THE THEORY OF TRANSFINITE NUMBERS AND ORDER TYPES.

29

ON SOME DIFFICULTIES IN THE THEORY OF TRANSFINITE

NUMBERS AND ORDER TYPES

By B. RUSSELL.

[Received November 24th, 1905.—Head December 14th, 1905.]

DR. HOBSON'S most interesting paper in the Proceedings of the London

Mathematical Society* raises a number of questions which must be
answered before the principles of mathematics can be considered to be
at all adequately understood. I do not profess to know the complete
answers to these questions, and most of the present paper will consist
only of tentative suggestions, made as possibly a step towards the true
solutions, not as themselves constituting solutions. With the greater
part of Dr. Hobson's paper I find myself in agreement ; my purpose,
therefore, will not be in the main polemical, but rather to carry the
discussion a stage further by introducing certain distinctions which
I believe to be relevant and important, and by generalizing as far as
possible the difficulties and contradictions hitherto discovered in the
theory of the transfinite.
There are two wholly distinct difficulties to be considered in the theory
of transfinite cardinal numbers, namely :
(1) The difficulty as to inconsistent aggregates (as they are
called by Jourdain);
(2) The difficulty as to what we may call Zermelo's axiom.!
These two difficulties do not seem to be clearly distinguished by
Dr. Hobson ; yet they are, so far as appears, largely independent and
of very different degrees of importance. The first leads to definite con-
tradictions, and renders all reasoning about classes and relations, prima
facie, suspect; while the second merely raises a doubt as to whether
a certain much used axiom is true, without showing that any fundamental
difficulties arise either from supposing it true or from supposing it false.
I shall consider these difficulties separately, beginning with the first,
because it is more fundamental.

* Series 2, Vol. 3, pp. 170-188.

t See his "Beweis, dass jede Menge wohlgeordnet werden kann," Math. Annaleii,
Vol. LIX., pp. 514-516. For statements of various forms of this axiom see the third part
of this paper.
80 MR. B. RUSSELL [Dec. 14,

I.
When Dr. Hobson speaks of the necessity of a norm for constituting
an aggregate, he appears sometimes to suppose that the norm is absent
or ill-defined in the case of inconsistent aggregates, at other times to
suppose it absent where Zermelo's axiom requires it. But the two
cases are, in reality, quite distinct. The doubt as to the truth of
Zermelo's axiom arises from the impossibility of discovering a norm by
which to select one term out of each of a set of classes, while the difficulty
of inconsistent aggregates arises from the presence of a perfectly definite
norm combined with the demonstrable absence of a corresponding
aggregate. This suggests that a norm is a necessary but not a sufficient
condition for the existence of an aggregate; if so, the complete solution
of our first set of difficulties would consist in the discovery of the precise
conditions which a norm must fulfil in order to define an aggregate.
Logical determinateness, it seems, is not sufficient, as Dr. Hobson supposes
(p. 173), and the meaning which he attaches to the term aggregate (ib.)
appears to be too wide. This is proved by a perfectly strict argument,
which I shall try to state after explaining some ways of generating
inconsistent aggregates.
In the first place, since the discussion belongs to symbolic logic,
which already possesses technical names for the ideas we require, it is
desirable to compare Dr. Hobson's terms with those in current use.
What he calls a norm is what I call a propositional function. A pro-
positional function of x is any expression <p\ x whose value, for every
value of x, is a proposition; such is " x is a man " or " sin x = 1."
Similarly we write <j>! (x, y) for a propositional function of two variables;
and so on.
In this paper I shall use the words norm, property, and propositional
function as synonyms.
The word aggregate is used sometimes with an implication of order,
sometimes without; I shall use class where there is no implication of
order, and where there is order I shall consider the relation of before and
after which generates the order. This last is necessary because every
class which can be ordered at all can be ordered in many ways; so that
only the ordering relation, not the class, determines what the order is to
be. A relation will be used in an extensional sense, i.e., so that two
relations are identical provided each holds whenever the other holds. We
shall find that a propositional function <f>\ x may be perfectly definite, in
the sense that, for every value of x, <p\ x is determinably true or determin-
ably false, while yet the values of x for which <p! x is true do not form a
class. And, similarly, we shall find that a propositknal function 0! (re, y)
1905.] THE THEORY OF TRANSFINITE NUMBERS AND ORDER TYPES. 31

may be in the same sense definite, without there being any relation R
which holds between x and y when and only when <j>\ (x, y) is true.
In order to eliminate at the outset a number of considerable but
irrelevant difficulties, I may point out that the argument we are about to
consider does not depend upon this or that view as to the nature of
classes and relations. The refutable assumption as to the nature of
classes and relations is only this: that a class is always uniquely
determined by a norm or property containing one variable, and that two
norms which are not equivalent {i.e., such that, for any value of the
variable, both are true or both false) do not determine the same class, with
a similar assumption as regards relations. It is in no way essential to
the argument to suppose that classes and relations are taken in extension,
i.e., that two equivalent norms determine the same class or relation.
Thus the argument proves that a norm itself is in general not an entity ;
that is, if we make statements of the form <p! x about a number of
different values of x, we cannot pick out an entity <f> which is the common
form of all these statements, or is the property assigned to x when we
state (j>\x. In other words, a statement about x cannot in general be
analyzed into two parts, x and what is said about x. There is no harm
in talking of norms or properties so long as we remember this fact; but, if
we forget it, we become involved in contradictions.
The two contradictions first discovered concerned respectively the
greatest ordinal and the greatest cardinal.* Of these the cardinal con-
tradiction is the simpler, and lends itself more readily to the removal from
arithmetic to logic which I wish to effect for both. I shall therefore
consider it first.
The cardinal contradiction is simply this: Cantor has a proof t that
there is no greatest cardinal, and yet there are properties (such as
" x = x ") which belong to all entities. Hence the cardinal number of
entities having such a property must be the greatest of cardinal numbers.
Hence a contradiction.
If every logically determinate norm defines a class, there is no escape
from the conclusion that there is a cardinal number of all entities. For,
in that case, the norm "x=x" defines a class, which contains all
entities: call this class V. Then the norm " u is similar to V" defines a
set of classes which may be taken as being the cardinal number of V,

