mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Wagstaff PRP Search (https://www.mersenneforum.org/forumdisplay.php?f=102)
-   -   New Wagstaff PRP exponents (https://www.mersenneforum.org/showthread.php?t=18569)

 ryanp 2013-09-08 20:01

New Wagstaff PRP exponents

Hello,

I believe I have found the two largest known Wagstaff primes, after Tony Reix's discovery of (2^4031399+1)/3. Here they are:

[code](2^13347311 + 1)/3 is 3-PRP! (16355.1659s+0.0028s)[/code]

and:

[code](2^13372531 + 1)/3 is 3-PRP! (34165.4750s+0.0029s)[/code]

Each is a probable prime with about 4 million decimal digits.

Can anyone with some spare cycles help verify these using PFGW or other primality testing software?

Thanks!

-- Ryan

 Batalov 2013-09-08 20:37

Congratulations!

You can:
* try the Vrba-Reix as implemented in LLR (you have to modify llr.ini)
* also run pfgw with -b5, -b7 (and another dozen bases).
* and run pfgw with -t, -tp and -tc.

I'll run a few of these for you, in parallel.

 paulunderwood 2013-09-08 21:16

Wowee! :w00t:

 paulunderwood 2013-09-08 21:18

What ranges did you search, Ryan?

 diep 2013-09-08 23:17

Finding 2 close to each other, similar like how things work in Mersenne sometimes,

Maybe the odds for finding a Wagstaff in range [n;2n] which seemed a diverging sequence, so odds getting slowly a tad less each doubling of n, maybe maybe this is odds it's converging towards a near similar chance like one has to find a Mersenne.

Note that TF and P-1 rates of Wagstaff are considerable better than for Mersenne, so when i say 'converging towards' i still mean a considerable worse chance in range [n;2n], yet not as bad as the real small odds it seemed like considering the previous 2 were just under a million and something in the 4 million bits.

Moving towards the 4 million is factor 4+, then suddenly 2 at 13 million is factor 3, yet there is 2, where there is 2 there could be more.

So that's pretty good news, of course assuming both are PRP!

 ATH 2013-09-10 00:20

Prime95 concur on the first one:

2^13347311+1/3 is a probable prime! We4: B33A699A,00000000

 philmoore 2013-09-10 00:59

A Hearty congratulations to the project! Thanks especially to Tony Reix, Paul Underwood, and Vincent Diepeveen for testing so many candidates over the years. I am surprised that it took until now for Five or Bust to lose its lock on the largest known probable prime (since January 2009), but apparently the Wagstaff search hit a fairly substantial gap.

If these were proven primes, they would come in at 11th and 12th largest known, just below the 39th Mersenne and just above the last discovery of Seventeen or Bust.

[QUOTE=diep;352449]
Note that TF and P-1 rates of Wagstaff are considerable better than for Mersenne[/QUOTE]

Do you have any statistics supporting this? I'm not saying it is wrong, just that I would expect the rates to be very similar.

 diep 2013-09-10 01:14

[QUOTE=philmoore;352540]A Hearty congratulations to the project! Thanks especially to Tony Reix, Paul Underwood, and Vincent Diepeveen for testing so many candidates over the years. I am surprised that it took until now for Five or Bust to lose its lock on the largest known probable prime (since January 2009), but apparently the Wagstaff search hit a fairly substantial gap.

If these were proven primes, they would come in at 11th and 12th largest known, just below the 39th Mersenne and just above the last discovery of Seventeen or Bust.

Do you have any statistics supporting this? I'm not saying it is wrong, just that I would expect the rates to be very similar.[/QUOTE]

Jeff Gilchrist did do lately really a lot of work.

Interesting is to know what range was checked and what TF was done to find these 2 and whether a similar algorithm was used that was used to find some of the largest Mersennes past few years, prioritizing which exponent to use based upon factorisation of N+1 and N-1 of the exponent.

As for the TF and P-1, i tend to remember posts that Mersenne TF'ed roughly 50% and that P-1 removed roughly 7.5%. Please correct that if it's different.

