mersenneforum.org > News Oops - New Prime! (M49 related)
 Register FAQ Search Today's Posts Mark Forums Read

2016-01-12, 07:49   #111
ewmayer
2ω=0

Sep 2002
República de California

3·53·31 Posts

Quote:
 Originally Posted by Batalov On the contrary, this is a ridiculously small jump by Poisson statistics.
Not necessarily - 1.4*57885161 ~= 81M, and 70M ~= 1.2*57885161, so at the bottom of the 7******* range we're still 1.2x larger than the last one - not at all extreme - and at the top of the range we're very close to the average ratio for the purported Poisson distribution.

On the other hand, going back 10 previous M-primes we get 13466917*1.4^10
= 389,536,843 which is much larger than the latest one.

OTOOH, such distributions are well-known to be 'clumpy', so much depends on the chosen starting point: going back 20 M-primes we get 110503*1.4^20 = 92,455,932, which is again quite close (given the largeness of the logarithmic interval) to the latest one.

Quote:
 Originally Posted by Dubslow Edit: For AVX/Prime95, 4000K (!= 4M) is for exponents 76210000-76790000, 4032K for exponents up to 77990000, then 4096K=4M for exponents through 85200000.
You sure about that last one? Based on my estimates p ~= 85M should require FFT of ~4500K.

Last fiddled with by ewmayer on 2016-01-12 at 07:55

 2016-01-12, 07:54 #112 Batalov     "Serge" Mar 2008 Phi(4,2^7658614+1)/2 41×229 Posts 1.4 is a rather dubious number. Is that a new conjecture? ;-) There is a large difference between 1.4^10 and 1.47576^10; that's roughly 28.93 and 49, so (quote:) going back 10 previous M-primes we get 13466917*1.47576^10 = 389,536,843 659,820,084 which would have been monstrous. Not much better going back 20 previous M-primes: 110503*1.47576^20 = 92,455,932, 265,270,382, which is again quite close monstrous. Last fiddled with by Batalov on 2016-01-12 at 08:02
2016-01-12, 07:58   #113
Dubslow

"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

3×29×83 Posts

Quote:
 Originally Posted by ewmayer You sure about that last one? Based on my estimates p ~= 85M should require FFT of ~4500K.
Considering I read those values from the jump table in George's mult.asm, yeah I'm pretty confident. Unless I somehow misinterpreted it, which honestly is probably far more likely than this prime.

Edit: I two different jump tables after the AVX comment. The top one is directly below the AVX comment, and I'm not sure what the second jump table is.
Code:
        PRCSTRT 76210000, 4096000, 0.028093
PRCENTRY2               yfft_r4dwpn_4000K_12800_2, 8005632
PRCENTRY2               yfft_r4dwpn_4000K_6400_2, 6982912, I7_64
DD                      0
PRCSTRT 76790000, 4128768, 0.027915
PRCENTRY2               yfft_r4dwpn_4032K_9216_2, 6665728
PRCENTRY2               yfft_r4dwpn_4032K_4608_2, 5633280
PRCENTRY2               yfft_r4dwpn_4032K_2304_4, 6786304
DD                      0
PRCSTRT 77990000, 4194304, 0.026160
PRCENTRY2               yfft_r4dwpn_4M_14_4, 9644800
PRCENTRY2               yfft_r4dwpn_4M_14_2, 9644800
PRCENTRY2               yfft_r4dwpn_4M_14_1, 9644800
PRCENTRY2               yfft_r4dwpn_4M_13_4, 8500736
PRCENTRY2               yfft_r4dwpn_4M_13_2, 8500736, I7 + FMA3_64
PRCENTRY2               yfft_r4dwpn_4M_13_1, 8500736
PRCENTRY2               yfft_r4dwpn_4M_12_4, 5320704
PRCENTRY2               yfft_r4dwpn_4M_12_2, 5320704
PRCENTRY2               yfft_r4dwpn_4M_12_1, 5320704
PRCENTRY2               yfft_r4dwpn_4M_11_4, 6458368
DD                      0
PRCSTRT 85200000, 4587520, 0.031930
PRCENTRY2               yfft_r4dwpn_4480K_10240_2, 7414272
PRCENTRY2               yfft_r4dwpn_4480K_5120_2, 6263040, I7
DD                      0
PRCSTRT 87400000, 4718592, 0.030538
PRCENTRY2               yfft_r4dwpn_4608K_12288_2, 8322304
PRCENTRY2               yfft_r4dwpn_4608K_12288_1, 8322304
PRCENTRY2               yfft_r4dwpn_4608K_9216_2, 9464320, I7_64
PRCENTRY2               yfft_r4dwpn_4608K_9216_1, 9464320
PRCENTRY2               yfft_r4dwpn_4608K_6144_2, 7122944, I7_32 + FMA3_64
PRCENTRY2               yfft_r4dwpn_4608K_4608_4, 5932032
PRCENTRY2               yfft_r4dwpn_4608K_3072_4, 8450560
PRCENTRY2               yfft_r4dwpn_4608K_2304_4, 7235072

...

