I I

C. H. B e n n e t t

Logical Reversibility of Computation*

Abstract: T h e usual general-purpose computing automaton (e.g., a Turing machine) is logically irreversible —its transition function
lacks a single-valued inverse. H e r e it is shown that such machines may be made logically reversible at every step, while retaining their
simplicity and their ability to do general computations. This result is of great physical interest because it makes plausible the existence
of thermodynamically reversible computers which could perform useful computations at useful speed while dissipating considerably
less than kT of energy per logical step. In the first stage of its computation the logically reversible automaton parallels the corre-
sponding irreversible automaton, except that it saves all intermediate results, thereby avoiding thè irreversible operation of erasure.
The second stage consists of printing out the desired output. T h e third stage then reversibly disposes of all the undesired intermediate
results by retracing the steps of the first stage in backward order ( a process which is only possible because the first stage has been car-
ried out reversibly), thereby restoring the machine (except for the now-written output tape) to its original condition. T h e final machine
configuration thus contains the desired output and a reconstructed copy of the input, but no other undesired data. T h e foregoing results
are demonstrated explicitly using a type of three-tape Turing machine. T h e biosynthesis of messenger R N A is discussed as a physical
example of reversible computation.

Introduction
T h e usual digital computer program frequently performs
operations that seem to throw a w a y information about
the computer's history, leaving the machine in a state
whose immediate predecessor is ambiguous. Such opera-
tions include erasure o r overwriting o f data, and entry
into a portion o f the program addressed by several dif-
ferent transfer instructions. In other w o r d s , the typical
computer is logically irreversible —its transition function
(the partial function that maps each whole-machine state
onto its successor, if the state has a successor) lacks a
single-valued inverse.
Landauer [ 1 ] has posed the question o f whether logi-
cal irreversibility is an unavoidable feature o f useful
computers, arguing that it is, and has demonstrated the
physical and philosophical importance o f this question
by showing that w h e n e v e r a physical computer throws
away information about its previous state it must gener-
ate a corresponding amount o f entropy. T h e r e f o r e , a
computer must dissipate at least A T l n 2 o f energy (about
3 x 10~ 21
joule at r o o m temperature) for each bit o f in-
formation it erases o r otherwise throws a w a y .
sufficient detail that the preceding state would be
uniquely determined by the present state and the last
record on the tape. H o w e v e r , as Landauer pointed out,
this would merely postpone the problem o f throwing
away unwanted information, since the tape would have
to be erased before it could be reused. It is therefore
reasonable to demand o f a useful reversible computer
that, if it halts, it should have erased all its intermediate
results, leaving behind only the desired output and the
originally furnished input. ( T h e machine must be al-
l o w e d to save its input —otherwise it could not be rever-
sible and still carry out computations in which the input
was not uniquely determined by the output.) W e will
show that general-purpose reversible computers (Turing
machines) satisfying these requirements indeed exist,
and that they need not be much more complicated than
the irreversible computers on which they are patterned.
Computations on a reversible computer take about t w i c e
as many steps as on an ordinary one and may require a
large amount o* temporary storage. Before proceeding
with the formal demonstration, the argument will be car-

A n irreversible computer can always be made reversi- ried through at the present heuristic level.

ble by having it save all the information it would other-

wise throw away. F o r e x a m p l e , the machine might be
given an extra tape (initially blank) on which it could *Much of Ihe work on physical reversibility reported in this paper was done under
the auspices of the U . S . Atomic F.nergy Commission while the author was employed
record each operation as it was being performed, in by the Argonne National Laboratory, Argonne, Illinois. 525

