Even faster integer multiplication

by fdej • 12 years ago

94 25 12 years ago

sambeau• 12 years ago

Can anyone out there explain this in pseudocode? I tried my best to understand the paper but was thoroughly defeated by it.

fdej2• 12 years ago

All asymptotically fast algorithms for integer multiplication basically reduce the problem to one or several polynomial multiplications.

Polynomial multiplication can be done in subquadratic time using an evaluation-interpolation strategy: namely, to multiply two polynomials of degree less than n, we evaluate both polynomials at 2n points, compute 2n pointwise products, and recover the polynomial product by interpolation.

When one chooses roots of unity as the evaluation points, the evaluation and interpolation steps reduce to discrete Fourier transforms (forward and inverse, respectively), which can be done in O(n log n) arithmetic operations by means of the fast Fourier transform (the pointwise products require O(n) arithmetic operations).

Now, there are a couple of difficulties. How do you decompose the integers into polynomials? Is it better to use short polynomials with large coefficients, or long polynomials with small coefficients? Since the coefficients asymptotically grow with the length, you have to account for the bit complexity of the arithmetic operations in the FFTs and pointwise products (here you get recursive integer multiplications).

Which FFT algorithm should you use (there are many to choose from)?

Also, the ring of integers does not have enough roots of unity to do FFTs, so you have to lift the computation to a larger ring, such as the field of complex numbers (with numerical approximations), an appropriate finite field, or (as in the Schönhage-Strassen algorithm) a polynomial extension ring. Which choice gives the best complexity?

The improvements to integer multiplication published in the last few decades have basically been tightened analyses of such tradeoffs. I have not read this paper in detail yet, but it seems to provide a clear description in the introduction. Quote:

The idea of the new algorithm is remarkably simple. Given two n-bit integers, we split them into chunks of exponentially smaller size, say around log n bits, and thus reduce to the problem of multiplying integer polynomials of degree O(n/log n) with coeffcients of bit size O(log n). We multiply the polynomials using discrete Fourier transforms (DFTs) over C, with a working precision of O(log n) bits. To compute the DFTs, we decompose them into "short transforms" of exponentially smaller length, say length around log n, using the Cooley-Tukey method. We then use Bluestein's chirp transform to convert each short transform into a polynomial multiplication problem over C, and finally convert back to integer multiplication via Kronecker substitution. These much smaller integer multiplications are handled recursively.

tobz2• 12 years ago

When I read "integer multiplication", I think of "123 * 456". Is that what is actually being said here? If so, why would you do anything but "123 * 456" to arrive at the actual answer?

I'm sure there's a reason, but reading your explanation sounds like the solution to something much harder than just multiplying two whole numbers.

ggchappell• 12 years ago

That's a very nice explanation. Thanks!

owlish• 12 years ago

Undergrad here, what's the best way to keep up with new papers in a field (e.g. distributed computing or machine learning)? What sources/methods are good for making the most of your reading time?

petercooper• 12 years ago

For the areas they cover, being a member of the relevant IEEE group can be useful (I guess ACM probably does something similar but I'm not a member). For one of the groups I'm in, I receive a pamphlet every few months that summarizes a ton of papers in the field and a CD with the full content on. (It's all online as well, but I know I'd never remember to actively check there.)

You may also find the relevant section of http://arxiv.org/ useful as well. For example, http://arxiv.org/list/cs.SE/recent

sampo• 12 years ago

In other sciences, you'd follow some leading journals. In computer science, you follow some of the leading conferences (and hope that the conference publishes the proceedings in some place that your university has access).

theschreon• 12 years ago

Once you have identified the most significant individuals on a field, you can use google scholar alerts to get notified when they publish something new. (http://scholar.google.com/scholar_alerts?view_op=list_alerts). In the case of machine learning that would be Geoffrey Hinton, Yann LeCun, Yoshua Bengio, Andrew Ng (this list is not exhaustive of course).

tlarkworthy• 12 years ago

get suggestions from more informed people (e.g reading groups, professors, news aggregators :p ).

work smart, not hard.

TazeTSchnitzel• 12 years ago

Is this something users of GNU MP will get to see the benefits of?

fdej2• 12 years ago

It's mostly a theoretical result. The previous FFT-based multiplication algorithms (including the one in GMP) all have almost linear complexity. For integers that can be stored within the observable universe, overhead and memory locality matter much more than nested logarithmic factors in the asymptotic complexity bound. It's conceivable that the new algorithm is superior in the real world, but nothing really suggests that it should be.

jekub2• 12 years ago

The improvement is only theoretical. The Fürer algorithm is quicker than Schönhage-Strassen only for impractical numbers and no arbitrary precision toolkit use it as far as I know.

The Fürer algorithm change the loglog term of the complexity to 2^log* which doesn't change anything in practice for workable numbers but it also add a lot to the constant term which is not counted by the big O notation but is very important in practice.

userbinator• 12 years ago

That reminds me of the fastest known (in terms of number of multiplications) matrix-multiplication algorithm, where the constant factor is so huge that it's of no practical use.

Even the more practical ones e.g. http://en.wikipedia.org/wiki/Strassen_algorithm trades 8 multiplications and 4 additions with 7 multiplications and 18 additions, won't be all that much faster on modern architectures where multiplication is often surprisingly fast (almost the same as addition) and how much data is read/written also matters (memory bandwidth and latency).

throwaway_yy2Di• 12 years ago

It's not likely; the authors checked,

    "We have implemented an unoptimised version of the algorithm
    from section 8 in the Mathemagix system [29] and found our
    implementation to be an order of magnitude slower than the
    Gmp library [23]. There is certainly room for improvement,
    but we doubt that even a highly optimised implementation of
    the new algorithm will be competitive in the near future."

As the other people point out, when an algorithm has asymptotic complexity improvements (like here), it's often much slower on small problem sizes (the constant factor), on real hardware (i.e. memory locality).

And check out the table of multiplication algorithms in the paper's introduction, and the table in the GMP manual [0]. The FFT (Schönhage–Strassen) GMP uses is O(n log n log log n), and the algorithm in this paper is O(n log n 8^(log* n)). Those functions [1] diverge a bit slowly: they have the same value at n=2^2^2^21 (2^2^2,097,152).

[0] https://gmplib.org/manual/Multiplication-Algorithms.html#Mul...

[1] https://en.wikipedia.org/wiki/Iterated_logarithm

taeric• 12 years ago

Having just gotten through the Knuth volume on this, I'm curious how this compares to many of the techniques he touches on.

In particular, I have to admit my mind was blown when he went over the "balanced ternary" method of representing numbers where multiplication was closer to addition.

Granted, I'm a complete outsider to this field, so I'm sure most of these techniques would blow my mind. Or, more likely, just go so far over what I'm used to seeing that it would be incomprehensible. Definitely exciting to grasp at, though.

Edit: I'm also more than a little curious/jealous for folks that have these sorts of concerns in what they do. The section on fast evaluation of polynomials, while ridiculously comprehensive and awesome, was so far from my field of work that it felt almost like I wasn't a programmer after a bit.

Edit2: I should say I did skim this paper, and I see they make references to Knuth's tightening of the bounds. So, I should clarify that I'm more interested in knowing roughly how this compares to all of methods given in the 3rd edition. And I'm mostly interested in whether this sees any usage and by whom.

Even faster integer multiplication

Related Stories