Does this prime function actually work?
Question:
Since I’m starting to get the hang of Python, I’m starting to test my newly acquired Python skills on some problems on projecteuler.net.
Anyways, at some point, I ended up making a function for getting a list of all primes up until a number ‘n’.
Here’s how the function looks atm:
def primes(n):
"""Returns list of all the primes up until the number n."""
# Gather all potential primes in a list.
primes = range(2, n + 1)
# The first potential prime in the list should be two.
assert primes[0] == 2
# The last potential prime in the list should be n.
assert primes[-1] == n
# 'p' will be the index of the current confirmed prime.
p = 0
# As long as 'p' is within the bounds of the list:
while p < len(primes):
# Set the candidate index 'c' to start right after 'p'.
c = p + 1
# As long as 'c' is within the bounds of the list:
while c < len(primes):
# Check if the candidate is divisible by the prime.
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed.
primes.pop(c)
# Move on to the next candidate and redo the process.
c = c + 1
# The next integer in the list should now be a prime,
# since it is not divisible by any of the primes before it.
# Thus we can move on to the next prime and redo the process.
p = p + 1
# The list should now only contain primes, and can thus be returned.
return primes
It seems to work fine, although one there’s one thing that bothers me.
While commenting the code, this piece suddenly seemed off:
# Check if the candidate is divisible by the prime.
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed from the list.
primes.pop(c)
# Move on to the next candidate and redo the process.
c += 1
If the candidate IS NOT divisible by the prime we examine the next candidate located at ‘c + 1’. No problem with that.
However, if the candidate IS divisible by the prime, we first pop it and then examine the next candidate located at ‘c + 1’.
It struck me that the next candidate, after popping, is not located at ‘c + 1’, but ‘c’, since after popping at ‘c’, the next candidate “falls” into that index.
I then thought that the block should look like the following:
# If the candidate is divisible by the prime:
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed from the list.
primes.pop(c)
# If not:
else:
# Move on to the next candidate.
c += 1
This above block seems more correct to me, but leaves me wondering why the original piece apparently worked just fine.
So, here are my questions:
After popping a candidate which turned out not be a prime, can we assume, as it is in my original code, that the next candidate is NOT divisible by that same prime?
If so, why is that?
Would the suggested “safe” code just do unnecessary checks on the candidates which where skipped in the “unsafe” code?
PS:
I’ve tried writing the above assumption as an assertion into the ‘unsafe’ function, and test it with n = 100000. No problems occurred. Here’s the modified block:
# If the candidate is divisible by the prime:
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed.
primes.pop(c)
# If c is still within the bounds of the list:
if c < len(primes):
# We assume that the new candidate at 'c' is not divisible by the prime.
assert primes[c] % primes[p] != 0
# Move on to the next candidate and redo the process.
c = c + 1
Answers:
If prime p divides candidate c, then the next larger candidate that is divisible by p is c + p. Therefore, your original code is correct.
However, it’s a rotten way to produce a list of primes; try it with n = 1000000 and see how slow it gets. The problem is that you are performing trial division when you should be using a sieve. Here’s a simple sieve (pseudocode, I’ll let you do the translation to Python or another language):
function primes(n)
sieve := makeArray(2..n, True)
for p from 2 to n step 1
if sieve[p]
output p
for i from p+p to n step p
sieve[i] := False
That should get the primes less than a million in less than a second. And there are other sieve algorithms that are even faster.
This algorithm is called the Sieve of Eratosthenes, and was invented about 2200 years ago by a Greek mathematician. Eratosthenes was an interesting fellow: besides sieving for primes, he invented the leap day and a system of latitude and longitude, accurately calculated the distance from Sun to Earth and the circumference of the Earth, and was for a time the Chief Librarian of Ptolemy’s Library in Alexandria.
When you are ready to learn more about programming with prime numbers, I modestly recommend this essay at my blog.
Your observation seems to be accurate, which is quite a good catch.
I suspect the reason that it works, at least in some cases, is because composite numbers are actually factored into multiple primes. So, the inner loop may miss the value on the first factor, but it then picks it up on a later factor.
For a small’ish “n”, you can print out values of the list to see if this is what is happening.
This method of finding primes, by the way, is based on the Sieve of Eratothenes. It is possible when doing the sieve that if “c” is a multiple of “p”, then the next value is never a multiple of the same prime.
