You are on page 1of 5

:-".\;..

,
'I'-

:.'~';':;"

,'" .f~j

Greeks pondered large values as well.


But historically, a large number was
whatever the prevailing culture dccmcJ
it to be-an intrinsically circular defini.tion. The Romans initially had no terms:
'or symbols for figures above 100.000.
:, And the Greeks usually stoPI'eJcoulltin~
..ata 11tvriad.a ,vorJ mcanillg " I ().Ol!()."

million-di~it
.

111deej,a popular idea in an~ientGrl:ece


was that no number was greater than
the total cmlnt of sanJ grai;1s tll:\.,JCJto
fill the universe.
4- 111
the thirJ century [\.(:., (;reek mathematician ArchimeJes s(wght tocorreL"t
this belict. In a letter to King Gekm of
Syral:use. he set (Jut to call:ulatc the JCtual number of sanJgrains in the llniverse. To LU)so. Archimedes Lieviseda
clever sl:hemc illvolvill~ successiver;ltios that woulLi effectively cxtcnLI the
prevailing C;reek number system,which
had no exponential scaling. His results,
which in current tt'rn1Splaced tht' numbcr somewhere between 1 O.~
I to 10"\

\ve can now characterize


numbers about which earlicr" mathematicians could only dream.
2.. Interest in large numbers datcs back
to ancient times. We kno~., for exampIc, that the early Hindus, who Jcveloped the decimal system, contemplateJ
them. In the 1l0Wcommollplace decimal
system the position of a digit (ls, lOs,
100s ;1ndso on} dern)tes its scale. Using
this shorthand, the Hindus n.lmcd many
were visionary; in'fact, a sphere having
large numbers; one having 1.S3digitsor as we might say today, a number of the radius of Pluto's orbit ,vould conorder 10J53_is mentioned in a myth ,Jain on thc orJer of 10.51grains.
J
Scholars in thl: I .11th
and 19th centurabout BudJha.
iescontemplated large nlllnbers that still
3 The ancient Egyptians, Romans and

,.!. :,

:7'

(.~:

'i,1'

...,-,

,,;J
,.;

.,. :'

:'~~
;

\:~

: ..\'i

"!;-~:i

:{;~t
"'.~~'.;tJ.~';~)
;,,;.~,

have practiL'al sl:ientifil: rell:vanl:e. ConsiJer AvogaJro's nunlbcr, nameJ after


the 19th-l:entury Italian I:ht=mist Amcdeo AvogaJro, It is rlllighly 6,02 X lOll
anJ represents tIle nulllhl'r of atl)l11Sill
12 grams of pure I:arbon. On~ way to
think about Avogallru's number, also
I:allell a mIlle, is as fllllll\vs: if jllst one
gram of I:aroon Wl'r~ e,xpallJI:J to the
size of planet l:arth, a single l'arbl)l1
atl)l11 woulJ loom sllml:thing like a
bowling b.ll.1.
b
i\l1Ilther intcrestillg way tll inla/!,ine a
mole is to I:onsiller the tlltalllilmbt=r of
I:omputer operatil)l1s-that is, the arithmetic operations ocl:urring within a
I:omputcr's I:irl:uits-cvt=r pt=rf(lrlllcll hy
all computers in histllry. Even a sm.\11
m.\l:hine I:an exel:utemillions of operations per second; mainframes can do
m.my more. Thus, tht= total oper.ltion
I:ount to Jate, though impossible to estim.\te precisely, must be I:IIJseto a mole.
It will undoubteJly have exl:eelleJ that
by the year 2000.1ToJay scientists Jt=al with numbers
much larger than the mole. The number of protons in tllC known universe,
for example, is thlJu/!,ht tll h~ ah(lut

