Entropy and perpetual computers

Somendra M. Bhattacharjee
Institute of Physics, Bhubaneswar 751 005, India email: [email protected] (Dated: February 2, 2008) A denition of entropy via the Kolmogorov algorithmic complexity is discussed. As examples, we show how the meaneld theory for the Ising model, and the entropy of a perfect gas can be recovered. The connection with computations are pointed out, by paraphrasing the laws of thermodynamics for computers. Also discussed is an approach that may be adopted to develop statistical mechanics using the algorithmic point of view.

arXiv:cond-mat/0310332v2 [cond-mat.stat-mech] 29 Oct 2003

I.

INTRODUCTION

The purpose of this lecture note is to illustrate a route for the denition of entropy using our experience with computers. In the process the connection between statistical physics and computations comes to the fore.
A. What is entropy?

by the angular brackets, but no such average for S. As a result, one cannot talk of free energy of a conguration at any stage of the simulation. All the denitions mentioned above associate S to the ensemble, or distributions over the phase space. They simply forbid the question what is the entropy of a conguration. Too bad!

B.

On computers

This is a question that plagues almost all especially beginning physics students. There are several correct ways to answer this. 1. It is the perfect dierential that one gets by dividing the heat transfered by a quantity T that gives us the hot-cold feeling (i.e. temperature). (thermodynamics) 2. It is the log of the number of states available. (Boltzmann) 3. It is something proportional to pi ln pi where pi is the probability that the system is in state i. (Gibbs) 4. It is just an axiom that there exists an extensive quantity S, obeying certain plausible conditions, from which the usual thermodynamic rules can be obtained. (Callen) But the colloquial link between disorder or randomness and entropy remains unexpressed though, agreeably, making a formal connection is not easy. Our plan is to establish this missing link a la Kolmogorov. Besides these conceptual questions, there is a practical issue that bugs many who do computer simulations where dierent congurations are generated by some set of rules. In the end one wants to calculate various thermodynamic quantities which involve both energy and entropy. Now, each conguration generated during a simulation or time evolution has an energy associated with it. But does it have an entropy? The answer is of course blowing in the wind. All thermodynamic behaviours ultimately come from a free energy, say, F = E T S where E, the energy, generally known from mechanical ideas like the Hamiltonian, enters as an average, denoted

Over the years we have seen the size of computers shrinking, speed increasing and power requirement going down. Centuries ago a question that tickled scientists was the possibility of converting heat to work or nding a perfect engine going in a cycle that would completely convert heat to work. A current version of the same problem would be: Can we have a computer that does computations but at the end does not require any energy. Or, we take a computer, draw power from a rechargeable battery to do the computation, then do the reverse operations and give back the energy to the battery. Such a computer is in principle a perpetual computer. Is it possible? What we mean by a computer is a machine or an object that implements a set of instructions without any intelligence. It executes whatever it has been instructed to do without any decision making at any point. At the outset, without loss of generality, we choose binary (0,1) as the alphabet to be used, each letter to be called a bit. The job of the computer is to manipulate a given string as per instructions. Just as in physics, where we are interested in the thermodynamic limit of innitely large number of particles, volumes etc, we would be interested in innitely long strings. The question therefore is can bit manipulations be done without cost of energy?

II.

RANDOMNESS

The problem that a conguration can not have an entropy has its origin in the standard statistical problem that a given outcome of an experiment cannot be tested for randomness. E.g., one number generated by a random number generator cannot be tested for randomness. For concreteness, let us consider a general model sys-

2 tem of a magnet consisting of spins si = 1 arranged on a square lattice with i representing a lattice site. If necessary, we may also use an energy (or Hamiltonian) E = J <ij> si sj where the sum is over nearest neighbours (i.e. bonds of the lattice). Suppose the temperature is so high that each spin can be in anyone of the two states 1 with equal probability. We may generate such a conguration by repeated tossing of a fair coin. If we get + + + + + + + (+:H,:T) is it a random conguration? Or Can the congurations of spins as shown in Fig. 1 be considered random? so that the size of the programme is same as the size of the string itself. This example shows that the size of the programme gives an expression to the naive feeling of randomness we have. We may then adopt it for a quantitative measure of randomness. Denition : Let us dene randomness of a string as the size of the minimal programme that generates the string. The crucial word is minimal. In computer parlance what we are trying to achieve is a compression of the string and the minimal programme is the best compression that can be achieved. Another name given to what we called randomness is complexity, and this particular measure is called Kolmogorov algorithmic complexity. The same quantity, randomness, is also called information, because the more we can compress a string the less is the information content. Information and randomness are then two sides of the same coin: the former expressing a positive aspect while the 2nd a negative one! Let K(c) be a programme for the string of conguration c and let us denote the length of any string by | ... |. The randomness or complexity is S(c) = min |K(c)|. (1)

