106
$\begingroup$

I've heard of some other paradoxes involving sets (i.e., "the set of all sets that do not contain themselves") and I understand how paradoxes arise from them. But this one I do not understand.

Why is "the set of all sets" a paradox? It seems like it would be fine, to me. There is nothing paradoxical about a set containing itself.

Is it something that arises from the "rules of sets" that are involved in more rigorous set theory?

$\endgroup$
3
  • 27
    $\begingroup$ Intuitively, you can put many things in a basket -- but one thing you can never put in that basket is the basket itself. $\endgroup$
    – littleO
    Aug 26, 2014 at 14:55
  • 5
    $\begingroup$ @AirMike: Please avoid making trivial edits to old posts. Let alone stylistic edits like this (which, frankly, doesn't make any sense). $\endgroup$
    – Asaf Karagila
    Sep 20, 2020 at 19:16
  • $\begingroup$ @littleO the set of all sets, though, would not be akin to a basket. A more appropriate concept might be a conception of God as being all things. In a sense God thus /contains/ all things: the set of all sets, but is not /larger/ than it... Applying, for example, 2 to the power God would presuppose that in the new paradigm we are working in 2 could even be conceived of as being raised to the power God, there was a higher power which was hithertofore unconceived... $\endgroup$ Oct 19, 2021 at 7:38

13 Answers 13

109
$\begingroup$

Let $|S|$ be the cardinality of $S$. We know that $|S| < |2^S|$, which can be proven with generalized Cantor's diagonal argument.


Theorem

The set of all sets does not exist.

Proof

Let $S$ be the set of all sets, then $|S| < |2^S|$. But $2^S$ is a subset of $S$. Therefore $|2^S| \leq |S|$. A contradiction. Therefore the set of all sets does not exist.

$\endgroup$
28
  • 15
    $\begingroup$ What does 2^S mean? $\endgroup$
    – Casebash
    Jul 21, 2010 at 0:51
  • 20
    $\begingroup$ @Casebash: 2^S is standard notation for the power set of S: That is, the set of all subsets of S. If S were {1, 2}, 2^S would be {{}, {1}, {2}, {1, 2}}. $\endgroup$
    – Larry Wang
    Jul 21, 2010 at 0:58
  • 9
    $\begingroup$ @Seamus Here S is defined as the set of all sets. Every element of 2^S is a set, implies they are also elements of S. $\endgroup$
    – Chao Xu
    Jul 21, 2010 at 12:22
  • 10
    $\begingroup$ How is that incorrect. We always consider the standard framework when we are not presented other ones. $\endgroup$
    – Chao Xu
    Mar 6, 2011 at 7:55
  • 8
    $\begingroup$ @GeorgeChen The notion of cardinality is not at fault here. You can carry out the proof without talking about cardinal numbers as follows: Observe that there is an injective function $S\to 2^S$ given by $x \mapsto \{x\}$. Observe that there is an injective function $2^S \to S$ given by $x\mapsto x$, since $S$ is the set of all sets. Now use the Cantor-Schroeder-Bernstein argument to show that there is a bijection between $2^S$ and $S$. Finish with Cantor's diagonal argument showing that the existence of such a bijection leads to a contradiction. $\endgroup$ Aug 26, 2014 at 16:56
70
$\begingroup$

Just by itself the notion of a universal set is not paradoxical.

It becomes paradoxical when you add the assumption that whenever $\varphi(x)$ is a formula, and $A$ is a preexisting set, then $\{x\in A\mid \varphi(x)\}$ is a set as well.

This is known as bounded comprehension, or separation. The full notion of comprehension was shown to be inconsistent by Russell's paradox. But this version is not so strikingly paradoxical. It is part of many of the modern axiomatizations of set theory, which have yet to be shown inconsistent.

We can show that assuming separation holds, the proof of the Russell paradox really translates to the following thing: If $A$ is a set, then there is a subset of $A$ which is not an element of $A$.

In the presence of a universal set this leads to an outright contradiction, because this subset should be an element of the set of all sets, but it cannot be.

But we may choose to restrict the formulas which can be used in this schema of axioms. Namely, we can say "not every formula should define a subset!", and that's okay. Quine defined a set theory called New Foundations, in which we limit these formulas in a way which allows a universal set to exist. Making it consistent to have the set of all sets, if we agree to restrict other parts of our set theory.

