# What is the complexity of depth first traversal that don't label nodes as discovered?

I've found an algorithm that acts like a depth first traversal that don't recognizes nodes that have been visited before.

  A
/ \
B   C
\ /
D
|
E


If run on this graph, the algorithm will traverse in this order:

A, B, D, E, D, B, A, C, D, E

Is there any way I can express the run-time complexity of this algorithm in terms of edges, nodes, or the depth of the graph? I expect it would be high, I have the feeling it will be exponential, but I'm struggling to explain why.

EDIT: The real algorithm that I am analyzing is computing the result of an expression tree, only it is not run at a tree, but at a DAG. This means, that nodes that have several parents get computed once for all parents.

Simple pseudo-code for the algorithm might be:

calculate(root){
if(root.isLeaf){
return root.value;
} else {
leftval = calculate(root.left);
rightval = calculate(root.right);
return root.operator(leftval, rightval);
}
}

• Problems have complexities, algorithms have costs. That aside, have you tried constructing a worst-case instance? Can you put into words what the algorithm does? – Raphael Jun 10 '14 at 8:31
• Regarding your pseudo code, I guess calculate and calculate_expression mean the same thing? Now it is no longer clear what the question is: are you asking about the runtime on arbitrary DAGs, or about the runtime on such DAGs that are the result of minimising an arithmetic expression? (The answers may differ.) – Raphael Jun 10 '14 at 10:36
• @Raphael: Yes, you are right, edited the pseudo code. Regarding your other question, the DAG is not built from a textual expression like "1+2*1", but by creating nodes and edges. So, yes, I guess it's an arbitrary DAG. – bobbaluba Jun 10 '14 at 10:51

For general DAGs, the runtime could indeed be exponential in the depth. Here's an example of a family of such graphs. Let $G_d$ be a digraph comprised of a vertex stacked on a collection of 4-cycles as below, where I've illustrated $G_3$: In this figure, the edges are directed from top to bottom and every vertex has two children. The number of function calls your algorithm makes (and hence the run time) for a digraph $G_d$ of depth $d$ will satisfy $T(0)=1, T(d)=2T(d-1)+1$, since every $G_d$ is comprised of a root vertex at the top with two overlapping copies of $G_{d-1}$ below it. This recurrence has solution $T(d) = 2^{d+1}-1$, exponential in the depth.