It does so by iteratively marking as composite (i.e., not prime) the multiples of each prime, starting with the first prime number, 2. The multiples of a given prime are generated as a sequence of numbers starting from that prime, with constant difference between them that is equal to that prime. This is the sieve's key distinction from using trial division to sequentially test each candidate number for divisibility by each prime. Once all the multiples of each discovered prime have been marked as composites, the remaining unmarked numbers are primes.
The earliest known reference to the sieve (Ancient Greek: κόσκινον Ἐρατοσθένους, kóskinon Eratosthénous) is in Nicomachus of Gerasa's Introduction to Arithmetic, an early 2nd cent. CE book, which describes it and attributes it to Eratosthenes of Cyrene, a 3rd cent. BCE Greek mathematician.
A prime number is a natural number that has exactly two distinct natural number divisors: the number 1 and itself.
To find all the prime numbers less than or equal to a given integer n by Eratosthenes' method:
The main idea here is that every value given to p will be prime, because if it were composite it would be marked as a multiple of some other, smaller prime. Note that some of the numbers may be marked more than once (e.g., 15 will be marked both for 3 and 5).
As a refinement, it is sufficient to mark the numbers in step 3 starting from p2, as all the smaller multiples of p will have already been marked at that point. This means that the algorithm is allowed to terminate in step 4 when p2 is greater than n.
Another refinement is to initially list odd numbers only, (3, 5, ..., n), and count in increments of 2p from p2 in step 3, thus marking only odd multiples of p. This actually appears in the original algorithm. This can be generalized with wheel factorization, forming the initial list only from numbers coprime with the first few primes and not just from odds (i.e., numbers coprime with 2), and counting in the correspondingly adjusted increments so that only such multiples of p are generated that are coprime with those small primes, in the first place.
To find all the prime numbers less than or equal to 30, proceed as follows.
First, generate a list of integers from 2 to 30:
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
The first number in the list is 2; cross out every 2nd number in the list after 2 by counting up from 2 in increments of 2 (these will be all the multiples of 2 in the list):
45 67 89 1011 1213 1415 1617 1819 2021 2223 2425 2627 2829 30
The next number in the list after 2 is 3; cross out every 3rd number in the list after 3 by counting up from 3 in increments of 3 (these will be all the multiples of 3 in the list):
45 67 8 9 1011 1213 14 15 1617 1819 20 21 2223 2425 26 27 2829 30
The next number not yet crossed out in the list after 3 is 5; cross out every 5th number in the list after 5 by counting up from 5 in increments of 5 (i.e. all the multiples of 5):
45 67 8 9 1011 1213 14 15 1617 1819 20 21 2223 24 25 26 27 2829 30
The next number not yet crossed out in the list after 5 is 7; the next step would be to cross out every 7th number in the list after 7, but they are all already crossed out at this point, as these numbers (14, 21, 28) are also multiples of smaller primes because 7 × 7 is greater than 30. The numbers not crossed out at this point in the list are all the prime numbers below 30:
2 3 5 7 11 13 17 19 23 29
algorithm Sieve of Eratosthenes is input: an integer n > 1. output: all prime numbers from 2 through n. let A be an array of Boolean values, indexed by integers 2 to n, initially all set to true. for i = 2, 3, 4, ..., not exceeding √ do if A[i] is true for j = i2, i2+i, i2+2i, i2+3i, ..., not exceeding n do A[j] := false return all i such that A[i] is true.
This algorithm produces all primes not greater than n. It includes a common optimization, which is to start enumerating the multiples of each prime i from i2. The time complexity of this algorithm is O(n log log n), provided the array update is an O(1) operation, as is usually the case.
As Sorenson notes, the problem with the sieve of Eratosthenes is not the number of operations it performs but rather its memory requirements. For large n, the range of primes may not fit in memory; worse, even for moderate n, its cache use is highly suboptimal. The algorithm walks through the entire array A, exhibiting almost no locality of reference.
If Δ is chosen to be √, the space complexity of the algorithm is O(√), while the time complexity is the same as that of the regular sieve.
For ranges with upper limit n so large that the sieving primes below √ as required by the page segmented sieve of Eratosthenes cannot fit in memory, a slower but much more space-efficient sieve like the sieve of Sorenson can be used instead.
An incremental formulation of the sieve generates primes indefinitely (i.e., without an upper bound) by interleaving the generation of primes with the generation of their multiples (so that primes can be found in gaps between the multiples), where the multiples of each prime p are generated directly by counting up from the square of the prime in increments of p (or 2p for odd primes). The generation must be initiated only when the prime's square is reached, to avoid adverse effects on efficiency. It can be expressed symbolically under the dataflow paradigm as
primes = [2, 3, ...] \ [[p², p²+p, ...] for p in primes],
Primes can also be produced by iteratively sieving out the composites through divisibility testing by sequential primes, one prime at a time. It is not the sieve of Eratosthenes but is often confused with it, even though the sieve of Eratosthenes directly generates the composites instead of testing for them. Trial division has worse theoretical complexity than that of the sieve of Eratosthenes in generating ranges of primes.