LARGE NUMBERS-such asthe 100-digit, or googol-size,oncs running acrossthe


tops of these pagcs-have become more
accessibleover time thanks to advancesin
computing. Archimedes, whose bust appears at the left, had to invent new mathematics to estimate the number of sand
grains required to fiIl the universe.His astonishingly accurate result, 10'., was by
ancient standards truly immense. Modern
machines,however,routinely handle vastly greater values. Indeed, any personal
computer with the right software can
completely factor a number of order 1051.
Th/?(;hl1///?/ll'/?(If 'fIr,,/! N"",h"ro

1080. But the hlllnan imagillatiOlI can


press furtl1er. It is legcnJary tllat thc
ninc-ycar-oIJ nephew of m;ltl1ematician
I:J\varJ Kasner JiJ coin, in I YJ8, the
g()()g()I,as I f(lllow~J hy 1(10;ll'!"()eS,()r
10100.With respect to some classcs of
C()lJlputati()lJal prohlems, thc go()gol
r()ughly Jem.lrcatcs the nulllher nlagnituJcs that begin seriously to (halleJ1ge
moJerll 111.lchinery.I~\'cn S(),m;1chincs
can evenanswcr sOlnc <.jucsti()lJS
about
gargantua as largc as the mighty googolplex, which is I f()lk)weJ hy a googIll of zcroes, or 1010100.E\'en if you
useJ a proton for every zero, you coulJ
m)t scribe the googolplex onto the
kll()\vlluniverse.
Manipulating the Mcrcly I_argc

8:
s omewhat above the googol lie num-

readers of Martin Gardner's column in


the August 1977 issue of Scielztific
A"zericulz to factor a 129-digit number,
uubbcu I~SA-129, and find a hiducn
l1Ie~~age.It wa~ not until 1994 that Arjen K. Lenstra of Bellcore, Paul Leyland
of the University of Oxford and then
graduate student Derek Atkins of M.I. T.
allu unuergraduate student Michael
C;raff of Iowa State Universit)" working
with hundreds of colleagues on the Internet, sul.:l.:eedeu.
(The secretencrypted
messagewas "T~IE MAGIC WORDSt\I~l:
S(~lJEAMISHOSSIFRAGE.")Current recommendations suggest that RSA encryption keys have at least 230 digits to
be sel.:lIre.
\0 Net\vork I.:ollaborationsare now commonplace, and a solid factoring culture
has sprung up. Samuel S. Wagstaff,jr.,
of Purdue University maintains a factoring newsletter listing recent factorizations. And along similar lines, Chris
K. Caldwell of the University of Tenncssee at Martin maintains a World
Wide Web site (http://www.utm.edu/
research/primes/largest.html) for prime
number records. Those who practice
factoring typically turn to three powerftli algorithms. The Quadratic Sieve (QS)
method, pioneered by Carl Pomerance
of the University of Georgia in the 1980s,
remains a strong, general-purpose attack for factoring numbers somewhat
.larger than it googol. (The QS, in fact,
conquered RSA-129.) To factor a mystery number, the QS attempts to factor
many smaller, related numbers generat:ed via a clever. sieving process. .:rhese.
smaller factorizations are combined to "

bers that present a sh.lrp challenge


to practitioners of tlle art of factoring:
the art of breaking numbers into their
prinle fa~tors, whcrc priml's arc thcmselvestlivisiblc only by I anJ tllcmst:lves.
for example, 1,799,257 factors illto
7,00 I X 257, but to decompose a sufficiently large number into its prime factors can be so problematic that computer scientistshave harnessedthis difficulty
to encrypt data. Indeed, one prevailing
el1crYptionalg<.'rithm,called I~SA,transforll1s the problem of cracking encrypted messagesinto that of fa~toring cer-;
rain large numbers, called public keys.
(RSA is nameJ after its inventors, RonalJ L. I~ivestof the Massachusetts InstItute' of Technology, Adi Shamir of the
Weizmann Institute 'of Science in Israel,
'
and Leonard M. Adleman o'f the Uni- yield a factorofthemysterynu~ber.:..:
versity of Southern California.)
,
,\~\ A newer strategy,.the Number Ficld ,'
I To demonstrate the sti,ength of RSA,
Sieve(NFS),toppled 3;.rS5-'digitR~~
;
Rivest, Sharnir and Adlem~n challenged the nin~.F_erI.I1atnl,lm~e.;F~~;!Na~~d ...':
,

f.,:

