Documentos de Académico
Documentos de Profesional
Documentos de Cultura
TECHNOLOGY
THE
LIMITSOF
Quantum
By Scott Aaronson
H
aggar Physicists Develop Quantum Slacks, read a headline
in the satirical weekly the Onion. By exploiting a bizarre
Schrdingers Pants duality, the article explained, these
non-Newtonian pants could paradoxically behave like formal wear and
casual wear at the same time. Onion writers were apparently spoong
the breathless articles about quantum computing that have lled the
popular science press for a decade.
A common mistake see for instance the February 15, 2007, issue of
the Economist is to claim that, in principle, quantum computers could
rapidly solve a particularly difcult set of mathematical challenges
called NP-complete problems, which even the best existing computers
cannot solve quickly (so far as anyone knows). Quantum computers
would supposedly achieve this feat not by being formal and casual at
the same time but by having hardware capable of processing every pos-
sible answer simultaneously.
If we really could build a magic computer capable of solving an NP-
ILLUSTRATIONS BY DUAN PETRICIC
62 March 2008
2008 SCIENTIFIC AMERIC AN, INC.
Computers
ity. Given any holy grail of mathematics such quantum computer and its environment, which
as Goldbachs conjecture or the Riemann hy- introduces errors). In particular, the bounds on
pothesis, both of which have resisted resolution what it is mathematically possible to program a
for well over a century we could simply ask our computer to do would apply even if physicists
computer to search through all possible proofs managed to build a quantum computer with no
and disproofs containing up to, say, a billion sym- decoherence at all.
bols. (If a proof were much longer than that, it is
not clear that we would even want to read it.) Hard, Harder, Hardest
If quantum computers promised such godlike How is it that a quantum computer could pro-
mathematical powers, maybe we should expect vide speedups for some problems, such as break- KEY CONCEPTS
them on store shelves at about the same time as ing codes, but not for others? Isnt a faster com-
Quantum computers would
warp-drive generators and antigravity shields. puter just a faster computer? The answer is no,
exploit the strange rules of
But although we should not accept the usual and to explain why takes one straight to the intel-
quantum mechanics to process
hype, in my view it is equally misguided to dis- lectual core of computer science. For computer information in ways that are
miss quantum computing as science ction. In- scientists, the crucial thing about a problem is impossible on a standard
stead we should nd out what the limits of quan- how quickly the time needed to solve it grows as computer.
tum computers are and what they could really the problem size increases. The time is measured
They would solve certain specif-
do if we had them. in the number of elementary steps required by
ic problems, such as factoring
In the 26 years since physicist Richard Feyn- the algorithm to reach a solution. For example, integers, dramatically faster
man rst proposed the idea of quantum comput- using the grade school method, we can multiply than we know how to solve
ing, computer scientists have made enormous two n-digit numbers in an amount of time that them with todays computers,
progress in guring out what problems quantum grows like the number of digits squared, n2 (an but analysis suggests that for
computers would be good for. According to our amount of time said to be a polynomial in n). most problems quantum com-
current understanding, they would provide dra- But for factoring a number into primes, even the puters would surpass conven-
matic speedups for a few specic problems most advanced methods known take an amount tional ones only slightly.
such as breaking the cryptographic codes that of time that grows exponentially with the num- Exotic alterations to the known
are widely used for monetary transactions on ber of digits (in particular, like 2 to the cube root laws of physics would allow
the Internet. For other problems, however such of n power). Thus, factoring seems intrinsically construction of computers that
as playing chess, scheduling airline ights and harder than multiplying and when we get up to could solve large classes of hard
proving theorems evidence now strongly sug- thousands of digits, this difference matters much problems efciently. But those
gests that quantum computers would suffer more than the difference between a Commodore alterations seem implausible. In
the real world, perhaps the im-
from many of the same algorithmic limitations 64 and a supercomputer.
possibility of efciently solving
as todays classical computers. These limitations The kind of problems that computers can
these problems should be taken
are completely separate from the practical dif- solve in a reasonable amount of time, even for
as a basic physical principle.
