The Fundamental Physical

Limits of Computation

What constraints govern the physical process o f computing? Is a

minim um amount of energy required, for example, per logic step?
There seems to be no minimum, but some other questions are open

by Charles H. Bennett and Rolf Landauer

A computation, whether it is per-

formed by electronic machin-
ery, on an abacus or in a biolog-
ical system such as the brain, is a physi-
ber of other workers at other institu-
tions have entered this field.
In our analysis of the physical lim-
its of computation we use the term "in-
In fact, computation as it is current-
ly carried out depends on many opera-
tions that destroy information. The so-
called and gate is a device with two
cal process. It is subject to the same formation" in the technical sense of input lines, each of which may be set at
questions that apply to other physical information theory. In this sense infor- 1 or 0, and one output, whose value
processes: How much energy must be mation is destroyed whenever two pre- depends on the value of the inputs. If
expended to perform a particular com- viously distinct situations become in- both inputs are 1, the output will be 1.
putation? How long must it take? How distinguishable. In physical systems If one of the inputs is 0 or if both are 0,
large must the computing device be? In without friction, information can nev- the output will also be 0. Any time the
other words, what are the physical lim- er be destroyed; whenever information gate's output is a 0 we lose informa-
its of the process of computation? is destroyed, some amount of ener- tion, because we d o not know which
So far it has been easier to ask these gy must be dissipated (converted into of three possible states the input lines
questions than to answer them. T o the heat). As an example, imagine two eas- were in (0 and 1, 1 and 0, or 0 and 0).
extent that we have found limits, they ily distinguishable physical situations, In fact, any logic gate that has more in-
are terribly far away from the real lim- such as a rubber ball held either one put than output lines inevitably dis-
its of modern technology. We cannot meter or two meters off the ground. If cards information, because we cannot
profess, therefore, to be guiding the the ball is dropped, it will bounce. If deduce the input from the output.
technologist or the engineer. What we there is no friction and the ball is per- Whenever we use such a "logically ir-
are doing is really more fundamental. fectly elastic, a n observer will always reversible" gate, we dissipate energy
We are looking for general laws that be able to tell what state the ball start- into the environment. Erasing a bit of
must govern all information process- ed out in (that is, what its initial height memory, another operation that is fre-
ing, n o matter how it is accomplished. was) because a ball dropped from two quently used in computing, is also fun-
Any limits we find must be based sole- meters will bounce higher than a ball damentally dissipative; when we erase
ly on fundamental physical principles, dropped from one meter. a bit, we lose all information about
not on whatever technology we may If there is friction, however, the ball that bit's previous state.
currently be using. will dissipate a small amount of ener- Are irreversible logic gates and era-
There are precedents for this kind gy with each bounce, until it eventual- sures essential t o computation? If they
of fundamental examination. In the ly stops bouncing and comes to rest are, any computation we perform has
1940's Claude E. Shannon of the Bell on the ground. It will then be impos- to dissipate some minimum amount
Telephone Laboratories found there sible to determine what the ball's ini- of energy.
are limits on the amount of informa- tial state was; a ball dropped from As one of us (Bennett) showed in
tion that can be transmitted through a two meters will be identical with a ball 1973, however, they are not essential.
noisy channel; these limits apply no dropped from one meter. Information This conclusion has since been demon-
matter how the message is encoded will have been lost as a result of ener- strated in several models; the easiest of
into a signal. Shannon's work repre- gy dissipation. these to describe are based on so-called
sents the birth of modern information reversible logic elements such as the
science. Earlier, in the mid- and late example of informa-
19th century, physicists attempting to Heretionis another
destruction: the expression
Fredkin gate, named for Edward Fred-
kin of the Massachusetts Institute
determine the fundamental limits on +
2 2 contains more information than of Technology. The Fredkin gate has
the efficiency of steam engines had cre- the expression =4. If all we know is three input lines and three outputs.
ated the science of thermodynamics. that we have added two numbers to The input on one line, which is called
In about 1960 one of us (Landauer) yield 4, then we do not know whether the control channel, is fed unchanged
and John Swanson at IBM began at- + +
we have added 1 3, 2 2, 0 4 or + through the gate. If the control channel
tempting to apply the same type of some other pair of numbers. Since the is set a t 0, the input on the other two
analysis to the process of computing. output is implicit in the input, no com- lines also passes through unchanged. If
Since the mid-1970's a growing num- putation ever generates information. the control line is a 1, however, the
CONVENTIONAL COMPUTING DEVICES, the abacus and the dissipative because of friction between its beads and rods. It could
logic chip, both dissipate energy when they are operated. The 'Vog- not be built of frictionless components: if there were no static fric-
ic gates" central to the design of a chip expend energy because they tion, the beadsypositions would change under the influence of ran-
discard information. A chip consumes energy for a less fundamen- dom thermal motion. Static friction exerts a certain minimum force
tal reason as well: it employs circuits that draw power even when no matter what the beads' velocity, and so there is some minium en-