The problem is that the restrictions given by Quine are much harder to work with naively and intuitively. So we prefer to keep the full bounded comprehension schema, in which case the set of all set cannot exist for the reasons above.

While we are at it, perhaps it should be mentioned that the Cantor paradox, the fact that the power set of a universal set must be strictly larger, also fails in Quine's New Foundation for the same reasons. The proof of Cantor's theorem that the power set is strictly larger simply does not go through without using "forbidden" formulas in the process.

Not to mention that the Cantor paradox fails if we do not assume the power set axiom, namely it might be that not all sets have a power set. So if the universal set does not have a power set, there is no problem in terms of cardinality.

But again, we are taught from the start that these properties should hold for sets, and therefore they seem very natural to us. So the notion of a universal set is paradoxical for us, for that very reason. We are educated with a bias against universal sets. If you were taught that not all sets should have a power set, or that not all sub-collections of a set which are defined by a formula are sets themselves, then neither solution would be problematic. And maybe even you'd find it strange to think of a set theory without a universal set!

$\endgroup$
8
  • $\begingroup$ Can you expand a bit on "If A is a set, then there is a subset of A which is not an element of A." should it not be "If A is a set, then there is a set of subsets of A which is not an element of A."so suppose A = { 1,2,3,{1,2}} then i mean the set { {1,3} , {2,3} , {1, {1,2}} , {3,{1,2}} } $\endgroup$
    – Willemien
    Jan 17, 2015 at 15:30
  • 1
    $\begingroup$ sorry I am just trying to understand it, could you give a me the subset you mean given the set A I gave? $\endgroup$
    – Willemien
    Jan 17, 2015 at 15:45
  • 2
    $\begingroup$ Well, $A$ itself is that subset in the case of $A=\{1,2,3,\{1,2\}\}$. It is true, $\varnothing$ could also work, and so could any singleton, or so on. But given any set, $A$ the set $\{x\in A\mid x\notin x\}$ is either not a set, or not an element of $A$. In presence of comprehension axioms which do not restrict the formulas, it is always a set, and therefore not an element of $A$. In Quine's New Foundations, and other more "type theoretic" set theories, we restrict the formulas that can define a set, which mean it might not be a set at all. $\endgroup$
    – Asaf Karagila
    Jan 17, 2015 at 15:54
  • 1
    $\begingroup$ In all the sets that you can write down, this set will be $A$ itself, and $A\notin A$. This is because naively we cannot construct a set $x$ such that $x\in x$ (otherwise we can naively disprove the axiom of foundation; but that's not true). If $B=\{x\in A\mid x\notin x\}\in A$, then we ask either $B\in B$ or $B\notin B$. If $B\in B$, then $B\in A$ and $B\in B$, so $B\notin B$; and if $B\notin B$ then $B\in A$ and $B\notin B$, so $B\in B$. Either case a contradiction. So the only solution is that $B\notin A$. The question is whether or not $\{x\in A\mid x\notin x\}$ is even a set. [...] $\endgroup$
    – Asaf Karagila
    Jan 17, 2015 at 17:26
  • 1
    $\begingroup$ [...] In the presence of separation axioms, like the ones in the $\sf ZF$ system, it immediately follows that $\{x\in A\mid x\notin x\}$ is a set, and therefore it has to be that case that this is a subset of $A$ which is not an element of $A$. In set theories like Quine's New Foundations, this is not a set since the formula is not stratified and therefore doesn't have to define a set. It follows, therefore, that in $\sf ZF$ there is no set of all sets, since $\sf ZF$ proves that given a set, it doesn't contain all sets. $\endgroup$
    – Asaf Karagila
    Jan 17, 2015 at 17:28
40
$\begingroup$

The existence of universal set is incompatible with the Zermelo–Fraenkel axioms of set theory. However, there are alternative set theories which admit a universal set. One such theory is Quine's New Foundations.

