RcppBigIntAlgos uses the C library GMP (GNU Multiple Precision Arithmetic) for efficiently factoring big integers. Links to RcppThread for factoring integers in parallel. For very large integers, prime factorization is carried out by a variant of the quadratic sieve algorithm that implements multiple polynomials. For smaller integers, a simple elliptic curve algorithm is attempted followed by a constrained version of the Pollard’s rho algorithm (original code from https://gmplib.org/… this is the same algorithm found in the R gmp package called by the function factorize). Finally, one can quickly obtain a complete factorization of a given number n via divisorsBig.
install.packages("RcppBigIntAlgos")
## Or install the development version
devtools::install_github("jwood000/RcppBigIntAlgos")The function quadraticSieve implements the multiple polynomial quadratic sieve algorithm. Currently, quadraticSieve can comfortably factor numbers with less than 70 digits (~230 bits) on most standard personal computers. If you have access to powerful computers with many cores, factoring 100+ digit semiprimes in less than a day is not out of the question. Note, the function primeFactorizeBig(n, skipECM = T, skipPolRho = T) is the same as quadraticSieve(n).
## Generate large semi-primes
semiPrime120bits <- prod(nextprime(urand.bigz(2, 60, 42)))
semiPrime130bits <- prod(nextprime(urand.bigz(2, 65, 1)))
semiPrime140bits <- prod(nextprime(urand.bigz(2, 70, 42)))
## The 120 bit number is 36 digits
nchar(as.character(semiPrime120bits))
[1] 36
## The 130 bit number is 39 digits
nchar(as.character(semiPrime130bits))
[1] 39
## The 140 bit number is 42 digits
nchar(as.character(semiPrime140bits))
[1] 42
## Using factorize from gmp package which implements pollard's rho algorithm
##**************gmp::factorize*********************
system.time(print(factorize(semiPrime120bits)))
Big Integer ('bigz') object of length 2:
[1] 638300143449131711 1021796573707617139
user system elapsed
125.117 0.139 125.113
system.time(print(factorize(semiPrime130bits)))
Big Integer ('bigz') object of length 2:
[1] 14334377958732970351 29368224335577838231
user system elapsed
1437.246 0.309 1437.505
system.time(print(factorize(semiPrime140bits)))
Big Integer ('bigz') object of length 2:
[1] 143600566714698156857 1131320166687668315849
user system elapsed
2239.374 0.299 2239.641
##**************quadraticSieve*********************
## quadraticSieve is much faster and scales better
system.time(print(quadraticSieve(semiPrime120bits)))
Big Integer ('bigz') object of length 2:
[1] 638300143449131711 1021796573707617139
user system elapsed
0.075 0.000 0.075
system.time(print(quadraticSieve(semiPrime130bits)))
Big Integer ('bigz') object of length 2:
[1] 14334377958732970351 29368224335577838231
user system elapsed
0.093 0.000 0.092
system.time(print(quadraticSieve(semiPrime140bits)))
Big Integer ('bigz') object of length 2:
[1] 143600566714698156857 1131320166687668315849
user system elapsed
0.153 0.000 0.152As of version 0.3.0, we can utilize multiple threads with the help of RcppThread. Below, are a few examples:
Below are my machine specs and R version info:
MacBook Pro (15-inch, 2017)
Processor: 2.8 GHz Quad-Core Intel Core i7
Memory; 16 GB 2133 MHz LPDDR3
sessionInfo()
R version 4.0.3 (2020-10-10)
Platform: x86_64-apple-darwin17.0 (64-bit)
Running under: macOS Catalina 10.15.7
.
.
.
