# Goldbach's conjecture algorithm

I did a Goldbach's Conjecture exercise and got it to work. It's pretty slow though and I was wondering how I could optimize it.

number = int(input("Enter your number >> "))
print("\nCalculating...")
if number % 2 == 0: #Only even numbers
primes = primenums(number) #returns all prime numbers <= input number

else:


Right now, up to 10,000 the algorithm is pretty fast, but at 100,000 it takes about 3 seconds to finish. Is that just how it is or could I make it faster?

• Have you profiled this? How is primenums implemented? – Graipher Feb 13 '19 at 8:23
• Profiled? What do you mean? – Kevin Feb 13 '19 at 13:46
• @EricLippert Don't worry about language/compiler/interpreter optimization before you do mathematical/complexity optimization first. Switching languages can only improve performance by a small factor (1-10). Changing the algorithm can give speedups of 1-infinity. – Vortico Feb 13 '19 at 18:10
• @Vortico are you seriously lecturing Eric Lippert on performance gains and optimization? – MikeTheLiar Feb 13 '19 at 19:57
• @MikeTheLiar I know he knows, but someone reading his comment may be misled. It's bad advice to recommend to a programming beginner to change the tool he's learning before even pointing out the issue with his algorithm. It's discouraging, doesn't solve the problem at hand (because maybe his goal is to simply learn Python, and remember this is a code review community), and doesn't teach how to think on one's own. It's not "The first problem" with his code as he said. It's the 26th or maybe 126th problem he should worry about while playing around with Goldbach's conjecture. – Vortico Feb 13 '19 at 20:31

One thing that makes your code slow is your repeated calls to primes.index. Each of these calls needs to linearly scan primes until it finds the value you are looking for, making this $$\\mathcal{O}(n)\$$. This is especially bad in the first if case since you do this twice with exactly the same argument. Instead just keep the index around in addition to the number:

def goldbach_keep_index(number):
if number % 2 == 1: #Only even numbers
raise ValueError("Goldbach conjecture is only defined for even numbers")
primes = primenums(number) #returns all prime numbers <= input number
i, j = 0, 0

i = j = i + 1
else:
j += 1


Note that I made this a function, so you can reuse it.

Your main calling code could look like this:

if __name__ == "__main__":
number = int(input("Enter your number >> "))
p1, p2 = goldbach_keep_index(number)
print(f"{p1} + {p2} = {number}")


However, what you are really doing is taking all combinations of prime numbers, without repeating yourself. So just use itertools.combinations_with_replacement:

from itertools import combinations_with_replacement

def goldbach_combinations(number):
if number % 2 == 0:
raise ValueError("Goldbach conjecture is only defined for even numbers")
raise Exception(f"Found a counter-example to the Goldbach conjecture: {number}")


In this case primes does not even need to be a list, it can just be a generator.

If you instead make primes a set, you can easily implement the idea suggested by @Josiah in their answer by just checking if number - p is in primes:

def goldbach_set(number):
if number % 2 == 0: #Only even numbers
raise ValueError("Goldbach conjecture is only defined for even numbers")
primes = set(primenums(number)) #returns all prime numbers <= input number
for p in primes:
k = number - p
if k in primes:
return p, k
raise Exception(f"Found a counter-example to the Goldbach conjecture: {number}")


And now some timing comparisons, where goldbach is your code put into a function:

def goldbach(number):
if number % 2 == 0: #Only even numbers
raise ValueError("Goldbach conjecture is only defined for even numbers")
primes = primenums(number) #returns all prime numbers <= input number

else:


And primenums is a simple sieve:

def prime_sieve(limit):
prime = [True] * limit
prime = prime = False

for i, is_prime in enumerate(prime):
if is_prime:
yield i
for n in range(i * i, limit, i):
prime[n] = False

return list(prime_sieve(limit)) And when pulling the generation of the primes outside of the function (and calculating them up to the largest number in the plot): Of course the first three functions are vastly slower than the set because they need to go through more combinations now. However, all functions are then constant time, so the increase in time comes solely from the fact that you have to consider more primes.

• I'm intrigued by the last graph. There is an obvious oscillating pattern where there's a sudden drop in run time even when the size of the input increases. I wonder what's going in there – DeepSpace Feb 14 '19 at 12:06
• I wondered that. My guess is fixed probe numbers are used, and some happen to have a larger small addend needed to satisfy the conjecture, so they take more loop iterations. – Josiah Feb 14 '19 at 20:35
• @Josiah: That is indeed what I did (x = np.logspace(1, 5, dtype=int); x[x % 2 == 1] += 1). And while larger numbers indeed tend to have more pairs of prime numbers that sum to them, there is quite a lot of variation: en.wikipedia.org/wiki/Goldbach%27s_conjecture#/media/… One might be able to argue that this rise is slightly visible above 10^4 in my plot for the first functions, but since the OP's code took too long I could no go up much higher in numbers... – Graipher Feb 15 '19 at 8:05

As suggested by Graipher in the comment, you're probably spending a lot of time generating a list of primes. That would be worth profiling, but I can't speculate as to how it might be improved without at least seeing the code.

