0707.3400v2

Upload: alan-rios-fukelman

Post on 04-Jun-2018

216 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/13/2019 0707.3400v2

    1/24

    arXiv:0707.340

    0v2

    [physics.hist-ph]5Aug2008

    The physics of Maxwells demon and information

    Koji Maruyama1,2, Franco Nori1,2,3, and Vlatko Vedral4,51Advanced Science Institute, RIKEN (The Institute of Physical and Chemical Research), Wako-shi 351-0198, Japan

    2CREST, Japan Science and Technology Agency (JST), Kawaguchi, Saitama 332-0012, Japan3Center for Theoretical Physics, Physics Department, Center for the Study of Complex Systems,

    University of Michigan, Ann Arbor, MI 48109-1040, USA4The School of Physics and Astronomy, University of Leeds, Leeds, LS2 9JT, UK

    5Quantum Information Technology Lab, National University of Singapore, 117542 Singapore, Singapore(Dated: August 5, 2008)

    Maxwells demon was born in 1867 and still thrives in modern physics. He plays important rolesin clarifying the connections between two theories: thermodynamics and information. Here, wepresent the history of the demon and a variety of interesting consequences of the second law ofthermodynamics, mainly in quantum mechanics, but also in the theory of gravity. We also highlightsome of the recent work that explores the role of information, illuminated by Maxwells demon, inthe arena of quantum information theory.

    Contents

    I. Introduction 1

    II. Maxwells demon 2A. The paradox 2B. Szilards engine 2C. Temporary solutions to the paradox 3

    III. Exorcism of Maxwells demon: Erasure ofclassical information encoded in classicalstates 4

    IV. Other derivations of the erasure entropy7

    V. Some interesting implications of thesecond law 8

    A. Wave nature of light from the second law 8B. Gibbs paradox and quantum superpositionprinciple 8

    C. Quantum state discrimination and thesecond law 10

    D. Linearity in quantum dynamics 11E. Second law and general relativity 11F. Einstein equation from thermodynamics 12

    VI. Erasure of classical information encodedin quantum states 13

    VII. Thermodynamic derivation of the Holevobound 14A. from the erasure principle 14B. from the second law 15

    VIII. Entanglement detection by Maxwellsdemon(s) 17A. Work deficit 17B. Quantum discord 19C. Thermodynamic separability criterion 19

    IX. Physical implementations of the demon 20

    X. Concluding remarks 21

    Acknowledgments 21

    References 21

    I. INTRODUCTION

    The main focus of this article is the second law of ther-modynamics in terms of information. There is a longhistory concerning the idea of information in physics, es-pecially in thermodynamics, because of the significantresemblance between the information theoretic Shannonentropyand the thermodynamic Boltzmann entropy, de-spite the different underlying motivations and origins ofthe two theories (See for example Ref. [1] and references

    therein). A definitive discovery in this context is Lan-dauers erasure principle, which clearly asserts the rela-tionship between information and physics. As we willsee below, having been extended to the quantum regimeby Lubkin [2], it has proven useful in understanding theconstraints on various information processing tasks froma physical point of view.

    Here we will review the background on the correspon-dence between information and physics, in particularfrom the point of view of thermodynamics. Starting withthe classic paradox of Maxwells demon, we will discussthe erasure principle from the perspectives that will beof interest to us in later sections. Then we shall reviewa variety of consequences of the second law, mainly inquantum mechanics, and also briefly in the theory ofgravity. These are thought-provoking because the sec-ond law of thermodynamics is a sort of meta-rule, whichholds regardless of the specific dynamics of the system welook at. Besides, the fundamental postulates of quantummechanics, as well as general relativity, do not presumethe laws of thermodynamics in the first place. After re-viewing these classic works and appreciating how power-ful and how universal the second law is, we will discusssome of recent progress concerning the intriguing inter-

    http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2http://arxiv.org/abs/0707.3400v2
  • 8/13/2019 0707.3400v2

    2/24

    2

    play between information and thermodynamics from theviewpoint of quantum information theory.

