View Single Post
Old 2019-04-06, 21:08   #5
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

140416 Posts
Default Why don't we use the information in the series of p values for the known Mp to predict the next?

Well, some of us do, or pretend to, but it is not productive. Put simply it does not work. It has not ever worked, despite hundreds of documented tries (with a few tries yet to be resolved). Even the best number theorists are not sure how many there are in the interval up to p~109, and are unsure where or when any future such discoveries will appear. See for example https://primes.utm.edu/mersenne/index.html

See also the attachments here, the predict M45, predict M50, predict M51 or predict M52 threads, or the one based on curve fitting.
https://www.mersenneforum.org/showthread.php?t=6334 predict M45
https://www.mersenneforum.org/showth...137#post422137 predict M50
https://www.mersenneforum.org/showthread.php?t=22879 predict M51
https://www.mersenneforum.org/showthread.php?t=23892 predict M52

https://www.mersenneforum.org/showthread.php?t=24256 "Hidden number" malarky thread, based on curve fitting numerous unspecified curves

https://www.mersenneforum.org/showpo...4&postcount=69 also demonstrates the reliable failure of fits to predict even the highest known Mp on which a fit is based.

There are numerous other threads about various dubious claims to be able to make Mersenne prime predictions.

The accumulated experience of predicting or guessing Mersenne primes, in the various Predict Mxx threads, and various other threads concerning predictions, over a combined total of hundreds of guesses and predictions, is: hundreds proven composite, 6 yet to be settled, and zero proven successful guesses or predictions. Another way to look at it is of over 290 guesses or predictions I've found in the forum, about 98% have been proven composite, about 2.% are yet to be determined, and 0.00% (NONE) have been proven prime. And of the as-yet-unresolved, all 6 are for exponents so large that they are not amenable to any P-1 factoring or to primality testing by PRP or LL, either within the limits of existing software capabilities or of probable hardware lifetime, so can only be attacked with trial factoring currently. Some exponents (above about 67 bits) would even have save file sizes that would exceed the capacity of currently available file systems!
Six very large exponents, in exponent size order:
Note these examples are well beyond the capabilities of prime95 and other primality testing or P-1 factoring software and mfaktx TF software, as well as beyond feasible primality test or P-1 run times of currently available hardware, and will remain so for a long period of time. Some will remain so forever, since their sizes dwarf the number of subatomic particles in the known universe. That makes constructing a memory of adequate size and speed for primality testing or P-1 factoring them, or sufficiently trial factoring them impossible, regardless of technical advances.

They would have extraordinarily large memory requirements. A huge extrapolation from recent gpuowl test results for stage 2 P-1 memory per buffer indicates 8.6E13 to 8.9E32 MB for 66 to 127 bit exponents. Compare to 1.6E4 MB for ram on a Radeon VII or Tesla P100 gpu.

Similarly CUDALucas file sizes are extrapolated at 8.7E12 to 2.1E31 MB for 66 to 127 bit exponents. https://www.mersenneforum.org/showpo...1&postcount=10
Gpuowl file sizes were estimated at 8.7E12 to 2.1E31 MB for 66 to 127 bit exponents. https://www.mersenneforum.org/showpo...7&postcount=22

Other software such as Ernst Mayer's Mfactor or Luigi Morelli's Factor5 can be used to continue TF, and in the case of MM127, George Woltman's mmff on CUDA gpus.
There are also sometimes guesses or predictions in the 300M to 900M range, that take weeks to months each to primality test on fast hardware.

The absence of correct predictions or guesses is consistent with the probability of an equal number of randomly selected prime exponents, <1ppm per guess at nontrivial exponent. Probability of primality of a number n is ~1/ln(n). Where n=2p - 1, primality probability is x~1 / ( ln(2) * p). So, for example, for p~108, x~14ppb; p~852M, x ~ 0.000,000,001,7 (1.7 chances in a billion); for MM127, p=2127-1, x~8.5E-39.


Not really a set of predictions or claims, but more of a playful computation, is George Woltman's computation of exponents from the Wagstaff expected mean incidence of Mersenne primes, in the second attachment. It does well at small values but quickly falls apart by p>127.


Top of reference tree: https://www.mersenneforum.org/showpo...22&postcount=1
Attached Files
File Type: pdf pix etc and limits of curve fits.pdf (30.7 KB, 140 views)
File Type: pdf predicting Mxx wagstaff expected.pdf (10.7 KB, 156 views)
File Type: pdf predicting Mxx.pdf (63.4 KB, 0 views)

Last fiddled with by kriesel on 2021-05-12 at 03:38 Reason: status update
kriesel is online now