This is an example of multiplying a big integer to a list of integers.

3792538124902347509274019274102947 * {2352, 4981, 6492, 3469, 2093, 1049, 5482, 5932, 1923, 7912}


Note that 3792538124902347509274019274102947 is 34-digit number.
Each element in the list is 4-digit number and there are 10 elements in the list.

To generate the output, we had done

1-digit multiply 1-digit : 34*4*10==1360 times
34~37-digit plus 34~37-digit : 3*10 times
(to add four numbers, we plus three times)

Suppose that we have following multiplication table before :


Then we need
1-digit multiply 1-digit : 0 times
34~37-digit plus 34~37-digit : 3*10 times

So we save 1360 1-digit multiplications in this case.

But I know there are things to be considered carefully :

  1. To use the multiplication table, we have to substitute.
    For example, we have to substitute
    3792538124902347509274019274102947*5 with 18962690624511737546370096370514735.
    Seeking for the multiplication table or substitution itself is also a kind of workload.

  2. It is possible that mathematica's built-in feature may already calculated so. I mean mathematica automatically uses optimized method for multiplying a single integer
    and a list of integers.

  3. CPU works in base 2. My theory is about base 10.. that difference is a little worrisome.

Q) Do you agree that when multiplying
a fixed very big integer and very long list of integers,
it is good to build multiplication table first?

  • $\begingroup$ For your particular application, instead of handling the algorithm yourself, have you tried CUDA? $\endgroup$
    – Syed
    Feb 21, 2022 at 9:35
  • $\begingroup$ Thank you, I haven't. But I heard from somewhere. I'll try! $\endgroup$
    – imida k
    Feb 21, 2022 at 10:59
  • 1
    $\begingroup$ The short answer is no. I'd believe it only if I saw a representative example of this alternative with timings to show it was faster. $\endgroup$ Feb 21, 2022 at 18:53

2 Answers 2

  1. In binary format you would not need a lookup table because one simply decomposes one of the factors into powers of 2. Then multiplying by powers of 2 is just about shifting digits to the left. The actually intense part is the accumulation of the intermediate results.

  2. You assume that the integer multiplication is carried out digit per digit on the machine. I don't know about the details, but I m pretty sure that quite a lot concurrency is built into the circuits for integer arithmetic so that the runtime depends only on the register length and not on the actual number of digits of the factors. (Well, if both factors are too large, then maybe the handling of integer overflow may induce a certain extra cost.)

  3. With 39 decimal digits the result is too large for a machine integer register (nowadays often 64 bit). That means that the multiplication has to be handled in software. But then again, I doubt that the lookup table would help anything because the most costly part remains: the addition. I guess a good strategy would be to represent the numbers by a couple of 64 bit or 32 bit (unsigned?) integers, use the hardware to accelerate the multiplication of these parts and somehow handle the carry-overs. But instead of rolling out my own multiplication routine for long integers, I'd rather rely on people at Wolfram Research having done a good job.

  • $\begingroup$ Indeed, biginteger math is nowhere near as bad as using decimal digits as the chunk size; usually it's the register width of the machine, e.g. 64 bits on machines like x86-64 and AArch64. See Did the 2019 discovery of O(N log(N)) multiplication have a practical outcome? - So this is several Nx1 chunk multiplications, since integers in the list are small. I'd assume Mathematica handles this efficiently internally; hopefully with just a couple mul / add instructions (and maybe an adc) each if it special cases 2-chunks times 1-chunk. $\endgroup$ Feb 21, 2022 at 23:42

I expect that Wolfram Research has put more effort into efficiently multiplying big integers than any of us will. It's always going to be quicker to do math with "small" integers (fewer than 64 bits) than large ones. So I suggest reducing the number of big integer operations as much as possible.

To do so, multiply the four-digit integers together until the result is close to 64 bits. In most cases, that will mean multiplying four or five four-digit integers. Given the numbers you listed, here is what I am suggesting:

2352 * 4981 * 6492 * 3469 = 263837589293376
2093 * 1049 * 5482 * 5932 = 71397809887768
1923 * 7912 = 15214776

Then multiply the big integer by 263837589293376, the result of that by 71397809887768, and the result of that by 15214776. This method yields an answer with only three big integer operations. I suspect that you can find two groups of five of the four-digit integers where the product of each group is less than 2**64. Doing so would reduce the number of big integer operations to two.

You could also pull out all the factors of 2 from each of the small integers. Do your multiplications with all the remaining odd numbers and, at the end, multiply by the appropriate power of 2. Multiplying by a power of 2 should be fast because all you're doing is shifting bits but it depends on the internal representation of big integers.

  • $\begingroup$ Your answer is about multiplying all the numbers altogether. My question is not about it, but thank you for the information. $\endgroup$
    – imida k
    Feb 22, 2022 at 5:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.