The question is: are there any cases where all values between p*x and p*(x+1) are divisible by some prime less than p and p*x+1). (This is where the algorithm would miss a value and it would not be caught later.) However, one of these values is even, so it would be eliminated on round “2”. So, the real question is whether there are cases where all values between p*x and p*(x+2) are divisible by numbers less than p.
Off hand, I can’t think of any numbers less than 100 that meet this condition. For p = 5, there is always a value that is not divisible by 2 or 3 between two consecutive multiples of 5.
There seems to be a lot written on prime gaps and sequences, but not so much on sequences of consecutive integers divisible by numbers less than p. After some (okay, a lot) of trial and error, I’ve determined that every number between 39,474 (17*2,322) and 39,491 (17*2,233) is divisible by an integer less than 17:
39,475 5
39,476 2
39,477 3
39,478 2
39,479 11
39,480 2
39,481 13
39,482 2
39,483 3
39,484 2
39,485 5
39,486 2
39,487 7
39,488 2
39,489 3
39,490 2
I am not sure if this is the first such value. However, we would have to find sequences twice as long as this. I think that is unlikely, but not sure if there is a proof.
My conclusion is that the original code might work, but that your fix is the right thing to do. Without a proof that there are no such sequences, it looks like a bug, albeit a bug that could be very, very, very rare.
- Given two numbers n, m in the consecutive sequence of possible primes such that n and m are not divisible by the last divisor p, then m – n < p
-
Given q (the next higher divisor) > p, then if n is divisible by q, then the next number divisible by q is n + q > n + p > m
so m should be skipped in the current iteration for divisibility test
Here n = primes[c]
m = primes[c + 1], i.e. primes[c] after primes.pop(c)
p = primes[p]
q = primes[p+1]
Here is an idea:
Triptych explained1 that the next number after c cannot be c + p, but we still need to show that it can also never be c + 2p.
If we use primes = [2], we can only have one consecutive “non-prime”, an number divisible by 2.
If we use primes = [2,3] we can construct 3 consecutive “non-primes”, a number divided by 2, a number divided by three, and a number divided by 2, and they cannot get the next number. Or
2,3,4 => 3 consecutive “non-primes”
Even though 2 and 3 are not “non-primes” it is easier for me to think in terms of those numbers.
If we use [2,3,5], we get
2,3,4,5,6 => 5 consecutive “non-primes”
If we use [2,3,5,7], we get
2,3,4,5,6,7,8,9,10 => 9 consecutive “non-primes”
The pattern emerges. The most consecutive non-primes that we can get is next prime – 2.
Therefore, if next_prime < p * 2 + 1, we have to have at least some number between c and c + 2p, because number of consecutive non-primes is not long enough, given the primes yet.
I don’t know about very very big number, but I think this next_prime < p * 2 + 1 is likely to hold very big numbers.
I hope this makes sense, and adds some light.
1 Triptych‘s answer has been deleted.
This doesn’t provide a remotely conclusive answer, but here’s what I’ve tried on this:
I’ve restated the required assumption here as (lpf stands for Least Prime Factor):
For any composite number, x, where:
lpf(x) = n
There exists a value, m, where 0 < m < 2n and:
lpf(x+m) > n
It can be easily demonstrated that values for x exist where no composite number (x+m), exists to satisfy the inequality. Any squared prime demonstrates that:
lpf(x) = x^.5, so x = n^2
n^2 + 2n < (n + 1)^2 = n^2 + 2n + 1
So, in the case of any squared prime, for this to hold true, there must be a prime number, p, present in the range x < p < x + 2n.
I think that can be concluded given the asymptotic distribution of squares (x^.5) compared to the the Prime Number Theorem (asymptotic distribution of primes approx. x/(ln x)), though, really, my understanding of the Prime Number Theorem is limited at best.
And I have no strategy whatsoever for extending that conclusion to non-square composite numbers, so that may not be a useful avenue.
I’ve put together a program testing values using the above restatement of the problem.
Test this statement directly should remove any got-lucky results from just running the algorithm as stated. By got-lucky results, I’m referring to a value being skipped that may not be safe, but that doesn’t turn up any incorrect results, due to a skipped value not being divisible by the number currently being iterated on, or being picked up by subsequent iterations. Essentially, if the algorithm gets the correct result, but either doesn’t find the LEAST prime factor of each eliminated value, or doesn’t rigorously check each prime result, I’m not satisfied with it. If such cases exist, I think it’s reasonable to assume that cases also exist where it would not get lucky (unusual though they may be), and would render an incorrect result.
