mersenneforum.org > Data Missed small factors
 Register FAQ Search Today's Posts Mark Forums Read

2003-11-18, 06:22   #1
dswanson

Aug 2002

3108 Posts
Missed small factors

k5gj, GP2, and I have been having a PM exchange about factors that were missed during trial factoring. This exchange was prompted by a message GP2 received from k5gj, who has been running P-1 on exponents that have long since been verified composite by LL testing:

Quote:
 k5gj wrote: GP2, I do not know if you are interested in anomalies, however, the following 16 digit factor is a great example of why we should run P-1 tests. [Sat Nov 15 16:25:13 2003] P-1 found a factor in stage #2, B1=262144, B2=26214400. UID: k5gj/4, M268813 has a factor: 1017111715779881 Here are three more 17 digit factors that I found last month. [Thu Oct 09 02:02:53 2003] P-1 found a factor in stage #2, B1=2048, B2=204800. UID: k5gj/1, M2441083 has a factor: 75649430225324231 [Thu Oct 09 15:58:17 2003] P-1 found a factor in stage #1, B1=2048. UID: k5gj/1, M2448961 has a factor: 95075864569432231 [Tue Oct 28 05:04:50 2003] P-1 found a factor in stage #2, B1=2048, B2=204800. UID: k5gj/1, M2623141 has a factor: 23522191782267041
All four of these factors should have been found during trial-factoring. The bit depth of these for factors is
Code:
                      Bit     Report TF
Exponent         Depth      Depth
1017111715779881    49.85        57
75649430225324231    56.07        59
95075864569432231    56.40        59
23522191782267041    54.38        59
where the final column is the depth to which the nofactor.txt file stated that the exponent had been trial-factored. Hence, as both k5gj and GP2 pointed out, these "easy" factors were apparently missed during trial factoring, resulting in two unnecessary LL tests each (first test plus doublecheck).

I got pulled into this exchange because GP2 had recently run across a two-year old posting of mine in a mailing list archive:
Quote:
 dswanson wrote: You either have an extremely good memory, or you've been rereading old postings from the mailing list recently. Yes, about two years ago I found a range in which a bunch of factors seemed to have been missed. I found the first number in this range the same way - a P-1 result with an anomalously low bit depth. There was a discussion of this in Digest issues 887 through 890. If a thread gets started, it would probably be worth pasting the relevant items from those Digests into the thread. If I recall correctly, I determined that the full range of incompletely factored exponents was 7019000 - 7055000. I subsequently factored a fraction of this range to 2^58, but didn't devote the resources to factor it any deeper. At the time the leading edge of doublechecks was within this range. To avoid poaching and unncessary work, I factored all exponents that were not yet assigned, plus any exponents that expired. I did not touch any exponents that were in progress, nor any that had already completed the second test. Unless someone else subsequently factored this range, I am sure there are still missing factors between 2^56 and 2^64 [sic - final number should be 2^62].
For those with access to the mailing list digests, the discussion was in digest #887 - #890. For those without, I am posting the exchange below. There are a lot of interesting points made, most of which are still quite relevant.