When testing each prime, the optimal trial division algorithm uses all prime numbers not exceeding its square root, whereas the sieve of Eratosthenes produces each composite from its prime factors only, and gets the primes "for free", between the composites. The widely known 1975 functional sieve code by David Turner is often presented as an example of the sieve of Eratosthenes but is actually a sub-optimal trial division sieve.
This section needs additional citations for verification.(June 2015)
The work performed by this algorithm is almost entirely the operations to cull the composite number representations which for the basic non-optimized version is the sum of the range divided by each of the primes up to that range or
where n is the sieving range in this and all further analysis.
By rearranging Mertens' second theorem, this is equal to n ( log log n + M ) as n approaches infinity, where M is the Meissel–Mertens constant of about 0.26149...
The optimization of starting at the square of each prime and only culling for primes less than the square root changes the "n" in the above expression to √ (or n1/2) and not culling until the square means that the sum of the base primes each minus two is subtracted from the operations. As the sum of the first x primes is 1/2x2 log x and the prime number theorem says that x is approximately x/log x, then the sum of primes to n is n2/2 log n, and therefore the sum of base primes to √ is 1/log n expressed as a factor of n. The extra offset of two per base prime is 2π(√), where π is the prime-counting function in this case, or 4√/log n; expressing this as a factor of n as are the other terms, this is 4/√ log n. Combining all of this, the expression for the number of optimized operations without wheel factorization is
For the wheel factorization cases, there is a further offset of the operations not done of
where x is the highest wheel prime and a constant factor of the whole expression is applied which is the fraction of remaining prime candidates as compared to the repeating wheel circumference. The wheel circumference is
and it can easily be determined that this wheel factor is
as p − 1/p is the fraction of remaining candidates for the highest wheel prime, x, and each succeeding smaller prime leaves its corresponding fraction of the previous combined fraction.
Combining all of the above analysis, the total number of operations for a sieving range up to n including wheel factorization for primes up to x is approximately
To show that the above expression is a good approximation to the number of composite number cull operations performed by the algorithm, following is a table showing the actually measured number of operations for a practical implementation of the sieve of Eratosthenes as compared to the number of operations predicted from the above expression with both expressed as a fraction of the range (rounded to four decimal places) for different sieve ranges and wheel factorizations (Note that the last column is a maximum practical wheel as to the size of the wheel gaps Look Up Table - almost 10 million values):
|n||no wheel||odds||2/3/5 wheel||2/3/5/7 wheel||2/3/5/7/11/13/17/19 wheel|
The above table shows that the above expression is a very good approximation to the total number of culling operations for sieve ranges of about a hundred thousand (105) and above.
The sieve of Eratosthenes is a popular way to benchmark computer performance. As can be seen from the above by removing all constant offsets and constant factors and ignoring terms that tend to zero as n approaches infinity, the time complexity of calculating all primes below n in the random access machine model is O(n log log n) operations, a direct consequence of the fact that the prime harmonic series asymptotically approaches log log n. It has an exponential time complexity with regard to input size, though, which makes it a pseudo-polynomial algorithm. The basic algorithm requires O(n) of memory.
The bit complexity of the algorithm is O(n (log n) (log log n)) bit operations with a memory requirement of O(n).
The normally implemented page segmented version has the same operational complexity of O(n log log n) as the non-segmented version but reduces the space requirements to the very minimal size of the segment page plus the memory required to store the base primes less than the square root of the range used to cull composites from successive page segments of size O(√/log n).
To show that the above approximation in complexity is not very accurate even for about as large as practical a range, the following is a table of the estimated number of operations as a fraction of the range rounded to four places, the calculated ratio for a factor of ten change in range based on this estimate, and the factor based on the log log n estimate for various ranges and wheel factorizations (the combo column uses a frequently practically used pre-cull by the maximum wheel factorization but only the 2/3/5/7 wheel for the wheel factor as the full factorization is difficult to implement efficiently for page segmentation):
|n||no wheel||odds||2/3/5 wheel||2/3/5/7 wheel||combo wheel||2/3/5/7/11/13/17/19 wheel|
The above shows that the log log n estimate is not very accurate even for maximum practical ranges of about 1016. One can see why it does not match by looking at the computational analysis above and seeing that within these practical sieving range limits, there are very significant constant offset terms such that the very slowly growing log log n term does not get large enough so as to make these terms insignificant until the sieving range approaches infinity – well beyond any practical sieving range. Within these practical ranges, these significant constant offsets mean that the performance of the Sieve of Eratosthenes is much better than one would expect just using the asymptotic time complexity estimates by a significant amount, but that also means that the slope of the performance with increasing range is steeper than predicted as the benefit of the constant offsets becomes slightly less significant.