* The contradiction concerning the greatest ordinal was first set forth by Burali Forti, " Una
questione sui numeri transfiniti," Rendieonti del Cireolo Matematico di Palermo, 1S97. The con-
tradiction concerning the greatest cardinal is discussed in my Principles of Mathematics,
§344/.
t Jahrexbericht der Deiitschen Mathematiker-Vereinigung, Vol. i., 1892, p. 77.
32 MR. B. RUSSELL [NOV. 14,

i.e., the greatest cardinal number.* Thus, if every logically determinate

norm defines a class, it is impossible to escape the conclusion that there
is a greatest cardinal.
The other horn of the dilemma yields more interesting results.
Cantor's proof that there is no greatest cardinal may be simplified into the
following:—Let u be any class, and B a one-one correlation of all the
members of u to some (or all) of the classes contained in u. There are
such correlations, since one of them is obtained by correlating each member
of u with the class whose only term is that member. Consider now the
following norm :—" x is a member of u, but is not a member of the class
with which B correlates it." Suppose this norm defines a class w. Then
w is omitted from the correlation ; for, if w were correlated with x, then,
if x is any member of w, it follows from the definition of w that x is not
a member of its correlate, i.e., is not a member of w ; while, conversely,
if x is not a member of w, it is a member of its correlate, i.e., of w.
Hence the supposition that w is the correlate of x leads to a contradic-
tion. Hence, in any one-one correlation of all the terms of u with classes
contained in u, at least one class contained in u is omitted. Therefore,
whatever class u may be, there are more classes contained in u than there
are members of u.
We may test this conclusion, in the case of the class of all entities, by
constructing, according to the method of the proof of the Schroder-Bernstein
theorem, an actual one-one correlation of all terms with all classes, and
then considering the class which Cantor shows to be omitted. This
process leads us to the consideration of the norm : " x is not a class which
is a member of itself." If this norm defines a class w, then the class w
is omitted from our correlation. But it is easy to see that this norm does
aot define a class at all. For, if it defined a class w, we should find that,
if w is a member of itself, then it is not a member of itself, and vice versa.
Hence there is no such class as w. Essentially the same argument may
be stated as follows:—If u be any class, then, when x = u, the statement
" x is not an x " is equivalent to " x is not a u." Hence, whatever class
u may be, there is one value of x—namely, u—for which " x is not an x "
is equivalent to " x is not a w " ; thus there is no class w such that " x is
not an x " is always equivalent to " x is a w." Hence, again, this norm
does not define a class.
We thus find that, quite apart from any view as to the nature of
cardinals, and without any considerations belonging to arithmetic, we can
prove that at least one perfectly determinate norm does not define a class.

* I shall consider later Dr. Hobson's objection to this definition.

1905.] THE THEOEY OF TRANSFINITE NUMBERS AND ORDER TYPES. 33

By the same method, we can easily construct other such norms. Take
any class u for which we can correlate all entities to some u's by a one-one
correlation. By the method of the proof of the Schroder-Bernstein
theorem, construct an actual one-one correlation of all the members of u
to all classes contained in u, and then consider the norm : " x is a member
of u which is not a member of its correlate according to the correlation in
question." This norm does not define a class. Thus from the class of all
relations we obtain the norm: " B is not a relation which is a member of
its own domain." From the class of all couples we obtain the norm :
" v is such that the couple whose members are (1) the class of all entities,
(2) v, is not a member of v." Thus it appears that the contradiction dealt
with in chapter x,. of my Principles of Mathematics is a special case of a
general type of contradictions which result from supposing that certain
propositional functions determine classes, when, in fact, they do not do so.
The above method of discovering such propositional functions is not re-
quired for proving, when they are discovered, that they are of the sort
that do not define classes. In each case it is easy to discover a definite
simple contradiction, analogous to that discussed in the above mentioned
chapter, which results from supposing that the propositional functions in
question do determine classes.
In like manner propositional functions of two variables do not always
determine relations. For example, " B does not have the relation B to
S " does not determine a relation T between B and S, i.e., it is not equi-
valent, for all values of B and S, to " B has the relation T to S." For, if
it were, substituting T for B and for S, we should have " Tdoes not have
the relation T to T," equivalent to " T has the relation T to T," which is
a contradiction.
The following contradiction, of an analogous type to those discussed
above, shows that a norm or property is not always an entity which can
be detached from the argument of which it is asserted. Consider the
norm " x does not have any property which it is." If this assigns to x
the property 6, then " x has the property 6 " is equivalent to " x does not
have any property which it is." Hence, substituting 6 for x, " 6 has the
property (9" is equivalent to " 6 does not have any property which it is,"
which is equivalent to " 6 does not have the property 0 " ; whence a con-
tradiction. The solution, in this case, is that properties are not always
(if ever) separable entities which can be put as arguments either to other
properties or to themselves. Thus, when we speak of properties we are
sometimes (if not always) employing an abbreviated form of statement,
which leads to errors if we suppose that the properties we are speaking of
are genuine entities.
8BR. 2 . VOL. 4 . NO. 9 1 4 . D
84 MR. B. KUSSELL [Dec. 14,

We have thus reached the conclusion that some norms (if not all) are
not entities which can be considered independently of their arguments,
and that some norms (if not all) do not define classes. Norms (containing
one variable) which do not define classes I propose to call non-predicative;
those which do define classes I shall call predicative. Similarly, by exten-
sion, a norm containing two variables will be called predicative if it
defines a relation; in the contrary case it will be called non-predicative.
Thus we need rules for deciding what norms are predicative and what are
not, unless we adopt the view (which, as we shall see, has much to re-
commend it) that no norms are predicative.
I come now to Burali-Forti's contradiction concerning the greatest
ordinal, and I shall show how this too reduces to a simple logical contra-
diction resulting from supposing that a certain non-predicative function is
predicative.
Burali-Forti's contradiction may be stated, after some modification, as
follows:—If u is any segment of the series of ordinals in order of magni-
tude, the ordinal number of u is greater than any member of u, and is, in
fact, the immediate successor of u (i.e., the limit if u has no last term,
or the immediate successor of the last term if u has a last term). The
ordinal number of u is always an ordinal number, and is never a member
of u. But now consider the whole series of ordinal numbers. This is
well ordered, and therefore should have an ordinal number. This must
be an ordinal number, and yet must be greater than any ordinal number.
Hence it both is, and is not, an ordinal number, which is a contradiction.
To generalize this contradiction, put <p\ x in place of " x is an ordinal,"
and f'u* in place of " the ordinal number of u." Then in the case of the
ordinals 0 and / are such that, if all the members of u satisfy <p, then f'u
satisfies <p and is not a member of u. Whenever these two conditions are
satisfied for all values of u, one or other of two conclusions follows:
namely, either (1) <j>\ x is not a predicative property; or (2), if <f>\x is
predicative and defines the class w, then there must be no such entity as
f'w. This is proved very simply as follows:—If there is such a class as
w, and such an entity as f'w, then, since every member of w satisfies <p,
it follows that/'w satisfies 0 ; but, conversely,/'w must be not a member
of w, and must therefore not have the property $, since w consists of all
terms having the property 0. In the special case of the ordinals, our two
alternatives are : (1) the ordinals do not form a class ; (2) although they
form a class, they have no ordinal number. The second alternative is