For Wagstaff even a quick TF already removes 60% and with gpu's add another 10% and P-1 removes also 10%. The overlap of deeper TF and P-1 is not so large. So in total you look at a 70-75% that gets removed pretty easily with far less computational effort than has been done for Mersenne with respect to the P-1.

More accurate statistics will be there in some months hopefully with gpu factorisation stats.

It's been some years i datamined through Mersenne statistics there. Can't remember how shallow those were compared to what we're doing.

Note that Mersenne is a -1 formula and that Wagstaff is a +1 formula which should already explain a lot.

Mersenne give a reasonable steady number of primes in each given range just like 3 * 2^n - 1 also does.

Wagstaff so far was pretty much a gamble whether it would be converging or diverging or constant in odds to the next PRP.

I would guess blindfolded doing factorisation attempts other than P-1 to remove some composites from Wagstaffs list to be tested might be more succesful than for Mersenne.

In all cases and with respect to any statement the important emphasis is on the word "similar". Even if something factors 1% better i would not consider that similar.

 Jeff Gilchrist 2013-09-10 01:29

I just finished both exponents with LLR and they also show as PRP:

[CODE](2^13347311+1)/3 is Vrba-Reix PRP! Time : 56958.781 sec.
(2^13372531+1)/3 is Vrba-Reix PRP! Time : 57032.491 sec.
[/CODE]

Good find.

 ryanp 2013-09-10 01:46

[QUOTE=diep;352543]Jeff Gilchrist did do lately really a lot of work.

Interesting is to know what range was checked and what TF was done to find these 2 and whether a similar algorithm was used that was used to find some of the largest Mersennes past few years, prioritizing which exponent to use based upon factorisation of N+1 and N-1 of the exponent.[/QUOTE]

Yes, I am indebted to Jeff's prior work here as well. For my part, I started with the first 25,000 prime exponents from each of q=10e6, 11e6, 12e6 and 13e6. A large fraction of these were weeded out by a very basic program I wrote to do simple trial factoring up to d=1000, then many more by PFGW's own trial factoring. I'll have to do a bit of work to determine exactly how many exponents were fully tested by PFGW.

 Batalov 2013-09-10 02:25

Note that for the pfgw trial factoring you would want to use something like this command line:
[B]pfgw -f{13372607*2,1} -q(2^13372607+1)/3[/B]
[B](2^13372607+1)/3 has factors: 50253508240009[/B]

(this will only look for factors of form 2*p*k+1)
Compare the running time of the above to this:
[B]pfgw -f -q(2^13372607+1)/3[/B]
____________________________________

P.S. The Vrba-Reix tests and some LLR and PFGW tests in a bunch of bases on the same two exponents are almost done here, too.

 Jeff Gilchrist 2013-09-10 02:53

[QUOTE=ryanp;352547]Yes, I am indebted to Jeff's prior work here as well. For my part, I started with the first 25,000 prime exponents from each of q=10e6, 11e6, 12e6 and 13e6. A large fraction of these were weeded out by a very basic program I wrote to do simple trial factoring up to d=1000, then many more by PFGW's own trial factoring. I'll have to do a bit of work to determine exactly how many exponents were fully tested by PFGW.[/QUOTE]

Your part? Are you part of a larger group of people working on this? Has anyone worked on anything less than 10e6? What ranges are you actively working on now and how much CPU power are you throwing at this?

Jeff.

 T.Rex 2013-09-10 08:19

Congratulations !

Waowwww ! Two new Wagstaff PRPs !! :smile: :bow: :smile: :razz: :smile: :w00t: :smile:
That's tremendous, wonderful, extraordinary, etc. I'm missing (english) words !

The sad side of this nice story is that my own Wagstaff PRP now looks quite small... :sad:
Moreover, Diep, Jeff, and Paul spent a lot of time looking for such a Wagstaff PRP with no luck up to now. :sad: I hope that they will have some success some day. :smile: They deserve it.

