized fragment ZFCfin of first-order ZFC set theory which con-
sists of Extensionality and finitely many further axioms 1 & . . . & n: Who
would think that CI*2 below is a good criterion of identity for sets, even
though CI*2 and NT2 taken together are logically equivalent to ZFCfin by
our Proposition from above?
(NT2) There are at least two sets.
ðCI 2 Þ For all sets x, y : x ¼ y iff ½if 1 & ::: & n then for all z ½z 2 x iff z 2 y:
But why exactly is this so? It could not be because CI*2 is so strong deduct-
ively, if the ‘enlightened’ metaphysical view of criteria of identity is right.
Accordingly, given the metaphysical view on criteria of identity again,
it could not be because Extensionality is analytic of membership while CI*2
is not. So which properties does Extensionality have, that CI*2 does not,
which make Extensionality good and CI*2 bad?
Section 5 of Horsten ([2010]) enumerates the following conditions that he
takes to be individually necessary and jointly sufficient for a criterion of iden-
tity to be acceptable (using Horsten’s term again): (i) The criterion should be
formally adequate, that is, its right-hand side must express an equivalence
relation; (ii) It should be materially adequate, that is, it should be true;
(iii) It should even be a necessary truth, and in order to be significant
it must follow from theoretical principles concerning the subject matter in
question; (iv) It should be non-trivial (or informative, as Horsten puts it),
that is, it should not be a truth of logic alone.6 Unfortunately, this list just
5
This is Horsten’s [2010] term; see his Section 5.
6
Horsten [2010] also argues convincingly that Predicativity or Non-Circularity is not a necessary
condition for a criterion of identity to be acceptable; see his Section 6.
Criteria of Identity 67
by itself will not answer our question from above. Clearly, the following
claims hold with generality: (i0 ) If ’[x,y] (as used in CI) expresses an equiva-
lence relation, then ’*[x,y] does so, too; additionally, if ’[x,y] expresses an
equivalence relation by logic alone, then so does ’*[x,y]; (ii0 ) If F is true,
then CI* is true; (iii0 ) If F is necessary, then CI* is necessary; (iv0 ) If CI is
not a truth of logic, neither is CI*. But now consider any good metaphys-
ical theory, F, about some ontological category, C, where F includes a
criterion of identity CI for entities of that very category. Maybe this
theory, F, might ultimately serve as the metaphysical foundation of some
scientific theory for entities of that category that would result from supply-
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
ing F with additional empirical hypotheses about entities of the category.
For instance, F might convey general principles of the ontology of events,
including a criterion of identity CI for events (such as Davidson’s or some-
one else’s), and F might ultimately be intended to be added to some finite
set of law hypotheses for physical events in space-time. Now, if F is a good
theory, then one might expect all of the antecedents in (i0 )-(iv0 ) to be satis-
fied, but then all the corresponding consequents will be satisfied too and
hence the criterion CI* that would result from applying the procedure
above to F would have all the virtues that were enumerated by Horsten.
Does this mean that, at least for any such F, the procedure would be
advisable after all? And that CI*2 from before would in fact be an accept-
able criterion of identity for sets, as ZFCfin satisfies all the antecedents in
question too?
We do not think so. But this leaves those who advocate a metaphysical
account of criteria of identity with the following interesting challenge:
What is it that makes CI* bad or unacceptable even when F and hence CI
are good and acceptable? Is it something ‘merely’ semantic or ‘merely’
epistemic-pragmatic rather than ontic? If so, even though criteria of identity
are supposed to be metaphysical principles, what is good about them would
then not be metaphysically salient, at least to a significant extent. In the old
days, the answer to the questions from above would have been simple: A good
criterion of identity is analytic. If 1 & . . . & n in F is synthetic, then normally
CI* is so too even when CI in F is analytic.7
7
Here is a tentative thought on what a metaphysically salient difference between CI and
CI* might be: The relation expressed by ’*[x,y] is disjunctive in nature even when ’[x,y] is
not. For ’*[x,y] is logically equivalent to: not (1 & . . . & n) or ’[x,y]. So maybe one should
ban disjunctive conditions from the right-hand sides of criteria of identity (as suggested, e.g.
by Lombard [1986], p. 33) for the same reasons as one does so commonly for specifications
of natural kinds, and just as one used to do traditionally for the definiens of a real definition?
But, of course, one would need to make this much more precise, which would be problematic in
itself.
68 Hannes Leitgeb
Acknowledgements
We would like to thank Leon Horsten, Øystein Linnebo, and an anonymous
referee for their comments and suggestions. We are grateful to the
Leverhulme Trust and the Alexander von Humboldt Foundation for support-
ing this work.
Faculty of Philosophy
Philosophy of Science and Study of Religion
Ludwig-Maximilians-University Munich
Geschwister-Scholl-Platz 1
Downloaded from http://bjps.oxfordjournals.org/ at University of Amsterdam on April 5, 2013
D-80539 Munich, Germany
[email protected] References
English, J. [1973]: ‘Underdetermination: Craig and Ramsey’, The Journal of Philosophy,
70, pp. 453–62.
Hilbert, D. and Bernays, P. [1934]: Grundlagen der Mathematik, Volume 1, Berlin:
Springer.
Horsten, L. [2010]: ‘Impredicative Identity Criteria’, Philosophy and Phenomenological
Research, 80, pp. 411–39.
Lombard, L. B. [1986]: Events: A Metaphysical Study, London: Routledge and Kegan
Paul.
Lowe, E. J. [1998]: The Possibility of Metaphysics, Oxford: Oxford University Press.
Williamson, T. [1990]: Identity and Discrimination, Oxford: Basil Blackwell.
