View on GitHub


An arbitrary-precision arithmetic library for Rust.

Iterators that generate everything, part 2

03 Jul 2022 — Mikhail Hogrefe


In Part 1 of this series, I described how Malachite provides iterators that generate all values of a type, or all values satisfying some condition. Now I’m going to show you how to generate tuples.

It’s not immediately obvious how to do this. One thing you might try is generating the tuples in lexicographic order. Malachite lets you do this:

for p in lex_pairs_from_single(exhaustive_unsigneds::<u64>()).take(10) {
    println!("{}", p);
(0, 0)
(0, 1)
(0, 2)
(0, 3)
(0, 4)
(0, 5)
(0, 6)
(0, 7)
(0, 8)
(0, 9)

(The “from_single” in lex_pairs_from_single indicates that both elements of the output pairs come from the same iterator. There’s also a lex_pairs function that accepts two iterators.)

As you can see, lexicographic generation only makes sense if the input iterator has a small number of elements, like exhaustive_bools. It doesn’t work so well for exhaustive_unsigneds::<u64>, since you’d have to wait forever to get to any pair that doesn’t start with 0. But Malachite does provide exhaustive_pairs and exhaustive_pairs_from_single, that work well for any input iterators. How do these functions work?

First attempt: the Cantor pairing function

For simplicity, let’s first consider exhaustive_pairs_from_single(exhaustive_naturals()). To think about generating every pair of Naturals, it helps to visualize a path through an infinite grid. One such path is this:

The path defined by the Cantor pairing function

You can think of it as first generating (0, 0), whose sum is 0; then (1, 0) and (0, 1), whose sum is 1; then (2, 0), (1, 1), and (0, 2), whose sum is 2; and so on. This path describes a bijection between the indices 0, 1, 2, …, and all pairs of natural numbers. The opposite direction of this bijection, that takes pairs to indices, is known as the Cantor pairing function.

This bijection has some nice properties; for example, the function from pairs to indices is a polynomial in the elements of the pair: the pair \((x, y)\) corresponds to index \(\tfrac{1}{2}(x + y)(x + y + 1)\). We can also look at the two functions from sequence index to the first and second elements of the pairs. These functions look like this:

The inverses of the Cantor pairing function

Although the functions jump around a lot (they have to, since they must evaluate to every natural number infinitely many times), there’s a sense in which they are nicely balanced. Both are \(O(\sqrt n)\), and if you take evaluate them at a random large input, they’re generally about the same size:

The Cantor unpairing of powers of 3

However, it isn’t obvious how generalize the Cantor pairing function to triples in a balanced way. The standard way to create a tripling function from a pairing function is \(g(x, y, z) = f(x, f(y, z))\), but this is no longer balanced. If \(f\) is the Cantor pairing function, then the three outputs of \(g^{-1}\) are \(O(\sqrt n)\), \(O(\sqrt[4] n)\), and \(O(\sqrt[4]n)\). Here’s what the corresponding table looks like:

An unbalanced Cantor untripling of powers of 3

More balance and flexibility with bit interleaving

Instead of the Cantor pairing function, Malachite uses bit interleaving. This idea is not new; it has been described by Steven Pigeon and others.

To generate the pairs corresponding to index 0, 1, 2, and so on, first write the indices in binary, with infinitely many leading zeros:

First step in un-interleaving bits

And then distribute the bits of each index into the two slots of the corresponding pair. The even-indexed bits (counting from the right) end up in the second slot, and the odd-indexed bits in the first:

Final steps in un-interleaving bits

I’ve color-coded the bits according to which slot they end up in. Malachite keeps track of this using an explicit bit map that looks like \([1, 0, 1, 0, 1, 0, \ldots]\), indicating that bit 0 goes to slot 1, bit 1 goes to slot 0, bit 2 goes to slot 1, and so on.

This is how this method walks through the grid of pairs of natural numbers:

The path defined by the bit-interleaving pairing function

This path is called a Z-order curve, although with this choice of coordinates it looks like it’s made up of N’s rather than Z’s.

And here’s what the functions from the index to the first and second elements of the pairs look like:

The inverses of the bit interleaving pairing function

Like the inverses of the Cantor pairing function, these functions are balanced in the sense that both are \(O(\sqrt n)\), although here the first element lags noticeably behind the second. And here’s a table showing them evaluated at large indices:

Bit un-interleaving of powers of 3

The advantage that this approach has over the Cantor pairing approach is flexibility. We can generate triples in a balanced way simply by changing the bit map to \([2, 1, 0, 2, 1, 0, 2, 1, 0, \ldots]\):

un-interleaving bits to form triples

Here are the three functions from index to output. They each are \(O(\sqrt[3]n)\).