So, now we need someone to provide a proof of Vrba-Reix conjecture ! :razz: so that PRPs become true primes ! I hope to see that before I die. Some years ago, I promissed 100€ for such a proof. Now, I would give 200€. Anyone to contribute and add some € or \$ to build a reward ? :wink:

On my side, I'm now making photographs:
[url]http://500px.com/Tony_Reix[/url]
[url]www.tonyfcr.book.fr/‎[/url]
that's fun too !

Regards

Tony

 ixfd64 2013-09-10 18:11

It looks like Ryan holds the current record for the ECM largest factor [I]and[/I] the largest Wagstaff PRP. Really impressive, to say the least. Now all he's missing is a new Mersenne prime. xD

Also, welcome back, Tony! :smile:

 ATH 2013-09-10 18:17

2nd exponent in Prime95:

2^13372531+1/3 is a probable prime! We4: E3ACC173,00000000

 paulunderwood 2013-09-10 18:57

Serge should be giving some PRP results soon.

I have started a GMP implementation of my test for the new Wagstaffs on a Q6600:
(L+2)^(N+1)==5 (mod N, L^2+1)

It will take a few weeks :smile:

 danaj 2013-09-10 19:08

Sounds like a race, Paul. I started my machine on SPSP-2, ES Lucas, and your Frobenius test yesterday for both numbers. As you say, it will take a long time as they're generic tests using GMP.

 paulunderwood 2013-09-10 19:15

I have a special version that calculates modulo 2^p+1 -- so no generic modulo. :smile:

 ATH 2013-09-10 21:44

[QUOTE=danaj;352638]Sounds like a race, Paul. I started my machine on SPSP-2, ES Lucas, and your Frobenius test yesterday for both numbers. As you say, it will take a long time as they're generic tests using GMP.[/QUOTE]

SPSP-2 will always pass for Wagstaff numbers?

2^p = -1 (mod 2^p+1) => 2^p = -1 (mod (2^p+1)/3)

 danaj 2013-09-11 04:18

[QUOTE=ATH;352655]SPSP-2 will always pass for Wagstaff numbers?[/quote]That is my understanding and I debated whether to start them or not, but I thought it wouldn't hurt and since it should finish first, will give me some idea of when the Lucas and Frobenius tests will be done. It'll be a long wait...

 Batalov 2013-09-11 16:31

[CODE](2^13347311+1)/3 is Base 27 - Strong Fermat PRP! Time : 132008.336 sec.
(2^13347311+1)/3 is Vrba-Reix PRP! Time : 131975.246 sec.
(2^13372531+1)/3 is Base 27 - Strong Fermat PRP! Time : 132322.407 sec.
(2^13372531+1)/3 is Vrba-Reix PRP! Time : 132513.794 sec.

(2^13347311+1)/3 is 5-PRP! (229303.3908s+398.4266s)
(2^13347311+1)/3 is 7-PRP! (277072.1587s+419.3222s)
(2^13347311+1)/3 is 11-PRP! (291376.7357s+419.8591s)
(2^13347311+1)/3 is 13-PRP! (291281.3062s+418.0968s)
(2^13347311+1)/3 is 17-PRP! (278100.7506s+382.4561s)

(2^13372531+1)/3 is 5-PRP! (249353.0860s+421.2026s)
(2^13372531+1)/3 is 7-PRP! (229612.6239s+399.1560s)
(2^13372531+1)/3 is 11-PRP! (230611.3771s+399.9541s)
(2^13372531+1)/3 is 13-PRP! (231212.1002s+401.6447s)
(2^13372531+1)/3 is 17-PRP! (230925.5544s+399.9358s)

#--- tests for false positives below ---
(2^13372309+1)/3 is not prime. RES64: 991A9DB47059EE26. OLD64: D0794CB28426C889 Time : 132114.403 sec.
(2^13372309+1)/3 is not prime. Vrba-Reix RES64: FE684FD81D62F060 Time : 132268.574 sec.
# -b5 (2^13372309+1)/3 is composite: RES64: [E9EDD6FC780DA64D] (225266.1438s+403.6897s)
# -b3 (2^13372309+1)/3 is composite: RES64: [FA860AF0EE31F99B] (230901.3211s+404.2113s)
# -b2 (2^13372309+1)/3 is 2-PRP! (225516.2154s+403.9142s)[/CODE]