* The inverted comma may be read " of." The notation ^'M means the same as f{u), but is
for several reasons more convenient.
1905.] T H E THEORY OF TRANSFINITE NUMBERS AND ORDER TYPES. 85

equivalent to the assumption that either the whole series of ordinals is

not well ordered, or, if it is well ordered, the dual property " a and /3 are
ordinal numbers, and a is less than /3 " is non-predicative; so that the
series as a whole has no definite type, i.e., no ordinal number. The
supposition that the whole series of ordinals is not well ordered can be
disproved* ; hence we are left with the alternatives that either (1) the
property " x is an ordinal number" is non-predicative, or (2), though "x
is an ordinal number " is predicative, " x and y are ordinal numbers and
x is less than y " is non-predicative.
We have seen that Burali-Forti's contradiction is a particular case of
the following:—
" Given a property <f> and a function / , such that, if 0 belongs to all
the members of u, f'u always exists, has the property <p, and is not a
member of u; then the supposition that there is a class to of all terms
having the property <p and that f'w exists leads to the conclusion that
f'w both has and has not the property <p."
This generalization is important, because it covers all the contradic-
tions that have hitherto emerged in this subject. In the case of the class
of terms which are not members of themselves, we put " x is not a member
of x " for <f>\x, and u itself lox f'u. In this case, owing to the fact that
f'u is u itself, we have only one possibility : namely, that " x is not a
member of x " is non-predicative. In other cases, we have two possibili-
ties, and it may often be difficult to decide which of them to choose.
When we have a pair such as <p and / above, we can define, in terms
of / alone, without introducing <p, a series ordinally similar to that of all
ordinals, and obtain, as regards this series, a contradiction analogous to
Burali-Forti's, provided / satisfies certain conditions. We do this as
follows :—Taking any class x, for which f'x exists, take f'x as the first
term of our series, take t h e / o f the class got by adding f'x to x as the
second term, and so on. Generally, the successor of any term
is the / of the class consisting of that term together with all
its predecessors and x, and the successor of a class it of terms
having no maximum is the / of the class consisting of the whole seg-
ment defined by the class u. This gives Cantor's two principles of
formation, and we can define the property of occurring in this series by

* This supposition can be disproved (by the generalized form of induction which applies
throughout any well-ordered series) by means of the theorem that every segment of the series of
ordinals is well ordered. It is not disproved by Jourdain's theorem, that every series which is
not well ordered must contain a part of type *w ; for this theorem depends upon Zermelo'a axiom,
of which the truth is doubtful.
D 2
36 MR. B. EUSSELL [Dec. 14,

the generalized form of induction.* We may then, subject to certain con-

ditions as to / , substitute for <p the property of occurring in the /-series
starting from x. If / has the property that, if u is composed of terms of
the above series, then f'u exists and is not a member of u, it will follow
that the whole series does not form a class; for, if it did, its / would both
be and not be a member of the series. In the particular case of the
ordinals, if u is a class of ordinals, f'u is their immediate successor ; the
whole series of ordinals can be generated by the above method, starting
from 0. In the case of " x is not an x," f'x is x itself: if we start from
any class which is not a member of itself, and proceed by the above method,
we obtain a series, like the series of all ordinals,! consisting entirely of
classes which are not members of themselves, and the series as a whole
does not form a class.
The above considerations point to the conclusion that the contradic-
tions result from the fact that, according to current logical assumptions,
there are what we may call self-reproductive processes and classes. That
is, there are some properties such that, given any class of terms all having
such a property, we can always define a new term also having the property
in question. Hence we can never collect all the terms having the said
property into a whole; because, whenever we hope we have them all, the
collection which we have immediately proceeds to generate a new term
also having the said property. It is probable, in view of the above
general form for all known contradictions, that, if <j> is any demonstrably
non-predicative property, we can actually construct a series, ordinally
similar to the series of all ordinals, composed entirely of terms having the
property <p. Hence, if the terms satisfying <f> can be arranged in a series
ordinally similar to a segment of the series of ordinals, it follows that no
contradiction results from assuming that <p is a predicative property. But
this proposition is of very little use, until we know how far the series of
ordinals goes; and at present it is not easy to see where this series begins
to be non-existent, if such a bull may be permitted.

* This is done as follows :—A property is inductive in the/-series if whenever it belongs to

a class N it belongs to the class got by adding f'u to u, and whenever it belongs to each of a set'
of classes it belongs to their logical sum, i.e., to the class of members of members of the set. A
term belongs to the "/-series starting from x " if it possesses every property which is possessed
by x and is inductive in the/-series.
t I owe the proof that this series is well ordered and ordinally similar to the series of all
ordinals to Mr. Gr. G-. Berry, of the Bodleian Library.
1905.] THE THEORY OF TRANSPINITE NUMBERS AND ORDER TYPES. 37

II.
We have now seen the nature of the contradictions which beset the
theory of the transfinite : we have seen that they are not an isolated few,
but can be manufactured in any required number by a recipe; we have
seen that all of them belong to a certain definite type, and we have seen
that none of them are essentially arithmetical, but all belong to logic, and
are to be solved, therefore, by some change in current logical assumptions.
I propose, in this section, to consider three different directions in which
such a change may be attempted. I shall endeavour to set forth the
advantages and disadvantages of each of the three, without deciding in
favour of any one of them.
What is demonstrated by the contradictions we have considered is
broadly this: " A propositional function of one variable does not always
determine a class."*
In view of this fact, it is open to us, prima facie, to adopt one or other
of two theories. We may decide that all ordinary straightforward propo-
sitional functions of one variable determine classes, and that what is needed
is some principle by which we can exclude the complicated cases in which
there is no class. In this view, the state of things is like that in the
differential calculus, where every commonplace continuous function has a
derivative, and only rather complicated and recondite functions have to
be excluded. The other theory which suggests itself is that there are no
such things as classes and relations and functions as entities, and that
the habit of talking of them is merely a convenient abbreviation.
The first of these two theories itself divides into two, according as we
hold that what classes have to avoid is excessive size, or a certain charac-
teristic which we may call zigzagginess. Of these, the second is the more
conservative, i.e., it preserves more of the theory of the transfinite than
the first. Both preserve more of it than does the theory that there are
no such things as classes. I shall consider these three theories in the
following order, and by the following names :—
A. The zigzag theory.
B. The theory of limitation of size.
C. The no classes theory.

* Here it is to be understood that the arguments which show that there is not always a
class also show that there is not always a separable entity which is the propositional function (as
opposed to its value) ; also that some propositional functions of two variables do not determine a
relation either in intension or in extension, if we mean by a relation a separable entity which
can be considered apart from related terms.
38 MR. B. EUSSELL [Dec. 14,

