46

How do I show that every partial order can be extended to a linear ordering?

I think that I manage to prove that claim for finite set, how can I prove it for infinite set?

Thank you.

17SI.34SA
  • 2,115
  • 2
  • 19
  • 28

3 Answers3

47

You have to use the axiom of choice in the infinite case. And this usage is essential because it is consistent that there are sets which cannot be linearly ordered (in which the case the identity relation cannot be extended to a linear order).

Let $(P,R)$ be a partially ordered set. We define the collection $\{\leq\mid (P,\leq)\text{ is a poset and }R\subseteq\leq\}$, order this by $\subseteq$ and show it satisfies the conditions of Zorn's lemma. The maximal element has to be a linear order.


It should be remarked that simply requiring that every partially ordered set can be extended to a linear order is strictly weaker than the axiom of choice.

Here is a very neat trick using the Compactness Theorem of first-order logic (which is weaker than the axiom of choice). Although you probably didn't cover that in the course, it's a nice trick to know:

Let $(P,\leq)$ be a partially ordered set. Let $\cal L$ be the language with one binary relation $\preceq$ and a constant symbol $c_p$ for every $p\in P$. We write the following $T$:

  1. $\preceq$ is a partial order;
  2. whenever $p\leq q$ we add the axiom $c_p\preceq c_q$;
  3. for every pair $p,q\in P$ we add the axiom $c_p\preceq c_q\lor c_q\preceq c_p$.

Now we will show that $T$ is finitely-satisfiable. If $S\subseteq T$ has only axioms of the form $1$ or from the schema described in $2$ then $(P,\leq)$ is a model for $S$. Otherwise it has a finite number of axioms from the schema in $3$. We can extend $\leq$ to some $\leq^\ast$ which satisfy these axioms.

We do this by induction: if there is only one axiom and it hasn't been satisfied then $p$ and $q$ are incomparable, and we can take $\leq^\ast=\leq\cup\{\langle p,q'\rangle\mid q\leq q'\}$.

Suppose we did this for $k$ axioms; simply repeat the argument for the $k+1$. This requires no axiom of choice because we only have to make finitely many choices.

(Alternatively one can replace schema $3$ by the axiom $3'$, $\forall x\forall y(x\preceq y\lor y\preceq x)$, now notice that any finite fragment talks only about finitely many constants, $c_{p_1},\ldots,c_{p_k}$. Restricting $\leq$ to $\{p_1,\ldots,p_k\}$ is a finite partially ordered set, and can be extended to a linear ordering which satisfies $3'$ as wanted.)

Therefore every finite fragment of $T$ has a model, and so by the compactness theorem $T$ itself has a model $(P',\leq')$. We define $\preceq$ on $P$ as follows: $p\preceq q\iff (P',\leq')\models c_p\preceq c_q$. The schema $2$ ensures this extends $\leq$, and schema $3$ (or axiom $3'$) ensures this is a linear ordering.

