I was hacking around with a program that needed tree-structured data, and I wanted to generate all possible trees of a certain size. I searched around and found nothing useful. After a few experiments involving Python generators, postfix expressions, and recursive trees, I've got some code to enumerate binary trees.

I started with postfix expressions because they are easy to work with, and are equivalent to trees, in that trees are like infix notation expressions, and they are equivalent to postfix. After some puzzling on it, I saw the recursive light, and came up with this to generate all possible postfix expressions:

# Enumerate postfix expressions

def _exprs(expr, stack_depth, vals, ops):
    """ Generate postfix expressions recursively.
        `expr` is the expression so far, a list of values and operators.
        `stack_depth` is the depth of the stack created by expr.
        `vals` is a list of values remaining to add to the expression.
        `ops` is a list of possible operators to choose from.
    """
    if stack_depth == 1 and not vals:
        # This is a valid expression.
        yield expr
    if stack_depth > 1:
        # Try using an operator
        for o in ops:
            for e in _exprs(expr+[o], stack_depth-1, vals, ops):
                yield e
    if vals:
        # Vals are available, push one on the stack
        for e in _exprs(expr+[vals[0]], stack_depth+1, vals[1:], ops):
            yield e

def exprs(vals, ops):
    """ Generate postfix expressions created from `vals`, the list of values
        to use, and `ops`, the possible operators to combine them with.
    """
    return _exprs([], 0, vals, ops)

def all_exprs(n, ops='+'):
    vals = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"[:n]
    return [ " ".join(e) for e in exprs(vals, ops) ]

print "\n".join(all_exprs(4, '+*'))

for i in range(2, 15):
    ndyck = len(all_exprs(i))
    print "%2d: %7d expressions" % (i, ndyck)

The nice thing here is the generator is a pretty clear expression of the enumeration technique: as you are building an expression:

• if the stack has one item, and there are no more values to squeeze in, you have a valid expression.
• if the stack is two or more deep, then you can add to expression by using an operator to combine those top two values, and the stack is now one shallower.
• if you have more values to use, you can append the next one to the expression, and the stack is now one deeper.

The nature of recursive generators requires the for/yield structure:

# When recurring in generators, you have to loop to pull the values,
# and yield each one:
for v in _recursive_generator(blah):
    yield v
    
# Wouldn't it be cool if you could do this instead:
yield _recursive_generator(blah)

# but that would yield a generator, a style of recursion I can't quite wrap
# my head around!

Running the code shows all the expressions of two operators on five values, and the counts of expressions of one operator over varying numbers of values:

A B + C + D +
A B + C + D *
A B + C * D +
A B + C * D *
A B + C D + +
A B + C D + *
A B + C D * +
A B + C D * *
A B * C + D +
A B * C + D *
A B * C * D +
A B * C * D *
A B * C D + +
A B * C D + *
A B * C D * +
A B * C D * *
A B C + + D +
A B C + + D *
A B C + * D +
A B C + * D *
A B C + D + +
A B C + D + *
A B C + D * +
A B C + D * *
A B C * + D +
A B C * + D *
A B C * * D +
A B C * * D *
A B C * D + +
A B C * D + *
A B C * D * +
A B C * D * *
A B C D + + +
A B C D + + *
A B C D + * +
A B C D + * *
A B C D * + +
A B C D * + *
A B C D * * +
A B C D * * *
 2:       1 expressions
 3:       2 expressions
 4:       5 expressions
 5:      14 expressions
 6:      42 expressions
 7:     132 expressions
 8:     429 expressions
 9:    1430 expressions
10:    4862 expressions
11:   16796 expressions
12:   58786 expressions
13:  208012 expressions
14:  742900 expressions

I still wanted to find other work on this stuff, and I figured anyone who had successfully enumerated trees would also have done a count like I did, so I did a search on the last two counts (208012 742900). I was directed to a few interesting pages, in particular the wikipedia page on Catalan numbers, which confirmed that my count of trees is correct, and led to the theory of Dyck languages, which is too much math for my purposes, and led to a few interesting papers that I couldn't understand.

To make trees, my first impulse was to use my expression generator and then execute the expressions to construct trees, but that seemed unnecessary, so I adapted the expression code to create trees directly:

# Enumerate binary trees

class Leaf:
    def __init__(self, val):
        self.val = val
    
    def __str__(self):
        return str(self.val)

class Node:
    def __init__(self, op, left, right):
        self.op = op
        self.left, self.right = left, right

    def __str__(self):
        return "(%s %s %s)" % (self.left, self.op, self.right)
    
def _trees(stack, vals, ops):
    if len(stack) == 1 and not vals:
        yield stack[0]
    if len(stack) > 1:
        for op in ops:
            left = stack[-2]
            right = stack[-1]
            if hasattr(right, 'op') and right.op == op:
                # Avoid duplicates: our operations are associative,
                # so we can skip trees like (A+(B+C)), since we'll
                # get ((A+B)+C) instead on another iteration.
                continue
            new_node = Node(op, left, right)
            new_stack = stack[:-2]
            new_stack.append(new_node)
            for t in _trees(new_stack, vals, ops):
                yield t
    if vals:
        stack.append(Leaf(vals[0]))
        for t in _trees(stack, vals[1:], ops):
            yield t
                     
def trees(vals, ops="+"):
    return _trees([], vals, ops)

def all_trees(n, ops='+'):
    vals = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"[:n]
    return trees(vals, ops)

print "\n".join(map(str, all_trees(4, '+*')))

Which produces:

(((A + B) + C) + D)
(((A + B) + C) * D)
(((A + B) * C) + D)
(((A + B) * C) * D)
((A + B) * (C + D))
((A + B) + (C * D))
(((A * B) + C) + D)
(((A * B) + C) * D)
(((A * B) * C) + D)
(((A * B) * C) * D)
((A * B) * (C + D))
((A * B) + (C * D))
((A * (B + C)) + D)
((A * (B + C)) * D)
(A * ((B + C) + D))
(A + ((B + C) * D))
((A + (B * C)) + D)
((A + (B * C)) * D)
(A * ((B * C) + D))
(A + ((B * C) * D))
(A + (B * (C + D)))
(A * (B + (C * D)))

One tricky thing about this generator: it avoids trees which are structurally different, but with associate operators produces the same result, so it won't return both ((A+B)+C) and (A+(B+C)). If you want both of those trees, remove the check at line 25 for right associativity.

The number of trees grows very large as the number of leaves becomes big enough to be interesting. I guess I'll have to dig into application-specific ways to limit the choices.