$\endgroup$
3
  • 4
    $\begingroup$ Possibly also useful: Elementary Set Theory With a Universal Set, by M. Randall Holmes. $\endgroup$
    – MJD
    Mar 18, 2012 at 18:12
  • $\begingroup$ Late to the party, but how does this relate to Cantor's proof about cardinals? $\endgroup$
    – Guido
    Mar 10, 2017 at 23:56
  • $\begingroup$ The stratified formulas used in New Foundations prevent the direct comparison of a set's size with that of its power set. If you try to make a comparison, the best you can do is that the size of the set of all singletons of S is less than the size of P(S), so there are fewer singletons than there are sets of any size. Interestingly, NBG en.wikipedia.org/wiki/Von_Neumann-Bernays-Gödel_set_theory states that the proper class of singletons is equinumerous to the proper class of all sets, which is the opposite of what NFU says. $\endgroup$
    – Zemyla
    Jun 17, 2019 at 11:18
27
$\begingroup$

The "set of all sets" is not so much a paradox in itself as something that inevitably leads to a contradiction, namely the well-known (and referenced in the question) Russell's paradox.

Given any set and a predicate applying to sets, the set of all things satisfying the predicate should be a subset of the original set. If the "set of all sets" were to exist, because self-containment and non-self-containment are valid predicates, the set of all sets not containing themselves would have to exist as a set in order for our set theory to be consistent. But this "set of all sets" cannot exist in a consistent set theory because of the Russel paradox.

So the non-existence of the "set of all sets" is a consequence of the fact that presuming it's existence would lead to the contradiction described by Russel's paradox.

This is in fact the origin of Russel's paradox.

In his work "The Basic Laws of Arithmetic", Gottlob Frege had taken as a postulate the existence of this "set of all sets". In a letter to Frege, Bertrand Russell essentially blew away the basis of Frege's entire work by describing the paradox and proving that this postulate could not be a part of a consistent set theory.

$\endgroup$
10
  • 3
    $\begingroup$ If the "set of all sets" were to exist, because self-containment and non-self-containment are valid predicates, the set of all sets not containing themselves would have to exist as a set in order for our set theory to be consistent - Not true, as Quine's New Foundations, cited by François, shows. It has a universal set, but not the Russell set. $\endgroup$ Jul 21, 2010 at 11:05
  • 1
    $\begingroup$ I guess this depends on what one considers a "set theory". The usual image of the universe of sets, given by the cumulative hierarchy, doesn't allow for the set of all sets. $\endgroup$ Mar 8, 2012 at 9:13
  • 3
    $\begingroup$ I don't see how you could not consider Quine's New Foundations a set theory; it has the usual universe of sets in it. It may in fact have a superset (superclass?) of ZF's sets, though it's possible stratified comprehension excludes some sets ZF includes. It is akin to NBG set theory; some definite weirdness if you're used to ZFC, but basically the same thing for practical purposes. $\endgroup$
    – prosfilaes
    Oct 17, 2012 at 6:48
  • 1
    $\begingroup$ NF doesn't necessarily have all of ZF's sets in it. While it's consistent that the von Neumann version of $\omega$ exists in NF, I believe it's not provable that it does. $\endgroup$ Jun 23, 2014 at 10:12
  • 1
    $\begingroup$ Stratification is the key difference. PM has universal set too, but its members are typically homogeneous. They are all from the type next below. When OP said the set of all set, his universe was flat. $\endgroup$ Aug 31, 2014 at 1:30
15
$\begingroup$

Russel's paradox arises if you consider the set $U=\left\{x:x\not\in x\right\}$. Ask yourself if $U\in U$. If you suppose so, then by the definition of unrestricted set comprehension $U\not\in U$. You have a contradiction, so it must be the opposite of what you supposed, that is, $U\not\in U$. But this is the same as saying $U$ belongs to the complement of itself, that is, $U\in U$. You now have another contradiction, but this is far worse, since you have no hypotheses. The whole theory is logically inconsistent.

In set theory there are two ways for getting rid of the Russel's paradox: either you disallow the set of all sets and other similar sets (see for example the Zermelo-Fraenkel set theory), or you allow them, but you also restrict the way they are used (see for example the Morse-Kelley set theory).

In the first case, set comprehension says if you have a set $A$ you can have $\left\{x\in A:\phi\left(x\right)\right\}$ (notice: writing $\left\{x:\phi\left(x\right)\right\}$ is just wrong in this case, because you have to have an initial set). If you now define $U=\left\{x\in A:x\not\in x\right\}$ and you repeat the same passages as before, it only follows that $U\not\in A$. There's no contradiction and the theory is consistent.