(A)

(B)

(C)

FIG. 1: Ising magnet. Spins 1 are represented by arrows pointing up or down. (A) A ferromagnetic state, (B) an antiferromagnetic state, and (C) a seemingly random conguration.

With N spins (or bits), under tossing of a fair coin, the probability of getting Fig. 1(A) is 2N and so is the probability of (B) or (C). Therefore, the fact that a process is random cannot be used to guarantee randomness of the sequence of outcomes. Still, we do have a naive feeling. All Heads in N coin toss experiments or strings like 1111111... (ferro state of Fig. 1(A)) or 10101010... (anti-ferro state of Fig 1(B)) are never considered random because one can identify a pattern, but a string like 110110011100011010001001... (or conguration of Fig 1(C)) may be taken as random. But what is it that gives us this feeling?

We now dene a string as random, if its randomness or complexity is similar to the length of the string, or, to be quantitative, if randomness is larger than a pre-chosen threshold, e.g, say, S(c) > |c| 13. The choice of 13 is surely arbitrary here and any number would do.

1.

Comments

A few things need to be mentioned here. (i) By denition, a minimal programme is random, because its size cannot be reduced further. (ii) It is possible to prove that a string is not random by explicitly constructing a The naive expectation can be quantied by a dierent small programme, but it is not possible to prove that type of arguments, not generally emphasized in physics. a string is random. This is related to Gdels incomo Suppose I want to describe the string by a computer propleteness theorem. For example, the digits of may gramme; or rather by an algorithm. Of course there look random (and believed to be so) until one realizes is no unique programming language nor there is a that these can be obtained from an ecient routine for, computer - but these are not very serious issues. We say, tan1 . We may not have a well-dened way of conmay choose, arbitrarily, one language and one computer structing minimal algorithms, but we agree that such an and transform all other languages to this language (by algorithm exists. (iii) The arbitrariness in the choice of adding translators) and always choose one particular language leads to some indeniteness in the denition computer. The two strings, the ferro and the anti-ferro of randomness which can be cured by agreeing to add states, can then be obtained as outputs of two very small a translator programme to all other programmes. This programmes, still leaves the dierences of randomness of two strings to be the same. In other words, randomness is dened (A) Print 1 5 million times (ferro state) (B) Print 10 2.5 million times (antiferro state) upto an arbitrary additive constant. Entropy in classical thermodynamics also has that arbitrariness. (iv) Such a denition of randomness satises a type of subadditivity In contrast, the third string would come from condition S(c1 + c2 ) S(c1 ) + S(c2 ) + O(1), where the O(1) term cannot be ignored. (C) Print 110110011100... (disordered state)
A. Algorithmic approach

3
B. Entropy

Accepting that this Kolmogorovian approach to randomness makes sense and since we connect randomness in a physical system with entropy, let us associate this randomness S(c) with the entropy of that string or conguration c. For an ensemble of strings or congurations with probability pi for the i-th string or conguration ci , the average entropy will be dened by SK =
i

we have not attempted to compress . For example, no matter what the various strings are, all of the N spin congurations can be generated by a loop (algorithm represented schematically) 10 i = 0 i = i+1 L = length of i in binary Print 0 (N-L) times, then If ( i < N ) go to 10 stop

"i" in binary

pi S(ci )

(2)

(taking the Boltzmann constant kB = 1). We shall claim that this is the thermodynamic entropy we are familiar with. Since the denition of entropy in Eq. (2) looks ad hoc, let us rst show that this denition gives us back the results we are familiar with. To complete the story, we then establish the equivalence with the Gibbs denition of entropy.
C. Example I: Mean led theory for the Ising model

Consider the Ising problem. Let us try to write the free energy of a state with n+ + spins and n spins with n+ + n = N . The number of such congurations is = N! . n+ ! n ! (3)