The inverses of the bit interleaving tripling function

Here they are evaluated at large indices:

Bit un-interleaving of powers of 3 to form triples

This generalizes straightforwardly to \(k\)-tuples for any \(k\). Notice that the first \(2^{ka}\) tuples are all the tuples with elements less than \(2^a\).

Customized balancing

By changing the bit map further, we can deliberately unbalance the tuples. For example, if we want to produce pairs where the second element of the pair is \(O(\sqrt[3]n)\) and the first element is \(O(n^{2/3})\), we can use the bit map \([1, 0, 0, 1, 0, 0, 1, 0, 0, \ldots]\). If we want the the first element to be exponentially larger than the second, we can use \([0, 1, 1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, \ldots]\), where there is a 1 at the \(2^i\)th position and 0 elsewhere.

To actually generate the pair sequences in the two preceding examples, you can use




respectively. See the documentation of exhaustive_pairs_custom_output and BitDistributorOutputType for more details.

Generating tuples of other iterators

So far I’ve only talked about generating tuples of Naturals. What about tuples of elements from some other iterator xs? If xs is infinitely long, then the process is straightforward. Instead of generating, say, (8, 3), generate the pair consisting of the 8th and 3rd elements of xs (indexing from 0, of course). Internally, Malachite uses a structure called IteratorCache, which stores the values produced by xs into a Vec for easy access.

For example, here’s how to generate all pairs of the characters 'a' to 'z' (in this case, there are additional complications due to xs being finite, but we’ll address those in a moment):

for p in exhaustive_pairs_from_single('a'..='z').take(10) {
    println!("{:?}", p);
(a, a)
(a, b)
(b, a)
(b, b)
(a, c)
(a, d)
(b, c)
(b, d)
(c, a)
(c, b)

Dealing with finite iterators

Finite iterators are a bit of a problem. You can’t generate the (8th, 3rd) pair if the input iterator has no 8th element. To take an extreme case, consider exhaustive_pairs(exhaustive_naturals(), exhaustive_bools()). There are only two bools, so any index pair \((i, j)\) where \(j \geq 2\) can’t be used for indexing. Here’s our grid of indices again, with valid indices green and invalid indices red:

Valid indices when one iterator has length 2

My original solution was just to skip over the invalid indices, with a bit of extra logic to determine when both input iterators are finished. This was very slow: out of the first \(n\) pairs, only about \(2 \sqrt n\) are valid, and the situation gets worse for higher-arity tuples. So what does exhaustive_pairs do instead?

Again, we can leverage the bit map! This scenario is actually what led me to make the bit map an explicit object in memory. What exhaustive_pairs does is notice when one of its iterators has completed, and then adjusts the bit map on the fly. In this particular case, it realizes that since exhaustive_bools produces two elements, it only needs a single bit for indexing; and then bit map is modified from \([1, 0, 1, 0, 1, 0, 1, 0, \ldots]\), to \([1, 0, 0, 0, 0, 0, 0, \ldots]\). In general, modifying the bit map on the fly is dangerous because you might end up repeating some output that you’ve already produced, but exhaustive_pairs only modifies the part of the map that hasn’t been used yet.

If the finite iterator’s length is a power of 2, as it was in this example, then updating the bit map results in all indices being valid. If it isn’t a power of 2, then some indices will remain invalid; for example, in exhaustive_pairs(exhaustive_naturals(), 0..6) the bit map will be updated from \([1, 0, 1, 0, 1, 0, 1, 0, \ldots]\), to \([1, 0, 1, 0, 1, 0, 0, 0, \ldots]\), so that the index of the second pair slot will vary from 0 to 7, and only \(3/4\) of all the indices will be valid. But this is a constant proportion; it’s much better than the earlier example where only \((2 \sqrt n)/n\) were valid.

(In this particular case, the best thing to do would be to use lex_pairs(exhaustive_naturals(), exhaustive_bools()), which would produce (0, false), (0, true), (1, false), (1, true), and so on. But in general, when you call exhaustive_pairs(xs, ys) you might not know ahead of time whether ys is short, long, or infinite.)

In the next part, I will discuss how Malachite generates all Vecs containing elements from some iterator.

Copyright © 2024 Mikhail Hogrefe