culties of building quantum computers, such as large values of n, are those for which we have an
decoherence (unwanted interaction between a algorithm that uses a number of steps that grows The Editors
w w w. S c i A m . c o m SCIENTIFIC AMERICAN 63
2008 SCIENTIFIC AMERIC AN, INC.
what better than trying every possible solution
Quantum Computing 101 are known for these problems, no algorithm is
known that is fundamentally better. Every known
P hysicists are hotly pursuing the construction of quantum computers, which would
harness the quirks of quantum mechanics to perform certain computations more
efciently than a conventional computer.
algorithm will take an amount of time that in-
creases exponentially with the problem size.
It turns out that the three problems I just list-
ed have a very interesting property: they are all
1 The fundamental feature of a quantum computer is that it
the same problem, in the sense that an efcient
uses qubits instead of bits. A qubit may be a particle such as an
electron, with spin up (blue) representing 1, spin down algorithm for any one of them
(red) representing 0, and quantum states called super- would imply efficient algo-
positions that involve spin up and spin down simultane- rithms for all the others. Ste-
ously (yellow). 1 phen A. Cook of the University
of Toronto, Richard Karp of the
2 A small number of particles in superposition University of California, Berkeley,
states can carry an enormous amount of informa- and Leonid Levin, now at Boston
tion: a mere 1,000 particles can be in a superpo- University, arrived at this remarkable
sition that represents every number from 1 to conclusion in the 1970s, when they de-
21,000 (about 10300 ), and a quantum comput-
2
veloped the theory of NP-completeness.
er would manipulate all those numbers in NP stands for nondeterministic poly-
parallel, for instance, by hitting the parti- nomial time. Do not worry about what that
cles with laser pulses. means. Basically, NP is the class of problems
for which a solution, once found, can be recog-
3 When the particles states are
nized as correct in polynomial time (something
measured at the end of the compu-
tation, however, all but one ran-
3 like n2 , and so on) even though the solution it-
dom version of the 10300 paral- self might be hard to nd. As an example, if you
lel states vanish. Clever are given a map with thousands of islands and
manipulation of the particles bridges, it may take years to nd a tour that vis-
could nonetheless solve its each island once. Yet if someone shows you a
certain problems very tour, it is easy to check whether that person has
rapidly, such as factoring succeeded in solving the problem. When a prob-
a large number. lem has this property, we say that it is in NP. The
class NP captures a huge number of problems of
as n raised to a xed power, such as n, n2 or n 2.5. practical interest. Note that all the P problems
Computer scientists call such an algorithm ef- are also NP problems, or to put it another way,
cient, and problems that can be solved by an ef- the class P is contained within the class NP. If
cient algorithm are said to be in the complexity you can solve a problem quickly you can also
class P, which stands for polynomial time. verify the solution quickly.
A simple example of a problem in P is: Given NP-complete problems are in essence the
a road map, is every town reachable from every hardest of the NP problems. They are the ones
other town? P also contains some problems with the property found by Cook, Karp and
whose efcient solutions are not so obvious, Levin: If an efcient algorithm for any one of
A good quantum such as: Given a whole number, is it prime (like them were found, it could be adapted to solve all
computer algo- 13) or composite (like 12)? Given a list of which the other NP problems as well.
men and women are willing to marry one anoth- An efcient algorithm for an NP-complete
rithm ensures that er, is it possible to pair everyone off with a will- problem would mean that computer scientists
computational ing partner? present picture of the classes P, NP and NP-com-
But now suppose you are given the dimen- plete was utterly wrong, because it would mean
paths leading to sions of various boxes and you want a way to that every NP problem (including all the NP-
a wrong answer pack them in your trunk. Or suppose that you complete ones) was actually a P problem. In oth-
are given a map and you want to color each er words, the class P would equal the class NP,
cancel out and country red, blue or green so that no two neigh- which is written P = NP.
that paths leading boring countries are colored the same. Or that Does such an algorithm exist? Is P equal to
you are given a list of islands connected by bridg- NP? That is literally a million-dollar question
to a correct es and you want a tour that visits each island ex- it carries a $1,000,000 reward from the Clay
answer reinforce. actly once. Although algorithms that are some- Math Institute in Cambridge, Mass. and it has
w w w. S c i A m . c o m SCIENTIFIC AMERICAN 65
20 08 SCIENTIFIC AMERIC AN, INC.