One should also note that in using the calculated operation ratios to the sieve range, it must be less than about 0.2587 in order to be faster than the often compared sieve of Atkin if the operations take approximately the same time each in CPU clock cycles, which is a reasonable assumption for the one huge bit array algorithm. Using that assumption, the sieve of Atkin is only faster than the maximally wheel factorized sieve of Eratosthenes for ranges of over 1013 at which point the huge sieve buffer array would need about a quarter of a terabyte (about 250 gigabytes) of RAM memory even if bit packing were used. An analysis of the page segmented versions will show that the assumption that the time per operation stays the same between the two algorithms does not hold for page segmentation and that the sieve of Atkin operations get slower much faster than the sieve of Eratosthenes with increasing range. Thus for practical purposes, the maximally wheel factorized Sieve of Eratosthenes is faster than the Sieve of Atkin although the Sieve of Atkin is faster for lesser amounts of wheel factorization.
Using big O notation is also not the correct way to compare practical performance of even variations of the Sieve of Eratosthenes as it ignores constant factors and offsets that may be very significant for practical ranges: The sieve of Eratosthenes variation known as the Pritchard wheel sieve has an O(n) performance, but its basic implementation requires either a "one large array" algorithm which limits its usable range to the amount of available memory else it needs to be page segmented to reduce memory use. When implemented with page segmentation in order to save memory, the basic algorithm still requires about O(n/log n) bits of memory (much more than the requirement of the basic page segmented sieve of Eratosthenes using O(√/log n) bits of memory). Pritchard's work reduced the memory requirement to the limit as described above the table, but the cost is a fairly large constant factor of about three in execution time to about three quarters the sieve range due to the complex computations required to do so. As can be seen from the above table for the basic sieve of Eratosthenes, even though the resulting wheel sieve has O(n) performance and an acceptable memory requirement, it will never be faster than a reasonably Wheel Factorized basic sieve of Eratosthenes for any practical sieving range by a factor of about two. Other than that it is quite complex to implement, it is rarely practically implemented because it still uses more memory than the basic Sieve of Eratosthenes implementations described here as well as being slower for practical ranges. It is thus more of an intellectual curiosity than something practical.
Euler's proof of the zeta product formula contains a version of the sieve of Eratosthenes in which each composite number is eliminated exactly once. The same sieve was rediscovered and observed to take linear time by Gries & Misra (1978). It, too, starts with a list of numbers from 2 to n in order. On each step the first element is identified as the next prime, is multiplied with each element of the list (thus starting with itself), and the results are marked in the list for subsequent deletion. The initial element and the marked elements are then removed from the working sequence, and the process is repeated:
 (3) 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 67 69 71 73 75 77 79 ...  (5) 7 11 13 17 19 23 25 29 31 35 37 41 43 47 49 53 55 59 61 65 67 71 73 77 79 ...  (7) 11 13 17 19 23 29 31 37 41 43 47 49 53 59 61 67 71 73 77 79 ...  (11) 13 17 19 23 29 31 37 41 43 47 53 59 61 67 71 73 79 ... [...]
Here the example is shown starting from odds, after the first step of the algorithm. Thus, on the kth step all the remaining multiples of the kth prime are removed from the list, which will thereafter contain only numbers coprime with the first k primes (cf. wheel factorization), so that the list will start with the next prime, and all the numbers in it below the square of its first element will be prime too.
Thus, when generating a bounded sequence of primes, when the next identified prime exceeds the square root of the upper limit, all the remaining numbers in the list are prime. In the example given above that is achieved on identifying 11 as next prime, giving a list of all primes less than or equal to 80.
Note that numbers that will be discarded by a step are still used while marking the multiples in that step, e.g., for the multiples of 3 it is 3 × 3 = 9, 3 × 5 = 15, 3 × 7 = 21, 3 × 9 = 27, ..., 3 × 15 = 45, ..., so care must be taken dealing with this.
primes = sieve [2..]; sieve (p:nos) = p:sieve (remove (multsof p) nos); remove m = filter (not . m); multsof p n = rem n p==0). But see also Peter Henderson, Morris, James Jr., A Lazy Evaluator, 1976, where we find the following, attributed to P. Quarendon:
primeswrt[x;l] = if car[l] mod x=0 then primeswrt[x;cdr[l]] else cons[car[l];primeswrt[x;cdr[l]]] ; primes[l] = cons[car[l];primes[primeswrt[car[l];cdr[l]]]] ; primes[integers]; the priority is unclear.
Edited: 2021-06-18 17:49:52