# Count the trees

A tree is a connected, undirected graph with no cycles. Your task is to count how many distinct trees there are with a given number of vertices.

Two trees are considered distinct if they are not isomorphic. Two graphs are isomorphic if their respective vertices can be paired up in such a way that there is an edge between two vertices in one graph if and only if there is an edge between the vertices paired to those vertices in the other graph. For a more complete description, see the link above.

To see what all of the distinct trees of sizes 1 to 6 look like, take a look here.

The series you are trying to output is A000055 at the OEIS.

Restriction: Your solution must take in the range of minutes or less to run on the input 6. This is not intended to eliminate exponential time algorithms, but it is intended to eliminate doubly-exponential time algorithms, such as brute forcing over all edge sets.

Input: Any non-negative integer.

Input may be by any standard means, including STDIN, command line parameter, function input, etc.

Output: The number of distinct trees with as many vertices as the input.

Output may be by any standard means, including STDOUT, function return, etc.

Examples: 0, 1, 2, 3, 4, 5, 6, 7 should return 1, 1, 1, 1, 2, 3, 6, 11.

Scoring: Code golf, by bytes. May the shortest code win!

Standard loopholes forbidden.

## CJam (69 bytes)

]qi:X,{1e|,:):N{N\f{1$%!*}W$.*:+}%1$W%.*:+N,/+}/W\+_1,*X=\_W%.*:+-Y/z  Online demo ### Explanation The basic idea is to implement the generating function described in OEIS. The input $0$ is a nasty special case, but the final tweaks I made ended up producing $-1$ for that case, so the $z$ (for absolute value) tidies it up. That's the weirdest trick in here. .*:+ is repeated three times, and looks like it could save a byte if extracted as {.*:+}:F~. However, this breaks with the special case $0$, because it doesn't execute the outer loop at all. We use the auxiliary generating function for A000081, whose terms have the recurrence a[0] = 0 a[1] = 1 For n >= 1, a[n+1] = (sum_{k=1}^n a[n-k+1] * sum_{d|k} d * a[d]) / n  I'm sure some languages have built-ins for the inverse Möbius transform $\sum_{d\mid k} d \times a[d]$, but CJam doesn't; the best approach I've found is to build an array mapping $d$ to k % d == 0 ? d : 0 and then do a pointwise multiplication with $a$ using .*. Note that here it is convenient to have built up $a$ starting at index 1, because we want to avoid division by zero when setting up the weights. Note also that if the two arrays supplied to the pointwise operation are not the same length then the values from the longer one are left untouched: therefore we have to either take the first $k$ terms of $a$ or make the array of weights go up to $n$. The latter seems shorter. So this inverse Möbius transform accounts for N\f{1$%!*}W$.*:+ If we call the result of the inverse Möbius transform M, we now have $$a[n+1] = \frac{1}{n} \sum_{k=1}^n a[n-k+1] \times M[k]$$ The numerator is obviously a term from a convolution, so we can handle it by reversing either a copy of $a$ or $M$ and then taking a pointwise multiplication and summing. Again, our index varies from $1$ to $n$, and in addition we want to pair up indices which sum to $n + 1$, so it's again convenient to index $a$ from 1 rather than 0. We've now accounted for  qi:X,{ ,:):N{N\f{1$%!*}W$.*:+}%1$W%.*:+N,/+}/


The point of the auxiliary generating function is given by the formula section of A000055:

G.f.: A(x) = 1 + T(x) - T^2(x)/2 + T(x^2)/2,
where T(x) = x + x^2 + 2*x^3 + ... is the g.f. for A000081.


In terms of $a$, this means that the output we seek is $$[x = 0] + a[x] + \frac{1}{2}\left(a[x/2] - \sum_{i=0}^n a[i] \times a[n-i]\right)$$

where for $a[x/2]$ with odd $x$ we get 0. The shortest way I've found to do that is to inflate with zeroes (1,*) and then take the Xth element (X=).