What Classical Computers called destructive interference. So a good quan-
tum computer algorithm would ensure that com-
March 2008
20 08 SCIENTIFIC AMERIC AN, INC.
achieve an exponential speedup by treating the it just produces a smaller exponential. And Gro- [THE AUTHOR]
problems as structureless black boxes, con- vers algorithm is as good as it gets for this kind Scott Aaronson is an assistant
sisting of an exponential number of solutions to of black box searching: in 1994 researchers had professor of electrical engineering
be tested in parallel. Some speedup can nonethe- shown that any black box quantum algorithm and computer science at the
Massachusetts Institute of Technol-
less be wrung out of this black box approach, needs at least S steps. ogy. A high school dropout, he
and computer scientists have determined just Over the past decade, researchers have went on to receive a bachelors
how good and how limited that speedup is. shown that similar modest speedups are the degree from Cornell University and
The algorithm that produces the speedup is the limit for many other problems besides search- a Ph.D. in computer science from
second major quantum algorithm. ing a list, such as counting ballots in an election, the University of California, Berke-
ley, advised by Umesh Vazirani.
The black box approach can be illustrated by nding the shortest route on a map, and playing Outside of research, Aaronson is
pretending that you are searching for the solu- games of strategy such as chess or Go. One best known for his widely read blog
tion to a difcult problem and that the only op- problem that presented particular difculty was (www.scottaaronson.com/blog), as
eration you know how to perform is to guess a the so-called collision problem, the problem of well as for creating the Complexity
solution and see if it works. Let us say there are nding two items that are identical, or that Zoo (www.complexityzoo.com), an
online encyclopedia of more than
S possible solutions, where S grows exponential- collide, in a long list. If there were a fast quan- 400 complexity classes.
ly as the problem size n increases. You might get tum algorithm to solve this problem, many of
lucky and guess the solution on your rst try, but the basic building blocks of secure electronic
in the worst case you will need S tries, and on av- commerce would be useless in a world with
erage you will need S/2. quantum computers.
Now suppose you can ask about all the pos- Searching a list for an item is like looking for
sible solutions in quantum superposition. In a needle in a haystack, whereas searching for a
1996 Lov Grover of Bell Laboratories developed collision is like looking for two identical pieces
an algorithm to nd the correct solution in such of hay, which provides the problem with a kind
a scenario using only about S steps. A speedup of structure that a quantum computer could
from S/2 to S is a useful advance for some potentially exploit. Nevertheless, I showed in
problems if there are a million possible solu- 2002 that within the black box model, any
tions, you will need around 1,000 steps instead quantum algorithm needs exponential time to
of 500,000. But the square root does not trans- solve the collision problem.
form an exponential time into a polynomial time; Admittedly, these black box limitations do
n n chess
PSPACE n n Go
BQP does not seem to t neatly with the other classes.)
The BQP class (the letters stand for bounded-error, quantum, Box packing
polynomial time) includes all the P problems and also a few other NP Map coloring
NP-
complete Traveling salesman
problems, such as factoring and the so-called discrete logarithm n n Sudoku
problem. Most other NP and all NP-complete problems are believed to
Efciently solved NP Graph isomorphism
be outside BQP, meaning that even a quantum computer would require
by quantum
Harder
outside the class known as PSPACE, which also contains all the
NP problems. PSPACE problems are those that a conventional
computer can solve using only a polynomial amount of memory
but possibly requiring an exponential number of steps.
w w w. S c i A m . c o m SCIENTIFIC AMERICAN 67
20 08 SCIENTIFIC AMERIC AN, INC.