An ordered list (say lexicographical) of all of these congurations is then made. If all of these states are equally likely to occur then one may specify a state by a string that identies its location in the list of congurations. The size of the programme is then the number of bits required to store numbers of the order of . Let S be the number of bits required. For general N, n+ , n , S is given by 2S = = S = log2 . (4)

By a suitable choice of N (e.g., N = 11.....1) the code for representation of N can be shortened enormously by compressing N . This shows that one may generate all the spin congurations by a small programme though there are several congurations that would require individually much bigger programmes. This should not be considered a contradiction because it produces much more than we want. It is fair to put a restriction that the programmes we want should be self delimiting (meaning it should stop without intervention) and should produce just what we want, preferably no extra output. Such a restriction then automatically excludes the above loop. Secondly, many of the numbers in the sequence from 1 to can be compressed enormously. However, what enumeration scheme we use, cannot be crucial for physical properties of a magnet, and therefore, we do need S bits to convey an arbitrary conguration. It is also reassuring to realize that there are random (i.e. incompressible) strings in 2N possible N -bit strings. The proof goes as follows. If an N -bit string is compressible, then the compressed length would be N 1. But there are only 2N 1 such strings. Now the compression procedure has to be one to one (unique) or otherwise decompression will not be possible. Hence, for every N , there are strings which are not compressible and therefore random. A related question is the time required to run a programme. What we have dened so far is the space requirement. It is also possible to dene a time complexity dened by the time required to get the output. In this note we avoid this issue of time altogether.

Stirlings approximation then gives S = n+ log2 n+ + n log2 n = N [p log2 p + (1 p) log2 (1 p)], (5)

2.

Free energy

with p = n+ /N , the probability of a spin being up. Resemblance of Eq. (4) with the Boltzmann formula for entropy (Sec. I) should not go unnoticed here. Eq. (5) is the celebrated formula that goes under the name of entropy of mixing for alloys, solutions etc.
1. Comments

In the Kolmogorov approach we can now write the free energy of any conguration, ci as Fi = Ei T Si with the thermodynamic free energy coming from the average over all congurations, F F = E T S . If we now claim that S obtained in Eq. (5) is the entropy of any conguration, and since no compression is used, it is the same for all (this is obviously an approximation), we may use S = S. The average energy may be approximated by assuming random mixture of up and down spins with an average value s = p (1 p). If q is

It is important to note that no attempt has been made for minimalizations of the algorithm or in other words

4 the number of nearest neighbours (4 for a square lattice), the free energy is then given by q F = J(2p 1)2 T [p log p + (1 p) log(1 p)]. (6) N 2 Note that we have not used the Boltzmann or the Gibbs formula for entropy. By using the Kolmogorov denition what we get back is the mean eld (or Bragg-Williams) approximation for the Ising model. As is well-known, this equation on minimization of F with respect to p, gives us the Curie-Weiss law for magnetic susceptibility at the ferro-magnetic transition. No need to go into details of that because the purpose of this exercise is to show that the Kolmogorov approach works.
D. Example II: Perfect gas III. A. COMPUTERS On computation

For the computer problem mentioned in the Introduction, one needs to ponder a bit about reality. In thermodynamics, one considers a reversible engine which may not be practical, may not even be implementable. But a reversible system without dissipation can always be justied. Can one do so for computers?

1.

Reversible computers?

A more elementary example is the Sckur-Tetrode fora mula for entropy of a perfect gas. We use cells of small sizes V such that each cell may contain at most one particle. For N particles we need = (V /V )N numbers to specify a conguration, because each particle can be in V one of V /V cells. The size in bits is S = N log2 V so that the change in randomness or entropy as the volume is changed from Vi to Vf is S = N log2 Vf . Vi (7)

To implement an algorithm (as given to it), one needs logic circuits consisting of say AND and NAND gates (all others can be built with these two) each of which requires two inputs (a,b) to give one output (c). By construction, such gates are irreversible: given c, one can not reconstruct a and b. However it is possible, at the cost of extra signals, to construct a reversible gate (called a Tooli gate) that gives AND or NAND depending on a third extra signal. The truth table is given in Appendix A. Reversibility is obvious. A computer based on such reversible gates can run both ways and therefore, after the end of manipulations, can be run backwards because the hardware now allows that. Just like a reversible engine, we now have a reversible computer. All our references to computers will be to such reversible computers.