Note that the sum is yet another convolution, but this time it's convenient to index from 0. The obvious solution is 0\+, but there is a nice little optimisation here. Since $a[0] = 0$, two terms of the convolution are guaranteed to be zero. We could take this as an opportunity to index from 1, but the special case $X = 0$ gets ugly. If we instead do W\+, the convolution gives us $-2a[x] + \sum_{i=0}^n a[i] \times a[n-i]$ and after subtraction and division by $2$ we've handled the $a[x]$ term outside the sum as well.

So we've explained

 qi:X,{   ,:):N{N\f{1$%!*}W$.*:+}%1$W%.*:+N,/+}/W\+_1,*X=\_W%.*:+-Y/  The remaining details relate to the special case. I originally followed the recurrence more strictly by starting out with 1] and iterating from $N = 1$ with 1]qi:X,1>{ ... }/  The result when $X = 0$ is that we compute $a$ as [-1 1]: one term more than we need. The inflation and convolution wind up giving a result of $0$. (Perhaps better than we deserve - it's what we should have, because we haven't done anything about the term $[x = 0]$). So we fix it up for three chars either as a final X!+ or using the standard "fallback" technique as 1e|. To get the "right" length of $a$ we need to avoid supplying that initial $1$, and instead to produce it from the main loop with $N = 0$. A completely straightforward ]qi:X,{ ... /+}/  obviously gives division by zero. But if we try ]qi:X,{1e| ... /+}/  then it works. We get  e# Stack: [] 0 1e| e# Stack: [] 1 ,:):N e# Stack: [] [1] { e# We only execute this loop once N\f{1$%!*} e#   1 divides 1, so stack: [] [1]
W$.* e# Remember: if the two arrays supplied to the pointwise operation e# are not the same length then the values from the longer one are e# left untouched. Stack: [] [1] :+ e# Fold over a singleton. Stack: [] 1 }% e# And that was a map, so stack: [] [1] 1$W%.*:+     e# Another [1] [] .*:+, giving the same result: 1
N,/          e# 1 / 1 = 1
+            e# And we append 1 to a giving [1]


which produces exactly the value we require.

Now if $X = 0$ that main loop never executes, so after we hack in a $-1$ at the bottom we get [-1]. We end up computing $(-1 - \frac12(-1 \times -1)) = -1$, which isn't correct. But whereas correcting $0$ to $1$ took three chars, correcting $-1$ to $1$ only takes one: z gives the absolute value.

# Pyth, 35 bytes

l{mSmSSMcXdUQk2.pQusm+L,dhHGhHtQ]Y


Demonstration.

This program can be divided into two parts: First, we generate all possible trees, then we remove duplicates.

This code generates the trees: usm+L,dhHGhHtQ]Y. Trees are represented as a concatenated list of edges, something like this:

[0, 1, 0, 2, 2, 3, 1, 4]

Each number stands for a vertex, and every two numbers is an edge. It is built by repeatedly adding an edge beteen each possible vertex that already exists and one which is newly created, and adding this on to each possible tree from the previous step. This generates all possible trees, since all trees can be generated by repeatedly adding one vertex and an edge from it to the existing tree. However, isomorphic trees will be created.

Next, for each tree, we perform every possible relabeling. This is done by mapping over all possible permutations of the vertices (m ... .pQ), and then transating the tree from the standard ordering to that ordering, with XdUQk. d is the tree, k is the permutation.

Then, we separate the edges into separate lists with c ... 2, sort the vertices within each edge with SM, sort the edges within the tree with S, giving a cannonical representation of each tree. These two steps are the code mSSMcXdUQk2.pQ.

Now, we have a list of lists composed of every possible relabeling of each tree. We sort these lists with S. Any two isomorphic trees must be able to be relabelled into the group of trees. Using this fact, we convert each list to a string with , then form the set of those lists with {, and output its length with l.