PRCSTRT 76050000, 4096000, 0.029047
PRCENTRY2               yfft_r4dwpn_4000K_ac_6400_2, 4583424, I7_64
DD                      0
PRCSTRT 77950000, 4194304, 0.027172
PRCENTRY2               yfft_r4dwpn_4M_ac_14_4, 9577984
PRCENTRY2               yfft_r4dwpn_4M_ac_14_2, 9577984
PRCENTRY2               yfft_r4dwpn_4M_ac_14_1, 9577984
PRCENTRY2               yfft_r4dwpn_4M_ac_13_4, 5068288
PRCENTRY2               yfft_r4dwpn_4M_ac_13_2, 5068288, I7 + FMA3_64
PRCENTRY2               yfft_r4dwpn_4M_ac_13_1, 5068288
PRCENTRY2               yfft_r4dwpn_4M_ac_12_4, 3998720
PRCENTRY2               yfft_r4dwpn_4M_ac_12_2, 3998720
PRCENTRY2               yfft_r4dwpn_4M_ac_12_1, 3998720
DD                      0
Either way I'm pretty certain this was (and still is by a decent margin) a Cat 4 LL assignment; quite lucky.

Last fiddled with by Dubslow on 2016-01-12 at 08:07

2016-01-12, 08:18   #114
ewmayer
2ω=0

Sep 2002
República de California

3×53×31 Posts

Quote:
 Originally Posted by Batalov 1.4 is a rather dubious number. Is that a new conjecture? ;-)
That's the approximate average-ratio as I recall it from way-back-when ... of course new data help us continue to refine it.

Perhaps someone with a handy-dandy stats best-fit suite can input the known M-prime exponents and compute the best-fit params for a Poisson distribution, and - if possible - 'goodness of fit' to the given distribution.

2016-01-12, 08:23   #115
axn

Jun 2003

491310 Posts

Quote:
 Originally Posted by Dubslow Considering I read those values from the jump table in George's mult.asm, yeah I'm pretty confident. Unless I somehow misinterpreted it, which honestly is probably far more likely than this prime.
These are the maximum exponents testable by the FFT.
If 77990000 is the maximum for 4096, then 77990000 * 4480/4096 = 85301562.5 would be the maximum for 4480 (with some reduction due to fewer bits/limb).

2016-01-12, 08:31   #116
Dubslow

"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

1C3516 Posts

Quote:
 Originally Posted by axn These are the maximum exponents testable by the FFT.