Asaf Karagila
  • 405,794
  • 4
    I think that as an extension of $\leq$ we should take $\mathord{\leq^\ast}=\mathord{\leq}\cup{\langle p,q'\rangle\mid q\leq q'}\cup{\langle p',q\rangle\mid p'\leq p}$ to guarantee transitivity of $\leq^\ast$. – Rafał Gruszczyński Aug 20 '20 at 11:54
  • Yes, you're right. – Asaf Karagila Aug 20 '20 at 12:05
  • When we have that p and q are incomparable what is the set $\leq^\ast=\leq\cup{\langle p,q'\rangle\mid p\leq q'}\cup{\langle p',q\rangle\mid p'\leq q}$. Does ${\langle p,q'\rangle\mid p\leq q'}$ represent all the elements less than or equal to $q$ ? Wouldn't we just have to append the tuple $(p, q)$ to the relation to fix it? – cuppajoeman Oct 27 '21 at 21:55
  • @cuppajoeman: Not necessarily. Consider the order which is two disjoint chains, each with three elements. Now say that $p$ and $q$ are the two middle ones. Say $p$ will be less than $q$. But we need to also add the minimal below $p$ to be less than $q$, and likewise the maximal above $q$ to be above $p$. – Asaf Karagila Oct 27 '21 at 21:59
  • So we have two disjoint chains, say ${x, y, z} \supseteq {x, y} \supseteq {x}$, and also ${a, b, c } \supseteq {a, b} \supseteq {a}$ so now we set $p = {x, y}$ and $q = { a, b }$, and in hopes to create a valid relation we add $(p, q)$ to our relation $\le^\ast$. I guess as Mad Hatter says, we would also need ${x } \le^\ast {a, b, c}$ due to transitivity, so their additions make sense. In this case would we not need: $\leq^\ast=\leq\cup{\langle p,q'\rangle\mid q\leq q'}\cup {\langle p', q\rangle\mid p'\leq q}\cup\langle p, q \rangle$ to be our relation?(appended tuple) – cuppajoeman Oct 28 '21 at 00:21
  • @cuppajoeman: But you'd also need to add ${x, y}\leq^*{a,b,c}$. – Asaf Karagila Oct 28 '21 at 00:24
  • Yes, exactly, we need it for all the ones above and all the ones below. I understand it now, but we still do need that extra tuple $\langle p,q \rangle$ to be added as well? – cuppajoeman Oct 28 '21 at 00:26
  • 1
    @cuppajoeman: That depends on how you write it out exactly. But yes, of course you need it in $\leq^*$. – Asaf Karagila Oct 28 '21 at 00:32
  • Don't we even need to take $\leq^\ast=\leq\cup{\langle p',q'\rangle\mid q\leq q' \text{ and } p'\leq p}$? For example, if we have ${a_1, a_2, a_3, b_1, b_2, b_3}$ with $a_1 \leq a_2 \leq a_3$ and $b_1 \leq b_2 \leq b_3$ and ${a_1, a_2, a_3}$ incomparable with ${b_1, b_2, b_3}$ and then we introduce the new relation $a_2 \leq b_2$, we need to introduce all of the following relations: $a_1 \leq b_2$, $a_1 \leq b_3$, $a_2 \leq b_2$, $a_2 \leq b_3$. I particular, I don't think either of the previous approaches include the relation $a_1 \leq b_3$. – Smiley1000 Dec 03 '23 at 20:01
21

You will have to use the axiom of choice (or one of its consequences) in some form. One possibility is to use Zorn’s lemma. Let $\langle P,\le\rangle$ be a partial order. Let $$\mathscr{L}=\big\{\langle X,\preceq_X\rangle:X\subseteq P\text{ and }\preceq_X\text{ is a linear order on }X\\\text{ extending }\le\upharpoonright(X\times X)\}\;.$$

For $\langle X,\preceq_X\rangle,\langle Y,\preceq_Y\rangle\in\mathscr{L}$ define $\langle X,\preceq_X\rangle\sqsubseteq\langle Y,\preceq_Y\rangle$ iff $X\subseteq Y$ and $\preceq_Y\upharpoonright(X\times X)=\preceq_X$; clearly $\langle\mathscr{L},\sqsubseteq\rangle$ is a partial order. Let $\mathscr{C}$ be a chain in $\langle\mathscr{L},\sqsubseteq\rangle$.

  • Show that $\mathscr{C}$ has an upper bound in $\langle\mathscr{L},\sqsubseteq\rangle$; the most obvious choice works.

  • Use Zorn’s lemma to conclude that $\langle\mathscr{L},\sqsubseteq\rangle$ has a maximal element $\langle M,\preceq_M\rangle$.

  • Show that $M=P$.