ZONES OF not rule out the possibility that efcient quan- both then at M.I.T., showed that if a small non-
tum algorithms for NP-complete or even harder linear term is added to the equations of quantum
THOUGHT problems are waiting to be discovered. If such mechanics, quantum computers would be able
Unlike the real world, in which
algorithms existed, however, they would have to to solve NP-complete problems efciently. Be-
computational limits are believed
to be the same everywhere, the exploit the problems structure in ways that are fore you get too excited, you should realize that
galaxy in Vernor Vinges 1992 unlike anything we have seen, in much the same if such a nonlinear term existed, then one could
science-ction novel A Fire Upon way that efcient classical algorithms for the also violate Heisenbergs uncertainty principle
the Deep is divided into concen- same problems would have to. Quantum magic and send signals faster than the speed of light. As
tric zones of thought having
by itself is not going to do the job. Based on this Abrams and Lloyd pointed out, perhaps the best
different inherent computational
and technological limits. insight, many computer scientists now conjec- interpretation of these results is that they help to
ture not only that P NP but also that quantum explain why quantum mechanics is linear.
In the Unthinking Depths, nearest computers cannot solve NP-complete problems Another speculative type of machine would
the galactic core, even simple in polynomial time. achieve extravagant computational abilities by
automation fails and IQs plummet. cramming an innite number of steps into a -
Magical Theories nite time. Unfortunately, according to physi-
The Slow Zone contains Earth and is
Everything we know is consistent with the pos- cists current understanding, time seems to de-
as limited as we know it.
sibility that quantum computers are the end of generate into a sea of quantum uctuations
In the Beyond, nearly sentient the line that is, that they are the most general something like a foam instead of a uniform
nanotechnology factories construct kind of computer compatible with the laws of smooth line on the scale of 10 43 second (the
wonders such as anti-gravity fabrics, physics. But physicists do not yet have a nal the- Planck time), which would seem to make this
and hypercomputation enables ory of physics, so one cannot rule out the possi- kind of machine impossible.
faster-than-light travel.
bility that someday a future theory might reveal If time cannot be sliced with arbitrary thin-
a physical means to solve NP-complete problems ness, then perhaps another way to solve NP-
The Transcend is populated by
dangerous, godlike ber- efciently. As you would expect, people specu- complete problems efciently would be to ex-
intelligences having technologies late about yet more powerful kinds of comput- ploit time travel. Physicists studying the issue
and thought processes ers, some of which would make quantum com- talk not about time machines but about closed
unfathomable to lower beings. puters look as pedestrian as vending machines. timelike curves (CTCs). In essence a CTC is a
All of them, however, would rely on speculative route through space and time that matter or en-
changes to the laws of physics. ergy could travel along to meet up with itself in
One of the central features of quantum me- the past, forming a closed loop. Current physi-
chanics is a mathematical property called linear- cal theory is inconclusive on whether CTCs can
ity. In 1998 Daniel S. Abrams and Seth Lloyd, exist, but that need not stop us from asking what
the consequences would be for computer science
if they did exist.
ber-Computers from Exotic Physics? It seems obvious how one could use a CTC to
speed up a computation: program your comput-
March 2008
20 08 SCIENTIFIC AMERIC AN, INC.
A New Physical Principle?
B ecause implausible kinds of physics (such as time travel) seem neces-
sary for constructing a computer able to solve NP-complete problems
quickly, I predict that scientists might one day adopt a new principle: NP-
complete problems are hard. That is, solving those problems efciently is impos-
sible on any device that could be built in the real world, whatever the nal laws of
physics turn out to be. The principle implies that time travel is impossible, because such
travel would enable creation of ber-computers that could solve NP-complete problems
efciently. Further, if a proposed theory were shown to permit such computers, that theo-
ry could be ruled out. Application of the principle would be analogous to applying the
laws of thermodynamics to conclude that perpetual-motion machines are
impossible (the laws of thermodynamics forbid them) and to deduce previ-
ously unknown features of physical processes. S.A.