NOVF.MDKR 1973 LOGICAL Rl.VF.RSIBll.ITY

 We begin with the reversible but untidy computer
mentioned earlier, which has produced, and failed to
erase, a long history o f its activity. N o w , a tape full o f
random data cannot be erased e x c e p t by an irreversible
process: h o w e v e r , the history tape is not random —there
exists a subtle mutual redundancy between it and the
machine that produced it, which may be exploited to
erase it reversibly. F o r example, if at the end o f the
computation a new stage o f computation w e r e begun
using the inverse o f the original transition function, the
machine would begin carrying out the entire computa­
tion backward, eventually returning the history tape to
its original blank c o n d i t i o n [ 2 ] . Since the forward c o m ­
putation was deterministic and reversible, the backward
stage would be also. Unfortunately, the backward stage
would transform the output back into the original input,
rendering the overall computation completely useless.
Destruction o f the desired output can be prevented sim­
ply by making an extra c o p y o f it on a separate tape, af­
ter the forward stage, but before the backward stage.
During this copying operation (which can be done re­
versibly if the tape used for the c o p y is initially b l a n k ) ,
the recording o f the history tape is suspended. T h e back­
ward stage will then destroy only the original and not the
c o p y . A t the end o f the computation, the computer will
contain the (reconstructed) original input plus the intact
c o p y o f the output; all other storage will have been re­
stored to its original blank condition. E v e n though no his­
tory remains, the computation is reversible and deter­
ministic, because each o f its stages has been so.
One disadvantage of the reversible machine would
appear to be the large amount o f temporary storage
needed for the history —for a v-step first stage, v rec­
ords o f history would have to be written. In a later sec­
tion it will be argued that by performing a j o b in many
stages rather than just three, the required amount o f
temporary storage can often be greatly reduced. T h e final
section discusses the possibility o f reversible physical
computers, capable o f dissipating less than kT o f energy
per step, using examples from the biochemical apparatus
o f the genetic c o d e .
Logically reversible Turing m a c h i n e s
T h i s section formalizes the argument o f the preceding
section by showing that, given an ordinary Turing
machine S, o n e can construct a reversible three-tape
Turing machine R , which emulates S on any standard
input, and which leaves behind, at the end o f its compu­
tation, only that innut and the desired output. T h e R
machine's computation proceeds by three stages as de­
scribed a b o v e , the third stage serving to dispose o f the
history produced by the first. T h e remainder o f this sec­
tion may be skipped by those uninterested in the details
526 o f the proof.
C. H. BENNETT
The ordinary type o f one-tape Turing machine [ 3 ]
consists o f a control unit, a read/write head, and an infi­
nite tape divided into squares. Its behavior is g o v e r n e d
by a finite set o f transition formulas ( c o m m o n l y called
quintuples) o f the read-write-shift type. T h e quintuples
have the form
AT -* T a A \ (1)
meaning that if the control unit is in state A and the head
scans the tape symbol T, the head will first write 7" in
place o f T; then it will shift left one square, right one
square, o r remain w h e r e it is, according to the value o f
<r(—, + , o r 0, r e s p e c t i v e l y ) ; finally the control unit will
revert to state A'. In the usual generalization to n-tape
machines, T, T', and o- are all w-tuples within the quin­
tuple.
Each quintuple defines a (partial) one-to-one mapping
o f the present whole-machine state (i.e., tape contents,
head positions, and control state) onto its successor and,
as such, is deterministic and reversible. T h e r e f o r e a T u r ­
ing machine will be deterministic if and only if its quintu­
ples have non-overlapping domains, and will be reversi­
ble if and only if they have non-overlapping ranges. T h e
former is customarily guaranteed by requiring that the
portion to the left o f the arrow be different for each quin­
tuple. O n the other hand, the usual Turing machine is
not reversible.
In making a Turing machine reversible, w e will need
to add transitions that closely resemble the inverses o f
the transitions it already has. H o w e v e r , because the
write and shift operations d o not commute, the inverse
o f a read-write-shift quintuple, though it exists, is o f a
different t y p e ; namely, shift-read-write. In constructing a
reversible machine it is necessary to include quintuples
o f both types, or else to use a formalism in which transi­
tions and their inverses have the same form. H e r e the
latter approach is taken —the reversible machine will use
a simpler t y p e o f transition formula in which, during a
given transition, each tape is subjected to a read-write o r
to a shift operation but no tape is subjected to both.
Definition: A quadruple ( f o r an «-tape Turing machine
having one head per t a p e ) is an expression o f the form
2
<*tV V " 0 - [ ' , ' . / , ' . • " C K . ( )
w h e r e A and A' are positive integers (denoting internal
states o f the control unit before and after the transition,
r e s p e c t i v e l y ) ; each t k may be either a positive integer
denoting a symbol that must be read on the Ath tape o r a
solidus ( / ) , indicating that the Ath tape is not read during
the transition; each t ' is either a positive integer denoting
k
the symbol to be written on the Ath tape or a member o f
the set ( - , 0, + ) denoting a left, null, or right shift o f the
Ath tape head. F o r each tape A, t 'E.(—, 0, + ) if and only if
k
IBM J . RES. DEVELOP.
I
 !