they merely hold information and d o not process it. The abacus is ergy that the abacus requires no matter how slowly it is operated.
outputs of the other two lines are if we erase them, it will cost us all it to its original state--without dissi-
switched: the input of one line be- the energy dissipation we have been pating any energy.
comes the output of the other and vice trying to avoid.
versa. The Fredkin gate does not dis- Actually these bits have a most im- o far we have discussed a set of log-
card any information; the input can al- portant use. Once we have copied S ic operations, not a physical device.
ways be deduced from the output. down the result of our computation, It is not hard, however, to imagine
Fredkin has shown that any logic de- which will reside in the normal output a physical device that operates as a
vice required in a computer can be im- bits, we simply run the computer in Fredkin gate. In this device the infor-
plemented by an appropriate arrange- reverse. That is, we enter the "gar- mation channels are represented by
ment of Fredkin gates. To make the bage bits" and output bits that were pipes. A bit of information is repre-
computation work, certain input lines produced by the computer's normal sented by the presence or absence of a
of some of the Fredkin gates must be operation as "input" into the "back ball in a particular section of pipe; the
preset at particular values [see lower il- end" of the computer. This is possible presence of a ball signifies a 1 and the
lustration below]. because each of the logic gates in the absence of a ball signifies a 0.
Fredkin gates have more output computer is itself reversible. Running The control line is represented by a
lines than the gates they are made to the computer in reverse discards no in- narrow segment of pipe that is split
simulate. In the process of computing, formation, and so it need not dissipate lengthwise down the middle. When a
what seem to be "garbage bits," bits of any energy. Eventually the computer ball enters the split segment of pipe, it
information that have no apparent use, will be left exactly as it was before the pushes the two halves of the pipe
are therefore generated. These bits computation began. Hence it is possi- apart, actuating a switching device.
must somehow be cleared out of the ble to complete a "computing cyclew- The switching device channels any in-
computer if we are to use it again, but to run a computer and then to return put balls that may be in the other two
pipes: when a ball is present in the con-
trol line, any ball that enters an input
AND GATE OR GAT pipe is automatically redirected to the
INPUT OUTPUT INPUT OUTPUT other pipe. To ensure that the switch is
A A closed when no control ball is present,
there are springs that hold the two
B halves of the split pipe together. A ball
entering the split pipe must expend en-
ergy when it compresses the springs,
but this energy is not lost; it can be
recovered when the control ball leaves
the split pipe and the springs expand.
All the balls are linked together and
pushed forward by one mechanism, so
that they move in synchrony; other-
wise we could not ensure that the vari-
CONVENTIONAL LOGIC GATES dissipate energy because they discard information. ous input and controlling balls would
For example, if the output of an and gate is 0, there is no way to deduce what the input was. arrive at a logic gate together. In a
sense the forward progress of the com-
putation is really nhotion along a sin-
REVERSIBLE gle degree of freedom, like the motion
FREDKIN GATE AND GATE -GARBAGE" of two wheels rigidly attached to one
INPUT OUTPUT axle. Once the computation is done we
push all the balls backward, undoing
all the operations and returning the
A computer to its initial state.
/ If the entire assembly is immersed in
PRE-SE~ BIT an ideal viscous fluid, then the friction-
al forces that act on the balls will be
proportional to their velocity; there
will be no static friction. The frictional
force will therefore be very weak if we
are content to move the balls slowly. In
any mechanical system the energy that
must be expended to work against fric-
tion is equal to the product of the fric-
tional force and the distance through
which the system travels. (Hence the
faster a swimmer travels between two
points, the more energy he or she will
expend, although the distance traveled
FREDKIN REVERSIBLE LOGIC GATE need not dissipate energy; the input can always is the same whether the swimmer is
be deduced from the output. The gate has a ucontrol" line, the value of which is not changed fast or slow.) If we move the balls
by the gate. If the bit on the control line is a 0, the values of the other two lines are also un-
touched; if it is a I, however, the input of line A becomes the output of line B and vice versa. through the Fredkin gates at a low
Reversible gates can be arranged to implement any function performed by an irreversible speed, then the energy expended (the
gate. To implement the and operation (right) one input is preset to equal 0, and two output product of force and distance) will be
bits, called garhge bits, are temporarily ignored. When the computation is complete, these very small, because the frictional force
bits are used to operate the gate in reverse, returning the computer to its original state. depends directly on the balls' speed. In
IDEALIZED PHYSICAL REALIZATION of a Fredkin gate sub- control channel closed when no ball is in it. This gate does not need
stitutes pipes for wires and the presence or absence of a ball for a static friction in order to operate; it could be immersed in a viscous
1 or 0. A narrow, split segment of pipe represents the control chan- fluid, and the frictional forces could be made to depend only on
nei. When a ball passes through it, the pipe spreads apart, operating the balls' velocity. Then the energy dissipation could be as small as
a switching mechanism; the mechanism in turn switches any input the user wished: to lower the amount of energy dissipated, it would
ball from line A to line B and vice versa. A pair of springs keeps the only be necessary to drive the balls through the device more slowly.