A. The Zigzag Theory.

Each of the three theories can be recommended as plausible by the
help of certain a priori logical considerations. In the zigzag theory, we
start from the suggestion that propositional functions determine classes
when they are fairly simple, and only fail to do so when they are com-
plicated and recondite. If this is the case, it cannot be bigness that makes
a class go wrong ; for such propositional functions as " x is not a man "
have an exemplary simplicity, and are yet satisfied by all but a finite
number of entities. In this theory, as well as in the theory of limitation
of size, we define a predicative propositional function as one which de-
termines a class (or a relation, if it contains two variables); thus in the
zigzag theory the negation of a predicative function is always predicative.
In other words, given any class u, all the terms which are not members of
u form a class which may be called the class not-w.
If now 0 ! x is a non-predicative function, it follows that, given any
class u, there must either be members of u for which <p! x is false, or
members of not-u for which <p\ x is true. (For, if not, 0! x would be true
when, and only when, a; is a member of u; so that <jt\ x would be predic-
ative.) It thus appears that <p\x fails to be predicative just as much by
the terms it does not include as by the terms it does. Again, given
any class u, the property <f>\x belongs either to some, but not all, of
the members of u, or to some, but not all, of the members of
not-u. This is the zigzag property which gives its name to the theory
we are considering. This theory is specially suggested by the argu-
ment of Cantor's proof that there is no greatest cardinal. This proof,
as we have already seen, constructs a would-be class w by the norm " x is
not a member of the class with which it is correlated by the relation H,"
where E is a relation which correlates individuals with classes. Such
would-be classes, as we saw, are very apt to be not classes, and they all
have a certain zigzag quality, in the fact that x is a w when x is not a
member of its correlate, and is not a w when a; is a member of its
correlate.
The full development of this theory requires axioms as to the kinds of
functions that are predicative. It has the great advantage that it admits
as predicative all functions which can be stated simply, and only excludes
such complicated cases as might well be supposed to have strange
properties.*
The principal objection to this theory, so far as it is at present de-

• For suggestions of a solution more or less on the above lines, see my Principles of Mathe-
matics, §§ 103, 104.
1905.] T H E THEORY OP TRANSFINITB NUMBERS AND ORDER TYPES. 89

veloped, is that the axioms as to what functions are predicative have to

be exceedingly complicated, and cannot be recommended by any intrinsic
plausibility. This is a defect which might be remedied by greater in-
genuity, or by the help of some hitherto unnoticed distinction. But
hitherto, in attempting to set up axioms for this theory, I have found no
guiding principle except the avoidance of contradictions; and this, by
itself, is a very insufficient principle, since it leaves us always exposed to
the risk that further deductions will elicit contradictions. The general
postulate, that predicative propositional functions must have a certain
simplicity, does not lend itself readily to the decision whether this or that
propositional function has the requisite degree of simplicity. Nevertheless,
since these difficulties are all such as further research might conceivably
remove, the theory is not to be rejected wholly, but is rather to be re-
tained as one of those that are possible. Speaking broadly, one may say
that it applies better to cardinal than to ordinal contradictions: it deals
more readily with such difficulties as that of the class of classes which are
not members of themselves than with such difficulties as that of Burali-
Forti.
The zigzag theory, in some form or other, is that assumed in the
definitions of cardinal and ordinal numbers as classes of classes (if
numbers are supposed to be entities). For all these classes of classes, if
they are legitimate, must contain as many members as there are entities
altogether; hence, if bigness makes classes go wrong, as we suppose in
the " limitation of size " theory, cardinals and ordinals so defined will
be illegitimate classes. Dr. Hobson has various criticisms on these defin-
itions of cardinals and ordinals; but on the zigzag theory his criticisms
can, I think, be all satisfactorily met.
Dr. Hobson says * : " I t has been seen that the assumptions that
an ordered aggregate possesses a definite order type and a definite cardinal
number, which can be treated as objects, lead to the contradiction pointed
out by Burali-Forti." This statement seems to me somewhat too sweep-
ing. It is quite open to us to hold every ordered aggregate possesses
a definite cardinal number, and that every ordered aggregate which is
ordinally similar to a segment of the ordinals in order of magnitude
possesses a definite ordinal number. All that Burali-Forti's contradiction
forces us to admit is that there is no maximum, ordinal, i.e., that the func-
tion " a and /3 are ordinal numbers, and a is less than /3 " and all other
functions ordinally similar to this one are non-predicative. In the same
way the difficulty of the greatest cardinal is met by denying that the

* P. 176, beginning of No. 5.

40 MR. B. RUSSELL [Dee. 14,

defining functions of Cantor's omitted classes are predicative in certain

cases. Thus we conclude that in this theory there is a greatest cardinal,
but there is no greatest ordinal; in each case contradictions are avoided
by regarding certain functions as not predicative.
Dr. Hobson distinguishes two methods of establishing the existence
of a class of mathematical entities: the genetic method and the msthod
by postulation. He rejects the former, as regards cardinals and ordinals;
but he seems not to perceive that this can only be done by recognizing
that there may be no class even where there is a perfectly definite norm.
From his No. 2 one would suppose that he regards the norm as a
sufficient condition for the class ; yet, later on, he refuses to admit
classes which are defined by unimpeachable norms. It seems hardly
correct to say, as he does: " In the genetic method, as applied to the
construction of the whole series of ordinal members, this notion of
correspondence plays no part." (No. 6, p. 177.) It is the notion of
correspondence which defines the class of relations constituting an
ordinal number; this class consists of all the relations which are like * a
certain given relation. " The existence of a number," he truly says, " is
constantly inferred from that of a single unique ordered aggregate." (Ib.)
But there can be no objection to this procedure, unless on the ground
that, when P is given, " Q is like P " is not predicative in respect of Q.
It is, of course, very easy to prove, when we have one series of a
certain type, that there are an infinite number of series of the same
type. To do this we need only substitute other terms for the terms of
our series. Suppose, e.g., our series is composed of numbers. We may
substitute Socrates for any terms of our series ; this will give as many
new series of the same type as there are terms in the given series. If
our series is infinite, we can obtain N0 series of the same type by
merely knocking off terms at the beginning ; and so on. Thus, if
multiplicity of series of a given type is desired, there is no difficulty
in obtaining as many series of the given type as there are entities
altogether, i.e., the maximum cardinal number of series of the given
type. (For, instead of Socrates, we may substitute any other term not
occurring in our series.) Thus it is not the case that the genetic method
involves " the setting up of a scale of standards, to which standards
no aggregates not consisting of the preceding nuribars conform " (No. 6,
p. 173), though I do not see what harm there woiald be if this were the
case.

• I use like to mean ordinally similar to. For the precise definition cf. Principles of Kathe-
matict, § 253.
1905.] THE THEORY OF TRANSFINITE NUMBERS AND ORDER TYPES. 41