t =l.
k Thus the machine writes on a tape if and only if it Definition: A n input or output is said to be standard
has just read it, and shifts a tape only if it has not just when it is on otherwise blank tape and contains no
read it. embedded blanks, when the tape head scans the blank
square immediately to the left o f it, and when it includes
L i k e quintuples, quadruples define mappings o f the
only letters belonging to the tape alphabet o f the ma­
whole-machine state which are one-to-one. A n y read-
chine scanning it.
write-shift quintuple can be split into a read-write and a
shift, both expressible as quadruples. F o r e x a m p l e , the Definition: A standard Turing machine is a finite set o f
quintuple ( 1 ) is equivalent to the pair o f quadruples one-tape quintuples

AT -+ V A" (3) AT —> T a A' (1)

A"[/l---l] - > a A', (4) satisfying the following requirements:

1 ) Determinism: N o t w o quintuples agree in both A

where A" is a new control-unit state different from A and
and T.
A'. When several quintuples are so split, a different con­
2) Format: I f started in control state/1, on any standard
necting state A" must be used for each, to avoid intro­
input, the machine, if it halts, will halt in control state
ducing indeterminacy.
A j ( / b e i n g the number o f control states), leaving its
Quadruples have the folldwing additional important
output in standard format.
properties, which can be verified by inspection. L e t
3) Special quintuples: T h e machine includes the follow­
ing quintuples
«^A[t,,--;t ] n - > [*,',••-.f/K (5)

and A, b - » b +A., (7)

A _,
f b -» b 0A ,f (8)
/8 = « [ « , , • • • . « „ ] - [ « , ' , • • • , u ']B'
n (6)

and control states A, and A appear in no other quintuple.

f

be t w o n-tape quadruples.
T h e s e t w o are thus the first and last executed respectively

1) a and B are mutually inverse (define inverse map­ in any terminating computation on a standard input. T h e

pings o f the whole-machine state) if and only if A = letter b represents a blank.

B' and B =A' and, for e v e r y k, either U = u = / k k

T h e phrase "machine M , given standard input string / ,

and t k — — u ') or (t # / and t
k k k = u and
k = u ).k

computes standard output string P" will be abbreviated

The inverse o f a quadruple, in other w o r d s , is o b ­
M: / - » P. F o r an «-tape machine this will become
tained by interchanging the initial control state with
M: (/,; • • • ; / „ ) - » ( / ' , ; P.,: • • • ; / > „ ) , where l k and P k
the final, the read tape symbols with the written, and
are the standard input and the standard output on the
changing the signs o f all the shifts.
A:th tape. A blank tape will be abbreviated B.
2 ) T h e domains o f a and B overlap if and only if A = B
T h e main theorem can now be stated:
and, for e v e r y k, (t = I or u = / or t = u ). N o n -
k k k k

overlapping o f the domains requires a differing initial
control state o r a differing scanned symbol on some
tape read by both quadruples.
3) T h e ranges o f a and B overlap if and only if A' = B'
and, for e v e r y k, (t — / or u = / o r t
k k
property is analogous to the previous o n e , but d e ­
pends on the final control state and the written tape
symbols.
A reversible, deterministic n-tape
may now be defined as a finite set o f n-tape quadruples,
no t w o o f which overlap either in domain or range. W e
now wish to show that such machines can be made to
emulate ordinary (irreversible) Turing machines. It is
convenient to impose on the machines to be emulated
certain format-standardization requirements, which, how­
e v e r , d o not significantly limit their computing p o w e r [ 4 ] .
Theorem: F o r every standard one-tape Turing machine
S, there exists a three-tape reversible, deterministic Tur­
ing machine R such that if / and P are strings on the al­
phabet o f S, containing no embedded blanks, then S halts
k = u ). T h e
k on / if and only if R halts on (l;B;B), and S: / - » P if
and only i f R : (I;B;B) -* (l;B;P).
Furthermore, if S has / control states, N quintuples
and a tape alphabet o f z letters, including the blank,
R will have If + IN + 4 states. 4N + 2c + 3 quadruples
Turing machine and tape alphabets o f z, N + 1, and z letters, respective­
ly. Finally, if in a particular computation S requires v
steps and uses ,v squares o f tape, producing an output o f
length X, then R will require 4i> + 4 \ + 5 steps, and use
s, v + I , and X + 2 squares on its three tapes, respec­
tively. ( I t will later be argued that where r » . v , the
total space requirement can be reduced to less than
l\fvs.) 527