 2003-11-18, 06:23 #2 dswanson     Aug 2002 C816 Posts Date: Sun, 30 Sep 2001 15:26:51 -0700 From: "Daniel Swanson" Subject: Mersenne: Factoring failure? Yesterday I was assigned M7027303 to doublecheck. The number came with the claim that it had no factors less than 62 bits. But my P-1 factorization of the number found this 55-bit factor: [Sun Sep 30 13:53:26 2001] P-1 found a factor in stage #1, B1=35000, B2=463750. UID: dswanson/nosnawsd, M7027303 has a factor: 31090234297428433 I thought this was mighty peculiar, so I stuck "Factor=7027303,0,0" into my worktodo file to force a complete refactorization of the number. Sure enough, about three minutes later out pops the result: [Sun Sep 30 15:06:10 2001] UID: dswanson/nosnawsd, M7027303 has a factor: 31090234297428433 Somebody wasted a lot of time doing a first-time LL test on this number. Does anybody have any idea how common factoring errors might be in the "exponents not factored" database, particularly since doublechecks are not commonly done on the factoring results? Dan
 2003-11-18, 06:24 #3 dswanson     Aug 2002 110010002 Posts Date: Sun, 30 Sep 2001 16:07:34 -0700 From: "Daniel Swanson" Subject: Mersenne: Re: Factoring Failure? A follow-on to my previous message: I went through the Cleared Exponents report looking for other examples of factors found during double-checks that should have been found during the initial factorization. I found the following 7 examples, including the one I just reported: 5977297 53 DF 6726544627832489 26-Aug-01 12:03 TempleU-DI C031EBA9B 6019603 57 DF 137024179940485697 18-Aug-01 18:55 TempleU-DI C031EBA9B 7019297 57 DF 160100125459121849 27-Sep-01 22:52 TempleU-DI TD01489_Cub1 7020641 58 DF 226230108157229263 30-Sep-01 02:05 RayPelzer Homebase 7025987 56 DF 74052063365823791 30-Sep-01 01:12 shaneamy P600A 7027303 55 DF 31090234297428433 30-Sep-01 20:58 dswanson nosnawsd 10159613 56 DF 68279769831982367 19-May-01 10:13 hornup lbe_pc Seems to be a big clump in the 701xxxx - 702xxxx range. Given that this is at the leading edge of numbers currently being assigned for double-checks, I'd be willing to bet several more will turn up in the next few days. Were numbers in this range all originally factored by the same user or computer? Another interesting note: Given the number of different computers working double-checks, what are the odds that TempleU-DI/C031EBA9B would pick up TWO of these?!? Dan
 2003-11-18, 06:25 #4 dswanson     Aug 2002 110010002 Posts Date: Mon, 01 Oct 2001 15:01:05 -0400 From: George Woltman Subject: Re: Mersenne: Re: Factoring Failure? Hi, At 04:07 PM 9/30/2001 -0700, Daniel Swanson wrote: >I went through the Cleared Exponents >report looking for other examples of factors found during double-checks that >should have been found during the initial factorization. > 5977297 53 DF 6726544627832489 > 6019603 57 DF 137024179940485697 > 7019297 57 DF 160100125459121849 > 7020641 58 DF 226230108157229263 > 7025987 56 DF 74052063365823791 > 7027303 55 DF 31090234297428433 >10159613 56 DF 68279769831982367 >Were numbers in this range all originally factored by the same user or >computer? My logfiles from that long ago have been zipped and stored on CDROM. It is possible that 7,010,000 - 7,030,000 were all factored by one person. It was not uncommon for me to hand out large blocks for factoring to users without Internet connections. While I no longer do this, there are a handful of users pre-factoring the 20,000,000 - 80,000,000 area. I hope their machines are reliable!! They probably are as they are finding the expected number of factors. Anyway, it doesn't appear to be a program bug as you were able to find the factor with trial factoring. I'm guessing either bad hardware or an older prime95 version had a bug. Either way, GIMPS has never considered missing a factor as a big deal. It only means some wasted effort running a LL test that could have been avoided. Thanks for the interesting findings! Regards, George
 2003-11-18, 06:26 #5 dswanson     Aug 2002 23×52 Posts Date: Mon, 1 Oct 2001 22:23:38 +0200 From: "Jean-Yves Canart" Subject: FW: Mersenne: Re: Factoring Failure? Hello all, I have browsed some logs I archived long time ago and I have found this: In may 1998, one user, "tomfakes", cleared around 80 exponents with factor found = "1" It was in the range 7013000-7055000. Regards, Jean-Yves
