2

I'm interested in all four combinations: directed and undirected, cyclic and acyclic.

I'm having trouble calculating how big the complexity gets as you add more nodes to a graph. Clearly, the number of possible graphs goes up with adding directability, and wildly (adding Ω(2n) complexity, roughly).

My best guess on a DAG is close to Ω(n!).

This question concerns itself with knowledge representation. How many neural networks are there with n neurons? Given that different knowledge must be encoded differently, it gives some sort of data about how knowledge can scale in the brain.

[Edit: "multigraphs", obviously, aren't part of the question, disconnected graphs should count as their lower order counterparts, and v1 is separate from v2 such that a set V containing both has 3 DAGs.]

[Edit2: Looks like for DCGs, it is about 23n. For DAGs, it's about 22n.]

[Note: I tagged this under "descriptive complexity" because it's not really a simulation. Let me know if this is wrong.]

Marcos
  • 334
  • Labeled or unlabeled? (For instance, are there $\frac12 n!$ path graphs on $n$ vertices, or only $1$?) – Misha Lavrov Nov 19 '18 at 19:06
  • In my understanding of "labeled" graphs, there would be an infinite number, so it's inapplicable. – Marcos Nov 19 '18 at 19:07
  • Labeled graphs in the sense that we name the vertices $v_1, v_2, \dots, v_n$ (or if you prefer just $1, 2, \dots, n$) and go from there. This is the standard thing to do when, for example, we count labeled trees. Another way to phrase my question: do you want to count isomorphic graphs multiple times or not? – Misha Lavrov Nov 19 '18 at 19:10
  • @MishaLavrov: I think we/I have to reconsider your question wrt labeled or not. A 2-vertex DAG from v1 to v2 is counted separately than a graph of v2 to v1, even though they are "isomorophic" without regarding labels. The reason is that a vertex is generally anchored to some meaning outside the graph even though a graph's definition doesn't care. – Marcos May 16 '19 at 18:10

1 Answers1

8

In general these counts do not have nice closed formulas, but some satisfy nice recurrence relations.

Graphs

  • Graphs on $n$ nodes is OEIS A000088: $$1, 1, 2, 4, 11, 34, 156, 1044, 12346, 274668, \ldots .$$ Flajolet & Sedgwick's Analytic Combinatorics, $\S~$II.5 gives that this sequence is asymptotic to $2^{\frac{1}{2} n (n - 1)}/n!$. (They cite Harary & Palmer's text Graphical Enumeration for this fact, but I haven't checked it myself. Probably that reference gives some of the below data, too.)
  • Acyclic graphs on $n$ nodes (forests) is A005195: $$1, 1, 2, 3, 6, 10, 20, 37, 76, 153, \ldots .$$ This is asymptotic to $c n^{-5/2} d^n$ for some $c > 0$ and $d > 1$.
  • Cyclic graphs on $n$ nodes is A286743: $$0, 0, 1, 5, 24, 136, 1007, 12270, 274515, 12004839, \ldots .$$ By definition this is just [A000088] - [A005195]. The latter has much smaller growth than the former, so this also asymptotic to $2^{\frac{1}{2} n (n - 1)}/n!$.

Directed graphs

  • Directed graphs on $n$ nodes is OEIS A000273: $$1, 1, 3, 25, 543, 29281, 3781503, 1138779265, 783702329343, 1213442454842881, \ldots .$$ This is asymptotic to $2^{n (n - 1)}/n!$
  • Directed acyclic graphs on $n$ nodes is OEIS A003087: $$1, 1, 2, 6, 31, 302, 5984, 243668, 20286025, 3424938010, \ldots .$$ OEIS doesn't give asymptotics for this sequence, but we can deduce from asymptotics for the case of labeled directed acyclic graphs that this sequence is asymptotic to some function in between $A p^{-n} 2^{\frac{1}{2} n (n - 1)}$ and $A p^{-n} 2^{\frac{1}{2} n (n - 1)} n!$ for some $A > 0$ and $p > 1$. In any case before this is much slower growth than the count for directed graphs, so graphs on $n$ nodes that have at least one cycle is also asymptotic to $2^{n (n - 1)} / n!$.
Travis Willse
  • 99,363
  • 1
    A compact Maple routine for OEIS A000088 may be found at the following MSE link (scroll to end of post, is not limited to question from title of link). – Marko Riedel Nov 19 '18 at 20:01
  • That is amazing. Thanks for the answer! I will evaluate further. – Marcos Nov 19 '18 at 21:27
  • You're welcome. I've added a reference to A286743. – Travis Willse Nov 19 '18 at 22:24
  • I haven't redone the calculation myself, but the cited rate for D.A.G. comes from the linked OEIS entry, which cites:

    M. D. McIlroy, Calculation of numbers of structures of relations on finite sets, Massachusetts Institute of Technology, Research Laboratory of Electronics, Quarterly Progress Reports 17 (1955), 14-22. A hand-annotated pdf copy can be found at http://oeis.org/A000088/a000088.pdf

    In that article the relevant sequence is denoted $\textrm{ref}_n$.

     – Travis Willse Nov 29 '18 at 13:32
  • I'm looking for a general equation. Something like (2^(n+1))-1, which is a sort of recursive definition in that the number of graphs up to degree n is the number of graph of degree n + the number of graphs up to degree n-1. – Marcos Dec 21 '19 at 20:18
  • I'm not sure what you mean: Like I wrote at the beginning of my answer there are no nice known closed forms for these sequences, and $2^{n + 1} - 1$ is not a recursive expression. You might see if the OEIS entries (e.g., https://oeis.org/A000088) contain what you're looking for. – Travis Willse Dec 21 '19 at 23:47