# Bounds on the gaps in a variant of polylog-smooth numbers.

Sorry for the long intro. I think the explanation motivates the question and puts it in context. But if you want to skip the story, then just move on to the grey boxes; they should contain enough information to make out my question.

This question is inspired by Sadeq Dousti's question on producing a factored integer from a small interval: Factoring some integer in the given interval. For convenience, I will assume that the problem is to produce a factored integer from the interval $[x-G; x]$, and our goal is to make $G = G(x)$ as small as possible. (Beware that Sadeq's version uses $N$ in place of my $x$, and fixes $G(x)$ concretely to be $O(\log x)$.)

One can consider several candidate easily-stated algorithms for this problem, all of which work in the following general way. Let $S$ be a set of integers such that

• Given a number $N$, testing if $N$ is in $S$ can be done efficiently. Moreover, any $N \in S$ can also be factored in time $\mathrm{poly}(\log N)$.
• The gaps in $S$ are "small": for all sufficiently large $x > 0$, there exists $N \in S$, such that $x - G(x) \leq N \leq x$. (Hence my notation $G$.)

Then the given problem can be solved for intervals of length $G(x)$: just search for an $S$-element in the interval and output its factorization. The whole problem then becomes how small can we make $G(x)$ to be. Here are some candidate sets:

1. Powers of $2$.
2. Primes.
3. The set of numbers $N$ such that all prime divisors of $N$ are at most $\log^A(N)$ for some absolute constant $A < \infty$. I'll call these polylog-smooth numbers. (See the fineprint at the end of the question.)
4. The set of numbers $N$ such that at most one of its prime divisors exceeds $\log^A(N)$ where $A$ is as before. I'll call these almost polylog-smooth numbers.

Ok, enough with the names :-). All these examples satisfy the first requirement (testing+factoring) by design. (For instance, in the last case, divide out all small prime divisors of $N$ by brute force, and we will be left with either $1$ or a large prime number, and so we'll be in good shape.) The big question then is to analyze the gaps in these sets.

1. The powers of two are very sparse. They have a gap $G(x) = O(x)$ and this is tight.
2. The primes are suspected to have better gaps (see the Consequences of Riemann hypothesis and Cramér's conjecture). In short, the gap might be as small as $O(\log^2 x)$. On the other hand, the best unconditional bound is $O(x^{\theta})$ for $\theta < .53$.
3. I do not have enough background to grok the question of gaps for the polylog-smooth numbers. Nevertheless, I found Andrew Granville's survey on smooth numbers (Smooth numbers: Computational Number Theory and beyond) says that even the average gap is around $x^{1/A}$. (I am quoting Theorem 1.14 from page 4/page 270 in the pdf I have linked to.) This might be quite bad compared to the conjectured gap for the primes example, but a great improvement over the powers of $2$.

Now, what about the set of almost polylog-smooth numbers? They are denser than the primes, so it's definitely at least as small as $O(\log^2 x)$ conditionally and $O(x^{\theta})$ unconditionally.

But taking a clue from the improvement we got in going from the powers of $2$ to the smooth-numbers, I believe that the actual gaps for the almost-smooth numbers should be even better than that of the primes. That is the motivation behind my question:

What are the current best bounds on the gaps on the almost polylog-smooth numbers (conditionally and unconditionally)? Specifically, is it possible that $G = O(\log x)$ for this case?

Any help?

Fineprint that morally shouldn't matter. A number is $B$-smooth if its prime divisors are at most $B$. It seems to me that when people analyze properties of the smooth numbers $\leq x$, they usually set the smoothness parameter $B$ to be a function of $x$, like $B = \log^A x$. I defined a smooth number more "intrinsically" by setting $B = \log^A N$. I presume that this is just a technical issue, and none of the bounds will really change that much. If I am wrong, then please let me know. In any case, I see no reason why one definition is better for this application than the other.

-
+1, very well formulated! – M.S. Dousti Jul 31 '11 at 13:26
After the linked question I had also thought of an extremely similar idea: does there exist a subset $S\subset\mathbb{N}$ such that (a) we can efficiently determine if a number is in $S$ and (b) there is an algorithm to efficiently factor numbers specified in $S$. You've added a third feature, (c) gaps asymptotically grow as slowly as possible. Interesting. – anon Jul 31 '11 at 14:46
@anon The growth in gaps requirement just comes from translating the requirement that you need to find a factored number as close to the input number as possible. Without this, the problem becomes trivial, no? – Srivatsan Jul 31 '11 at 14:50
Yes, I understood the motivation behind the condition. – anon Jul 31 '11 at 15:01
a little bit more motivation. – Kaveh Aug 4 '11 at 15:18

I just want to point out that you can generalize your almost polylog-smooth numbers a bit further: You do not require that the remaining factor is prime, you just require that it is easy to factor. So, for example, if your time bound is $ln^A x$, you can have the remaining factor be of size M where $ln^A x = exp( (\frac{4}{9^{1/3}} + o(1)) (ln M)^{1/3} (ln ln M)^{2/3} )$ or $M = o( exp(\frac{9}{64} A^3 (ln ln x)^3) )$. The average distance of this almost polylog-smooth set S in a neighborhood of x is then approximately the average distance of the polylog-smooth set S' in a neighborhood of x' = x/M. Note that this only gives a lower bound; $G_S(x) \geq G_{S'}(x')$.