Running my test, however, shows no counter-examples in the values from 2 – 2,000,000. So, for what it’s worth, values from the algorithm as stated should be safe up to, at least, 2,000,000, unless my logic is incorrect.
That’s what I have to add. Great question, Phazyck, had fun with it!
It fails for much bigger numbers. The first prime is 71, for that the candidate can fail. The smallest failing candidate for 71 is 10986448536829734695346889 which overshadows the number 10986448536829734695346889 + 142.
def primes(n, skip_range=None):
"""Modified "primes" with the original assertion from P.S. of the question.
with skipping of an unimportant huge range.
>>> primes(71)
[2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71]
>>> # The smallest failing number for the first failing prime 71:
>>> big_n = 10986448536829734695346889
>>> primes(big_n + 2 * 71, (72, big_n))
Traceback (most recent call last):
AssertionError
"""
if not skip_range:
primes = list(range(2, n + 1))
else:
primes = list(range(2, skip_range[0]))
primes.extend(range(skip_range[1], n + 1))
p = 0
while p < len(primes):
c = p + 1
while c < len(primes):
if(primes[c] % primes[p] == 0):
primes.pop(c)
if c < len(primes):
assert primes[c] % primes[p] != 0
c = c + 1
p = p + 1
return primes
# Verify that it can fail.
aprime = 71 # the first problematic prime
FIRST_BAD_NUMBERS = (
10986448536829734695346889, 11078434793489708690791399,
12367063025234804812185529, 20329913969650068499781719,
30697401499184410328653969, 35961932865481861481238649,
40008133490686471804514089, 41414505712084173826517629,
49440212368558553144898949, 52201441345368693378576229)
for bad_number in FIRST_BAD_NUMBERS:
try:
primes(bad_number + 2 * aprime, (aprime + 1, bad_number))
raise Exception('The number {} should fail'.format(bad_number))
except AssertionError:
print('{} OK. It fails as is expected'.format(bad_number))
I solved these numbers by a complicated algorithm like a puzzle by searching possible remainders of n modulo small primes. The last simple step was to get the complete n (by chinese remainder theorem in three lines of Python code). I know all 120 basic solutions smaller than primorial(71) = 2 * 3 * 5 * 7 * 11 * 13 * 17 * 19 * 23 * 29 * 31 * 37 * 41 * 43 * 47 * 53 * 59 * 61 * 67 * 71 repeated periodically by all multiples of this number. I rewrote the algorithm many times for every decade of tested primes because for every decade was the solution much slower than for the previous. Maybe I find a smaller solution with the same algorithm for primes 73 or 79 in acceptable time.
Edit:
I would like to find also a complete silent fail of the unsafe original function. Maybe exists some candidate composed from different primes. This way of solution would only postpone the final outcome for later. Every step would be much more and more expensive for time and resources. Therefore only numbers composed from one or two primes are attractive.
I expect that only two solutions the hidden candidate c are good: c = p ** n or c = p1 * p ** n or c = p1 ** n1 * p ** n where p and p1 are primes and n is a power greater than 1. The primes function fails if c - 2 * p is divisible by no prime smaller than p and if all number between c-2n and c are divisible by any prime smaller than p. The variant p1*p**n requires also that the same c had failed before for p1 (p1 < p) as we already know infinite number of such candidates.
EDIT: I found a smaller example of failure: number 121093190175715194562061 for the prime 79. (which is about ninety times less than for 71) I can’t continue by the same algorithm to find smaller examples because all 702612 basic solutions took more than 30 hours for the prime 79 on my laptop.
I also verified it for all candidates smaller than 400000000 (4E10) and for all relevant primes, that no candidate will fail the assertion in the question. Until you have terabytes of memory and thousands years of time, the assertion in the algorithm will pass, because your time complexity is O((n / log(n)) ^2) or very similar.
This program does not work correctly, i.e., it incorrectly reports a composite number as prime. It turns out to have the same bug as a program by Wirth. The details may be found in Paul Pritchard, Some negative results concerning prime number generators, Communications of the ACM, Vol. 27, no. 1, Jan. 1984, pp. 53–57. This paper gives a proof that the program must fail, and also exhibits an explicit composite which it reports as prime.