In the second case, you consider classes, not just sets. Sets are classes that belong to some other class, while proper classes are classes that belong to no class. Set comprehension, in this case, says you can have $\left\{x:\phi\left(x\right)\right\}$, but all its members are sets by definition. If try to reproduce Russel's paradox, you get that $U\not\in U$. If you then suppose that $U$ is a set, then you have a contradiction, so $U$ must be a proper class. This is all you get. No contradictions. The theory is consistent.

$\endgroup$
3
  • $\begingroup$ There are other ways to avoid Russell's paradox, too. For example, the axiom schema of comprehension in NF doesn't allow you to construct $\{x: x \notin x\}$ (or, AIUI, even $\{x \in A: x \notin x\}$ in general) because $x \notin x$ is not a stratified formula. It does let you construct the universal set $\{x: x=x\}$, though. $\endgroup$ Mar 8, 2012 at 8:57
  • $\begingroup$ Another way, don't assume LEM on $\in$. $\endgroup$
    – DanielV
    Mar 10, 2018 at 1:48
  • $\begingroup$ Or you could forbid ∉ and other negations in the axiom of comprehension, like in positive set theory. The universal set exists in this theory, as does { x : x ∈ x }, but not { x : x ∉ x }. $\endgroup$
    – Zemyla
    Oct 11, 2020 at 3:49
9
$\begingroup$

The set of all sets is not a paradox.

When set theory was invented, it was assumed that given any predicate, there was a set containing all the things that satisfied the predicate. This assumption is called naive comprehension. Unfortunately, this allowed paradoxes like the set of all sets not containing themselves.

So people invented restricted axioms of comprehension. These are rules that say that only certain predicates give rise to sets. There are different kinds of set theory with different comprehension restrictions.

One idea for limiting comprehension is to say that if the class of things satisfies the predicate is too big, the class is not a set. All the commonly used set theories use this idea. The set of all sets is very large indeed, so it is not a set in these theories.

There are other ways of restricting comprehension that get rid of the paradoxical sets but do allow a universal set. One such set theory was invented by Quine and called New Foundations. There is a book by Holmes, Set Theory with a Universal Set.

$\endgroup$
1
  • $\begingroup$ That is essentially what I wrote in my answer. $\endgroup$
    – Asaf Karagila
    May 5, 2015 at 3:57
6
$\begingroup$

Perhaps the following image will help take from here:

enter image description here

$\endgroup$
5
$\begingroup$

Some of the comments in the previous answers make a subtle mistake, and I think it may be worth clarifying some issues. I am assuming the standard sort of set theory in what follows.

Cantor's diagonal theorem (mentioned in some of the answers) gives us that for any set $X$, $|X|<|\mathcal P(X)|$. Unlike what some comments claim, this really has nothing to do with cardinalities. All it says is that no map $f\!:X\to\mathcal P(X)$ is onto. The usual proof proceeds by noting that $A=\{x\in X:x\notin f(x)\}$ is not in the range of $f$, because if $A=f(a)$, then $a\in A$ if and only if $a\notin f(a)=A$.

The usual argument for Russell's paradox (also mentioned in some of the answers) proceeds by considering $A=\{a:a\notin a\}$. If $V$ is a set, $A$ would be a set as well (by comprehension, if you wish), and we reach a contradiction by noting that $A\in A$ if and only if $A\notin A$.

I think it is misleading to think (as some of the comments suggest) that the two proofs are (fundamentally) different. They are essentially the same.

The point is that if there is a set of all sets (let's call it $V$), then $\mathcal P(V)$ is a set as well (by comprehension) and in fact $\mathcal P(V)=V$ because, on the one hand, any subset of $V$ is a set, and therefore a member of $V$, and on the other hand, any member of $V$ is itself a subset of $V$ (since the members of any set are sets themselves), and therefore a member of $\mathcal P(V)$.

Now, the identity function is a map from $V$ to $\mathcal P(V)$. Let's call it $f$. The set not in the range of $f$ given to us by the standard proof of Cantor's diagonal theorem recalled above is $A=\{x\in V:x\notin f(x)\}$, which in this case reduces to $\{x:x\notin x\}$, which in turn is precisely the set given by the standard proof of Russell's paradox.