,-

s,

..~
.~

"

:~.
'."- .:::..;~ ;;.:~:..:.
;':;~~~:",~i:~1i"-'::.~"'NTIFI~
AM"RT~AIIi:;,:~~r;ri","1.99'l.~:~9

f,'\

:u:,:~

for the great French theorist Pierre de


Fermat, the nth Fermat number is F" =
22"+1.) In 1990 F9 fell to Arjen Lenstra, Hendrik w: Lenstra,Jr., of the University of California at Berkeley,Mark
Manasse of Digital Equipment Corporation and British mathematician John
Pollard, again aidcd by a substantial
machine network. This spectacular factorization dcpended on F9's spccial
form. But Joseph Buhlcr of Rccd ColIcge, Hendrik Lenstra and Pomerance
have since developcd a variation of the
NFS for factoring arbitrary numbers.
This general NFS can, today, comfortably factor numlJer~of 130 digits. In rctrospcct. RSA-129 could have bccn factored in lesstimc this w:\y.
\1... Thc third common factoring tactic,
the Elliptic Curve Method (ECM), developed by Hendrik Lenstra, can take
apart much larger nllmlJers, provided
that at least one of the number's prime
factors is sufficiently small. For exampic, Richard P. Brcnt of thc Allstralian
National University rccently factored
Plo using ECM. after first finding a single prime factor "only" 40 digits long.
It is difficult to find factors having more
than 40 digits using ECM. For arbitrary numbers betwecn, say, 101.~O
and

101.000.000,
ECM stands as the method is the Fast Fourier Transform (FFT).
of choice, although ECM cannot be ex- T11eFi'-r is most often thought of as a
pected to find all factors of such gar- meansfor ascertaining some spectrum,
as is done in analyzing birdsongs or hugantuan numbers.
13 Even for numbers that truly dwarf the man voices or in properly tuning an
googol, isolated factors can sometimes acoustic auditorium. It turns out that
be found using a centuries-old sieving ordinary multiplication-a fundamenmcthod. "n1eidea is to usewhat is called tal opcrationbetween numbers-can be
modular arithmctic, which keeps the dramatically enhancedvia FFT [seebox
sizes of numbers under control so that be/olv]. Arnold Schonageof the Univermachine memory is not exceeded,and sity of Bonn and others refined this asadroitly scan ("sieve") overtrial factors. tute observation into a rigorous theory
A decadeago Wilfrid Keller of the Uni- ~ring the 1970s.
versity of Hamburg useda sieving tech- }'5 FFT multiplication has beenused in
niquc to find a factor for the awesome cclebrated calculations of 7tto a great
F23471.
which has roughly 107.000deci- many Jigits. Granted 7t is not a bona
mal digits. Keller's factor itselfhas "only" ,fide large number, but to compute 7tto
about 7,000 digIts. And Robert ;J.Har- millions of digits involvesthe same'kind
lev, then at the California Institute of of arithmetic used in large-numberstudT~chnology,turned to sieving to find a ies. In 1985 R. William Gosper,Jr., of
Inc., in Palo Alto, Calif., com36-digit factor for the stultifying (goo- Syn1bolics,
golplex + 1); the factor is 316,912,650, puteJ 17 million digits of 7t. A year lat057,057,350,374,175,801,344,000,001. er David Bailey of the National Aeronautics nnd SpaccAdministration Ames
Rescarch <'~cntcrcomputed 7tto more
Algorithmic Advancemcnts
than 29 million digits. More recently,
l\f
Baileyand Grcgory Chudnovskyof ColM
any modern results on large numumbia Univcrsity renched one billion
bers have dcpended on algorithms
from seemingly unrelated fields. One digits. And Yasumnsa Kanada of the
example that could fairly be called the University of Tokyo has reported five
workhorse of all engincering-algorithms billion Jigits. In caseanyone wants to