 2003-11-18, 06:26 #6 dswanson     Aug 2002 C816 Posts Date: Tue, 2 Oct 2001 19:52:42 -0000 From: bjb Subject: Re: FW: Mersenne: Re: Factoring Failure? On 1 Oct 2001, at 22:23, Jean-Yves Canart wrote: > I have browsed some logs I archived long time ago and I have found > this: > > In may 1998, one user, "tomfakes", cleared around 80 exponents with > factor found = "1" It was in the range 7013000-7055000. Well, (s)he's not lying - 0 = n (mod 1) is a property of integers ;-) In any event: (a) this is the _opposite_ of the reported problem - what seems to have happened is that "no factor found" was being reported, sometimes erroneously; (b) this won't get through now PrimeNet validates submitted factors; the code I wrote for this purpose rejects as garbage any single-digit factor, after stripping off any leading zeroes as well as white space. (Obviously a Mersenne number with a prime exponent p > 5 cannot have any factors less than 10, and we know pretty much all there is to know about exponents up to and including 5, so excluding these is not a practical problem). > > At 04:07 PM 9/30/2001 -0700, Daniel Swanson wrote: > > >I went through the Cleared Exponents > > >report looking for other examples of factors found during > > double-checks that > > >should have been found during the initial factorization. > > > 5977297 53 DF 6726544627832489 > > > 6019603 57 DF 137024179940485697 > > > 7019297 57 DF 160100125459121849 > > > 7020641 58 DF 226230108157229263 > > > 7025987 56 DF 74052063365823791 > > > 7027303 55 DF 31090234297428433 > > >10159613 56 DF 68279769831982367 > > >Were numbers in this range all originally factored by the same user > > >or computer? > > > > My logfiles from that long ago have been zipped and stored on CDROM. > > It is possible that 7,010,000 - 7,030,000 were all factored by one > > person. It was not uncommon for me to hand out large blocks for > > factoring to users without Internet connections. While I no longer > > do this, there are a handful of users pre-factoring the 20,000,000 - > > 80,000,000 area. I hope their machines are reliable!! They > > probably are as they are finding the expected number of factors. The primes from that block of 20,000 numbers represents quite a bit of work and maps poorly onto the "missed" factors reported. A few mistakes are inevitable but, since testing a factor takes of the order of a microsecond on current systems, hardware glitches shouldn't be much of a risk. (? Unless they get into the code stream used to generate potential factors?) Reports of two "missed" factors of exponents within spitting distance of 6,000,000 and no less than four just over 7,000,000 looks high for random glitches to be responsible, even on really ropy hardware. Remember that P-1 (which found the factors missed by trial factoring) can only find a small proportion of the "small" factors, especially when it's being run with "double checking" limits. > > > > Anyway, it doesn't appear to be a program bug as you were able to > > find the factor with trial factoring. I'm guessing either bad > > hardware or an older prime95 version had a bug. If it _was_ Prime95. There are other factoring programs out there; maybe there was a higher incidence of use about 3.5 years ago when these exponents would have been the subject of factoring assignments. > > Either way, GIMPS > > has never considered missing a factor as a big deal. It only means > > some wasted effort running a LL test that could have been avoided. True enough - though I'm concerned that the "no factors below 2^N" database may be seriously flawed, from the point of view of GIMPS it would seem to be a waste of time to go round redoing trial factoring just to fix this problem. However if it could be established that all the "missed" factors reported were the work of one user, perhaps it would be worth fixing the database to force rerunning of trial factoring for those factoring assignments run by that user when the exponents are reassigned for double checking (or LL testing). Regards Brian Beesley
 2003-11-18, 06:27 #7 dswanson     Aug 2002 23×52 Posts Date: Tue, 2 Oct 2001 23:09:31 -0500 From: "Steve Harris" Subject: Re: FW: Mersenne: Re: Factoring Failure? > > Either way, GIMPS > > has never considered missing a factor as a big deal. It only means > > some wasted effort running a LL test that could have been avoided. >True enough - though I'm concerned that the "no factors below 2^N" >database may be seriously flawed, from the point of view of GIMPS >it would seem to be a waste of time to go round redoing trial >factoring just to fix this problem. Yes, from the point of view of GIMPS (that is, searching for Mersenne primes) it's not a huge deal... but there also exists an effort to fully factor the candidates that are not prime, and this throws a big problem into that project. Someone could be trial factoring an exponent from 2^59 to 2^65 and find a factor in that range after a smaller factor had been missed, and it will go into the database as the smallest factor when it actually is not. Might be decades before the smaller factor is discovered. Oh well, Steve