fact, we can expend as little energy as scribed arrangements of mirrors that slightly wrong angle, the balls' trajec-
we wish, simply by taking a long time correspond to different types of logic tories will go astray. One or more balls
to carry . o u t the operation. There is gate, and they have shown that billiard- will deviate from their intended paths,
thus no minimum amount of energy ball models can be built to simulate and in due course errors will combine
that must be expended in order to per- any logic element that is necessary for to invalidate the entire computation.
form any given computation. computing. Even if perfectly elastic and friction-
T o start the computation we fire a less billiard balls could be manufac-
Themodel
energy lost t o friction in this
will be very small if the ma-
billiard ball into the computer wherev-
er we wish to input a 1. The balls must
tured, the small amount of random
thermal motion in the molecules they
chine is operated very slowly. Is it pos- enter the machine simultaneously. are niade of would be enough to cause
sible to design a more idealized ma- Since they are perfectly elastic, they errors after a few dozen collisions.
chine that could compute without any d o not lose energy when they collide; Of course we could install some kind
friction? Or is friction essential to the they will emerge from the computer of corrective device that would return
computing process? Fredkin, together with the same amount of kinetic ener- any errant billiard ball to its desired
with Tommaso Toffoli and others at gy we gave them at the beginning. path, but then we would be obliterat-
M.I.T., has shown that it is not. In operation a billiard-ball comput- ing information about the ball's ear-
They demonstrated that it is possible er produces "garbage bits," just as a lier history. For example, we might
to do computation by firing ideal, fric- computer built of Fredkin gates does. be discarding information about the
tionless billiard balls a t one another. In After the computer has reached an an- extent to which a mirror is tilted incor-
the billiard-ball model perfect reflect- swer we reflect the billiard balls back rectly. Discarding information, even
ing "mirrors," surfaces that redirect into it, undoing the computation. They to correct an error, can be done only in
the balls' motion, are arranged in such will come out of the machine exactly a system in which there is friction and
a way that the movement of the balls where we sent them in, and at the same loss of energy. Any correctional device
across a table emulates the movement speed. The mechanism that launched must therefore dissipate some energy.
of bits of information through logic them into the computer can then be Many of the difficulties inherent in
gates [see illustration on next page]. As used to absorb their kinetic energy. the billiard-ball computer can be made
before, the presence of a ball in a par- Once again we will have performed a less extreme if microscopic or submi-
ticular part of the computer signifies a computation and returned the com- croscopic particles, such as electrons,
1, whereas the absence of a ball signi- puter to its initial state without dissi- are used in place of billiard balls. As
fies a 0. If two balls arrive simulta- pating energy. Wojciech H. Zurek, who is now at the
neously at a logic gate, they will collide The billiard-ball computer has one Los Alamos National Laboratory, has
and their paths will change; their new major flaw: it is extremely sensitive to pointed out, quantum laws, which can
paths represent the output of the gate. slight errors. If a ball is aimed slightly restrict particles to a few states of mo-
Fredkin, Toffoli and others have de- incorrectly or if a mirror is tilted at a tion, could eliminate the possibility
that a particle might go astray by a (corresponding to a binary 1) and down chine: a Turing machine that does not
small amount. (corresponding to a 0). The interac- discard information and can therefore
Although the discussion so far has tions between particle spins can be pre- be run with as small an expenditure of
been based primarily on classical dy- scribed in such a way that the value of energy as the user wishes.
namics, several investigators have pro- one particle's spin changes depending A Turing machine has several com-
posed other reversible computers that on the spin of nearby particles; the spin ponents. There is a tape, divided into
are based on quantum-mechanical of the particle would correspond to discrete frames or segments, each of
principles. Such computers, first pro- one of the outputs of a logic gate. which is marked with a 0 or a 1; these
posed by Paul Benioff of the Argonne bits represent the input. A "read/write
National Laboratory and refined by
others, notably Richard P. Feynman Soedfaronthisinformation
discussion has concentrat-
processing. A
head" moves along the tape. The head
has several functions. It can read one
of the California Institute of Technol- computer must store information as bit of the tape at a time, it can print one
ogy, have so far been described only in well as process it. The interaction be- bit onto the tape and it can shift its
the most abstract terms. Essentially tween storage and processing is best position by one segment to the left or
the particles in these computers would described in terms of a device called a right. In order to remember from one
be arranged so that the quantum-me- Turing machine, for Alan M. Turing, cycle to the next what it is doing, the
chanical rules governing their interac- who first proposed such a machine head mechanism has a number ofedis-
tion would be precisely analogous to in 1936. A Turing machine can per- tinct "states"; each state constitutes a
the rules describing the predicted out- form any computation that can be per- slightly different configuration of the
puts of various reversible logic gates. formed by a modern computer. One of head's internal parts.
For example, suppose a particle's spin us (Bennett) has shown that it is possi- In each cycle the head reads the bit
can have only two possible values: up ble to build a reversible Turing ma- on the segment it currently occupies;