NOVKMHKR 1973 I.OCilCAL RHVr.KSIBIl.ITY

I I

Proof: T o construct the machine R w e begin by arrang­ Each quintuple is n o w broken into a pair o f quadruples
ing the N quintuples o f S in some order with the stan­ as described earlier. T h e mth quintuple b e c o m e s
dard quintuples first and last:
AjT-*rA '
m

1) /1 , / ? - > • +/l 2 A ' l ^ < r A . k (10)

m) AT —» T a A,. The newly added states A 'm are different from the old
states and from each other: each A' appears in only one
pair o f quadruples.
N) A _ b->bOA .
f t f (9)

Table 1 Structure and operation of a three-tape reversible Turing machine. T h e computation proceeds in three stages using different
sets of quadruples and control states, linkage occurring through states A &nd C . O n the right the contents of the tapes are shown sym­
f f

bolically at the beginning and end of each stage. T h e underbar denotes the position of the head. The initial state is/f, and, for a termi­
nating computation, C is the final state.

Contents of tape

Working History Output

Stage Quadruples tape tape tape

INPUT
A,[b I l>] - » [b + b]A,'
- V [ / M ] - [ + 1 0]A 2

1
[T'+b]A„
Compute'
[Ibi] [a-m 0]A k

Ah lb] [b + b]A ' N

N)
ft A„'Ubl] [0 N Q]A f

OUTPUT HISTORY
A [b N 6 ] - > [b N b]B '
r t

«,'[///]-* [+0+]B,
x * b: { ß,[jc N b]-*[x N x]B,' }
C o p y output" B [b N b]-+[b
t N b]B ' 2

x*b:{ B [x N x] -> [x N x]B '

2 2 }
B [b N b] —> [b N b]C
2 f

OUTPUT HISTORY OUTPUT

f C///V/] [0 b 0 ] C y
N)
\c;[bib] [b - b]C,_ t

f C„[lm
C„[l ml]->[-<rbO]C ' m
Retrace m)
Ic.-rr/ b] -* [T-h]C J

C.,11 I / ] - [ - f c O ] C , '
C '[b t I h] -> [b-h]C,
INPUT OUTPUT

"The labels I ) . . . m) . . N) are not part of the machine. They indicate correspondence to the quintuples of the original irreversible machine, which the reversible machine
emulates.
528 ''In the second stage the small braces indicate sets of quadruples, with one quadruple for each nonblank tape letter*.