The same remark applies to the criticism (No. 7, p. 179) of the

definition of a cardinal as a class of similar classes.* It is very easy
to show that the number of classes similar to the class of numbers
from 0 to n is as great as the total number of entities; and, even
if no other class were similar to this class, that would not, so far
as I can see, constitute any objection. The number n-\-\ would, in
that case, be the class whose only member is the class of numbers
from 0 to n.
Dr. Hobson explains that his opinion and mine are at variance
as to the definition of cardinals because I, unlike him, " regard the
activities of the mind as irrelevant in questions of existence of entities "
(No. 7, p. 180). This is a philosophical difference, and, like all philo-
sophical differences, it ought not to be allowed to affect the detail of
mathematics, but only the interpretation. Mathematics would be in
a bad way if it could not proceed until the dispute between idealism and
realism has been settled. When a new entity is introduced, Dr. Hobson
regards the entity as created by the activity of the mind, while I regard
it as merely discerned; but this difference of interpretation can hardly
affect the question whether the introduction of the entity is legitimate
or not, which is the only question with which mathematics, as opposed
to philosophy, is concerned.
There is another passage in No 7 (p. 179) which calls for explanations,
namely, the following :—" Russell objects to the conception of a number
as the common characteristic of a family of equivalent aggregates on
the ground that there is no reason to think that such a single entity
exists with which the aggregates have a special relation, but that there
may be many such entities. The mind does, however, in point of fact,
in the case of finite aggregates at least, recognize the existence of such
single entity, the number of the aggregates; and this is a valid result of
our mental activity, subject to the law of contradiction."
In the first place, it is not merely the case that " there may be many
such entities," but that there demonstrably are as many as there are
entities altogether. Given any many-one relation having the property
that when, and only when, u and v are similar classes, there is an
entity a to which both u and v have the relation S, the converse domain
of S (i.e., the terms to which classes have the relation S) will have all
the formal properties of cardinal numbers.! Now, if there is one such

• This definition is due to Frege. See his Grundlagm der Arithmetik, Breslau, 1S84,
pp. 79, 85.
t For a development of this point of view Bee §2 of " L a Logique des Relations,"
Jievue de Mathematiques, Vol. vn.
42 MR. B. RUSSELL [Dec. 14,

relation as S, it is very easy to prove that there are as many as there

are entities altogether ; and, if there is no such relation as S, then
there are no such entities as cardinal numbers. (There might be'cardinal
numbers for some classes and not for others, if there was a relation
such as S which had some classes in its domain, but no relation such
as S which had all classes in its domain.)
The supposition that there is no such relation as S is disproved by
the fact that the relation of a class to the class of all classes similar
to it has the properties we wish S to have. This disproof is rejected by
Dr. Hobson, since he considers that it involves improper classes. His
position seems to be that, at least in the case of finite aggregates, " the
mind " immediately recognizes a certain relation of the sort required.
The simplest formal statement of this point of view is, roughly, as
follows :—
In beginning cardinal arithmetic we introduce a new indefinable S,
concerning which we lay down the indemonstrable properties: *
(1) S is a many-one relation ;
(2) Every finite class (and, presumably, some infinite classes)
have the relation S to some term ;
(3) When, and only when, two finite classes (and, presumably,
some pairs of infinite classes) are similar they both have the
relation S to the same term ;
(4) Things which are not classes do not have the relation S to
anything.
The reason that S has to be indefinable and the above propositions
indemonstrable is that, if we regard the above propositions as giving
a definition of S " by postulates," they do not determine S, since an
infinite number of relations (if any) fulfil the above conditions, and
every entity will, for a suitable S, be the cardinal number (in respect
of that S) of some class which has a cardinal number. Moreover, the
recognition by " the mind," which Dr. Hobson speaks of, is precisely
the process of introducing an indefinable. It is a process of which, in
certain cases, I fully recognize the validity and the necessity ; but
indennables and indemonstrables are to be diminished in number as much
as possible.! Moreover, in the case supposed, where Dr. Hobson says

* It is probably possible to simplify the statement of these indemonstrables.

t This is merely the truism with which Dedekind begins " Was sind und was sollen die
Zahlen," namely: "Was beweisbariat, soil in der Wissensohaft nicht ohne Beweia geglaubt
werden."
1905.] THE THEORY OF TRANSPINITE NUMBERS AND ORDER TYPES. 43

that " the mind " recognizes such entities, I am unable to agree: if he
said "my mind," I should have taken his word for it; but, personally,
I do not perceive such entities as cardinal numbers, unless as classes of
similar classes.

B. The Theory of Limitation of Size.

This theory is naturally suggested by the consideration of Burali-
Forti's contradiction, as well as by certain general arguments tending
to show that there is not (as in the zigzag theory) such a thing as the
class of all entities. This theory naturally becomes particularized into
the theory that a proper class must always be capable of being arranged
in a well-ordered series ordinally similar to a segment of the series of
ordinals in order of magnitude ; this particular limitation being chosen
so as to advoid Burali-Forti's contradiction.* We still have the dis-
tinction of predicative and non-predicative functions ; but the test of
predicativeness is no longer simplicity of form, but is a certain limit-
ation of size. In this theory, if u is a class, " x is not a member of u "
is always non-predicative ; thus there is no such class as not u.
The reasons recommending this view are, roughly, the following :—We
saw, in the first part of this paper, that there are a number of processes,
of which the generation of ordinals is one, which seem essentially in-
capable of terminating, although each process is such that the class of
all terms generated by it (or a function of this class) ought to be the
last term generated by that process. Thus it is natural to suppose that
the terms generated by such a process do not form a class. And, if so,
it seems also natural to suppose that any aggregate embracing all the
terms generated by one of these processes cannot form a class. Con-
sequently there will be (so to speak) a certain limit of size which no
class can reach ; and any supposed class which reaches or surpasses this
limit is an improper class, i.e., is a non-entity. The existence of self-
reproductive processes of this kind seems to make the notion of a totality
of all entities an impossible one ; and this tends to discredit the zigzag-
theory, which admits the class of all entities as a valid class.
This theory has, at first sight, a great plausibility and simplicity,
and I am not prepared to deny that it is the true solution. But the
plausibility and simplicity tend rather to disappear on examination.
Let us first recall the generalizations of Burali-Forti's contra-

* This view has been advocated by Jourdain, "On the Transfinite Cardinal Numbers of
Well-ordered Aggregates," No. 4, Phil. Mag., January, 1904 ; also " On Transfinite Cardinal
Numbers of the Exponential Form," Phil. Mag., January, 1905.
44 MR. B. RUSSELL [Dec. 14,

diction which we obtained in the first part of this paper. The funda-
mental proposition is : " Given a property <j> and a function / such
that, if <f> belongs to all the members of any class u, then f'u always
exists and has the property <p, but is not a member of u, it follows that
either <p is non-predicative or, if <p is predicative and determines the
class w, then f'w does not exist."
The theory of limitation of size neglects the second alternative (that
f'w may not exist), and decides for the first (that <f> is not predicative).
Thus, in the case of the series of ordinals, the second alternative is
that the whole series of ordinals has no ordinal number, which is
equivalent to denying the predicativeness of " a and /3 are ordinal
numbers, and a is less than /3." The adoption of this alternative would
enable us to hold that all ordinals do form a class, and yet there is
no greatest ordinal. But the theory in question rejects this alternative,
and decides that the ordinals do not form a class. The only case in
which this is the only alternative is when flu is u itself; otherwise we
always have a choice.
A great difficulty of this theory is that it does not tell us how far up
the series of ordinals it is legitimate to go. It might happen that w
was already illegitimate: in that case all proper classes would be finite.
For, in that case, a series ordinally similar to a segment of the series
of ordinals would necessarily be a finite series. Or it might happen that
co2 was illegitimate, or co" or (ax or any other ordinal having no immediate
predecessor. We need further axioms before we can tell where the series
begins to be illegitimate. For, in order that an ordinal a may be
legitimate, it is necessary that the propositional function " /3 and y are
ordinal numbers less than a, and (3 less than y " should be predicative.
(Here, of course, " less than a " must be replaced by some property not
involving a, but such that, if a is legitimate, the property is equivalent
to being less than a.) But our general principle does not tell us under
what circumstances such a function is predicative.
It is no doubt intended by those who advocate this theory that all
ordinals should be admitted which can be defined, so to speak, from
below, i.e., without introducing the notion of the whole series of ordinals.
Thus they would admit all Cantor's ordinals, and they would only avoid
admitting the maximum ordinal. But it is not easy to see how to state
such a limitation precisely : at least, I have not succeeded in doing so.
The merits of this theory, therefore, would seem to be less than they
at first appear to be.
1905.] T H E THEORY OF TRANBFINITE NUMBERS AND ORDER TYPES. 45