flB AND NOT A

'4A AND NOT B

B A AND B

BILLIARD-BALL COMPUTER employs the movement of bii- function of any logic chip. For example, a billiard-ball computer
liard balls on a table to simulate the movement of bits through logic could be made to test whether a number is prime. One such comput-
gates. In billiard-ball logic gates (left) the balls' paths are redirected er (right) accepts as input any fivebit number (in this case 01101,
by collisions with one another or with reflecting "mirrors."In addi- or 13) and the fixed input sequence 01. Like a Fredkin gate, a bil-
tion to their role in gates, mirrors can deflect a ball's path (a), s h i i liard-ball computer typically returns more output bits than its user
the pathsideways (b),delay the ball'smotion withoutchangingits final needs. In the case shown, the computer returns the original input
direction or position (c) or allow two l i e s to crass (d). It is possible number itself (which is the "extra" output), and an "answer" se-
to arrange mirrors so that the resulting Lbcomputer" implements the quence: 10 if the input number is prime and 01 if it is composite.
then it prints a new bit onto the tape,
changes its internal state and moves
one segment to the left or right. The bit A READMIRITE
it prints, the state it changes into and HEAD TAPE
the direction in which it moves are de- /
termined by a fixed set of transition
rules. Each rule specifies a particular
set of actions. Which rule the machine
follows is determined by the state of
the head and the value of the bit that it
reads from the tape. For example, one
rule might be: "If the head is in state A
and is sitting on a segment of tape that
is printed with a 0,it should change
that bit to a 1, change its state to state B
and move one segment to the right." It TURING MACHINE can be constructed in such a way that it can perform any computa-
may happen that the transition rule in- tion a computer can. An infinitely long tape is divided into discrete segments, each of which
structs the machine not to change its bears either a 0 or a 1. A "read/write head," which can be in any of several internal states
internal state, not to print a new bit (here there are only two states, A and B), moves along the tape. Each cycle begins as the
onto the tape or to halt its operation. head reads one bit from a segment of the tape. Then, in accordance with a fixed set of transi-
Not all Turing machines are revers- tion rules, it writes a bit onto that segment, changes its own internalstate and moves one seg-
ible, but a reversible Turing machine ment to the left or right. This Turing machine, because it has only two head states, can do
only trivial computations; more complicated machines with more head states are capable
can be built to perform any possible of simulating any computer, even one much more complicated than themselves. To do s o
computation. they store a representation of the larger machine's complete logical state on the unlimited
The reversible Turing-machine tape and break down each complex cycle into a large number of simple steps. The ma-
models have an advantage over such chine shown is logically reversible: it is always possible to deduce the machine's previous
machines as the frictionless billiard- configuration. Other Turing machines, with different transition rules, are not reversible
ball computer. In the billiard-ball
computer random thermal motion
causes unavoidable errors. Reversible In a state of equilibrium a backward constituting a gene. A single strand of
Turing-machine models actually ex- reaction is just as likely to occur as a D N A is much like the tape of a Turing
ploit random thermal motion: they forward one. machine. At each position along the
are constructed in such a way that ther- In order to keep a reaction moving in strand there is one of four "bases":
mal motion itself, with the assistance the forward direction, we must supply adenine, guanine, cytosine or thymine
of a very weak driving force, moves reactant molecules and remove prod- (abbreviated A, G, C and T). RNA is
the machine from one state to the uct molecules; in effect, we must pro- a similar chainlike molecule whose
next. The progress of the computa- vide a small driving force. When the four bases, adenine, guanine, cytosine
tion resembles the motion of an ion (a driving force is very small, the reaction and uracil (A, G, C and U) bind to
charged particle) suspended in a solu- will take nearly as many backward "complementary" D N A bases.
tion that is held in a weak electric field. steps as forward ones, but on the aver- The RNA polymerase catalyzes this
The ion's motion, as seen over a short age it will move forward. In order to pairing reaction. The D N A helix is
period of time, appears to be random; provide the driving force we must ex- normally surrounded by a solution
it is nearly as likely to move in one pend energy, but as in our ball-and- containing a large number of nucleo-
direction as in another. The applied pipe realization of the Fredkin gate side triphosphate molecules, each con-
force of the electric field, however, the total amount of energy can be as sisting of an RNA base linked to a
gives the net motion a preferred direc- small as we wish; if we are willing to sugar and a tail of three phosphate
tion: the ion is a little likelier to move allow a long time for an operation, groups. The RNA-polymerase enzyme
in one direction than in the other. there is no minimum amount of ener- selects from the solution a single R N A
It may a t first seem inconceivable gy that must be expended. The reason base that is complementary to the base
that a purposeful sequence of opera- is that the total energy dissipated de- about to be copied on the DNA strand.
tions, such as a computation, could be pends on the number of forward steps It then attaches the new base to the end
achieved in an apparatus whose direc- divided by the number of backward of the growing RNA strand and releas-