The indistinguishability factor can also be taken into account in the above argument, but since it does not aect Eq. (7), we do not go into that. Similarly momentum contribution can also be considered.

2.

Laws of computation

Let us try to formulate a few basic principles applicable to computers. These are rephrased versions of laws familiar to us. Law I: It is not possible to have perpetual computation.
FIG. 2: Perfect gas: space divided into cells. The cells are occupied by the particles

It may be noted here that the work done in isothermal expansion of a perfect gas is
Vf

In other words, we cannot have a computer that can read a set of instructions and carry out computations to give us the output without any energy requirement. Proving this is not straight forward but this is not inconsistent with our intuitive ideas. We wont pursue this. This type of computer may be called perpetual computer of type I. First law actually forbids such perpetual computers. Law II: It is not possible to have a computer whose sole purpose is to draw energy from a reversible source, execute the instructions to give the output and run backward to deliver the energy back to source, and yet leave the memory at the end in the original starting state. A computer that can actually do this will be called a perpetual computer of second kind or type II.

P dV = N kB T ln
Vi

Vf = (kB ln 2)T S. Vi

(8)

Where P is the pressure satisfying P V = N kB T and S is dened in Eq. (7). Both Eqs. (7) and (8) are identical to what we get from thermodynamics. The emergence of ln 2 is because of the change in base from 2 to e. It seems logical enough to take this route to the denition of entropy and it would remove much of the mist surrounding entropy in the beginning years of a physics student.

5
3. What generates heat?

In order to see the importance of the second law, we need to consider various manipulations on a le (which is actually a string). Our interest is in long strings (length going to innity as in thermodynamic limit in physics). Now suppose we want to edit the le and change one character, say, in the 21st position. We may then start with the original le and add an instruction to go to that position and change the character. As a result the edit operation is described by a programme which is almost of the same length (at least in the limit of long strings) as the original programme giving the string. Therefore there is no change in entropy in this editing process. Suppose we want to copy a le. We may attach the copy programme with the le. The copy programme itself is of small size. The copy process therefore again does not change the entropy. One may continue with all the possible manipulations on a string and convince oneself that all (but one) can be performed at constant entropy. The exceptional process is delete or removal of a le. There is no need of elaboration that this is a vital process in any computation. When we remove a le, we are replacing the entire string by all zeros - a state with negligible entropy. It is this process that would reduce the entropy by N for N characters so that in conventional units the heat produced at temperature T is N kB T ln 2 (see Eq. (8)). We know from physics that entropy reduction does not happen naturally (we cannot cool a system easily).

phase to a very low temperature ferromagnetic phase. If the process of deletion on a computer occurs very fast in a very small region of space, this heat generation can create problem. It therefore puts a limit on miniaturization or speed of computation. Admittedly this limit is not too realistic because other real life processes would play major roles in determining speed and size of a computer. See Appendix C for an estimate of heat generated.
B. Communication 1. The problem

Let us now look at another aspect of computers namely transmission of strings (or les) or communication. This topic actually predates computers. To be concrete, let us consider a case where we want to transmit images discretized into small cells of four colours, RGBY with probabilities p(R) = 1/2, p(G) = 1/4, p(B) = p(Y) = 1/8. The question in communication is: What is the minimal length of string (in bits) required to transmit any such image?
2. Kolmogorov and Shannons theorem

4.

Memory as fuel

We can have a reversible computer that starts by taking energy from a source to carry out the operations but to run it backward (via Tooli gates) it has to store many redundant information in memory. Even though the processes are iso-entropic and can be reversed after getting the output to give back the energy to the source, we no longer have the memory in the same blank state we started with. To get back to that blank state, we have to clear the memory (remove the strings). This last step lowers the entropy, a process that cannot be carried out without help from outside. If we do not want to clear the memory, the computer will stop working once the memory is full. This is the second law that prohibits perpetual computer of second kind. The similarity with thermodynamic rules is apparent. To complete the analogy, a computer is like an engine and memory is the fuel. From a practical point of view, this loss of entropy is given out as heat (similar to latent heat on freezing of water). Landauer in 1961 pointed out that the heat produced due to this loss of entropy is kB T ln 2 per bit or N kB T ln 2 for N bits. For comparison, one may note that N kB ln 2 is the total amount of entropy lost when an Ising ferromagnet is cooled from a very high temperature paramagnetic