C. The No Classes Theory.

In this theory classes and relations are banished altogether.* It is
not necessary to the theory to assume that no functions determine
classes and relations ; all that is essential to the theory is to abstain
from assuming the opposite. This is the strong point of the theory we
are now to consider: the theory is constituted merely by abstinence
from a doubtful assumption, and thus whatever of mathematics it permits
us to obtain is indubitable in a way which anything involving classes
or relations cannot be. The objections to the theory are (1) that it
seems obvious to common sense that there are classes; (2) that a great
part of Cantor's theory of the transfinite, including much that it is hard
to doubt, is, so far as can be seen, invalid if there are no classes or
relations ; (3) that the working out of the theory is very complicated,
and is on this account likely to contain errors, the removal of which
would, for aught we know, render the theory inadequate to yield the
results even of elementary arithmetic.
To explain fully how this theory is to be developed would take too
much space. Some of its main points may, however, be briefly set
forth.
Instead of a function 0! x, where the notation inevitably suggests
the existence of something denoted by "<p," we proceed as follows:—
Let p be any proposition, and a a constituent of p. (We may say broadly
that a is a constituent of p if a is mentioned in stating p.) Then let
" p — " denote what p becomes when x is substituted for a in the place
a *
or places where a occurs in p. For different values of x this will give us
what we have been accustomed to call different values of a propositional
function. In place of <f> we have now two variables, p and a: in respect
to the different values of p —, we may call p the prototype and a the
origin or initial subject. (For a may be taken as being, in a generalized
sense, the subject of p.) Consider now such a statement as "p— is
true for all values of x." Let b be an entity which is not a constituent
b x
of p, and put q = p—; then " q -r- is true for all values of x" is
a o

* It must be understood that the postulate of the existence of classes and relations is exposed
to the same arguments, pro and eon., as the existence of propositional functions as separable
entities distinct from all their values. Thus, in the theory we are considering, anything said
about a propositional function is to be regarded as a mere abbreviation for a statement about
some or all of its values.
46 MB, B. RUSSELL [Dec. 14,

equivalent to "p— is true for all values of x" Thus, subject to

a
a certain reservation, the statement " p — is true for all values of x "
a
is independent of the initial subject a, and thus may be said to depend
only upon the form of p.* Statements of this sort replace what would
otherwise be statements having propositional functions for their argu-
ments. For example, instead of " <f> is a unit function" (i.e., " There
is one, and only one, x for which <f>\ x is true"), we shall have " There
is an entity b such that p — is true when, and only when, x is identical
Cb
with 6." There will not now be any such entity as the number 1 in
isolation ; but we shall be able to define what we mean by " One, and
only one, proposition of the type p — (for a given p and a) is true."
Cb

Instead of saying " The class u is a class which has only one member,"
we shall say (as above) " There is an entity b such that p — is true when,
and only when, x is identical with 6." Here the values of x for which p —
is true replace the class u; but we do not assume that these values
collectively form a single entity which is the class composed of them.
There is not much difficulty in re-wording most definitions so as
to fit in with the new point of view. But now the existence theorems
become hard to prove. We can manufacture enough different propositions
to show what is now equivalent to the existence of w and N0, though
the process is cumbrous and artificial. We shall be able, by con-
tinuing a similar process, to prove the existence of various transfinite
ordinal types. But we shall not be able to prove the existence of all
the usual ordinal types. I do not know at what point the series begins
to be non-existent; but I cannot at present, in this method, prove the
existence of c^ or NX, which must therefore be considered for the
moment as undemonstrated.
I hope in future to work out this theory to the point where it will
appear exactly how much of mathematics it preserves, and how much it
forces us to abandon. It seems fairly clear that ordinary arithmetic,
analysis, and geometry, and, indeed, whatever does not involve the later

* The reservation is merely that the initial subject must not occur in the prototype except
in the places which we wish to be variable. For example, if our prototype is " 3 > 2 , " and our
initial subject is 3, the substitution of x for 3 gives " x > 2 . " But, if we now take 2 as our
initial subject, so that our prototype is " 2 > 2 , " the substitution of x gives iix>x," which is
not the propositional function we want.
1905.] THE THEORY OF TRANSFINITE NUMBERS AND ORDER TYPES. 47

transfinite numbers, can be stated, though in a roundabout and difficult

way, without the use of classes and relations as independent entities.
A certain amount, also, of transfinite arithmetic can be preserved ;
but it is not easy to discover how much. The theory is safe, but
drastic; and, if, in fact, there are classes and relations, it is unnecessarily
difficult and complicated. For the present, therefore, it may be accepted
as one way of avoiding contradictions, though not necessarily the
way.

III.
I come now to the second of our difficulties, namely, the doubt as
to the truth of Zermelo's axiom. This is dealt with by Dr. Hobson in
his Nos. 10 and 11, with which I find myself in complete agreement.*
All that I wish to do is to state the question in various forms, and to
point out some of its bearings. I shall assume the existence of classes
and relations, for the sake of simplicity of statement. The difficulty is
of a different kind, and is more easily apprehended by this form of state-
ment.
Zermelo's axiom asserts the possibility of picking out one from each
of the classes contained in a given class (excepting the null class). It has
hitherto been commonly assumed by mathematicians, and Zermelo has
the merit of explicitly mentioning the assumption. The axiom may be
stated as follows :—" Given any class w, there is a function f'u such that,
if u is an existentt class contained in w, then f'u is a member of u."
That is, the axiom asserts that we can find some rule by which to pick out
one term from each existent class contained in w. The axiom may also be
stated : " Given a set k of all existent classes contained in a certain class
w, there is a many-one relation B, whose domain is k, which is such that,
if u is a member of k, the term to which u has the relation B is a member
of u." The axiom can be stated in a form which does not involve classes,
functions, or relations, but I shall not give this form of statement, as its
complication makes it almost unintelligible.
A simple illustration may serve to show the nature of the difficulty as
regards this axiom, and to introduce the analogous "multiplicative axiom."
Given N0 pairs of boots, let it be required to prove that the number
of boots is even. This will be the case if all the boots can be divided into