Using Fast Fourier Transforms for SpeedyMultiplication


O

rdinary multiplication is a long-winded process by any account, even for relatively small numbers: To multiply two
numbers, x and y, each having D digits, the usual, Hgrammar
schoolH method involves multiplying each successive digit of x
by every digit of y and then adding columnwise, for a total of
roughly oJ.operations. During the 1970s,mathematicians developed means for hastening multiplication of two D-digit numbers
byway of the FastFourier Transform (FFT). The FFT reduces the
number of operations down to the order of D log D. (For exam-

it signals as bipolar, meaning both positive and negative digits


are allowed. Another is to HweightHthe signals by first multiplying each one by some other special signal. These enhancements
have enabled mathematicians to discover new prime numbers
and prove that certain numbers are prime or composite (not
prime).
-R.E.C

x = 18986723467242

+ I ~b~an~d ~gn;{~:
-,v--vv ""
V\

pie, for two 1,000-digit numbers, the grammar school method


may take more than 1,000,000 operations, whereas an FFT might

"b I

.+

take

only

50,000

operations.)

A full discussion of the FFT algorithm for multiplication is be-

yond the scope of this article. In brief, the digits of two numbers,
x and y (actually, the digits in some number base most convenient for the computing machinery) are thought of as signals.
The FFTis applied to each signal in order to decompose the signal into its spectral components. This is done in the same way
that a biologist might decompose a whale song or some other
meaningful signal into frequency bands. These spectra are
kl
I I d
h f
b f
Th
.-pro
qulC y mu tIp Ie toget er, requency y requency.
en an Inverse FFT and some final manipulations are performed to yield
the digits of the product of x and y.
There are various, powerful modern enhancements to this bax.
sic FFTmultiplication. One such enhancement is to treat the dig-

60

. .

SCIENTIFIC AMERICAN

February 1997

=1001345787777

A..~":~~A~ance

.
sIgna

I"
0

ComputeFFTspectraof x andY
signalsand multiply spectratogether
I
..
+ I A..II
~.
T:raduct .npctrum "
d uct spectrum
Takeinversespectrumand
unbalance,to obtain x timesy /
-
y = 19012260610305440490