There are two things that stand out to me as very clear performance sinks.

The first, and easiest to change, is index. If you call index on a list, python has to check every element in the list one at a time to see whether it's the thing you're after. If you have millions of primes, every one of those index calls is secretly a hidden loop that runs millions of times. The good news is you can resolve this easily enough: just keep another variable which stores where in the list of primes you're up to. Then you don't have to search for your place again every time you go round the loop, because you have a bookmark to it.

Second, and slightly more subtle, is the way that this is looping. You are looking for two numbers that add up to a target. If your target is 100, you might first say "what if it's 3?" and then try all primes with 3. Then you say "what if it's 5?" and try all the primes with 5. And then you do the same with 7, and so on. This means your trying all the primes with all the primes.

Instead, if you are checking whether the first prime is 3, you know that the second prime has to be 97. So all you have to do is check whether 97 is prime. If so, it's a match, if not, you can move on to checking 5 and 95, and then 7 and 93, and so on. In this way your algorithm checks each prime against one other number instead of each prime against all the others, and should be massively faster.

• Your second suggestion is great indeed. I will implement it that way! Thanks – Kevin Feb 13 '19 at 14:03
• You can make this even more efficient by keeping track of "near misses" to your target which eliminate other integers. Suppose you were testing 102. You start by looking for 3 + some prime. 99 isn't a prime, but the nearest prime >= 99 is 101, so if you organize the search carefully, you can remember that you just found a prime pair for 104 = 3 + 99 while looking for a pair for 102, and the next even number you need to test is not 104 but 106 (unless you already found a pair for that as well. of course). – alephzero Feb 13 '19 at 17:10
• I would default to using a set as in @Graipher's implementation, which is very fast at discovering whether something is present, but does not tell you what the nearest miss was. – Josiah Feb 13 '19 at 20:18
• If I had to use a regular list, and if the list was sorted, then there is indeed a clever trick available. The idea is that as the small prime being tested grows, the number it needs to be added to shrinks. That means that you can keep two bookmark variables pointing at the list of primes, one at the start and moving forwards, and one at the end and moving backwards. As long as sum is too small, you move the first bookmark along. If the sum is too big, you bring the second bookmark backwards. This avoids having to scan the whole list each time. However, it's quite fiddly to get right. – Josiah Feb 13 '19 at 20:25

Let's take n = 1,000,000,000 for example and see what your code does.

It calculates all primes from 1 to 1 billion. That takes a while, but it gives you an array of all primes in sorted order.

It then calculates 2 + 2, 2 + 3, 2 + 5, 2 + 7, 2 + 999,999,xxx to check if one of these numbers equals 1,000,000,000. But obviously when addend1 = 2, addend2 has to be 999,999,998 to add to one billion, so you are checking tens of millions of sums unnecessarily.

It then calculates 3 + 3, 3 + 5, 3 + 7 etc., and again, addend2 would have to be 999,999,997 million. And then again for addend1 = 5, 7, 11 etc. So you are checking a huge number of sums needlessly.

Start with addend1 = first prime, addend2 = last prime in your array. If their sum is too small, replace addend1 with the next prime to make the sum larger by the smallest possible amound. If their sum is too large, replace addend2 with the previous prime. If the sum is right, you have found a solution. And once you reached addend1 > addend2, you know there is no solution. This will be very quick, since usually you don't need to check too many values for addend1.

That makes the search a lot quicker, but doesn't help with the sieve trying to find all the primes. You usually don't need all the primes, only a very small number. In the example with n = 1,000,000,000 you probably find two primes adding up to a billion with addend1 ≤ 1000 and addend2 ≥ 999,999,000.

So here's what you do: To find all primes say in the range 999,999,000 to 1,000,000,000, you need to check if these numbers are divisible by any prime up to the square root of 1,000,000,000 which is a bit less than 32,000. So first you use a sieve to find all numbers up to 32,000. Then you create a sieve that finds all primes from 999,999,000 to 1,000,000,000. You run your search algorithm until it tries to examine primes that are not in this range. This likely doesn't happen, but if it happens, you create another sieve for the numbers 999,998,000 to 999,999,000 and so on. So instead of maybe 50 million primes, you look only for maybe 50 or 100. Again, that makes it a lot faster.

• I didnt learn about sieves yet, gotta look into it – Kevin Feb 14 '19 at 5:52
• It might be worth testing the primality of the larger number as you hit it, rather than sieving a whole range. – Josiah Feb 14 '19 at 7:54