Since I’m starting to get the hang of Python, I’m starting to test my newly acquired Python skills on some problems on projecteuler.net.
Anyways, at some point, I ended up making a function for getting a list of all primes up until a number ‘n’.
Here’s how the function looks atm:
def primes(n):
"""Returns list of all the primes up until the number n."""
# Gather all potential primes in a list.
primes = range(2, n + 1)
# The first potential prime in the list should be two.
assert primes[0] == 2
# The last potential prime in the list should be n.
assert primes[-1] == n
# 'p' will be the index of the current confirmed prime.
p = 0
# As long as 'p' is within the bounds of the list:
while p < len(primes):
# Set the candidate index 'c' to start right after 'p'.
c = p + 1
# As long as 'c' is within the bounds of the list:
while c < len(primes):
# Check if the candidate is divisible by the prime.
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed.
primes.pop(c)
# Move on to the next candidate and redo the process.
c = c + 1
# The next integer in the list should now be a prime,
# since it is not divisible by any of the primes before it.
# Thus we can move on to the next prime and redo the process.
p = p + 1
# The list should now only contain primes, and can thus be returned.
return primes
It seems to work fine, although one there’s one thing that bothers me.
While commenting the code, this piece suddenly seemed off:
# Check if the candidate is divisible by the prime.
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed from the list.
primes.pop(c)
# Move on to the next candidate and redo the process.
c += 1
If the candidate IS NOT divisible by the prime we examine the next candidate located at ‘c + 1’. No problem with that.
However, if the candidate IS divisible by the prime, we first pop it and then examine the next candidate located at ‘c + 1’.
It struck me that the next candidate, after popping, is not located at ‘c + 1’, but ‘c’, since after popping at ‘c’, the next candidate “falls” into that index.
I then thought that the block should look like the following:
# If the candidate is divisible by the prime:
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed from the list.
primes.pop(c)
# If not:
else:
# Move on to the next candidate.
c += 1
This above block seems more correct to me, but leaves me wondering why the original piece apparently worked just fine.
So, here are my questions:
After popping a candidate which turned out not be a prime, can we assume, as it is in my original code, that the next candidate is NOT divisible by that same prime?
If so, why is that?
Would the suggested “safe” code just do unnecessary checks on the candidates which where skipped in the “unsafe” code?
PS:
I’ve tried writing the above assumption as an assertion into the ‘unsafe’ function, and test it with n = 100000. No problems occurred. Here’s the modified block:
# If the candidate is divisible by the prime:
if(primes[c] % primes[p] == 0):
# If it is, it isn't a prime, and should be removed.
primes.pop(c)
# If c is still within the bounds of the list:
if c < len(primes):
# We assume that the new candidate at 'c' is not divisible by the prime.
assert primes[c] % primes[p] != 0
# Move on to the next candidate and redo the process.
c = c + 1
If prime p divides candidate c, then the next larger candidate that is divisible by p is c + p. Therefore, your original code is correct.
However, it’s a rotten way to produce a list of primes; try it with n = 1000000 and see how slow it gets. The problem is that you are performing trial division when you should be using a sieve. Here’s a simple sieve (pseudocode, I’ll let you do the translation to Python or another language):
function primes(n)
sieve := makeArray(2..n, True)
for p from 2 to n step 1
if sieve[p]
output p
for i from p+p to n step p
sieve[i] := False
That should get the primes less than a million in less than a second. And there are other sieve algorithms that are even faster.
This algorithm is called the Sieve of Eratosthenes, and was invented about 2200 years ago by a Greek mathematician. Eratosthenes was an interesting fellow: besides sieving for primes, he invented the leap day and a system of latitude and longitude, accurately calculated the distance from Sun to Earth and the circumference of the Earth, and was for a time the Chief Librarian of Ptolemy’s Library in Alexandria.
When you are ready to learn more about programming with prime numbers, I modestly recommend this essay at my blog.
Your observation seems to be accurate, which is quite a good catch.
I suspect the reason that it works, at least in some cases, is because composite numbers are actually factored into multiple primes. So, the inner loop may miss the value on the first factor, but it then picks it up on a later factor.
For a small’ish “n”, you can print out values of the list to see if this is what is happening.
This method of finding primes, by the way, is based on the Sieve of Eratothenes. It is possible when doing the sieve that if “c” is a multiple of “p”, then the next value is never a multiple of the same prime.
