# A question on linear extensions of partial orders

If you're given a collection of partial orders, topological sort will tell you if there's an extension of the collection to a total order (an extension in this case is a total order consistent with each of the partial orders).

I've come across a variation:

Fix a set $V$. You're given sequences $\sigma_1, \ldots \sigma_k$ of elements drawn from $V$ without repetition (the sequences are of length between 1 and $|V|$).

Is there a way to fix orientations for each of the sequences (either forward or reverse) so that the resulting collection of chains (viewed as a partial order) admits an extension ?

Is this problem well-known ?

Note: The orientation is chosen for an entire sequence. So if the sequence is $1-2-4-5$, you can either keep it that way, or flip it to $5-4-2-1$, but you can't do anything else.

• If each of the sequences are of length $2$ then one can think of each sequence as an undirected edge and we are asking whether an undirected graph can be oriented to be a DAG - iff if there is no cycle. But a greedy algorithm also works. Start with an edge and orient it arbitrarily and keep going as long as you can and if you get stuck you know it is not possible. Did you try that for your variation? Seems like it may work. – Chandra Chekuri Jun 29 '12 at 19:32
• Er, every undirected graph can be oriented to be a DAG. Just choose an ordering of the vertices and use that ordering to orient the edges. – David Eppstein Jun 29 '12 at 19:52
• You are right of course, me not thinking straight. – Chandra Chekuri Jun 29 '12 at 19:56
• In my variation each subsequence has length exactly 4, so Yury's answer kicks in. My only hope at this point is that the subsequences have very special structure and are related to each other, so maybe something specific to the problem would help. But there's no general hammer. – Suresh Venkat Jun 29 '12 at 20:41

If every sequence has length 3, the problem is known as Betweenness. Even the Betweenness problem is NP-hard. In this problem, we are given a set of vertices and a set of constraints of the form $u$ lies between $v$ and $w$. Our goal is to order all vertices so as to maximize the number of satisfied constraints. Opantry [1] proved that the decision version of this problem is NP-hard. Chor and Sudan [2] proved that it is SNP-hard.

The best known approximation algorithm for the problem, by Chor and Sudan, satisfies 1/2 of all constraints if the instance is completely satisfiable.

[1] J. Opantry. Total Ordering Problem, SIAM Journal on Computing, 8(1):111—114, Feb. 1979.

[2] B. Chor and M. Sudan. A geometric approach to betweenness, SIAM Journal on Discrete Mathematics, 11(4):511-523, Nov. 1998.

Edits: clarified that the decision version of the problem is NP-hard.

• Yury, does that mean that the decision problem of whether all constraints can be satisfied is also hard? – Chandra Chekuri Jun 29 '12 at 19:58
• Yes, the decision problem is NP-hard. Moreover, for some $\epsilon > 0$ it is NP-hard to satisfy even a $1-\epsilon$ fraction of all constraints (i.e. the corresponding promise problem is NP-hard). – Yury Jun 29 '12 at 20:00
• If the instance is not completely satisfiable, the problem is very hard: of course, you can satisfy $1/3$ of all constraints by taking a random order; but it is UGC-hard to satisfy $1/3+ \varepsilon$ of all constraints if $OPT = 1 -\varepsilon$ for every constant $\varepsilon > 0$ [Charikar, Guruswami, Manokaran — CCC 2009]. – Yury Jun 29 '12 at 20:04
• my question might be stupid. but does the 3-regular ($|\sigma_i| =3$ for all $i$) hardness extends to 4-regular naturally? – Yixin Cao Jun 30 '12 at 16:24
• Yes, here is a reduction. Consider a 3-regular instance ${\cal I}$. Introduce a new variable $y_i$ for every sequence $\sigma_i$. Let $\sigma_i'$ be a sequence obtained by appending $y_i$ to the end of $\sigma_i$. We get a 4-regular instance ${\cal I}'$ on vertices $V \cup\{y_i\}$ with sequences $\{\sigma_i'\}$. It is easy to see that ${\cal I}'$ is satisfiable if ${\cal I}$ was satisfiable — take the solution for ${\cal I}$, put each $y_i$ either before all vertices in $V$ or after all vertices in $V$ depending on the orientation of $\sigma_i$ (the relative order of $\{y_i\}$ is irrelevant). – Yury Jun 30 '12 at 18:54