There are two possible ways to answer this question. The rst is given by the Kolmogorov entropy (= randomness = complexity) while the second is given by a dierent powerful theorem called Shannons noiseless coding theorem. Given a long string Cj of say N characters, if we know its Kolmogorov entropy Sj then that has to be the smallest size for that string. If we now consider all possible N character strings with Pj as the probability of the jth string, then SK = j Pj Sj is the average number we are looking for. Unfortunately it is not possible to compute Sj for all cases. Here we get help from Shannons theorem. The possibility of transmitting a signal that can be decoded uniquely is guaranteed with probability 1, if the average number of bits per character = pi log2 pi where pi s are the probabilities of individual characters. A proof of this theorem is given in Appendix B. Since the two refer to the same object, they are the same with probability 1, i.e., SK = N pi log2 pi .

3.

Examples

The applicability of the Shannon theorem is now shown for the above example. To choose a coding scheme, we need to restrict ourselves to prex codes (i.e. codes that do not use one code as the prex of another code. As

6 an example, if we choose R 0, G 1, B 10, Y 11, decoding cannot be unique. E.g. what is 010? RGR or RB? Nonuniqueness here came from the fact that B (10) has the code of G (1) as the rst string or prex. A scheme which is prex free is to be called a prex code. For our original example, we may choose R 0, G 10, B 110, Y 111 as a possible coding scheme to nd that the average length required to transmit a colour is l 1 1 1 1 7 +2 +23 = . 2 4 8 4 (9) gas showed, this entropy can be used to get the entropy of individual strings. Given a set, like the colours in the above example, we can have dierent probability distributions for the elements. The Shannon entropy would be determined by that distribution. In the Kolmogorov case, we are assigning an entropy Sj to the jth long string or state but SK is determined by the probabilities Pj s of the long strings which are in turn determined by the ps of the individual characters. Since both refer to the best compression on the average, they have to be equivalent. It should however be noted that this equivalence is only in the limit and is a probability 1 statement meaning that there are congurations which are almost not likely to occur and they are not counted in the Shannon entropy. Instead of the full list to represent all the congurations (as we did in Eqs. (3) and (4)), it suces to consider a smaller list consisting of the relevant or typical congurations. They are 2N p log2 p in number (see Appendix B for details), typically requiring S bits per character. A physical example may illustrate this. Even though all conguration of molecules in a gas are allowed and should be taken into account, it is known that not much harm is done by excluding those congurations where all the molecules are conned in a small volume in one corner of a room. In fact giving equal weightage to all the congurations in Eq. (4) is one of the sources of approximations of meaneld theory.

It is a simple exercise to show that any other method would only increase the average size. What is remarkable is that
i

pi log2 pi = 7/4,

an expression we are familiar with from the Gibbs entropy and also see in the Shannon theorem. In case the source changes its pattern and starts sending signals with equal probability p(R) = p(G) = p(B) = p(Y) = 1/4, we may adopt a dierent scheme with R 00, G 10, B 01, Y 11, for which the average length is l =2=
i

pi log2 pi .

IV.

STATISTICAL MECHANICS

This is less than what we would get if we stick to the rst scheme. Such simple schemes may not work for arbitrary cases as, e.g., for p(R) = 1/2, p(G) = 1 1 2, p(B) = p(Y) = + . 4 8

In the rst scheme we get l = 7 + 2 while the second 4 scheme would give l = 2. In the limit of = 1/8, we can opt for a simpler code R 0, B 10, Y 11, with l = 3/2.

One way to reduce this length is then to make a list of all possible 2N S strings, where S = p log2 p in some particular order and then transmit the item number of the message. This cannot require more than S bits per character. We see the importance of the Gibbs formula but it is called the Shannon entropy.
4. Entropy

It is to be noted that the Shannon theorem looks at the ensemble and not at each string independently. Therefore the Shannon entropy S = i pi ln pi is ensemble based, but as the examples of magnet or noninteracting