    II. MAXWELLS DEMON

    A. The paradox

    A character who has played an important role in thehistory of physics, particularly in thermodynamics andinformation, is Maxwells demon. It was first introducedby Maxwell in 1871 [3] to discuss the limitations of thesecond law of thermodynamics, which is also the titleof a section in his book. The second law (in Clausiussversion) states[4]: It is impossible to devise an enginewhich, working in a cycle, shall produce no effect other

    than the transfer of heat from a colder to a hotter body.

    Maxwell devised his demon in a thought experiment todemonstrate that the second law is only a statistical prin-ciple that holds almostall the time, and not an absolutelaw set in stone.

    The demon is usually described as an imaginary tinybeing that operates a tiny door on a partition which sep-arates a box into two parts of equal volumes, the left andthe right. The box contains a gas which is initially inthermal equilibrium, i.e. its temperature T is uniformover the whole volume of the box. LetvT denote theaverage speed of molecules that form the gas. The de-mon observes the molecules in the left side of the box,and if he sees one approaching the door with a speed lessthanvT, then he opens the door and lets the moleculego into the right side of the box. He also observes themolecules in the right, and if he sees one approachingwith a speed greater thanvT, then he opens the doorto let it move into the left side of the box.Once he has induced a small difference in temperaturesbetween the right and the left, his action continues totransfer heat from the colder part (right) to the hotterpart (left) without exerting any work, thus he is breakingClausiuss form of the second law. This type of demon isreferred to as the temperature demon.

    There is another type of Maxwells demon, who is lessintelligent than the temperature demon. Such a demonmerely allows all molecules moving in one direction togo through, while stopping all those moving the otherway to produce a difference in pressure. This pressuredemon runs a cycle by making the gas interact with aheat bath at a constant temperature after generating apressure inequality. The sole net effect of this cycle isthe conversion of heat transferred from the heat bath towork. This is also a plain violation of the second law,which rules out perpetuum mobile (in Kelvins form):It is impossible to devise an engine which, working in

    a cycle, shall produce no effect other than the extraction

    of heat from a reservoir and the performance of an equal

    amount of mechanical work.

    The second law can also be phrased as in any cyclicprocess the total entropy of the physical systems involved

    in the process will either increase or remain the same.Entropy is, in thermodynamics, a state variable Swhosechange is defined as dS=Q/Tfor a reversible processat temperatureT, whereQ is the heat absorbed. Thus,irrespective of the type of demon, temperature or pres-sure, what he attempts to do is to decrease the entropyof the whole system for the cyclic process.

    Historically, a number of physical mechanisms that

    might emulate the demon without any intelligent beingshave also been proposed. One notable example shouldbe the trap-door model by Smoluchowski [5]. Instead ofan intelligent demon operating the door, he considereda door that is attached to the partition by a spring sothat it only opens to one side, the left, say. Then fastmoving molecules in the right side can go into the leftside by pushing the door, but slow ones are simply re-flected as the door is shut tightly enough for them andno molecules can go into the right from the left. Aftera while, the temperature (as well as the density) of theleft side should become higher and the right side lower.Useful work would be extracted from this spontaneouslygenerated temperature difference. Smoluchowski pointedout that what prevents the trap-door mechanism fromachieving the demonic work are thermal fluctuations, i.e.Brownian motion, of the door. The door might be kickedsometimes to let a fast molecule in; however, the thermalfluctuations will lift the door up and let molecules go backto the right side, resulting in no net temperature differ-ence. This scenario was numerically analyzed in detailin Ref. [6] to confirm the above reasoning. While therehave been many other similar mechanisms proposed, amore sophisticated model was discussed by Feynman asthe ratchet-and-pawl machine[7], which, again, does notwork as a perpetual engine due to the thermal fluctua-tions.

    The demon puzzle, which had been a cardinal questionin the theory of thermodynamics, is now why a demoncan never operate beyond the apparently fundamentallimits imposed by the second law, no matter how intel-ligent he is and no matter what type (temperature orpressure) he is. An ingenious idea by Szilard treated thedemons intelligence as information and linked it withphysics.