around inside the CTC to solve hard problems. impossibility of faster-than-light communica- More about nonlinear
Indeed, by using a CTC, we could efciently tion two earlier limitations on technology that quantum mechanics,
hypercomputing, use
solve not only NP problems but even problems over time earned the status of physical princi-
of time travel, and
in an apparently larger class called PSPACE. ples. Yes, the second law might be experimen- another scheme called anthropic
PSPACE is the class of problems that could be tally falsied tomorrow but until that hap- computing can be found at
solved on a conventional computer using a poly- pens, physicists nd it vastly more useful to as- www.SciAm.com/ontheweb
nomial amount of memory but possibly taking sume it is correct and then use that assumption
an exponential amount of time. In effect, a CTC for studying everything from car engines to
would make time and space interchangeable as black holes. I predict that the hardness of NP-
computational resources. (I did not have to men- complete problems will someday be seen the
tion the polynomial memory constraint until same way: as a fundamental principle that de-
now, because for P and NP problems it makes no scribes part of the essential nature of our uni-
difference if the computer has access to more verse. There is no way of telling what theoreti-
than polynomial memory.) Recently John Wa- cal enlightenment or what practical conse-
trous of the University of Waterloo in Ontario quences might come from future application of MORE TO
and I showed that using a quantum computer in this kind of fundamental principle. EXPLORE
a CTC instead of a conventional one does not In the meantime, we know not to expect mag- Quantum Computation and
enable anything beyond PSPACE to be efcient- ic from quantum computers. To some, the ap- Quantum Information. Michael A.
ly solved. In other words, if CTCs exist, then parent limitations of quantum computers might Nielsen and Isaac L. Chuang.
quantum computers are no more powerful than come as a letdown. One can, however, give those Cambridge University Press, 2000.
classical ones. same limitations a more optimistic spin. They
NP-Complete Problems and
mean that although certain cryptographic codes Physical Reality. Scott Aaronson in
Computational Kryptonite could be broken in a world with quantum com- ACM SIGACT News, Complexity
Physicists do not know if future theories will per- puters, other codes would probably remain se- Theory Column, Vol. 36, No. 1,
mit any of these extraordinary machines. Yet cure. They increase our condence that quan- pages 3052; March 2005. Avail-
without denying our ignorance, we can view that tum computing will be possible at all because able at www.scottaaronson.com/
papers/npcomplete.pdf
ignorance from a different perspective. Instead the more a proposed technology sounds like a
of starting from physical theories and then ask- science-ction caricature, the more skeptical we Quantum Computer Science: An
ing about their computational implications, we should be. (Who would you be more inclined to Introduction. N. David Mermin.
could start by assuming that NP-complete prob- believe: the salesperson offering a device that Cambridge University Press, 2007.
lems are hard and then study the consequences produces unlimited free energy from the quan-
Shor, Ill Do It. (An explanation of
of that assumption for physics. For instance, if tum vacuum or the one offering a refrigerator Shors algorithm for the layperson.)
CTCs would let us solve NP-complete problems that is more efcient than last years model?) And Scott Aaronson. Available at www.
efciently, then by starting from the assumption last, such limitations ensure that computer sci- scottaaronson.com/blog/?p=208
that NP-complete problems are intractable, we entists will continue to have their work cut out
could conclude that CTCs cannot exist. for them in designing new quantum algorithms. Quantum Computing since Dem-
ocritus. Lecture notes from course
To some, such an approach will seem overly Like Achilles without his heel or Superman with- PHYS771, University of Waterloo,
dogmatic. To me, it is no different from assum- out kryptonite, a computer without any limita- Fall 2006. Available at www.
ing the second law of thermodynamics or the tions would get boring pretty quickly. scottaaronson.com/democritus/
w w w. S c i A m . c o m SCIENTIFIC AMERICAN 69
20 08 SCIENTIFIC AMERIC AN, INC.