We now try to argue that statistical mechanics can also be developed with the above entropy picture. To do so, we consider the conventional canonical ensemble, i.e., a system dened by a Hamiltonian or energy H in contact with a reservoir or bath with which it can exchange only energy. In equilibrium, there is no net ow of energy from one to the other but there is exchange of energy going on so that our system goes through all the available states in phase space. This process is conventionally described by appropriate equations of motions but, though not done generally, one may think of the exchange as a communication problem. In equilibrium, the system is in all possible states with probability pi for the ith state and is always in communication with the reservoir about its conguration. The communication is therefore a long string of the states of the system each occurring independently and identically distributed (thats the meaning of equilibrium). It seems natural to make the hypothesis that nature picks the optimal way of communication. We of course assume that the communication is noiseless. The approach to equilibrium is just the search for the optimal communication. While the approach process has a time dependence where the time complexity would play a role, it has no bearing in equilibrium and need not worry us. With that in mind, we may make the following postulates:

7 (1) In equilibrium, the energy E = i pi Ei remains constant. (2) The communication with the reservoir is optimal with entropy S = i pi ln pi . (3) For a given average energy, the entropy is maximum to minimize failures in communication. The third postulate actually assures that the maximum possible number of congurations (= 2S ) are taken into account in the communication process. No attempt has been made to see if these postulates can be further minimized. With these sensible postulates, we have the problem of maximizing S with respect to pi s keeping E =constant and i pi = 1. A straight forward variational calculation shows that pi = exp(Ei )/Z with Z = exp(Ei ) being the standard partition function. The parameter is to be chosen properly such that one gets back the average energy. The usual arguments of statistical mechanics can now be used to identify with the inverse temperature of the reservoir. a 0 0 1 1 0 0 1 1 b 0 1 0 1 0 1 0 1 c 0 0 0 0 1 1 1 1 a 0 0 1 1 0 0 1 1 b 0 1 0 1 0 1 0 1 c 0 0 0 1 1 1 1 0

a/ 11 00 11 00 b/ 11 00 11 00 gate c / Fig. A1: Tooli a b c

APPENDIX B: PROOF OF SHANNONS THEOREM

The statement of Shannons noiseless coding theorem is : If l is the minimal average code length of an optimal code, then S l S+1 where S =
j

pj log2 pj .

V.

SUMMARY

The adjective noiseless is meant to remind us that there is no error in communication. A more verbose statement would be If we use N l bits to represent strings of N characters with Shannon entropy S, then a reliable compression scheme exists if l > S. Conversely, if l < S, no compression scheme is reliable. The equivalence of the two statements can be seen by recognizing that S need not be an integer but l better be.

We have tried to show how the Kolmogorov approach to randomness may be fruitfully used to dene entropy and also to formulate statistical mechanics. Once the equivalence with conventional approach is established, all calculations can then be done in the existing framework. What is gained is a conceptual framework which lends itself to exploitation in understanding basic issues of computations. This would not have been possible in the existing framework. This also opens up the possibility of replacing engines by computers in teaching of thermodynamics.

1.

Simple motivation

Acknowledgments This is based on the C. K. Majumdar memorial talks given in Kolkata on 22nd and 23rd May 2003. I was fortunate enough to have a researcher like Prof. Chanchal Kumar Majumdar as a teacher in Science College. I thank the CKM Memorial Trust for organizing the memorial talk in Science College, Kolkata. Let us rst go through a heuristic argument to motivate Shannons coding theorem. Suppose a source is emitting signals {ci } independently and identically distributed with two possible values ci = 0 with probability p1 = p, and ci = 1 with probability p2 = 1 p. For a long enough string C c1 c2 c3 c4 ...cN the probability is P(C) = p(c1 )p(c2 )p(c3 )p(c4 )...p(cN ) p = 2
APPENDIX A: TOFFOLI GATE
Np

(B1a) (B1b) (B1c)

(1 p)

N (1p)

N [p log2 p+(1p) log2 (1p)]

The truth table of the Tooli gate is given below. With three inputs a,b,c, the output in c is the AND or NAND operation of a and b depending on c=0 or 1.

because for large N the number of expected 0 is N p and 1 is N (1 p). This expression shows that the probability of a long string is determined by S({pj }) = [p log2 p + (1 p) log2 (1 p)], (B2)

8 the entropy for this particular problem. Note the subtle change from Eq. (B1a) to Eq. (B1b). This use of expectation values for large N led to the result that most of the strings, may be called the typical strings, belong to a subset of 2N S strings (out of total 2N strings).
2. What is Typical?