other attached packages:
[1] RcppBigIntAlgos_1.0.0 gmp_0.6-0
loaded via a namespace (and not attached):
[1] compiler_4.0.3 tools_4.0.3 Rcpp_1.0.5
## Maximum number of available threads
stdThreadMax()
[1] 8mostWanted1983 <- as.bigz(div.bigz(sub.bigz(pow.bigz(10, 71), 1), 9))
quadraticSieve(mostWanted1983, showStats=TRUE, nThreads=8)
Summary Statistics for Factoring:
11111111111111111111111111111111111111111111111111111111111111111111111
| MPQS Time | Complete | Polynomials | Smooths | Partials |
|--------------------|----------|-------------|------------|------------|
| 19s 561ms | 100% | 16183 | 4089 | 4345 |
| Mat Algebra Time | Mat Dimension |
|--------------------|--------------------|
| 4s 591ms | 8301 x 8434 |
| Total Time |
|--------------------|
| 24s 435ms |
Big Integer ('bigz') object of length 2:
[1] 241573142393627673576957439049 45994811347886846310221728895223034301839rsa79 <- as.bigz("7293469445285646172092483905177589838606665884410340391954917800303813280275279")
quadraticSieve(rsa79, showStats = TRUE, nThreads = 8)
Summary Statistics for Factoring:
7293469445285646172092483905177589838606665884410340391954917800303813280275279
| MPQS Time | Complete | Polynomials | Smooths | Partials |
|--------------------|----------|-------------|------------|------------|
| 2m 40s 597ms | 100% | 91221 | 5651 | 7096 |
| Mat Algebra Time | Mat Dimension |
|--------------------|--------------------|
| 14s 964ms | 12625 x 12747 |
| Total Time |
|--------------------|
| 2m 56s 215ms |
Big Integer ('bigz') object of length 2:
[1] 848184382919488993608481009313734808977 8598919753958678882400042972133646037727rsa99 <- "256724393281137036243618548169692747168133997830674574560564321074494892576105743931776484232708881"
quadraticSieve(rsa99, showStats=TRUE, nThreads=8)
Summary Statistics for Factoring:
256724393281137036243618548169692747168133997830674574560564321074494892576105743931776484232708881
| MPQS Time | Complete | Polynomials | Smooths | Partials |
|--------------------|----------|-------------|------------|------------|
| 7h 2m 48s 595ms | 100% | 7351308 | 9203 | 15846 |
| Mat Algebra Time | Mat Dimension |
|--------------------|--------------------|
| 2m 2s 479ms | 24929 x 25049 |
| Total Time |
|--------------------|
| 7h 4m 55s 252ms |
Big Integer ('bigz') object of length 2:
[1] 4868376167980921239824329271069101142472222111193 52733064254484107837300974402288603361507691060217quadraticSieve(rsa100, showStats=TRUE, nThreads=8)
Summary Statistics for Factoring:
1522605027922533360535618378132637429718068114961380688657908494580122963258952897654000350692006139
| MPQS Time | Complete | Polynomials | Smooths | Partials |
|--------------------|----------|-------------|------------|------------|
| 15h 26m 53s 383ms | 100% | 14999534 | 9230 | 16665 |
| Mat Algebra Time | Mat Dimension |
|--------------------|--------------------|
| 2m 22s 560ms | 25799 x 25895 |
| Total Time |
|--------------------|
| 15h 29m 26s 293ms |
Big Integer ('bigz') object of length 2:
[1] 37975227936943673922808872755445627854565536638199 40094690950920881030683735292761468389214899724061primeFactorizeBigAs of version 1.0.0, we can take advantage of the power of Lenstra’s elliptic curve method. This method is particularly useful for quickly finding smaller prime factors of very large composite numbers. It is automatically utilized in the vectorized prime factorization function primeFactorizeBig. This function should be preferred in most situations and is identical to quadraticSieve when both skipECM and skipPolRho are set to TRUE.
It is optimized for factoring big and small numbers by dynamically using different algorithms based off of the input. It takes cares to not spend too much time in any of the methods and avoids wastefully switching to the quadratic sieve when the number is very large.
For example, using the defaults on mostWanted1983 above only adds a few seconds.
primeFactorizeBig(mostWanted1983, showStats=TRUE, nThreads=8)
Summary Statistics for Factoring:
11111111111111111111111111111111111111111111111111111111111111111111111
| Pollard Rho Time |
|--------------------|
| 154ms |
| Lenstra ECM Time | Number of Curves |
|--------------------|--------------------|
| 4s 517ms | 4181 |
| MPQS Time | Complete | Polynomials | Smooths | Partials |
|--------------------|----------|-------------|------------|------------|
| 19s 639ms | 100% | 16183 | 4089 | 4345 |
| Mat Algebra Time | Mat Dimension |
|--------------------|--------------------|
| 4s 644ms | 8301 x 8434 |
| Total Time |
|--------------------|
| 29s 216ms |
Big Integer ('bigz') object of length 2:
[1] 241573142393627673576957439049 45994811347886846310221728895223034301839Performing a few iterations of the Pollard’s rho algorithm followed by the quadratic sieve can prove to be a terrible solution when the input size is incredibly large. More than likely, your constrained Pollard’s rho algorithm won’t find any prime factors, and eventually will pass an enormous composite number to the quadratic sieve. The quadratic sieve isn’t optimized for finding smallish factors in large composites. It will treat the input similarly to a semiprime of the same size. For example, the number ecmExpo1 below is 428 digits and if passed to the quadratic sieve algorithm, it could take years to factor as not even the most sophisticated semiprime algorithms can easily factor numbers of this size (See https://en.wikipedia.org/wiki/RSA_Factoring_Challenge). However, with the ECM, this is light work.