 2003-11-18, 06:28 #8 dswanson     Aug 2002 110010002 Posts Date: Wed, 3 Oct 2001 02:30:10 -0400 From: "Carleton Garrison" Subject: Re: FW: Mersenne: Re: Factoring Failure? > At 04:07 PM 9/30/2001 -0700, Daniel Swanson wrote: > >I went through the Cleared Exponents > >report looking for other examples of factors found during > double-checks that > >should have been found during the initial factorization. > > 5977297 53 DF 6726544627832489 > > 6019603 57 DF 137024179940485697 > > 7019297 57 DF 160100125459121849 > > 7020641 58 DF 226230108157229263 > > 7025987 56 DF 74052063365823791 > > 7027303 55 DF 31090234297428433 > >10159613 56 DF 68279769831982367 > >Were numbers in this range all originally factored by the same user > >or computer? > Either way, GIMPS > has never considered missing a factor as a big deal. It only means > some wasted effort running a LL test that could have been avoided. I wonder if by using configuration settings, people are able to skip as many, if not all, factoring stages? I believe the idea of trying to skip P-1 factoring was talked about within the last 3 or 4 months. Apparently there are people who would just prefer to get credit for doing LL work than to find factors. Until factoring time (while one is LL testing) is at credited at the same same rate as LL testing, let alone getting credited not at all (most of the time no factors are found), results like the above could become common place.
 2003-11-18, 06:28 #9 dswanson     Aug 2002 23×52 Posts Date: Wed, 3 Oct 2001 02:41:27 -0500 (CDT) From: ribwoods Subject: Mersenne: Re: Factoring Failure? Daniel Swanson wrote: > 5977297 53 DF 6726544627832489 > 6019603 57 DF 137024179940485697 > 7019297 57 DF 160100125459121849 > 7020641 58 DF 226230108157229263 > 7025987 56 DF 74052063365823791 > 7027303 55 DF 31090234297428433 >10159613 56 DF 68279769831982367 I'm pretty sure the second column is the previous trial-factoring limit, in units of power-of-2. For M7019297, the found factor 160100125459121849 is larger than 2^57 (= 144115188075855872), and for M7025987, factor 74052063365823791 is greater than 2^56 (= 72057594037927936). Unless I'm misinterpreting something, those two cases don't indicate a trial-factoring failure. [dswanson comment on this posting: ribwoods is correct about the bit depth of the found factor. For example, the bit depth of 160100125459121849 is 57.15, which rounds down to the "57" listed. But ribwoods missed the point, which is that nofactor.txt had claimed that this exponent had no factor below 2^62. So it is indeed a trial-factoring failure.]
 2003-11-18, 06:29 #10 dswanson     Aug 2002 23×52 Posts Date: Wed, 3 Oct 2001 12:40:29 -0400 From: "Carleton Garrison" Subject: Re: FW: Mersenne: Re: Factoring Failure? > > I believe the idea of trying to skip P-1 factoring was talked about within > > the last 3 or 4 months. Apparently there are people who would just prefer > > to get credit for doing LL work than to find factors. > > This is why I like that you loose credit for a LL-test if someone else > finds a factor later, or if two other independant checks prove your > result to be wrong. Me too. I understand that George's top producer page does this, while the PrimeNet stat page does not. PrimeNet really needs this capability. Carleton Garrison LL#163 F#295 G#253 www.teamprimerib.com
 2003-11-18, 06:30 #11 dswanson     Aug 2002 23·52 Posts Date: Wed, 03 Oct 2001 18:29:52 -0700 From: Gerry Snyder Subject: Re: FW: Mersenne: Re: Factoring Failure? Carleton Garrison wrote: > > [? wrote:] > > > > This is why I like that you lose credit for a LL-test if someone else > > finds a factor later, or if two other independant checks prove your > > result to be wrong. > > Me too. I understand that George's top producer page does this, while the > PrimeNet stat page does not. PrimeNet really needs this capability. To me, there is no question that an LL test that is shown to be wrong should not count for anything. The number still required two more LL tests, so that it as if the erroneous one had not been done. But, at least in theory, every Mersenne number proven non-prime will eventually be factored. Again, to me, so what? At least the LL test showed that further factoring activity would eventually succeed. I have nothing against George doing things that way. (When I play ball with him, I play by his rules or I don't play at all. You know why? Because it's his ball, that's why.) Seriously, I can see some point to doing things that way, but I would do probably do it differently. But even more seriously, I'm just glad to be in the game, and I am grateful to George and all the others who have made it easy and fun to participate. Gerry PS I just got a chuckle from imagining a very competitive team tearing down an opponent by finding what numbers the opponent had done LL tests on, and factoring them.

 Similar Threads Thread Thread Starter Forum Replies Last Post lycorn Data 76 2015-04-23 06:07 TheMawn Information & Answers 7 2014-01-10 10:23 patrickkonsor GMP-ECM 16 2010-09-28 20:19 petrw1 Lone Mersenne Hunters 17 2009-11-20 03:40 Kees PrimeNet 6 2006-11-16 00:12

All times are UTC. The time now is 20:04.

Mon May 10 20:04:45 UTC 2021 up 32 days, 14:45, 1 user, load averages: 3.20, 3.56, 3.47