Tb" C/7al/e/lgeof Large Nli/nbcrs

check this at home, the one-billionth dec- \" Just as with factoring problems, provmore than one million decimal digits.
imal place of x, Kanada says, is nine.
ing that a large number is prime is much Almost all the work to resolve the char\"
FFf has also been used to find large more complicated if the number is arbiacter of F22 depended on yet another
prime numbers. Over the past decade or trary-that is, if it is not of some special modification of FFf multiplication. This
so, David Slowinski of Cray Research form, as are the Mersenne primes. For proof has beencalled the longest calcuhas made a veritable art of discovering
primes of certain special fornls, "large"
lation ever perfornled for a "one-bit,"
record primes. Slowinski and his co- falls somewhere in the range of 2 I,O(JO,()()().
or yes-no, answer, and it took ahout
worker Paul Gage uncovered the prime
But currently it takes considerable com1016 computer operations. That is
21,257,787
-1 in mid-1996. A few months
putational effort to prove that a "ranroughly the same amount that went into
later, in November, programmers Joel dom" prime having only a few thousand generating the revolutionary Pixar-DisArmengaud of Paris and George F. digits is indeed prime. For example, in ney movie Toy StOI}',with its gloriously
Woltman of Orlando, Fla., working as 1992 it took severalweeks for Franl;ois rendered surfaces and animations.
part of a network project run by WoltMorian of the University of Claude Ber- z. \ Although it is natural to suspect the
man, found an even larger
prime: 21,398,269_1. This
number, which has over
400,000 decimal digits, is
the largest known prime
number as of this writing.
It is, like most other record
holders, a so-called Mersenne prime. These numbers take the form 2Q-1,
where q is an integer, and
are named after the 17thcentury French mathematician Marin Mersenne.\1For this latest discovery,
W91trpan optimized an algorithm called an irrational-base discrete weighted transform, the theory of
which I developed in 1991
with Barry Fagin of Dartmouth College and Joshua, I
Doenias of NeXT Software
in Redwood City, Calif. ~
This method was actually a ~
by-product of cryptography ..
~
research at NeXT.
COLOSSI become somewhat easierto contemplate-and compare-if one adopts a statistical view.
years before a parrot, pecking randomly at a
18 Blaine Garst, Doug For instance, it would take approximately 103,00(),OOO
Mitchell, Avadis Tevanian, keyboard, could reproduce by chanceThe Hound of the Baskervilles.This time span, though enorJr. and I implemented at mous, pales in comparison to the 101033
years that would elapsebefore fundamental quantum fIucN;XT what is one of the tuations might topple a beer can on a levelsurface. .
strongest-if not the strongest-encryption schemesavailable today, nard, using techniques developed joint-ly
based on Mersenne primes. This patentwith A.O.L. Atkin of the Universityof
ed scheme, termed Fast Elliptic EncrypIllinois, and others, to prove by comtion (FEE), uses the algebra of elliptic
puter that a particularl,505-digit numcurves, and it is very fast. Using, for exber,termed a partition number, is prime.
ample, the newfound Armengaud-Woltman prime 21,398,269
-1 as a basis, the
Colossal Composites
FEE system could readily encrypt this 1.0
issue of Scielttific Allterican into seem- I tis quite a bit easierto prove tilat some
ing gibberish. Under current numbernumber is not prime (that it is comtheoretical beliefs about the difficulty of posite, that is, made up of more than
cracking FEE codes, it would require, one prime factor). In 1992 Doenias,
without knowing the secret key, all the Christopher Norrie of Amdahl Corpocomputing power on earth more than
ration and I succeeded in proving by
1010,000years to decrypt the gibberish machine that the 22nd Fermat number,
back into a meaningful magazine.
2222+ 1, is composite. This number has

TI,prhn/fp"up
n{1~rupN"...I",r.

validity of any machine proof, there is a


happy circumstance connected with this
one. An independent team of Vilmar
Trevisan and Joao B. Carvalho, working at the Brazilian Supercomputer Center with different machinery and software (they used, in fact, Bailey's FFf
software) and unaware of our completed proof, also concluded that F22is composite. Thus, it seemsfair to say,without
doubt, that F22is composite. Moreover,
F22 is also now the largest "genuine"
composite known-which means that
even though we do not know a single
explicit factor for F22 other than itself
and 1, we do know that it is not prime.

1-'L1ust as with Archimcdcs' sand grains


dcnt that great dctcctive epic, by Sir Arin his time, therc will always hc colossal
thur Conan Doyle, The HOI",d of the
numbers that transccnd thc prcvailin~
Rt/s/~"'l'ill's. To witncss a pcrfcctly
tools. Neverthclcss, thesc numhcrs can
spcllcd manuscript, one would expect
still be imagined anJ studied. m particto w:ltch thc bird work for approxiular, it is often hclpful tl) envisil)l1 statism:ltcly I OI,()()O,OQ()
yc:lrs. The proh:lhlc
tical or biolo~iC;l1 sl:elmrios. I:or in:lgr: of thc tmivcrse is morc likc a paltry
I () 10vr:ar~,
stance, the numocr I() to the thrcc-millionth powcr oegins to make somc ~l3 Rtlt I o-',IX)O,()()O
i~ as nothing compared
intuitive sense if we ask how long it
with thc time needed in other scenarios.
would take a lahoratory parrot, peckIm:l~inc :I full br:cr call, sitting on a level,
ing randomly and tirclcssly at a keystcady, rough-surfaccd tablc, suddenly
board, with a talon occasionally pumptoppling over on its side, an event made
ing the shift key, say, to rcnocr by 3ccipossihle by fund:lmcntal qu:lntum fluc-