    B. Szilards engine

    In 1929, the Hungarian scientist Leo Szilard presenteda classical (non-quantum) analysis of Maxwells demon(pressure demon), formulating an idealized heat enginewith one-molecule gas [8]. Szilards work was epoch-making in the sense that he explicitly pointed out, forthe first time, the significance of information in physics.

    The process employed by Szilards engine is schemati-cally depicted in Fig.1. A chamber of volumeV containsa gas, which consists of a single molecule (Fig. 1(a)). Asa first step of the process, a thin, massless, adiabaticpartition is inserted into the chamber quickly to divide

  • 8/13/2019 0707.3400v2

    3/24

    3

    FIG. 1: Schematic diagram of Szilards heat engine. A cham-ber of volume V contains a one-molecule gas, which can befound in either the right or the left part of the box. (a) Ini-tially, the position of the molecule is unknown. (b) Maxwellsdemon inserts a partition at the centre and observes themolecule to determine whether it is in the right or the lefthand side of the partition. He records this information in his

    memory. (c) Depending on the outcome of the measurement(which is recorded in his memory), the demon connects a loadto the partition. If the molecule is in the right part as shownin the figure, he connects the load to the right hand side ofthe partition. (d) The isothermal expansion of the gas doeswork upon the load, whose amount is kTln 2 which we call 1bit. (Adapted from Fig. 4 in Ref. [9].)

    it into two parts of equal volumes. The demon measuresthe position of the molecule, either in the right or in theleft side of the partition (Fig. 1(b)). The demon recordsthis result of the measurement for the next step. Then,

    he connects a load of a certain mass to the partition onthe side where the molecule is supposed to be in, accord-ing to his recorded result of the previous measurement(Fig.1(c)). Keeping the chamber at a constant temper-ature T by a heat bath, the demon can let the gas dosome work Wby quasistatic isothermal expansion (thepartition now works as a piston). The gas returns to itsinitial state, where it now occupies the whole volume V,when the partition reaches the end of the chamber. Dur-ing the expansion, heatQis extracted from the heat bathand thus W = Q as it is an isothermal process. Hence,Szilards engine completes a cycle after extracting heatQ and converting it to an equal amount of mechanical

    work.As the gas is expanded isothermally[106], the amount

    of extracted work W is kTVV/2

    V1dV = kTln2. An

    immediate question here might be if it is appropriateto assume the one-molecule gas as a normal ideal gasin discussing thermodynamic/statistical properties. Tofill this conceptual gap, we can consider an ensemble ofone-molecule gases. Then by taking averages over theensemble we can calculate various quantities as if it is anideal gas with a large number of molecules. In a sense,

    this can been seen as the origin of the intersection be-tween thermodynamics and information theory: lookingat the (binary) position of the molecule leads to its dualinterpretations, i.e., in terms of thermodynamics and in-formation theory.

    Naturally, the factor kT ln 2 appears often in the fol-lowing discussions on thermodynamic work, so we willtake it as a unit and call it 1 bit when there is no risk

    of confusion[107]. This will be especially useful when wecoordinate discussions of the information theoretic bitwith the thermodynamic work.

    The demon apparently violates the second law. Asa result of the perfect conversion of heat Q into workW, the entropy of the heat bath has been reduced byQ/T=W/T =k ln 2. According to the second law, theremust be an entropy increase of at least the same amountsomewhere to compensate this apparent decrease. Szi-lard attributed the source of the entropy increase to mea-surement. He wrote The amount of entropy generatedby the measurement may, of course, always be greaterthan this fundamental amount, but not smaller[8]. He

    referred to k ln 2 of entropy as a fundamental amountwell before Shannon founded information theory in 1948[10, 11]. Although he regarded the demons memory asan important element in analyzing his one-molecule en-gine, Szilard did not reveal the specific role of the mem-ory in terms of the second law. Nevertheless, his work isvery important as it was the first to identify the explicitconnection between information and physics.