Let us now get a lower bound for Ntyp . We have just established that the probability for a string to be typical is 1 . Using the other limit from Eq. (B3) we have 1

2N (S) = Ntyp 2N (S) , (B9)

Let us dene a typical string more precisely for any distribution. A string of N symbols C = c1 c2 c3 c4 ...cN will be called typical (or better -typical) if 2N (S+) P(C) 2N (S), (B3)

which gives Ntyp (1)2N (S). The nal result is that the total number of typical strings satises 2N (S+) Ntyp (1 )2N (S) where > 0 can be chosen small for large N . Hence, in the limit Ntyp 2N S .
4. Coding scheme

(B10)

for any given > 0. Eq. (B3) may also be rewritten as N 1 log2 P(C) S
3. How many typical strings?

(B4)

Now, for random variables ci , Xi s, dened by Xi = log2 p(ci ), are also independent identically distributed 1 random variables. It is then expected that X = N i Xi , the average value of Xi s, averaged over the string for large N , should approach the ensemble average, namely, X = j pj log2 pj = S. This expectation comes from the law of large numbers that Prob N 1
i

Now let us choose a coding scheme that requires N l number of bits for the string of N characters. Our aim is to convert a string to a bit string and decode it - the whole process has to be unique. Representing the coding and decoding by operators C and D respectively, and any string by c|, what we want can be written in a familiar form c|C|D = c| for all c|, cat myfile|gzip|gunzip gives myfile where the last line is the equivalent pipeline in a UNIX or GNU/Linux system. Lets take l > S. We may choose an such that l > S+ . It is a trivial result that Ntyp 2N (S+) < 2N l . Here 2N l is the total number of possible bit strings. Hence all the typical strings can be encoded. Nontypical strings occur very rarely but still they may be encoded. If l < S, then Ntyp > 2N l and obviously all the typical strings cannot be encoded. Hence no coding is possible. This completes the proof of the theorem.
APPENDIX C: HEAT GENERATED IN A CHIP

log2 p(ci ) S < 1,

(B5)

for any > 0. This means that given an we may nd a > 0 so that the above probability in Eq. B5 is greater than 1 . Recognizing that log2 p(ci ) = log2
i i

p(ci ) = log2 P(C),

(B6)

Eq. (B5) implies Prob N 1 log2 P(C) S < 1 . (B7)

We conclude that the probability that a string is typical as dened in Eqs. (B3) and (B4) is 1 . Let us now try to estimate the number Ntyp , the total number of typical strings. Let us use a subscript for the typical strings with going from 1 to Ntyp . The sum of probabilities P s of the typical strings has to be less than or equal to one, and using the denition of Eq. (B3), we have one inequality 1

2N (S+) = Ntyp 2N (S+) .

(B8)

This gives Ntyp 2N (S+) .

As per a report of 1988, the energy dissipation per logic operation has gone down from 103 joule in 1945 to 1013 joule in 1980s. (Ref: R. W. Keyes, IBM J. Res. Devel. 32, 24 (1988) URL: http://www.research.ibm.com/journal/rd/441/keyes.pdf) For comparison, thermal energy kB T at room temperature is of the order of 1020 joule. If one can pack 1018 logic gates in one cc operating at 1 gigahertz with minimal dissipation of kB T , it would release 3 megawatts of energy. Can one cool that? A more recent example. For a pentium 4 at 1.6GHz, if the cpu fan (that cools the CPU) is kept o, then during operations the cpu temperature may reach 107C (yes Celsius) as monitored by standard system softwares on an HCL made PC (used for preparation of this paper).

[***] Following references have been extensively used here and should be consulted for details . [1] A. N. Kolmogorov, Logical basis for information theory and probability theory, IEEE Trans. Information Theory, IT14, 662 (1968). [2] Ming Li and Paul Vitnyi, An introduction to Kolmogorov a Complexity and its applications (Springer-Verlag, New York, 1997). [3] M. A. Nielsen and I. L. Chuang, Quantum Computation

and Quantum Information (Cambridge U Press, Cambridge, UK, 2000). [4] W. H. Zurek, Algorithmic randomness and physical entropy, Phys. Rev. A40, 4731 (1989). [5] C. H. Bennett, The thermodynamics of computation - a review, Int. J. Theo. Phys. 21, 905 (1982). [6] G. J. Chaitin, Randomness and mathematical proof, Sci. Am. 232, 47 (1975).