tuations. Indeed,a physicistmight grant


that the quantum wave function of the
can does extend, ever so slightly, away
from the can so that toppling is not impossible. Calculations show that one
would expcctto wait about 101033
years
for the surprise event. Unlikely as the
can toppling might be,one can imagine
more staggeringodds. What is the probability, for example, that sometime in
your life you will suddenlyfind yourself
standing on planet Mars, reassembled
and at least momentarily alive?Making
sweeping assumptions about the reassembly of living matter, I estimate the
odds against this bizarre event to be
101051to 1. To write theseodds in decimal form, you would needa 1 followed
by a zero for everyone of Archimedes'
sand grains. To illustrate how unlikely
Mars teleportation is, consider that the
great University of Cambridge mathematician John Littlewood once estimated the odds against a mouse living on
the surface of the sun for a week to be
101042to 1.2.1f.
These doubly exponentiated numbers
pale in comparison to, say, Skewes's
number, 10101034,
which has actually
beenused in developing a theory about
the distribution of prime numbers. To
show the existence of certain difficultto-compute functions, mathematicians
have invoked the Ackermann numbers
(named after Wilhelm Ackermann of
the Gymnasien in Luedenscheid,Germany), which composea rapidly growing sequencethat runs: 0, 1,22,3333..,.
The fourth Ackenn3nn number,involving exponenti3ted 3's, is approximately
103,638,334,640,024.
The fifth one is so
13rgeth3t it could not be written on a
universe-sizesheetof paper, evenusing
exponenti31notation! Compared with
the fifth Ackerm3nn number,the mighty
googolplex is but a spit in the proverbi31bucket.
~

Ine Al/thor

Fltrther Readil1g

RICHARD E. CltANOALL
is chicf scicl1tist at NcXT Softwarc. Hc is also Vollum Adjunct Profcss{)r of 5cicncc and dircctor of thc Ccntcr for AJvanccd Computation at Rccd Collcgc. Crandall is thc
author of scven patcnts, {)II sul,jccts ran~in~ from clcctrc)llics tCt thc r;1st Elli"tic
l:ncryptu)11 systcm. 1111973 hc reccivcJ
his Ph.D. in physil."s from thc Mass;1chusctts Institutc of lcchl1olc)gy.

TIIF. \~'c)RKS01' AR(;JIIMEI>ES. Editcd hy T. L. Heath. Cambridge University Press, 1897.


'Ii II'. \Y/ORI.D 01' MAll11'.MATICS. I:dward Kasner and James R. Newman. Simon & Schuster,

62

SCIENTIFIC AMl'.J{I(;,\N

Fchrllary

1997

, 9.5(,.
TJIE FI\8RIC 01' 11iE HEI\VEN~: THE DEVELOPMENTOF ASTRONOMY AND DYNAMICS. Stephen
limlmin and.June Goodfield. Harper & Row,1961.
AN INTRODUCTION
\"(!right.

(;)arclldc)n

TO TilE

TJIEORY

OF NUMBERS.

Fifth

edition.

G. H.

Hardy

and E. M.

Press, 1978.

I.rrru,\,'ooo's
MISCELlANY. Editcd hy Bcla Bollohas. Cambridge Uni,'ersity Press, 1986.
I.IIRI' OJ' 11IE INTEGFR~.J. Rohcrts. M:lthematical Association of America, 1992.
PRC).JI;(:TS
IN SCIENTII'IC COMI'UTATION. Richard E. Crandall. TELOS/Springer-Verlag, 1994.

The Cl,allengeof Large N'tlnbers

.--"..""..,.~~~,~:

..:.c',

You might also like