The question is: are there any cases where all values between p*x and p*(x+1) are divisible by some prime less than p and p*x+1). (This is where the algorithm would miss a value and it would not be caught later.) However, one of these values is even, so it would be eliminated on round “2”. So, the real question is whether there are cases where all values between p*x and p*(x+2) are divisible by numbers less than p.
Off hand, I can’t think of any numbers less than 100 that meet this condition. For p = 5, there is always a value that is not divisible by 2 or 3 between two consecutive multiples of 5.
There seems to be a lot written on prime gaps and sequences, but not so much on sequences of consecutive integers divisible by numbers less than p. After some (okay, a lot) of trial and error, I’ve determined that every number between 39,474 (17*2,322) and 39,491 (17*2,233) is divisible by an integer less than 17:
39,475 5
39,476 2
39,477 3
39,478 2
39,479 11
39,480 2
39,481 13
39,482 2
39,483 3
39,484 2
39,485 5
39,486 2
39,487 7
39,488 2
39,489 3
39,490 2
I am not sure if this is the first such value. However, we would have to find sequences twice as long as this. I think that is unlikely, but not sure if there is a proof.
My conclusion is that the original code might work, but that your fix is the right thing to do. Without a proof that there are no such sequences, it looks like a bug, albeit a bug that could be very, very, very rare.
- Given two numbers n, m in the consecutive sequence of possible primes such that n and m are not divisible by the last divisor p, then m – n < p
-
Given q (the next higher divisor) > p, then if n is divisible by q, then the next number divisible by q is n + q > n + p > m
so m should be skipped in the current iteration for divisibility testHere n = primes[c] m = primes[c + 1], i.e. primes[c] after primes.pop(c) p = primes[p] q = primes[p+1]
Here is an idea:
Triptych explained1 that the next number after c cannot be c + p, but we still need to show that it can also never be c + 2p.
If we use primes = [2], we can only have one consecutive “non-prime”, an number divisible by 2.
If we use primes = [2,3] we can construct 3 consecutive “non-primes”, a number divided by 2, a number divided by three, and a number divided by 2, and they cannot get the next number. Or
2,3,4 => 3 consecutive “non-primes”
Even though 2 and 3 are not “non-primes” it is easier for me to think in terms of those numbers.
If we use [2,3,5], we get
2,3,4,5,6 => 5 consecutive “non-primes”
If we use [2,3,5,7], we get
2,3,4,5,6,7,8,9,10 => 9 consecutive “non-primes”
The pattern emerges. The most consecutive non-primes that we can get is next prime – 2.
Therefore, if next_prime < p * 2 + 1, we have to have at least some number between c and c + 2p, because number of consecutive non-primes is not long enough, given the primes yet.
I don’t know about very very big number, but I think this next_prime < p * 2 + 1 is likely to hold very big numbers.
I hope this makes sense, and adds some light.
1 Triptych‘s answer has been deleted.
This doesn’t provide a remotely conclusive answer, but here’s what I’ve tried on this:
I’ve restated the required assumption here as (lpf stands for Least Prime Factor):
For any composite number, x, where:
lpf(x) = n
There exists a value, m, where 0 < m < 2n and:
lpf(x+m) > n
It can be easily demonstrated that values for x exist where no composite number (x+m), exists to satisfy the inequality. Any squared prime demonstrates that:
lpf(x) = x^.5, so x = n^2
n^2 + 2n < (n + 1)^2 = n^2 + 2n + 1
So, in the case of any squared prime, for this to hold true, there must be a prime number, p, present in the range x < p < x + 2n.
I think that can be concluded given the asymptotic distribution of squares (x^.5) compared to the the Prime Number Theorem (asymptotic distribution of primes approx. x/(ln x)), though, really, my understanding of the Prime Number Theorem is limited at best.
And I have no strategy whatsoever for extending that conclusion to non-square composite numbers, so that may not be a useful avenue.
I’ve put together a program testing values using the above restatement of the problem.
Test this statement directly should remove any got-lucky results from just running the algorithm as stated. By got-lucky results, I’m referring to a value being skipped that may not be safe, but that doesn’t turn up any incorrect results, due to a skipped value not being divisible by the number currently being iterated on, or being picked up by subsequent iterations. Essentially, if the algorithm gets the correct result, but either doesn’t find the LEAST prime factor of each eliminated value, or doesn’t rigorously check each prime result, I’m not satisfied with it. If such cases exist, I think it’s reasonable to assume that cases also exist where it would not get lucky (unusual though they may be), and would render an incorrect result.