ecmExpo1 <- pow.bigz(prod(nextprime(urand.bigz(10, 47, 13))), 3) * nextprime(urand.bigz(1, 47, 123))
system.time(primeFactorizeBig(ecmExpo1, skipPolRho=TRUE, nThreads=8))
user system elapsed
17.911 0.081 4.743
ecmExpo2 <- pow.bigz(prod(nextprime(urand.bigz(10, 40, 42))), 3) * pow.bigz(prod(nextprime(urand.bigz(10, 45, 42))), 5) * nextprime(urand.bigz(1, 80, 123)) * nextprime(urand.bigz(1, 90, 123))
primeFactorizeBig(ecmExpo2, nThreads=8, showStats=TRUE)
Summary Statistics for Factoring:
215590243996472403826986190959172968924582673399860040722565967228161406366421528160220759715319487748182459438846599849086479492406371823155743767776724831907786824847947911279318978691074995051134928403920242678621702805852530836549215396392496547012632510254095452820560897056877208839289455859596425985161415486708317667524364034985007498553043810662487802224636724842739870398138687583364019516727138808818929150447675931521160760469848585726320428719422689614460288917155690252498420348768667307236940996727406088066077174182630522224530495122361513443818343671556622023320404680434511134662288782134897895034832405624062328931996299632052035105693354967480119521413889950571647670948619957219258020882926218179673783234886719528724327466781424377562596208305773866403070803284721517038640418360492984835882655310830736014378303508492181368455405610072971016964625940147583833553074577246736323251622391837090678894432642161330574148459946482734705537584444739753686240736188950594612747228733363204312823260997792251300802876878725604381077823
| Pollard Rho Time |
|--------------------|
| 19s 281ms |
| Lenstra ECM Time | Number of Curves |
|--------------------|--------------------|
| 26s 65ms | 24565 |
| MPQS Time | Complete | Polynomials | Smooths | Partials |
|--------------------|----------|-------------|------------|------------|
| 1s 32ms | 100% | 1482 | 1013 | 1450 |
| Mat Algebra Time | Mat Dimension |
|--------------------|--------------------|
| 313ms | 2425 x 2463 |
| Total Time |
|--------------------|
| 46s 868ms |
Big Integer ('bigz') object of length 82:
[1] 66642484459 66642484459
[3] 66642484459 385217678221
. . .
. . .
. . .
[79] 34936543827863 34936543827863
[81] 1130548241045557299883517 1051687085486158310119550449If you want to interrupt a command which will take a long time, hit Ctrl + c, or esc if using RStudio, to stop execution. When you utilize multiple threads with a very large number (e.g. 90 digit semiprime), you will be able to interrupt execution once every ~30 seconds.
## User hits Ctrl + c
## system.time(quadraticSieve(prod(nextprime(urand.bigz(2, 100, 42)))))
## Seed initialisation
## Error in PrimeFactorization(n, FALSE, showStats, TRUE, TRUE, nThreads, :
## C++ call interrupted by the user.
## Timing stopped at: 2.164 0.039 2.167divisorsBigThis function generates the complete factorization for many (possibly large) numbers. It is vectorized and can also return a named list.