* Though I do not agree with his special criticism of Mr. G. H. Hardy in No. 12,
according to which the second figure in Mr. Hardy's sequences " would have indefinitely great
values for all numbers fi of the second class, and thus that for sufficiently great ordinal numbers
of the second class the corresponding sequences can have no existence."
t An existent class is a class having at least one member.
4.8 MR. B. RUSSELL [Dec. 14,

two classes which are mutually similar. If now each pair has the right
and left boots different, we need only put all the right boots in one class,
and all the left boots in another: the class of right boots is similar to the
class of left boots, and our problem is solved. But, if the right and left
boots in each pair are indistinguishable, we cannot discover any property
belonging to exactly half the boots. Hence we cannot divide the boots
into two equal parts, and we cannot prove that the number of them is
even. If the number of pairs were finite, we could simply choose one out
of each pair ; but we cannot choose one out of each of an infinite number
of pairs unless we have & rule of choice, and in the present case no rule
can be found.
The problem involved in the above illustration raises grave difficulties
in regard to many elementary theorems about multiplication of cardinals.
Multiplication has been defined as follows by Mr. A. N. Whitehead :—*
Let k be a set of classes no two of which have any common terms.
Then we define the " multiplicative class of k " (denoted by X lk) as the
class formed by picking one and only one term out of each class belonging
to k, and doing this in all possible ways. That is, one member of X lk is
a class consisting of one member of each class belonging to k. Then the
number of terms in X'k is defined to be the product of the numbers
of the various classes belonging to k. This definition is perfectly
satisfactory when the number of classes which are members of k is finite,
and also when each class which is a member of k has some peculiar term
(for example, if each is given as a well-ordered series, and we can pick out
the first term). But in other cases it is not obvious that there is any rule
liy which we can pick out just one term of each member of k, and there-
tore it is not obvious that X 'k has any members at all. Hence, as far as
the definition shows, the product of an infinite number of factors none
of which is zero might be zero. Thus, in the case of the boots, we
wished to pick out one boot from each pair, but we could find no rule by
which this was to be done.
AVhat is required is not that we should actually be able to pick out one
term from each class which is a member of k, but that there should be
(whether we can specify it or not) at least one class composed of one term
from each member of k. It there is one, there must be many, unless all
the members of k are unit classes ; for, given one such class, if u is a
member of k, and x is the member of u which is picked out, we can
substitute for x any other member of u—say y—and we still have a
member of X'k. Thus the axiom we need may be stated: " Given a

American Journal of Mathematics, October, 1902.

1905.] THE THEORY OF TRANSFINITB NUMBERS AND ORDER TYPES. 49

mutually exclusive set of classes k, no one of which is null, there is at

least one class composed of one term out of each member of k."
This axiom is more special than Zermelo's axiom. It can be deduced
from Zermelo's axiom ; but the converse deduction, though it may turn out
to be possible, has not yet, so far as I know, been effected. I shall call
this the multiplicative axiom.
The multiplicative axiom has been employed constantly in proofs of
theorems concerning transfinite numbers. It is open to everybody, as yet,
to accept it as a self-evident truth, but it remains possible that it may
turn out to be capable of disproof by reductio ad absurdum. It may also,
of course, be capable of proof, but that is far less probable. A class of
classes of which this axiom holds may conveniently be called a inulti-
pliable class of classes.
The above axiom is required for identifying the two definitions of the
finite. We may define a finite cardinal number
(a) As a cardinal number which obeys mathematical induction
starting from 0 ;
(&) As a cardinal number such that any class which has that
number contains no part similar to itself.
We will for the present call any number of the kind (a) an inductive
number, and any number of the kind (6) a finite number. Then it is easy
to prove that all inductive numbers are finite; that every class whose
number is infinite contains a part whose number is N0 (where N0 is defined
as the number of inductive numbers), and vice versa; and that, if the
number of classes contained in a finite class is always finite, then all finite
numbers are inductive numbers. But, so far as I know, we cannot prove
that the number of classes contained in a finite class is always finite, or
that every finite number is an inductive number.*
The multiplicative axiom is also required for proving that the number
of terms in a sets of /3 terms is aX/3, i.e., for connecting addition and
multiplication. We cannot even prove, without this axiom, that the
number of terms in a sets of /3 terms is always the same. Similarly, we
cannot prove that the product of a factors each equal to /8 is /3a (taking

* Burali-Forti has shown that the two definitions of the finite can be identified if we assume
the following axiom:— " If u is any class of existent classes, the number of members of w is lees
than or equal to the number of members of members of « . " (" Le Classi finite," Proceedings
of t)i£ Accademia Beale delle Scienze di Torino, 1896-7.) This axiom leads at once to the result that
the number of classes contained in a finite class must be finite, whence the conclusion follows, as
above indicated. The axiom as it stands is untrue : it is necessary to assume that the classes are
mutually exclusive, or something analogous. Whether it will then give the desired result I do
not know.
SEB. 2. VOL. 4. NO. 915. E
50 MR. B. RUSSELL [Dec. 14,

Cantor's definition of exponentiation),* or even that it is always the same

number.
And in the case of the **0 pairs of boots we cannot prove that the
number of boots is N0 (i.e., N O +N O ), except in the case where we can
distinguish right and left boots.
The existence of X'A; can be proved whenever any method exists of
picking out one term from each member of k. If, for example, all terms
which are members of members of k belong to some one well-ordered
series, we get a member of X'A; by picking out the first terms of the
various members of k (k being assumed to be a set of mutually exclusive
existent classes). It does not follow that X 'k exists when every member
of k can be well ordered : for there will always be many ways of well
ordering each member of k, and we need some rule for picking out
one, in each case, of the various possible ways of well ordering each
member. That is, we need a term of the multiplicative class of the
class of which a single member is the class of relations by which
a single member of k is well ordered.
If k is any set of mutually exclusive existent classes, and if we
form another class k' by substituting for every member u of k the
class (u1) of all existent classes contained in u, then k' is a set of
mutually exclusive existent classes, and X'A;' exists, since A; is a mem-
ber of X 'k' (because each u is a member of its u').
Assuming that A; is a set of mutually exclusive existent classes, there
are certain cases in which the existence of X'A; can be proved, because
there is some structure which enables us to pick out particular terms
from members of k. Such, for example, is the following case :—
Suppose there is some one-many relation P, such that each term of
k consists of all the terms to which some term of the domain of P has
the relation P, and suppose further that every term of the domain
of P has the relation P to itself: then the domain of P is a member
of the multiplicative class of k.
If A; is a set of mutually exclusive existent classes, X 'k exists when,
and only when, there is a one-one relation S whose domain is A;, and
which relates each class u, belonging to k, to a member of u; for, when
this condition is satisfied, the converse domain of S is a member of
X lk ; and, given a term of X 'k, the relation of members of k to the
corresponding members of the given term of X'A; is an S fulfilling the
above conditions. Another way of stating the same thing is that X'A;
exists when, and only when, there is some function f'u such that,

* Math. Annalen, Vol. XLVI., § 4.