    C. Temporary solutions to the paradox

    As Szilard did, many generations believed for decades

    that the paradox of Maxwells demon could be solved byattributing the entropy increase to measurement. Note-worthy examples include those by Brillouin Ref. [12] andGabor[13]. They considered light to measure the speedof the molecules and (mistakenly) assumed this to bethe most general measurement setting. Inspired by thework of Demers [14], who recognized in the 1940s thata high temperature lamp is necessary to illuminate themolecules so that the scattered light can be easily distin-guished from blackbody radiation, Brillouin showed thatinformation acquisition via light signals is necessarily ac-companied by an entropy increase, which is sufficient tosave the second law[12]. Interestingly, in his speculation,Brillouin linked the thermodynamic and the informationentropies directly. Information entropy is a key functionin the mathematical theory of information, which wasfounded by Shannon only a few years before Brillouinswork, and although its logical origin is very different fromthermodynamics, Brillouin dealt with two entropies onthe same footing by putting them in the same equationto link the gain of information with the decrease of phys-ical entropy. This led to the idea of negentropy, whichis a quantity that behaves oppositely to the entropy[108][15,16]. The negentropy is usually defined as the differ-

  • 8/13/2019 0707.3400v2

    4/24

    4

    ence between the maximum possible entropy of a systemunder a given condition and the entropy it actually has,i.e. N :=Smax S.

    Brillouin distinguished two kinds of information, freeandbound. Free informationIfis an abstract and math-ematical quantity, but not physical. Bound informationIb is the amount of information that can be acquiredby measurement on a given physical system. Thus,

    roughly speaking, free information is equivalent to (ab-stract) knowledge in our mind and bound informationcorresponds to the information we can get about a phys-ical system, which encodes the information to be sentor stored. Bound information is then subject to environ-mental perturbations during the transmission. When theinformation carrier is processed at the end of the channel,it is transformed into free information. In Brillouins hy-pothesis, the gain in bound information by measurementis linked to changes in entropy in the physical system as

    Ib = Ipost-measb Ipre-measb

    = k(ln Ppre-meas ln Ppost-meas)= Spre-meas Spost-meas > 0, (1)

    wherePpre-measand Ppost-meas denote the numbers of pos-sible states of the physical system before and after themeasurement, and similarly Spre-meas and Spost-meas arethe entropies of the system[109]. The conversion coeffi-cient between physical entropy and bound information ischosen to be Boltzmanns constant to make the two quan-tities comparable in the same units. Equation(1) meansthat gaining bound information decreases the physicalentropy. This corresponds to the process (a) to (b) inFig.1.

    As bound information is treated with the physical en-tropy of the system on the same basis, the second lawneeds to be expressed with bound information as wellas the physical entropy. If no information on the phys-ical system is available initially, that is, if Iinitialb = 0,the final entropy of the system after obtaining (bound)information Ib is Sf =Si Ib. The second law of ther-modynamics says that in an isolated system the physicalentropy does not decrease[110]: Sf 0. By using thechange in negentropy N :=S, the second law maynow be written as

    Sf= (Si Ib) = Si Ib =Ni Ib0,which means

    (Ni+ Ib)0. (2)Naturally, if there is no change in the information avail-able to us, that is, Ib = 0, Eq. (2) is nothing but thestandard inequality for entropy, S 0. However, inEq. (2), information is treated as part of the total en-tropy, and it states that the quantity (negentropy + in-formation) never increases. This is a new interpretationof the second law of thermodynamics, implied by Bril-louins hypothesis.

    Following Brillouins hypothesis, Lindblad comparedthe entropy decrease in the system with the informationgain an observer can acquire [17]. He analyzed measure-ments of thermodynamic quantities in a fluctuating sys-tem and showed that the information gain by the ob-server is greater than or equal to the entropy reductionin the system. Hence, the total entropy never decreases,as expected.

    Brillouins idea of dealing with information and phys-ical entropy on an equal basis has been widely accepted.All discussions below about the physical treatment of in-formation processing tacitly assume this interpretation,which presupposes the dualityof entropy, i.e. both infor-mation theoretic and thermodynamic aspects.