 ATH 2013-09-19 22:58

[QUOTE]Running N+1 test using discriminant 2, base 1+sqrt(2)

Generic modular reduction using all-complex AVX FFT length 768K, Pass1=256, Pass
2=3K on (2^13372531+1)/3
Calling Brillhart-Lehmer-Selfridge with factored part 0.00%
(2^13372531+1)/3 is Lucas PRP! (851290.8327s+0.0578s)
[/QUOTE]

I lost the other test on (2^13347311+1)/3 on the way, 10 days is a long time without any intermediate savefiles. Several other people seems to have these test running on faster computers than mine though, so I didn't restart it.

 paulunderwood 2013-09-19 23:57

[QUOTE=ATH;353497]I lost the other test on (2^13347311+1)/3 on the way, 10 days is a long time without any intermediate savefiles. Several other people seems to have these test running on faster computers than mine though, so I didn't restart it.[/QUOTE]

:tu: I've got 'em covered. One more day to go on a 4770k at 4GHz.

 paulunderwood 2013-09-20 22:10

[CODE]Running N+1 test using discriminant 2, base 1+sqrt(2)
(2^13347311+1)/3 is Lucas PRP! (442900.4907s+0.0204s)
[/CODE]

and confirming ATH's result:

[CODE]Running N+1 test using discriminant 2, base 1+sqrt(2)
(2^13372531+1)/3 is Lucas PRP! (443577.1559s+0.0284s)
[/CODE]

The tests for (L+2)^(n+1)==5 (mod n, L^2+1) will take another 3 to 4 weeks :smile:

 paulunderwood 2013-10-17 21:27

[QUOTE=paulunderwood;353645]
The tests for (L+2)^(n+1)==5 (mod n, L^2+1) will take another 3 to 4 weeks :smile:[/QUOTE]

The tests passed for both PRPs. :smile:

I have emailed Henri Lifchitz requesting that the annotations for the PRPs in [URL="http://www.primenumbers.net/prptop/prptop.php"]his database[/URL] lists all our PRP tests.

 ATH 2013-10-18 00:58

[QUOTE=paulunderwood;353645]The tests for (L+2)^(n+1)==5 (mod n, L^2+1) will take another 3 to 4 weeks :smile:[/QUOTE]

Sorry for my ignorance but n is the Wagstaff PRP right, but what is L ? and what is meant by "n, L^2 +1" in the modular expression? My guess is that (L+2)^(n+1)==5 for both (mod n) and (mod L^2+1) ?

(and yes I did spend a bit of time looking for an answer myself, before I get told to study myself :) But not too much since it is not [I]that[/I] important)

 paulunderwood 2013-10-18 01:33

The modular reduction is done over "n" and "L^2+1". Here is a worked example for n=11 where L^2==-1 (mod 11, L^2+1)

[CODE]
(L+2)^2==L^2+4*L+4==4*L+3
(L+2)^3==(4*L+3)*(L+2)==4*L^2+11*L+6==0*L+6-4==2
(L+2)^6==2^2==4
(L+2)^12==4^2==16==5
[/CODE]

This will work for all prime n==3 (mod 4), but is not proven that all composites will fail. Wagstaff are 3 (mod 4)

To cover all n, I take minimal x>=0 such that jacobiSymbol(x^2-4,n)==-1, and check:

[CODE]
(L+2)^(n+1)==5+2*x (mod n, L^2-x*L+1)
[/CODE]

This is the basis of my [URL="http://www.mersenneforum.org/showpost.php?p=343535&postcount=73"]JavaScript program[/URL], which has the running time of 2 Miller-Rabin rounds. It is explained in [URL="http://www.mersenneforum.org/showpost.php?p=298027&postcount=44"]my pre-print "Quadratic Composite Tests"[/URL] where I used [TEX]\lambda[/TEX] instead of "L". The general "minimal x" method is good for n < 10^13 -- I am currently testing n < 7*10^13. :smile:

 All times are UTC. The time now is 08:40.