1905.] THE THEORY OF TRANSFINITE NUMBERS AND ORDER TYPES. 51

if u is a member of k, fit is a member of u; for the / ' s of the various

members of k make up a term of k. A sufficient condition for the
existence of X'k whenever k is a set of mutually exclusive existent
classes is that there should exist a function f'u such that, if u is any
existent class, f'u is a member of u. This is equivalent to Zermelo's
axiom,* and it is not, so far as I know, a necessary condition for the
existence of X lk in all such cases.
Zermelo's axiom is a generalized form of the multiplicative axiom,
and is interesting because he has shown t that, if it is true, then every
class can be well ordered. Since it is doubtful whether all classes obey
Zermelo's axiom, we may define a Zermelo class as one which does
obey the axiom; that is, w is a Zermelo class if there is at least one
many-one relation B such that the domain of B consists of all existent
classes contained in to, and if u has the relation B to x, then x is a
member of to. That is, a class w is a Zermelo class if there is a method
of correlating each existent class contained in 10 with one of its mem-
bers. Zermelo proves that any class w for which this holds can be well
ordered. The converse is obvious; for, if 10 is well ordered, we correlate
each existent class u, contained in 10, with the first term of u, which
gives a relation B satisfying the above conditions. Hence Zermelo's
axiom holds of those classes which can be well ordered, and of no others.
By applying his axiom to the class of all entities, we find that, if it
holds universally, there must be a function f'u such that, if u is any
existent class, then f'u is a member of u. Conversely, if there is such a
function, Zermelo's axiom is obviously always satisfied. Hence, if there
is a class of all entities, his axiom is equivalent to: " There is a func-
tion/'M such that, if u is any existent class, f'u is a member of «."
I think that Zermelo's axiom, applied in its functional form, and
without the assumption that there are classes or relations, leads to the
result that any propositional function only satisfied by terms of one type
is sach that all the terms satisfying it can be well ordered. If it should
appear, on other grounds, that this is not always true, it would follow
that Zermelo's axiom, in its functional form, is false. Whether or not it
is true in the form in which it applies only to classes is a question which
requires for its answer a previous decision as to what propositional
functions are predicative : the more we restrict the notion of class the
more likely this form of Zermelo's axiom is to be true, and the less

* Assuming that there is a class of all entities. But, if there is no such class, we only have
to adopt the statement of Zermelo's axiom which does not assume that there are classes,
f Math. Ann., Vol. LIX., pp. 514-516.
E 2
52 ME. B. RUSSELL [Dec. 14,

information it gives us. To discover the conditions subject to which

Zermelo's axiom and the multiplicative axiom hold would be a very im-
portant contribution to mathematics and logic, and ought not to be
beyond the powers of mathematicians.
It is easy to see that, if w is any Zermelo class, and k is a set of
mutually exclusive existent classes which between them contain all the
terms w and no more, then k is a multipliable class of classes. For
every member of k is an existent class contained in w; hence, if we
pick out one term from each existent class contained in w, we incident-
ally pick out one term from each member of k. Thus the universal
truth of Zermelo's axiom involves the universal truth of the multiplicative
axiom. The converse, so far as I know, has not been proved, and may or
may not be true.
It should be observed that, both in the case of Zermelo's axiom and in
that of the multiplicative axiom, what we are primarily in doubt about is
the existence of a norm or property such as will pick out one term from
each of our aggregates ; the doubt as to the existence of a class which
will make this selection is derivative from the doubt as to the existence
of a norm.
The problem concerned in such cases is like that of the " lawless"
decimal, which reduces to the problem of the " lawless" class of finite
integers. If we consider all the classes that can be formed of finite
integers, it seems at first sight obvious that many will consist of a
perfectly haphazard collection, not definable by any formula. But this
is open to doubt. It would seem that, as Dr. Hobson urges, an infinite
aggregate requires a norm, and that such haphazard collections as seem
conceivable are really non-entities. In the case of Cantor's " proof " that
there are more classes of finite numbers than there are finite numbers, it
is shown that no one denumerable set of formulse will cover all classes of
finite numbers; but the class shown to be left over in each case is defined
by a formula in the process of showing that it is left over. Thus this
process gives no ground for thinking that there are classes of finite
numbers which are not definable by a formula.
To sum up: there are two analogous axioms—Zermelo's and the
multiplicative axiom—which have been habitually employed by mathe-
maticians in reasonings about the transfinite, but which, most likely,
are not true without some restriction. Without them, we cannot, so far
as at present appears, identify the two definitions of the finite, or establish
the usual relations of addition, multiplication, and exponentiation. If
Zermelo's axiom were true, every class would be well ordered, and also, I
think, every aggregate of terms possessing some property. But in this
1905.] THE THEORY OF TRANSFINITE NUMBERS AND ORDER TYPES. 53

respect the problem considered in our second part is dependent upon that
considered in our first part.
The general position advocated in the foregoing paper may be briefly
stated as follows : —
When we say that a number of objects all have a certain property, we
naturally suppose that the property is a definite object, which can be
considered apart from any or all of the objects which have, or may be
supposed to have, the property in question. We also naturally suppose
that the objects which have the property form a class, and that the class
is in some sense a new single entity, distinct, in general, from each
member of the class. Both these natural suppositions can be proved, by
arguments so short and simple that they scarcely admit a possibility of
error, to be at any rate not universally true. We may, in view of this
fact, adopt one of two courses: we may either decide that the assumptions
in question are always false, or endeavour to find conditions subject to
which they are true, these conditions being such as to exclude the cases
where the falsehood of the assumptions can be proved. The latter course
has the advantage of being more consistent with common sense, and of
preserving more of Cantor's work ; but it has, as yet, the disadvantage of
great uncertainty and artificiality in detail, owing to the absence of any
broad principle by which to decide aa to which functions are predicative.
The former course, in practice, merely involves abstaining from the
doubtful assumptions, and does not commit us to the view that they are
false; it is therefore, so long as any doubt subsists, the prudent plan to
pursue the former course as far as possible. It appears on examination
that, without supposing either of the suspected assumptions to be ever
true, we can construct ordinary mathematics and most of the theory of
the transfinite ; and in this development we meet with no contradictions,
so far as is known at present. Whether it is possible to rescue more of
Cantor's work must probably remain doubtful until the fundamental
logical notions employed are more thoroughly understood. And whether,
in particular, Zermelo's axiom is true or false is a question which, while
more fundamental matters are in doubt, is very likely to remain un-
answered. The complete solution of our difficulties, we may surmise, is
more likely to come from clearer notions in logic than from the technical
advance of mathematics ; but until the solution is found we cannot be
sure how much of mathematics it will leave intact.

[Note added February 5th, 1906.—From further investigation I now

feel hardly any doubt that the no-classes theory affords the complete
solution of all the difficulties stated in the first section of this paper.]