## Get all divisors of a given number:
divisorsBig(1000)
Big Integer ('bigz') object of length 16:
[1] 1 2 4 5 8 10 20 25 40 50 100 125 200 250 500 1000
## Or, get all divisors of a vector:
divisorsBig(urand.bigz(nb = 2, size = 100, seed = 42), namedList = TRUE)
Seed initialisation
$`153675943236425922379228498617`
Big Integer ('bigz') object of length 16:
[1] 1 3
[3] 7 9
[5] 21 27
[7] 63 189
[9] 813100228764158319466817453 2439300686292474958400452359
[11] 5691701601349108236267722171 7317902058877424875201357077
[13] 17075104804047324708803166513 21953706176632274625604071231
[15] 51225314412141974126409499539 153675943236425922379228498617
$`261352009818227569107309994396`
Big Integer ('bigz') object of length 12:
[1] 1 2
[3] 4 155861
[5] 311722 623444
[7] 419206873140534785974859 838413746281069571949718
[9] 1676827492562139143899436 65338002454556892276827498599
[11] 130676004909113784553654997198 261352009818227569107309994396It is very efficient as well. It is equipped with a modified merge sort algorithm that significantly outperforms the std::sort/bigvec (the class utilized in the R gmp package) combination.
hugeNumber <- pow.bigz(2, 100) * pow.bigz(3, 100) * pow.bigz(5, 100)
system.time(overOneMillion <- divisorsBig(hugeNumber))
user system elapsed
0.364 0.029 0.390
length(overOneMillion)
[1] 1030301
## Output is in ascending order
tail(overOneMillion)
Big Integer ('bigz') object of length 6:
[1] 858962534553352218394101882942702121170179203335000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
[2] 1030755041464022662072922259531242545404215044002000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
[3] 1288443801830028327591152824414053181755268805002500000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
[4] 1717925069106704436788203765885404242340358406670000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
[5] 2576887603660056655182305648828106363510537610005000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
[6] 5153775207320113310364611297656212727021075220010000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000Another benefit is that it will return correct orderings on extremely large numbers when compared to sorting large vectors in base R. Typically in base R you must execute the following: order(asNumeric(myVectorHere)). When the numbers get large enough, precision is lost which leads to incorrect orderings. Observe:
set.seed(101)
testBaseSort <- do.call(c, lapply(sample(100), function(x) add.bigz(pow.bigz(10,80), x)))
testBaseSort <- testBaseSort[order(asNumeric(testBaseSort))]
myDiff <- do.call(c, lapply(1:99, function(x) sub.bigz(testBaseSort[x+1], testBaseSort[x])))
## Should return integer(0) as the difference should always be positive
## NOTE that the result will be unpredictable because of lack of precision
which(myDiff < 0)
[1] 1 3 4 7 9 11 14 17 19 22 24 25 26 28 31 32 33 36 37 38 40 42 45 47 48
[26] 50 51 54 57 58 59 63 64 65 66 69 70 72 75 78 81 82 85 87 89 91 93 94 97 98
## N.B. The first and second elements are incorrect order (among others)
head(testBaseSort)
Big Integer ('bigz') object of length 6:
[1] 100000000000000000000000000000000000000000000000000000000000000000000000000000038
[2] 100000000000000000000000000000000000000000000000000000000000000000000000000000005
[3] 100000000000000000000000000000000000000000000000000000000000000000000000000000070
[4] 100000000000000000000000000000000000000000000000000000000000000000000000000000064
[5] 100000000000000000000000000000000000000000000000000000000000000000000000000000024
[6] 100000000000000000000000000000000000000000000000000000000000000000000000000000029Credit to primo (Mike Tryczak) and his excellent answer to Fastest semiprime factorization.
Factoring large numbers with quadratic sieve on MSDN Archive.
A really nice concise example is given here: Factorization of n = 87463 with the Quadratic Sieve
Smooth numbers and the quadratic sieve by Carl Pomerance
Implementing the Elliptic Curve Method of Factoring in Reconfigurable Hardware by Gaj K. et al.
Integer Factorization using the Quadratic Sieve by Chad Seibert
In the stackoverflow question and answer What is the most efficient factoring algorithm for quadratic sieve extraction phase? by Ilya Gazman, an efficient method for checking divisibility is sketched out that utilizes built-in types. You can see more on a video Ilya put on youtube: E15: Quadratic Sieve Running on Java - Receiving. While mpz_divisible_ui_p is very efficient, we found better performance using this method.
Currenlty, our main focus for version 1.1.0 will be implementing the self initiallizing quadratic sieve.
I welcome any and all feedback. If you would like to report a bug, have a question, or have suggestions for possible improvements, please file an issue.