tion of motion at any one time is near- steps. (It is actually proportional to the es two of the phosphates into the sur-
ly random. This style of operation is logarithm of this ratio, but as the ratio rounding solution as a free pyrophos-
quite common, however, in the micro- increases or decreases so does its loga- phate ion. Then the enzyme shifts for-
scopic world of chemical reactions. rithm.) The slower the reaction moves ward one notch along the strand of
There the trial-and-error action of forward, the smaller the ratio will be. D N A in preparation for attaching the
Brownian motion, or random thermal (The apalogy of the faster and slower next RNA base. The result is a strand
motion, suffices to bring reactant mol- swimmers is valid once again: it re- of R N A that is complementary to the
ecules into contact, to orient and bend quires less total energy to go the same template strand of DNA. Without
them into the specific conformation re- net number of reaction steps ,forward R N A polymerase this set of reactions
quired for them to react, and to sepa- if the.reaction moves slowly.) would occur very slowly, and there
rate the product molecules after the would be little guarantee that the RNA
reaction. All chemical reactions are in and D N A molecules would be com-
principle reversible: the same Brown- Wemachine
can see how a Brownian Turing
might work by examin- plementary.
ian motion that accomplishes the for- ing a Brownian tape-copying machine The reactions are reversible: some-
ward reaction sometimes brings prod- that already exists in nature: R N A times the enzyme takes up a free pyro-
uct molecules together and pushes polymerase, the enzyme that helps to phosphate ion, combines it with the
them backward through the transition. construct RNA copies of the D N A last base on the RNA strand and re-
leases the resulting nucleoside triphos- the Turing machine's head. There are cule is now a type B, and it is attached
phate molecule into the surrounding several different types of "head mole- to the base that is one notch to the right
solution, meanwhile backing up one cule," each type representing a differ- of the previous head position.
notch along the D N A strand. At equi- ent machine state. During the operation of a Brownian
librium, forward and backward steps The machine's transition rules are Turing machine the tape would have
would occur with equal frequency; represented by enzymes. Each enzyme to be immersed in a solution contain-
normally other metabolic processes is capable of catalyzing one particular ing many enzyme molecules, as well
drive the reaction forward by remov- reaction. The way these enzymes work as extra O's, l's, A's and B's. T o drive
ing pyrophosphate and supplying the is best demonstrated by an example. the reaction forward there would have
four kinds of nucleoside triphosphate. Suppose the head molecule is type to be some other reaction that cleaned
In the laboratory the speed with which A (indicating that the machine is in the enzyme molecules of detached
RNA polymerase acts can be varied by state A) and is attached to a 0 base. heads and bases. The concentrations
adjusting the concentrations of the re- Also suppose the following transition of the reactants that clean the enzyme
actants (as Judith Levin and Michael rule applies: "When the head is in state molecules represent the force that
J. Chamberlin of the University of A and reads a 0, change the 0 to a 1, drives the Turing machine forward.
California at Berkeley have shown). change state to B and move to the Again we can expend as little energy
As the concentrations are brought right." A molecule of the enzyme rep- as we wish simply by driving the ma-
closer to equilibrium the enzyme resenting this rule has a site that fits a chine forward very slowly.
works more slowly and dissipates less type-A head molecule attached to a 1 The enzymatic Turing machine
energy to copy a given section of base. It also has one site that fits a 0 would not be error-free. Occasionally
DNA, because the ratio of forward to base and one site that fits a B head [see a reaction that is not catalyzed by any
backward steps is smaller. illustration on opposite page]. enzyme might occur; for example, a 0
T o achieve the transition, the en- base could spontaneously detach itself
lthough RNA polymerase merely zyme molecule first approaches the from the backbone molecule and a 1
A copies information without proc- tape molecule at a location just to the base could be attached in its place.
essing it, it is relatively easy to imagine right of the base on which the A head Similar errors do indeed occur during
how a hypothetical chemical Turing resides. Then it detaches from the tape RNA synthesis.
machine might work. The tape is a sin- both the head molecule and the 0 base In principle it would be possible to
gle long backbone molecule to which to which the head was attached, put- eliminate errors by building a Brown-
two types of base, representing the bi- ting in their place a 1 base. Next it ian Turing machine out of rigid, fric-
nary 0 and 1, attach at periodic sites. A attaches a B head to the base that is to tionless clockwork. The clockwork
small additional molecule is attached the right of the 1 base it has just added Turing machine involves less idealiza-
to the 0 or 1 group at one site along the to the tape. At this point the transition tion than the billiard-ball computer
chain. The position of this additional is complete. The head's original site is but more than the enzymatic Turing
molecule represents the position of changed from a 0 to a 1, the head mole- machine. On the one hand, its parts
need not be manufactured to perfect
tolerances, as the billiard balls would
have to be; the parts fit loosely togeth-
er, and the machine can operate even
FREE PYROPHOS- in the presence of a large amount of
thermal noise. Still, its parts must be