[What to make of this result from a foundational point of view is another matter. In $\mathsf{ZF}$ the conclusion is just that there is no set of all sets, although it is perhaps more accurate to say that the result is used to justify dismissing unrestricted comprehension and adopting instead the version of bounded comprehension used in $\mathsf{ZF}$. In $\mathsf{NF}$ the solution is instead to limit unbounded comprehension by only allowing stratified instances. Other foundational solutions may and have been adopted as well.]

$\endgroup$
5
$\begingroup$

This turns out to be the same as @donroby's answer. The scheme is inspired by the proof of Cantor's theorem, the use of which is implicit in the accepted answer.

As the following example shows, if all is construed as commonsense does, the set of all sets contains a member that simultaneously is and is not a member of itself.

Let $T = \{ \alpha \ |\ \alpha $ is a set and $ \alpha \notin \alpha \}$,

$T \in T \Rightarrow T \notin T$ --- (1)

$T \notin T \Rightarrow T \in T$ --- (2)

(1).(2) $\Rightarrow$ Paradox.

Note: This answer dispenses with such higher notions as "power set," "cardinality" and "less than" whose properties cannot be ascertained before the properties of set are ascertained.

Further analysis leads to Type Theory which uses meaning as a guarding criterion. In Type Theory, all sets, as commonsense understands it, is meaningless; propositions like $T\in T$ and $T \notin T$ are also meaningless.

$\endgroup$
5
  • $\begingroup$ @Elliott - Cleaned up. Thanks for the feedback. $\endgroup$ Oct 16, 2020 at 14:28
  • 2
    $\begingroup$ You're welcome. I read lots of the answers here, but it wasn't until I read yours that I really understood it - so thank you. =) $\endgroup$
    – Elliott
    Oct 16, 2020 at 18:42
  • 2
    $\begingroup$ Re: the last paragraph, type theory isn't the only place further analysis can lead to. $\mathsf{ZFC}$-style set theory is also a response, in which "$T\in T$" is meaningful but doesn't hold of any object, as are things like $\mathsf{NF(U)}$ or positive set theories such as $\mathsf{GPK^+_\infty}$ in which there is a universal set (and hence some sets do contain themselves!). $\endgroup$ Oct 16, 2020 at 19:47
  • $\begingroup$ Different people will prefer different approaches. For myself, ultimately the $\mathsf{ZFC}$-style approach seems most faithful to intuition. $\endgroup$ Oct 16, 2020 at 19:49
  • $\begingroup$ @NoahSchweber - Sure I don't think this is a good answer now but I can't recall what epiphany I had in mind back then. There is a more direct reason against the conception of the set of all sets: a totality is not determined until each of its constituents are determined; if one of the constituents is the totality itself, a vicious circle ensues. A set of all sets contains itself and thus involves a vicious circle. $\endgroup$ Oct 17, 2020 at 0:14
2
$\begingroup$

An informal explanation is Russel's Paradox. The wiki page is informative, here's the relevant quote:

Let us call a set "abnormal" if it is a member of itself, and "normal" otherwise. For example, take the set of all squares. That set is not itself a square, and therefore is not a member of the set of all squares. So it is "normal". On the other hand, if we take the complementary set that contains all non-squares, that set is itself not a square and so should be one of its own members. It is "abnormal".

Now we consider the set of all normal sets, R. Attempting to determine whether R is normal or abnormal is impossible: If R were a normal set, it would be contained in the set of normal sets (itself), and therefore be abnormal; and if it were abnormal, it would not be contained in the set of normal sets (itself), and therefore be normal. This leads to the conclusion that R is both normal and abnormal: Russell's paradox.

$\endgroup$
3
  • $\begingroup$ I mentioned this paradox in my question; it however is not the same one that I am asking about. $\endgroup$
    – Justin L.
    Jul 20, 2010 at 23:17
  • 3
    $\begingroup$ You're right, sorry. Probably you can say that, since "the set of all sets" is not so well-defined, it must necessarily contain the set from Russel's paradox. Since the set from Russel's Paradox can't exist, neither can any set that contains it. But I feel this is cheating somehow. $\endgroup$
    – Edan Maor
    Jul 20, 2010 at 23:37
  • $\begingroup$ That's not cheating, that's exactly correct. I always use Russel's Paradox for the "set of all sets" question. The main point in both cases anyway is remembering that Cantor's definition of a set is a little to vague if you dig deep enough. $\endgroup$
    – balpha
    Jul 21, 2010 at 5:10
2
$\begingroup$

"Why is "the set of all sets" a paradox? It seems like it would be fine, to me."