C. H. M i N N E T T IHM J . RES. DEVF.LOP.


T w o extra tapes are then added, o n e for the history
and one for the duplicate c o p y o f the output. T h e ouput
(third) tape is left blank and null-shifted for the present,
but the history ( s e c o n d ) tape is used to record the index
» 1 as each transition pair is executed.
T h e mth pair o f quadruples n o w has the form
U[77 6] - [T'+b]A '
m
,
[A [lbl]-*[<rmO}A .
m k
N o t i c e that the history ( s e c o n d ) tape is out o f phase
with the other t w o — it is written on while they are being
shifted and vice versa. T h i s phasing is necessary to as-
sure reversibility — it serves to capture the information
that would otherwise be thrown away when the specific
control state A ' passes to the more general state A . T h e
m k
+ shifting o f the history tape assures that a blank square
will always be ready to r e c e i v e the next m value. I f the
computation o f S does not halt, neither will that o f R ,
and the machine will continue printing on the history
tape indefinitely. O n the other hand, if ( o n a standard
input) S halts, R will eventually execute the Nth pair o f
quadruples, finding itself in state A , f with the output in
standard format on tape 1. T h e history head will be scan-
ning the number N which it has just written at the ex-
treme right end o f the history on tape 2. Control then
passes to the second stage o f computation, which copies
the output onto tape 3 ( s e e T a b l e 1 ) . T h e control states
for this stage are denoted by B ' s and are distinct from all
the /4-type control states. N o t i c e that the copying pro-
cess can be done reversibly without writing anything
more on the history tape. T h i s shows that the generation
(or erasure) o f a duplicate c o p y o f data requires no
throwing a w a y o f information.
T h e third stage undoes the work o f the first and con-
sists o f the inverses o f all first-stage transitions with C ' s
substituted f o r k ' s . In the final state C , , the history tape
is again blank and the other tapes contain the recon-
structed input and the desired output.
A s T a b l e 1 shows, the total number o f control states
is IN + 2f + 4, the number o f quadruples 4N + 2z + 3,
and the space and time requirements are as stated at the
beginning o f the proof. T h e non-overlapping o f the d o -
mains and ranges o f all the quadruples assures determin-
ism and reversibility o f the machine R . In the first stage,
the upper transitions o f each pair d o not overlap in their
domains because o f the postulated determinacy o f the
original Turing machine S, whose quintuples also began
AjT -*. T h e ranges o f the upper quadruples (as well as
the domains o f the l o w e r ) are kept from overlapping by
the uniqueness o f the states A '. m Finally, the ranges o f
the lower quadruples are saved from overlapping by the
unique output m on the history tape. T h e state A f causes
no trouble, even though it occurs in both stage 1 and
NOVEMBER 1973
I
stage 2, because by the definition o f the machine S it
does not occur on the left in stage 1 ; similarly for state
C.
f T h e non-overlapping o f the stage 2 quadruples can
be verified by inspection, while the determinism and
reversibility o f stage 3 follow from those o f stage 1.
Discussion
T h e argument developed a b o v e is not limited to three-
tape Turing machines, but can be applied to any sort o f
(11)
deterministic automaton, finite or infinite, p r o v i d e d it
has sufficient temporary storage to record the history.
One-tape reversible machines exist, but their frequent
shifting between the working and history regions o n the
1
tape necessitates as many as v steps to emulate a v-
step irreversible computation.
In the case that S is a universal Turing machine, R
becomes a machine for executing any computer program
reversibly. F o r such a general-purpose machine it seems
highly unlikely that w e can avoid having to include the
input as part o f the final output. H o w e v e r , there are
many calculations in which the output uniquely deter-
mines the input, and for such a problem one might hope
to build a specific reversible computer that would simply
map inputs onto outputs, erasing everything else. T h i s is
indeed possible, provided w e have access to an ordinary
Turing machine which, given an output, computes the
corresponding input. L e t S, be the (irreversible) Turing
machine that computes the output from the input and S 2
be the o n e that computes the input from the output. T h e
reversible computation proceeds by seven stages as
shown in T a b l e 2, o f which the first three e m p l o y a re-
versible form o f the S, computer and, as in T a b l e I . serve
to map the input onto the input and output. Stage four
interchanges input and output. Stages five and seven use
a reversible realization o f the S computer; stage five has
2
the sole purpose o f producing a history o f the S, compu-
tation ( i . e . , o f the input from the output) which, after the
extra c o p y o f the input has been erased in stage six, is
used in stage seven to destroy itself and the remaining
c o p y o f the input, while producing only the desired output
W e shall now return to the more usual situation, in
which the input must be saved because it is not a known,
computable function o f the output. Performing a compu-
tation reversibly entails only a modest increase in c o m -
puting time and machine complexity; thè main drawback
of reversible computers appears thus to be the large
amount o f temporary storage they require for the history
in any long, compute-bound j o b (i.e., o n e whose number
of steps, v, greatly exceeds the number o f squares o f
memory used, s). Fortunately, the temporary storage
requirement can be cut d o w n by breaking the j o b into a
sequence o f n segments, each one o f which would be
performed and retraced (and the history tape thereby
erased and made ready for reuse) before proceeding to
LOGICAL REVERSIBILITY
 Table 2 Reversible computer for a specific problem in which the input is a known, computable function of the output.

Slaye Action Tape I Tape 2 Tape 3

INPUT

Forward S computation

OUTPUT HISTORY 1

2. Copy output

OUTPUT HISTORY 1 OUTPUT

Retraced S, computation

INPUT OUTPUT

Interchange output with input

OUTPUT INPUT

Forward S computation
2

INPUT HISTORY 2 INPUT

6. Reversible erasure of extra copy of input

INPUT HISTORY 2
Retraced S, computation

OUTPUT

the next. Each segment would leave on the working tape
(tape 1) a restart dump that would be used as the input
of the next segment; but to preserve reversibility it
would also have to leave ( o n tape 3, s a y ) a c o p y o f its
own input, which would in most cases simply be the
preceding restart dump. A t the end o f the computation
w e would have, in addition to the original input and de­
sired output, all the n — 1 intermediate dumps (concate­
nated, e.g., on tape 3 ) . T h e s e intermediate results, which
would not have been produced had the j o b not been
segmented, either can be accepted as permanent (but
unwanted) output, in exchange for the n-fold reduction
of the history tape, or can themselves be reversibly
erased by first making an extra c o p y o f the desired final
output (putting it, say, on a previously unused part o f
tape 3 ) , then reversing the whole n-segment computa­
tion. T h i s reversal is possible because each segment has
been performed reversibly. T h e sequence o f restart
dumps thus functions as a kind o f higher-level history,
and it is erased by a higher-level application o f the same
technique used to erase the primary histories. A t the end
of the computation, the machine will contain only the
original input and the desired nth segment output, and
e v e r y step o f the original irreversible computation will
have been performed twice forward and twice backward.
F o r a j o b with v steps and a restart dump o f size .v, the
total temporary storage requirement (minimized by
530 choosing n = V~pjs) is 2\^vs squares, half on the history
tape and half on the dump tape. A (iv»V7)-fold reduc­
tion in space can thus be bought by a twofold increase in
time (ignoring the time required to write and read restart
d u m p s ) without any unwanted output. By a systematic
reversal o f progressively larger nested sequences o f
segments one might hope to reach an absolute minimum
temporary storage requirement growing only as log v, for
sufficiently large v, with the time increasing perhaps as
v~, because o f the linearly increasing number o f times
each segment would have to be retraced.
•It thus appears that every j o b o f computation can be
done in a logically reversible manner, without inordinate
increases in machine c o m p l e x i t y , number o f steps, un­
wanted output, or temporary storage capacity.
Physical reversibility
T h e existence o f logically reversible automata suggests
that physical computers might be made thermodynami-
cally reversible, and hence capable o f dissipating an ar­
bitrarily small amount o f energy per step if operated
sufficiently slowly. A full treatment o f physically reversi­
ble computers is beyond the scope o f the present paper
[ 5 ] , but it is worthwhile to give a brief and non-rigorous
introduction to how they might w o r k .
A n obvious approach to the minimizing the energy
dissipation is to design the computer so that it can oper­
ate near thermodynamic equilibrium. A l l moving parts
would then, at any instant, have near-thermal velocity.

п. HI.NNK I Г IBM J . R I . S . DF.VILLOP.


and the desired logical transitions would necessarily be
accomplished by spontaneous thermally activated mo­
tion o v e r free energy barriers not much higher than kT.
A t first sight this might seem impossible —in existing
electronic computers, for e x a m p l e , e v e n when a c o m p o ­
nent being switched is itself nondissipative (e.g., a mag­
netic c o r e ) , the switching process depends on temporari­
ly applying a strong external force to push the c o m p o ­
nent irreversibly o v e r a high free energy barrier. H o w ­
ever, nature provides a beautiful example o f a thermally
activated "computer" in the biochemical apparatus re­
sponsible for the replication, transcription and translation
of the genetic c o d e [ 6 ] . Each o f these processes in­
v o l v e s a long, deterministic sequence o f manipulations
of coded information, quite analogous to a computation,
and yet, so far as is known, each is simply a sequence o f
coupled, thermally activated chemical reactions. In bio­
chemical systems, e n z y m e s play the essential role o f
selectively lowering the activation barriers for the de­
sired transitions while leaving high barriers to obstruct
all undesired transitions —those which in a computer
would correspond to errors. A l t h o u g h the environment
in which e n z y m e s normally function is not at chemical
equilibrium, many enzyme-catalyzed reactions are freely
reversible, and one can find a set o f equilibrium reactant
concentrations at which both forward and reverse reac­
tions occur equally rapidly, while competing uncatalyzed
reactions have negligible rates. It is thus not unreasona­
ble to postulate a thermally activated computer
which, at equilibrium, e v e r y logically allowed transition
occurs equally often forward and backward, while illogi­
cal transitions hardly e v e r occur. In the following dis­
cussion chemical terminology will be used, without
implying that thermally activated computers musj be
1
chemical systems.
T h e chemical realization o f a logically reversible c o m ­
putation is a chain o f reactions, each coupled only to the
preceding one and the following one. It is helpful to
think o f the computing system as comprising a major
reactant (analogous to D N A ) that encodes the logical
state, and minor reactants that react with the major one
to change the logical state. O n l y one molecule o f the
major reactant is present, but the minor reactants are all
present at definite concentrations, which may be manip­
ulated to drive the computation forward or backward.
If the minor reactants are in equilibrium, and the major
reactant initially corresponds to the initial state o f a v-
step computation, the system will begin a random walk
through the chain o f reactions, and after about v steps
will briefly visit the final state. This does not d e s e r v e to
be called a computation; it would be legitimate to insist
that the system proceed through the chain o f reactions
with some positive drift velocity and, after sufficient time,
have a high probability o f being in the final state ( i f
NOVEMBER 1973
I
the computation has o n e ) . T h e former requirement can
be met by adjusting the chemical potentials o f the minor
reactants so that each forward step dissipates a little
energy e; the latter can be met by dissipating a trivial
extra amount during the last step. ( I f all steps had equal
dissipation, e < A T , the final state occupation probabili­
ty would be only about dkT. By dissipating an extra
kT In (3 kT/c) o f energy during the last step, this proba­
bility is increased to about 9 5 % . ) G i v e n a uniform rate T
for all forward reactions, an energy dissipation e < kT
per step will buy a drift velocity (i.e., computation
s p e e d ) o f Fz/kT steps per second. O n the other hand, for
£ > kT, backward steps will be effectively suppressed
and the computation speed will approach the forward
reaction rate F. T h e chemical system is thus a thermo-
dynamically reversible computer o f the type w e have
been seeking.
I f w e attempt to apply the preceding argument to a
logically irreversible computer, w e can see that here the
reactions form a branching structure, with a main trunk
corresponding to the desired computation path, and side
branches corresponding to incorrect or "extraneous"
reverse computations. T h e states on the side branches
are valid predecessors o f the final state but not valid
successors o f the initial state. A f e w such extraneous
states would pose no problem —a small driving force
would still suffice to push the system into the desired
final state. T e m p o r a r y backward excursions onto the
in
side branches would occur, but would not lead to errors,
contrary to what one might expect. Since no state o f a
deterministic computer can have more than one logical
successor, the erroneously reversed operations would be
corrected as soon as the computation proceeded forward
again, and the desired path would be rejoined. T h e real
problem comes from the enormous number o f extra­
neous predecessors; typically they outnumber the states
on the intended computation path by hundreds o f orders
of magnitude. This is because, in irreversibly pro­
grammed computations, one can usually proceed back­
ward along an extraneous path for many steps, making
further w r o n g choices along the w a y , before arriving at a
state that has no predecessors.
I f a thermally activated computer with many extra­
neous states is operated close to equilibrium, the system
will spend only a minuscule fraction o f its time on the
desired path o f computation, let alone in the desired final
state. A n acceptable computation rate requires 1) that
finite (but time-consuming) backward excursions be
largely suppressed, and 2 ) that infinite ones be c o m ­
pletely suppressed. This in turn means (roughly
speaking) that the dissipation per step must e x c e e d kT
In m, w h e r e m is the mean number o f immediate prede­
cessors I ) averaged o v e r states near the intended path,
or 2 ) averaged o v e r all accessible states, whichever is
LOGICAL REVERSIBILITY
I !

greater. For a typical irreversible computer,
throws a w a y about one bit per logical operation, m is
approximately t w o , and thus A T In 2 is, as Landauer has
argued [ 1 ] , an approximate l o w e r bound on the energy
dissipation o f such machines. F o r a logically reversible
computer, h o w e v e r , m is exactly one by construction.
The biosynthesis and biodégradation o f messenger
R N A may be v i e w e d as convenient examples o f logical­
ly reversible and irreversible computation, respectively.
M e s s e n g e r R N A , a linear p o l y m e r i c informational mac-
romolecule like D N A , carries the genetic information
from one o r m o r e genes o f a D N A molecule, and serves
to direct the synthesis o f the proteins encoded by those
genes. M e s s e n g e r R N A is synthesized by the e n z y m e
R N A polymerase in the presence o f a double-stranded
D N A molecule and a supply o f R N A monomers (the
four nucleotide pyrophosphates A T P , G T P , C T P , and
UTP) [ 7 ] . T h e e n z y m e attaches to a specific site on the
D N A molecule and m o v e s along, sequentially incorpo­
rating the R N A monomers into a single-stranded R N A
molecule w h o s e nucleotide sequence exactly matches
that o f the D N A . T h e pyrophosphate groups are re­
leased into the surrounding solution as free pyrophos­
phate molecules. T h e e n z y m e may thus be compared to
a simple tape-copying Turing machine that manufactures
its output tape rather than merely writing on it. T a p e
c o p y i n g is a logically reversible operation, and R N A
polymerase is both thermodynamically and logically
reversible. In the cellular environment the reaction is
driven in the intended forward direction o f R N A synthe­
sis by other reactions, which maintain a l o w concentra­
tion o f free pyrophosphate, relative to the concentrations
o f nucleotide pyrophosphates [ 8 ] . A high pyrophosphate
concentration would drive the reaction backward,
and the enzyme would carry out a sequence-spe­
cific degradation o f the R N A , comparing each nucleo­
tide with the corresponding D N A nucleotide before
splitting it off. T h i s process, which may be termed logi­
cally reversible erasure o f R N A , d o e s not normally o c ­
cur in biological systems —instead, R N A is degraded by
other e n z y m e s , such as polynucleotide phosphorylase
[9], in a logically irreversible manner ( i . e . , without
checking its sequence against D N A ) . Polynucleotide
phosphorylase catalyzes the reaction o f R N A with free
phosphate (maintained at high concentration) to form
nucleotide phosphate monomers. L i k e the polymerase
reaction, this reaction is thermodynamically reversible;
h o w e v e r , because o f its logical irreversibility, a fourfold
greater phosphate concentration is needed to drive it
forward than would be required for a logically reversible
which phosphorolytic degradation. T h e extra driving force is
necessary to suppress the undesired synthesis o f non­
sense R N A by random polymerization.
In biological systems, apparently, the speed and flexi­
bility o f irreversible erasure outweigh its extra cost in
free energy ( A T In 4 per nucleotide in this c a s e ) . Indeed,
throughout the genetic apparatus, energy is dissipated at
a rate o f roughly 5 to 50 A T per step; while this is ten
orders o f magnitude l o w e r than in an electronic c o m ­
puter, it is considerably higher than what would theoreti­
cally be possible if biochemical systems did not need to
run at speeds close to the kinetic maximum —presum­
ably to escape the harmful effects o f radiation, uncata-
l y z e d reactions, and competition from other organisms.
Acknowledgment
I thank R o l f Landauer for raising the question o f re­
versibility o f computation in the first place and for stim­
ulating discussions o f m y models.
References and notes
1. R. Landauer, IBM J. Res. Develop. 3, 183 ( 1 9 6 1 ) . R. W .
Keyes and R. Landauer, IBM J. Res. Develop. 14, 152
( 1 9 7 0 ) , investigate a specific model computer whose energy
dissipation per step is about AT.
2. R. W . Keyes [Science 168, 796 ( 1 9 7 0 ) ] , in summarizing
Landauer's argument [ 1 ] , commented that a saved history
might be used to reverse the steps of the original computa­
tion, but that this was not practical in a general purpose com­
puter. H e did not explicitly point out that a reversible ma­
chine can be made to erase its o w n history (an ability which,
w e have argued, allows it to be useful as a general purpose
computer).
3. For a good informal exposition of Turing machines see Chap­
ter 6 of M . L . Minsky, Compulation: Finite and Infinite Ma­
chines, Prentice-Hall, Inc., Englewood Cliffs, N . J., 1967.
4. By the addition of a f e w extra tape symbols and quintuples,
an arbitrary Turing machine can be made to obey these for­
mat requirements while computing essentially the same func­
tion as it did before. See M . Davis, Computability and Un-
solvabilitv, M c G r a w - H i l l Book C o . , Inc., N e w Y o r k , 1958.
. pp. 2 5 - 2 6 .
5. T h e author is currently preparing a paper on physically re­
versible model computers.
6. F o r a good introduction to this subject see James D . Watson
Molecular Biology of the Gene (2nd e d . ) , W . A . Benjamin,
Inc., N e w Y o r k , 1970.
7. Ibid., p. 336 ff.
8. Ibid., p. 155 ff.
9. Ibid., p. 403.
Received April 12, 1973
C. H. Bennett is located at the IBM Thomas J. Watson
Research Center, Yorktown Heights, New York 10598.

532

C. II. BENNETT II1M J . R E S . D E V E L O P .