@GE : I
perfectly rigid and free of static fric-
tion, properties not found in any mac-
roscopic body.
Because the machine's parts fit to-
gether loosely, they are held in place
not by friction but by grooves or
notches in neighboring parts [see illus-
tration on page 561. Although each part
of the machine is free to jiggle a lit-
tle, like the pieces of a well-worn
wood puzzle, the machine as a whole
can only follow one "computational
path." That is, the machine's parts in-
terlock in such a way that at any time
the machine can make only two kinds
of large-scale motion: the motion cor-
/ responding to a forward computation-
RNA POLYMERASE al step and that corresponding to a
RNA POLYMERASE, an enzyme, acts as a reversible tape-copying machine; it catalyzes backward step.
the reaction that constructs RNA copies of segments of DNA. As the enzyme moves along The computer makes such transi-
a strand of DNA, it selects from the surrounding solution a nucleoside triphosphate mole- tions only as the accidental result of
cule (an RNA base bound to a sugar and a "tail" of three phosphate groups) that is comple- the random thermal motion of its parts
mentary to the DNA base about to be copied. It then attaches the new base to the end of the biased by the weak external force. It is
RNA strand and releases a free pyrophosphaie ion consisting of two phosphates. The reac-
nearly as likely to proceed backward
tion is reversible: sometimes the enzyme takes up the last link of RNA, attaches it to a pyro-
phosphate ion and returns the resulting molecule to the solution, backing up a notch on the
along the computational path, undoing
DNA strand. When the reaction is close to chemical equilibrium, the enzyme takes almost the most recent transition, as it is to
as many backward as forward steps and the total energy needed to copy any segment of proceed forward. A small force, pro-
DNA is very small. The reaction can be made less dissipative by being run more slowly; vided externally, drives the computa-
there is no minimum amount of energy that must be expended to copy a segment of DNA. tion forward. This force can again be
as small as we wish, and so there is no
minimum amount of energy that must
be expended in order to run a Brown-
ian clockwork Turing machine. I 4