Last fiddled with by Dubslow on 2016-01-12 at 08:33

 2016-01-12, 10:55 #117 airsquirrels     "David" Jul 2015 Ohio 10000001012 Posts The Xeon run with a larger FFT and a separate CUDA run also came up positive. I think we have a winner here. The clLucas run and mLucas runs are both on track to back those up right now.
 2016-01-12, 10:59 #118 Batalov     "Serge" Mar 2008 Phi(4,2^7658614+1)/2 222558 Posts some statistics for the weary eyes If we take the first 40 M-primes and do a linear regression of log2 log2 Mn ~ n, then the 95% confidence interval for the slope's estimate is (0.557, 0.587) and both Wagstaff's and Eberhart's conjectured slopes are within this CI. Recall that Wagstaff's $\hat{\beta_1} = {1 \over {e ^ \gamma }} = 0.561458...$, while Eberhart's = log2 3/2 = 0.5849625... If we take the first 48 known M-primes, then the 95% confidence interval for the slope's estimate is (0.5366, 0.5659) and Wagstaff's is in 95% CI, but Eberhart's is not (in other words, Eberhart's, of "K-Y", conjecture can be rejected at 95% confidence level). If we take the first 49 known M-primes, then the confidence interval slides slightly lower still, but Wagstaff's is still in the 95% CI, and Eberhart's can be rejected with slighly higher confidence level. (Note that if there exist "missed" M-primes, the slope will be lower still.) The elementary implementation in R (following B.Caffo's class notation, if someone wants to check): Code: x <- 1:48 y <- c(1,1.584962501,2.321928095,2.807354922,3.700439718,4.087462841,4.247927513,4.95419631,5.930737338,6.475733431,6.741466986,6.988684687,9.025139562,9.245552706,10.32080055,11.10525378,11.15545073,11.65150022,12.05426514,12.11080953,13.24213206,13.27917527,13.4528847,14.28316072,14.4054739,14.50239674,15.44142045,16.39611974,16.75372601,17.01071385,17.72127946,19.52962691,19.71302565,20.2624562,20.4152105,21.50505023,21.52677478,22.73326384,23.68291627,24.32361192,24.51872848,24.63006217,24.85768459,24.95760092,25.14711776,25.3458327,25.36160654,25.78669022) fit <- lm(y ~ x) sumCoef <- summary(fit)$coefficients sumCoef[2,1] + c(-1, 1) * qt(.975, df = fit$df) * sumCoef[2, 2] [1] 0.5365694 0.5658668 > summary(fit) Call: lm(formula = y ~ x) Residuals: Min 1Q Median 3Q Max -1.71476 -0.45148 -0.07631 0.55816 1.24279 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 1.042977 0.204827 5.092 6.45e-06 *** x 0.551218 0.007277 75.743 < 2e-16 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 0.6985 on 46 degrees of freedom Multiple R-squared: 0.992, Adjusted R-squared: 0.9919 F-statistic: 5737 on 1 and 46 DF, p-value: < 2.2e-16
2016-01-12, 11:47   #119
axn

Jun 2003

173 Posts

Quote:
 Originally Posted by Batalov If we take the first 49 known M-primes, then the confidence interval slides slightly lower still, but Wagstaff's is still in the 95% CI, and Eberhart's can be rejected with slighly higher confidence level. (Note that if there exist "missed" M-primes, the slope will be lower still.)
If M50 turns out to be the next unverified exponent after M49 (and there are no other smaller primes), will that pull down the CI enough to reject Wagstaff's as well. Alternately, how high should M50 be (with enough precision to not compromise the identity of M49) to keep Wagstaff in play?

 2016-01-12, 12:35 #120 henryzz Just call me Henry     "David" Sep 2007 Cambridge (GMT/BST) 5,857 Posts Code: > fit <- lm(y - 0.5849625*x ~ x) > summary(fit) Call:lm(formula = y - 0.5849625 * x ~ x) Residuals: Min 1Q Median 3Q Max -1.71476 -0.45148 -0.07631 0.55816 1.24279 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 1.042977 0.204827 5.092 6.45e-06 *** x -0.033744 0.007277 -4.637 2.94e-05 *** ---Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 0.6985 on 46 degrees of freedom Multiple R-squared: 0.3185, Adjusted R-squared: 0.3037 F-statistic: 21.5 on 1 and 46 DF, p-value: 2.94e-05 I get a p-value of 2.94e-05.
2016-01-12, 13:06   #121
aketilander

"Åke Tilander"
Apr 2011
Sandviken, Sweden

10668 Posts

Quote:
 Originally Posted by Madpoo I crunched some data myself because I wondered if it were easily guessable from the known info. In the 4M FFT size (AVX2 table) there are 60,100 currently unverified exponents or with no known factors. Not only that but over 22K are unassigned, so it's kind of a big pool. Had it not been for the 4M FFT slip, it'd be even more obscure. LOL And besides, the method(s) used to hide it on the website seem to be just fine. I did uncover one odd quirk that may have tipped someone off if they were *really* diligent, which I'll bring up with George to make sure that's not the case next time, but otherwise, yeah, I think it's doing a good job of keeping it under wraps. I don't think anyone would stumble on it, but you're a bunch of smart folks and some of you are very persistent. Well, not too much longer before it's public anyway.
Aha, I think I know what you mean. In that case the new M-prime exponent is 78,xxx,xxx ? But I may be wrong as so many times before!

 Similar Threads Thread Thread Starter Forum Replies Last Post a1call Miscellaneous Math 26 2021-03-18 14:18 Prime95 News 505 2020-01-18 01:03 ltd Prime Sierpinski Project 21 2006-01-04 14:50 Uncwilly News 132 2005-05-10 19:47 ixfd64 Puzzles 2 2003-09-23 12:53

All times are UTC. The time now is 12:24.

Thu Apr 15 12:24:54 UTC 2021 up 7 days, 7:05, 0 users, load averages: 1.80, 1.80, 1.84