Added: However, it’s not necessary to use the full axiom of choice: the result follows, for instance, from the Boolean prime ideal theorem, which is strictly weaker than AC. For example, one of the equivalents of the BPI is the compactness theorem for first-order logic, which can be used to give a very simple proof. Let $R$ be a binary relation symbol, and for each $p\in P$ let $c_p$ be a constant symbol. Let $\Phi$ be the set of all sentences $R(c_p,c_q)$ such that $p,q\in P$ and $p\le q$, together with axioms making $R$ a linear order. Then $\Phi$ is finitely satisfiable, so by the compactness theorem $\Phi$ has a model. That model yields a linear extension of $\le$.

Added2: Okay, I like ultrafilters. For $F\in[P]^{<\omega}$ let $U_F=\big\{G\in[P]^{<\omega}:F\subseteq G\big\}$; clearly $\big\{U_F:F\in[P]^{<\omega}\big\}$ is centred (= has the finite intersection property), so it can be extended to an ultrafilter $\mathscr{U}$ on $[P]^{<\omega}$. This requires the ultrafilter extension theorem, which is equivalent to the BPI, but the next step uses AC itself: for each finite $F\subseteq P$ let $\preceq_F$ be a linear order on $F$ that extends $\le\upharpoonright\!\!(F\times F)$, let $\le_F\,=\,\le\cup\preceq_F$, and denote by $P_F$ the partial order $\langle P,\le_F\rangle$.

For $p,q\in P$ write $p\preceq q$ iff $\big\{F\in[P]^{<\omega}:p\le_F q\big\}\in\mathscr{U}$. If $p\le q$, then $p\le_F q$ for all $F\in[P]^{<\omega}$, so $p\preceq q$. Now let $p,q\in P$ with $p\ne q$. Then exactly one of the sets $\big\{G\in U_{\{p,q\}}:p\le_G q\big\}$ and $\big\{G\in U_{\{p,q\}}:q\le_G p\big\}$ belongs to $\mathscr{U}$, so exactly one of $p\preceq q$ and $q\preceq p$ holds. Finally, if $p,q,r\in P$ with $p\preceq q\preceq r$, then

$$\big\{G\in U_{\{p,q,r\}}:p\le_G r\big\}\supseteq\\\big\{G\in U_{\{p,q,r\}}:p\le_G q\big\}\cap\big\{G\in U_{\{p,q,r\}}:q\le_G r\big\}\in\mathscr{U}\;,$$

so $p\preceq r$. Thus, $\preceq$ is a linear order on $P$ extending $\le$.