A ccording to classical thermodynam-

ics, then, there is no minimum
amount of energy required in order to
perform a computation. Is the classi-
cal thermodynamical analysis in con-
flict with quantum theory? After all,
the quantum-mechanical uncertainty
principle states there must be an in-
verse relation between our uncertain-
ty about how long a process takes and
our uncertainty about how much ener-
gy the process involves. Some investi-
gators have suggested that any switch-
ing process occurring in a short period
of time must involve a minimum-ex-
penditure of energy.
In fact the uncertainty principle
does not require any minimum energy
expenditure for a fast switching event.
The uncertainty principle would be ap-
plicable only if we attempted to meas-
ure the precise time at which the event
took place. Even in quantum mechan-
ics extremely fast events can take
place without any loss of energy. Our
confidence that quantum mechanics
allows computing without any mini-
mum expenditure is bolstered when
we remember that Benioff and oth-
ers have developed models of reversi-
ble quantum-mechanical computers,
which dissipate no energy and obey
the laws of quantum mechanics.
Thus the uncertainty principle does
not seem to dace a fundamental limit
on the proce'ss of computation; neither
does classical thermodynamics. Does
this mean there are no physical limita-
tions to computing? Far from it. The
real limitations are associated with
questions that are much harder to an-
swer than those we have asked in this
article. For example, do elementary
logic operations require some mini-
mum amount of time? What is the
smallest possible gadgetry that could
accomplish such operations? Because
scales of size and time are connected
by the velocity of light, it is likely that
these two questions have related an-
swers. We may not be able to find these
answers, however, until it is deter-
mined whether or not there is,some
ultimate graininess in the universal HYPOTHETICAL ENZYMATIC TURING MACHINE could perform a computation
scales of time and length. with no minimum expenditure of energy. Molecules representing 0 and 1 bits are attached
At the other extreme, how large can at periodic intervals to a backbone molecule. A small additional molecule, representing the
we make a computer memory? How Turing machine's head, is attached to the 0 or 1 group at one site on the chain (I).There
many particles in the universe can we are several types of- head molecule, each type representing a different internal machine
bring and keep together for that pur- state. Transition rules are represented by enzymes. In each cycle an enzyme attaches itself
pose? The maximum possible size of a to the head molecule and the bit molecule to which the head is attached (2); then it detaches
them from the chain, putting in their place the appropriate bit molecule (3). As it does so it
computer memory limits the precision rotates, so that it attaches the appropriate head molecule to the bit that occupies the site one
with which we can calculate. It will notch to the right or left of the hit it has just changed. Now the cycle is complete (4): the
limit, for example, the number of deci- value of a bit has been changed, and the head has changed state and shifted its position.
mal places to which we can calculate Esch kind of enzyme is able to catalyze one such set of reactions. As in the case of RNA
pi. The inevitable deterioration proc- synthesis, these reactions can be made to dissipate an arbitrarily small amount of energy.
esses that occur in real computers pose are there certain calculations that can- mately be based, are themselves ex-
another, perhaps related, question: not be completed before the comput- pressed in terms of such mathematical
Can deterioration, at least in principle, er's hardware decays into uselessness? operations. Thus we are asking about
be reduced to any desired degree, or Such questions really concern limi- the ultimate form in which the laws of
does it impose a limit on the maximum tations on the physical execution of physics can be applied, given the con-
length of time we shall be able to de- mathematical operations. Physical straints imposed by the universe that
vote to any one calculation? That is, laws, on which the answers must ulti- the laws themselves describe.