The lesson of this and other set theory paradoxes is that when you define a set in a proof, you are obliged to demonstrate that the set exists and is well-defined. To me, a paradox of Russell's Paradox is that he fell into this trap since he pretty much wore himself out trying to generate a rigorous foundation for Mathematics.

$\endgroup$
1
  • 2
    $\begingroup$ The set of all sets is well-defined in the set theory Cantor used: to any logical predicate there is a set of all objects satisfying it. The problem is that Cantor's set theory is inconsistent. $\endgroup$
    – user14972
    Mar 8, 2012 at 9:36
2
$\begingroup$

How about a set that contains everything with the exception of itself?

$\endgroup$
6
  • 2
    $\begingroup$ You can still make the cardinality argument that the set cannot contain all its subsets. $\endgroup$ Mar 8, 2012 at 9:11
  • $\begingroup$ Insightful! @MichaelGreinecker. In type theory, the set of all the subsets resides on upstairs. $\endgroup$ Aug 22, 2014 at 23:50
  • $\begingroup$ A set cannot have any of its subset as members because a subset of a set involves the totality also, and therefore must not be a member of the larger set; otherwise there will be a vicious circle. $\endgroup$ Aug 23, 2014 at 0:42
  • 2
    $\begingroup$ @GeorgeChen There is something in set theory called a transitive set, and for such a set, every element is a subset. Here is an example: $\{\emptyset\}$. $\endgroup$ Aug 23, 2014 at 7:25
  • 1
    $\begingroup$ Suppose $S$ is such a set. Then $S \cup \{ S \}$ is the set of all sets. $\endgroup$
    – user14972
    Aug 27, 2014 at 9:10
0
$\begingroup$

I think this paradoxon is not so complicated and can be explained simply.

If you put all existing sets S(1) to S(k) together in a set S(all), you create a new set S(k+1)={S(1),...,S(k)} and this set has to be in the set of all sets as well (because it should contain all sets). Thus, you put the new set S(k+1) in the set of all sets S(all). But this in turn creates a new set S(k+2)={S(1),...,S(k+1)} wich has to be inserted in the set of all sets as well. Putting S(k+2) in S(all), however, creates a new set and so on and so on.

In conclusion, everytime you put the new set into the set of all sets you create a new set. As a consequence, you can never build a set S(all) that contains all sets without creating a set which is not contained in S(all).

$\endgroup$
5
  • 1
    $\begingroup$ Building a set inductively like this is not paradoxical. The natural numbers are built in exactly this way starting with the empty set. You can't talk about 'putting a set in' to the set of all sets. $\endgroup$
    – Dan Rust
    May 3, 2015 at 22:56
  • $\begingroup$ The building process is not the paradoxon, but it shows that a set cannot contains all sets without producing a new set that is not contained in this set and hence a set that is assumed to contain all sets actually does not contain all sets. The latter is the paradoxon. $\endgroup$
    – user237596
    May 4, 2015 at 1:05
  • $\begingroup$ But ok, I am not a mathematician and this solution was the first that comes to my mind. Maybe, it is completely bullshit ;-) $\endgroup$
    – user237596
    May 4, 2015 at 1:11
  • $\begingroup$ I am just curious. You mentioned the natural numbers. In my opinion, you have a similar case if you say there is a natural number that is maximal. You can simply show that this number cannot be maximal because you can add +1 to this number. Isn't this case similar? $\endgroup$
    – user237596
    May 4, 2015 at 1:16
  • $\begingroup$ with "there is a natural number that is maximal" I meant the assumption that there is a natural number that is greater than all other natural numbers. 1) set that contains all set => this set is larger than all other sets 2) maximal number => this number is greater than all other natural numbers. I see there some similarities. $\endgroup$
    – user237596
    May 4, 2015 at 2:23

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .