2 Information and Computation are Physical

An operation is “logically reversible" if it can be run backwards, that is, if its inputs can always be deduced from the outputs. Most logical gates are irreversible; a typical example is the NAND gate


(1)
which has two input bits and only one output bit. We cannot recover a unique input from the output bit because the result 1 can be obtained from three distinct inputs: .

Assume we operate the gate NAND with two Boolean variables, , and suppose that the four initial states, , have the same probability distribution, . Then, the initial entropy, which is calculated with Shannon’s formula


is then

The result will be a system with only two possible states, and , the outcome appearing with probability and the outcome appearing with probability . Consequently, the final entropy is

which means a loss of

Assume now that we operate the gate

and, again, suppose that the four initial states of the Boolean variables have the same probability distribution, . This gate has finally only
three final states, namely , two of them with probability and one with probability . Consequently, the final entropy is

In this case, the gate decreases the entropy by bits.

The first gate is “more irreversible" than the second one, since it decreases the entropy more.

In thermodynamics the entropy is defined by

Von Neumann noticed that the two entropies are related by some constant factor, so they are in fact the same notion. Hence a natural question arises:
what is the minimum energy required to carry out a computation? In 1961 Landauer [81] (see also [84] and also Feynman [61]) had a first answer. To operate a computer we have to ensure that distinct logical states are represented by distinct physical states. Each bit has two values, 0, 1, so it has one degree of freedom; it corresponds to one or more degrees of freedom of physical states. In general, a set of bits has degrees of freedom; they correspond to physical states. If we erase bits (which could be in any of the possible logical states), say we reset all to 0, then we have compressed logical states into a single state, a loss of entropy. The irreversible loss information increases the temperature of the system, which means heat dissipation. Consequently, operations which are not one-to-one cost energy. This cost is expressed by Landauer’s principle: erasure of information is a dissipative process.

Here is a simple “home" example. We need two basketballs to design a system of representing information. Put one on the floor by your left foot and hold the other in your (right) hand. Zero (0) is represented by the ball on the floor; one (1) is represented by the ball in your hand. Assume that we want to erase the bit 1, that is the bit in your hand. To do this you have to drop the ball. Simple? Not really, as the ball does not get directly into the floor (to become a 0), but in fact bounces for a while. With a perfectly elastic basketball and a good hard floor the ball may bounce close to your hand, i.e., to the 1 position! To settle down into 0 the ball has to encounter friction, with the air molecules and the floor. Eventually friction slows down the ball, so 1 has been erased. We could do it because the energy from bouncing the ball has been transmitted to the floor and the air. In a vacuum with a perfect frictionless floor erase would be impossible! Energy is consumed in the process of erasure.

Here is another example: the gate dissipates at least joules of energy if we assume that the process is isothermal at temperature .


The energy dissipation has been reduced by approximately a factor of ten every five years, so a rough extrapolation suggests that a reduction of the energy dissipation per logic operation below (thermal noise, that is of the order of picojoule at room temperature) may become relevant in about 10-15 years. This issue may cause a variety of problems for classical computers, e.g. cooling may be difficult (according to current day knowledge/technology).

Irreversible operations as the NAND gate (1 ), the binary addition (sum and carry) and the real addition , dissipate energy. Is logical reversibility dissipation free?

First, let us note that the above irreversible operations can be easily simulated by reversible ones. A reversible version of the NAND gate is, for example, Toffoli’s gate


(2)
Indeed, (2) is a reversible 3-bit gate that flips the third bit if the first two take the value 1 and does nothing otherwise. Hence, the third output bit becomes the NAND of and in case . The price paid to get reversibility consisted of adding a new variable .

Similar tricks can be used to produce reversible versions of the binary addition, and real addition . In the first case we replicated the first variable ; in the second case we added a new component storing some additional value.

A computer may be fully reversible and yet dissipate energy! The important point is that the laws of physics allow for technologies to make reversible computers operate with negligible dissipation. To build a reversible computer one needs only two types of logical gates, say AND and NOT (because any other gate can be constructed from these two types – they are
universal). Clearly, the NOT gate is reversible as its composition with itself gives the initial input. However, the AND gate is irreversible. Reason: it has two inputs and only one output, so it has to lose information (it is impossible to tell exactly what inputs must have been if all one is told is the output 0: any of the three combinations , could have been the “real input").

To make a reversible variant of the gate AND we need to make sure that we have the same number of output lines as input ones, so, in principle, we can just add some “garbage" output lines to solve the problem. However, this may not be enough, as we want to guarantee also universality! One possibility is Toffoli’s [63] reversible 3-bit gate which uses in addition to a control bit . Input bits and do not change their states; the control bit, however, will change its state, but only when . Toffoli’s truth table is given in Table 1.
input output
0 0 0 0 0 0
0 1 0 0 1 0
1 0 0 1 0 0
1 1 0 1 1 1
input output
0 0 1 0 0 1
0 1 1 0 1 1
1 0 1 1 0 1
1 1 1 1 1 0
                       
                                                                                   Table 1: Toffoli’s gate.


Fredkin’s [
63] reversible 3-bit gate also uses in addition to a control bit in the following way: (a) if , then the values of are transmitted unaltered, i.e. the output is the pair , (b) if , then the values of are switched to the opposite output, i.e. the output is the pair . Its truth table is the following:
input output
0 0 0 0 0 0
0 1 0 0 1 0
1 0 0 1 0 0
1 1 0 1 1 0
input output
0 0 1 0 0 1
0 1 1 1 0 1
1 0 1 0 1 1
1 1 1 1 1 1
                   
                                                                                    Table 2: Fredkin’s gate.


Fredkin’s gate is universal in the sense that it can be used to construct reversible variants of all Boolean gates, and satisfies one additional requirement: the number of 1s and 0s never changes. (Toffoli’s gate does not satisfy this condition; for example, is transformed into .) It is not difficult to show that the gates NOT and AND can be represented using Fredkin’s gate FREDKIN with particular inputs:


(3)
and



(4)
so FREDKIN is universal.

Fredkin’s gate has often been used for photon based gates where a 1 represents a photon and a 0 simply denotes the absence of a photon; nonlinear optics is used to control the output of an interferometer (see Milburn [
101]). The number of 1s cannot change as the number of photons cannot change – absorption is not allowed for reversible gates.

In both formulae (3) and (4) there are more outputs than are required for the computed functions (one for NOT and two for AND): these outputs, called
garbage bits, are a necessary consequence of reversible logic. Consequently, one may wonder whether we have only postponed the energy cost; garbage bits can be irreversibly erased, but that would require to pay Landauer’s price …

Bennett [11] found in 1973 that any computation can be performed using only reversible steps, and so “in principle" requires no dissipation and no power expenditure:
we do not need to erase the garbage bits! By pointing out that a reversible computer can run forward to the end of a computation, print out a copy of the answer (a logically reversible operation) and then reverse all of its steps to return to its initial configuration, Bennett invented a procedure to remove the garbage without any energy cost. Consequently, in principle, we need not pay any “power bill" to compute reversibly. The moral we draw from the above discussion is, once again, that, to use Landauer’s [84] words, information is inevitably physical. So, let us explore what physics, especially quantum physics, has to tell us about information.



Here is the logarithm in base 2.
2Here k is Boltzmann’sconstant.
Recall that is 1 only if and have different values, i.e. or .
A single NAND gate is as good as having both AND and NOT: NAND, ANDNAND NANDNAND.