MASTER CAMSHAFT

OBSTRUCTIVEREAD

BROWNIAN CLOCKWORK TURING MACHINE, made out right). In an obstructive read the reader moves past the block, fol-
of rigid, frictionless parts, relies on random jiggling of its loosely lowing a high or a low path; one of the paths will be obstructed by
fitted parts to change from state to state. When a part is held in the knob on the end of the block, and so there will be only one patb
place, it is not by friction but by grooves or notches in neighboring for the reader to follow. At the point on the master camshaft that
parts. Parts interlock in such a way that they can follow only one corresponds to this "decision" the grooves branch; each groove
"computational patb"; although they are free to jiggle a little, the splits into two, and the stylus is guided into the groove that corre-
only large-scale motions they can make correspond to forward or sponds to the bit's value (2). Then the master camshaft turns until
backward computational steps. The operation of the machine is the stylus is in the "write"segment (3). In this segment each groove
driven slowly forward by a very weak force; a t any instant the ma- contains a different set of "instructions" for the machine t o follow;
chine is almost as likely to move backward as forward. On the aver- the instructions are transmitted by a complex linkage between the
age, however, the machine will move forward and the computation stylus and the rest of the mechanim. If the instructions call for the
will eventually end. The machine can be made to dissipate as small bit's value to change, the manipulator moves over and grasps the
a n amount of energy as the user wishes, simply by employing a knob; then the screwdriver rotates the disk until the block is free to
force of the correct weakness. Segments of tape are represented by move, the manipulator moves the block up or down and the screw-
grooved disks; bits are represented by E-shaped blocks, which are driver rotates the disk again to hold the block in place. After the
locked onto the disks in either the up (1)or the down (0) position. stylus passes through the L'write" segment of the master camshaft it
The head consists of a rigid framework and a complicated mecha- enters the "shift"segment (4). Each groove in this segment contains
nism (most of which is not shown) from which are suspended a read- instructions to move the head one segment t o the left or right. Then
er, a manipulator and a screwdriver-shaped rod. The machine's the stylus enters the "change state" segment of the camshaft (9,
operation is controlled by a grooved "master camshaft," which re- where grooves merge in such a way that the stylus falls into the
sembles a phonograph record (top left and far right); different groove representing the next head state. The cycle is now complete
grooves correspond to different head states. At the beginning of a (6). Disks adjacent to the one being read are held in place by the
cycle the head is positioned above one of the disks and a "stylus" is head's framework. Disks that are farther away are held by "locking
in the "read"segment of the groove in the master camshaft that cor- tabs." The locking tab on each disk is coupled to a special bit, called
responds to the machine's current head state. During the "read" the Q bit, on a n adjacent disk. The linkages between Q bits and
part of the cycle ( I ) the reader determines whether the disk's "bit* locking tabs are constructed so that the disk currently being read
block is up or down by a process called an obstructive read (center is free to move, while disks far to the right or left are held still.