Brian M. Scott
  • 631,399
  • The last part is not necessarily true, you can write one axiom asserting that $R$ is a linear order. In order to use the compactness theorem you should add a schema writing explicitly for every pair of elements that they are comparable. – Asaf Karagila Jan 05 '13 at 16:14
  • @Asaf: That’s part of the axiom making $R$ a linear order. (Or wasn’t it clear that I meant a linear order of the ‘universe’?) – Brian M. Scott Jan 05 '13 at 16:23
  • $\forall x\forall y(x\mathrel{R}y\lor y\mathrel{R}x)$ is just one axiom. How can you show that the theory $T$ is finitely satisfiable if your original partial order is not linear? – Asaf Karagila Jan 05 '13 at 16:25
  • @Asaf: That part’s trivial, since we already know that finite partial orders can be extended to linear orders. – Brian M. Scott Jan 05 '13 at 16:27
  • Oh right. I forgot about that when I wrote my revised answer! I had to work so much harder :-) – Asaf Karagila Jan 05 '13 at 16:28
  • I think you meant that $\prec_X$ is a partial order, not a linear order. Otherwise your proof is circular. – Asaf Karagila Dec 12 '20 at 16:49
  • @AsafKaragila: You mean $\preceq_X$ in the first part of the answer? I definitely meant that it’s a linear order; that’s the whole point. – Brian M. Scott Dec 12 '20 at 19:30
  • So I'm guessing that you mean $\leq\restriction X$, then? – Asaf Karagila Dec 12 '20 at 19:30
  • @AsafKaragila: Yep. That was so obvious to me that I didn’t even notice that I hadn’t said it. I’ll fix that now. – Brian M. Scott Dec 12 '20 at 19:31
  • Then you might also want to take up writing an answer to https://math.stackexchange.com/questions/3945589/a-partial-order-can-be-extended-to-a-linear-ordering – Asaf Karagila Dec 12 '20 at 19:32
  • 1
    (By the way, I found a very recent paper with an incredibly nice proof: well-order $P$, then embed $\leq$ in the usual way to the power set with inclusion, then show that the embedding respects the lexicographic ordering given from the well-order. Very nice.) – Asaf Karagila Dec 12 '20 at 19:34
  • How do you choose $\leq_F$ for each finite set? (I mean, you don't, and you effectively use $\cal U$ to make that choice, but the current formulation seems to imply that you fix a choice for each finite set in advance.) – Asaf Karagila Sep 15 '21 at 11:21
  • @Asaf: AC does have a way of sneaking up on a person unnoticed, doesn’t it? I may be missing something — I’ve done very little math since I got terminally fed up — but it appears that I really was picking a fixed $\le_F$ for each set, and I don’t at the moment see any easy way to salvage the argument so as to use only BPI or the equivalent. I’ll think about it a bit more if I have time, but for now I’m just going to rewrite it to acknowledge that I’m using full choice at that point. – Brian M. Scott Sep 16 '21 at 07:34
  • @BrianM.Scott You write that the most obvious choice works for the upper bound of $\mathscr C$. I assume that you are pointing at $\bigcup\mathscr C$. If so, then I don't think that it works for $\mathscr C = \emptyset$ for $\emptyset\notin \mathscr L$ most likely. Can you please explain? – Atom Jan 20 '22 at 11:48
  • @BrianM.Scott If we intend to deal with just nonempty chins, then don't we need to ensure that our set $\mathscr L$ is nonempty in the first place? And is this case, we are just proving the nonemptiness of $\mathscr L$... – Atom Jan 20 '22 at 12:11
  • @Atom: We deal only with non-empty chains. $\mathscr{L}$ is certainly non-empty: it contains $\left\langle{x},{\langle x,x\rangle}\right\rangle$ for each $x\in P$. – Brian M. Scott Jan 20 '22 at 18:56
  • @BrianM.Scott So $\le\upharpoonright(X\times X)$ means restriction of $\preceq_X$ to $X\times X$? – Atom Jan 20 '22 at 22:51
  • 1
    @Atom: No, it’s the restriction of $\le$ to $X\times X$. – Brian M. Scott Jan 21 '22 at 07:35
  • If anybody's having trouble proving M = P, this resource explains how it's possible to extend a partial order R into another partial order S that contains $(x,y)$ that may be incomparable in $R$: .https://webspace.maths.qmul.ac.uk/p.j.cameron/csgnotes/posets.pdf (Proposition 6.2. It speaks of finite orders only, but that particular proposition applies regardless) – jsmith Apr 08 '22 at 02:07
1

Here's another way to find such a linear order without resorting to Zorn's lemma directly (though it still requires the Axiom of Choice. There's no way around that). This is essentially the same procedure as Asaf Karaglia commented, about embedding the poset in the power set, but it is somewhat more "concrete".

Let $(X,\le)$ be a poset and let $\preceq$ be any well-order on $X$.

For any $x\in X$ let $x^\downarrow := \{y\in X:y\le x\}$. Then we may let $\le_l$ be the order in $X$ given by $$x\le_ly \Longleftrightarrow x=y\text{ or }\min_\preceq (x^\downarrow\bigtriangleup y^\downarrow) \le y,$$ where $\bigtriangleup$ stands for symmetric difference. It should be easy to see that $\le_l$ extends $\le$, and showing that it is a linear order is a good exercise.

Castor
  • 355