Running my test, however, shows no counter-examples in the values from 2 – 2,000,000. So, for what it’s worth, values from the algorithm as stated should be safe up to, at least, 2,000,000, unless my logic is incorrect.
That’s what I have to add. Great question, Phazyck, had fun with it!
It fails for much bigger numbers. The first prime is 71, for that the candidate can fail. The smallest failing candidate for 71 is 10986448536829734695346889 which overshadows the number 10986448536829734695346889 + 142.
def primes(n, skip_range=None):
"""Modified "primes" with the original assertion from P.S. of the question.
with skipping of an unimportant huge range.
>>> primes(71)
[2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71]
>>> # The smallest failing number for the first failing prime 71:
>>> big_n = 10986448536829734695346889
>>> primes(big_n + 2 * 71, (72, big_n))
Traceback (most recent call last):
AssertionError
"""
if not skip_range:
primes = list(range(2, n + 1))
else:
primes = list(range(2, skip_range[0]))
primes.extend(range(skip_range[1], n + 1))
p = 0
while p < len(primes):
c = p + 1
while c < len(primes):
if(primes[c] % primes[p] == 0):
primes.pop(c)
if c < len(primes):
assert primes[c] % primes[p] != 0
c = c + 1
p = p + 1
return primes
# Verify that it can fail.
aprime = 71 # the first problematic prime
FIRST_BAD_NUMBERS = (
10986448536829734695346889, 11078434793489708690791399,
12367063025234804812185529, 20329913969650068499781719,
30697401499184410328653969, 35961932865481861481238649,
40008133490686471804514089, 41414505712084173826517629,
49440212368558553144898949, 52201441345368693378576229)
for bad_number in FIRST_BAD_NUMBERS:
try:
primes(bad_number + 2 * aprime, (aprime + 1, bad_number))
raise Exception('The number {} should fail'.format(bad_number))
except AssertionError:
print('{} OK. It fails as is expected'.format(bad_number))
I solved these numbers by a complicated algorithm like a puzzle by searching possible remainders of n modulo small primes. The last simple step was to get the complete n (by chinese remainder theorem in three lines of Python code). I know all 120 basic solutions smaller than primorial(71) = 2 * 3 * 5 * 7 * 11 * 13 * 17 * 19 * 23 * 29 * 31 * 37 * 41 * 43 * 47 * 53 * 59 * 61 * 67 * 71 repeated periodically by all multiples of this number. I rewrote the algorithm many times for every decade of tested primes because for every decade was the solution much slower than for the previous. Maybe I find a smaller solution with the same algorithm for primes 73 or 79 in acceptable time.
Edit:
I would like to find also a complete silent fail of the unsafe original function. Maybe exists some candidate composed from different primes. This way of solution would only postpone the final outcome for later. Every step would be much more and more expensive for time and resources. Therefore only numbers composed from one or two primes are attractive.
I expect that only two solutions the hidden candidate c are good: c = p ** n or c = p1 * p ** n or c = p1 ** n1 * p ** n where p and p1 are primes and n is a power greater than 1. The primes function fails if c - 2 * p is divisible by no prime smaller than p and if all number between c-2n and c are divisible by any prime smaller than p. The variant p1*p**n requires also that the same c had failed before for p1 (p1 < p) as we already know infinite number of such candidates.
EDIT: I found a smaller example of failure: number 121093190175715194562061 for the prime 79. (which is about ninety times less than for 71) I can’t continue by the same algorithm to find smaller examples because all 702612 basic solutions took more than 30 hours for the prime 79 on my laptop.
I also verified it for all candidates smaller than 400000000 (4E10) and for all relevant primes, that no candidate will fail the assertion in the question. Until you have terabytes of memory and thousands years of time, the assertion in the algorithm will pass, because your time complexity is O((n / log(n)) ^2) or very similar.
This program does not work correctly, i.e., it incorrectly reports a composite number as prime. It turns out to have the same bug as a program by Wirth. The details may be found in Paul Pritchard, Some negative results concerning prime number generators, Communications of the ACM, Vol. 27, no. 1, Jan. 1984, pp. 53–57. This paper gives a proof that the program must fail, and also exhibits an explicit composite which it reports as prime.