    III. EXORCISM OF MAXWELLS DEMON:ERASURE OF CLASSICAL INFORMATION

    ENCODED IN CLASSICAL STATES

    Although the exorcism of Maxwells demon by at-

    tributing an entropy increase to the acquisition of infor-mation had been widely accepted by physicists for morethan a decade, the demon turned out to have surviveduntil Landauer and Bennett put an end to the demonslife by reconsidering the role of memory, which Szi-lard barely overlooked. Landauer examined the processof erasure of information, introducing a new concept oflogical irreversibility[18].

    Indeed, O. Penrose independently discovered essen-tially the same result about information erasure as Lan-dauers. Penrose argued, in his book published in 1970,Foundations of Statistical Mechanics, that the paradoxof Maxwells demon could be solved by considering the

    entropy increase due to memory erasure. This was evenearlier than Bennetts 1982 analysis of the demon, how-ever, it was left virtually unnoticed by physicists. Pen-roses treatment was rather abstract and it did not go asfar as Bennetts work, which investigated the possibilityof measurement with arbitrarily little entropy increase.Here we focus on the viewpoint by Landauer and Ben-nett.

    Since information processing must be carried out by acertain physical system, there should be a one-to-one cor-respondence between logical and physical states. Logicalstates may be described as an abstract set of variableson which some information processing can be performed.Then, a reversible logical process, which means an injec-tive (one-to-one) mapping for logical states, correspondsto a reversible physical process. By implicitly assuminga correspondence between logical and physical entropies,as Brillouin proposed, this implies that a reversible log-ical process can be realized physically by an isentropicprocess, i.e. an entropy-preserving process.

    However, a logically irreversible process is non-injective, i.e. many-to-one, mapping. Such a processdoes not have a unique inverse as there may be manypossible original states for a single resulting state. The

  • 8/13/2019 0707.3400v2

    5/24

    5

    key here is that memory erasure is a logically irreversibleprocess because many possible states of memory shouldbe set to a single fixed state after an erasing procedure. Itis impossible to determine the state prior to erasure with-out the aid of further information, such as the particulartask of a computer programme or knowledge about thestates of other memory registers that are correlated tothe memory in question. This certain fixed state after

    erasure is analogous to a white or blank sheet of pa-per, on which no information is recorded. After erasingstored information, the state of memory should be in onespecific state, in order not to carry any information (bydefinition of erasure). We will refer to the specific stateafter erasure as a standard state.

    In terms of physical states, a logically irreversible pro-cess reduces the degrees of freedom of the system, whichimplies a decrease in entropy. In order for this process tobe physically legitimate, the energy must be dissipatedinto the environment. Landauer then perceived that logi-cal irreversibility must involve dissipation, hence erasinginformation in memory entails entropy increase (in the

    environment). This point will be the final sword to ex-orcize Maxwells demon and is referred to as Landauerserasure principle.

    Another important observation regarding the physicsof information was given by Bennett[19]. He illustratedthat measurement can be carried out reversibly, i.e. with-out any change in entropy, provided the measuring ap-paratus is initially in a standard state, so that recordinginformation in the memory does not involve the erasureof information previously stored in the same memory.The rough reason for this is that measurement can be re-garded as a process that correlates the memory with thesystem (in other words, a process that copies the memorystate to another system in a standard state), which canbe achieved reversibly, at least in principle.

    Bennett exemplified the reversible correlating processby a one bit memory consisting of an ellipsoidal pieceof ferromagnetic material. The ferromagnetic piece issmall enough so that it consists of only a single domainof magnetization. The direction of the magnetizationrepresents the state of the memory. Suppose that there isa double well potential with respect to the direction of themagnetization in the absence of an external field: Paralleland antiparallel to the major axis of the ellipsoid are themost stable directions. The central peak between the twowells is considerably higher than kTin the absence of anexternal magnetic field, so that thermal fluctuations do

    not allow the state to climb over the peak. Figure2showsa sketch of the potential that was illustrated in Ref. [19].

    Two minima of the potential represent the state ofmemory, either 0 or 1, and the blank memory isassumed to be in one standard state, e.g. 0, before in-formation is copied onto it from another memory. Let usconsider the process that correlates the state of a blankmemory B with that of a memory A, which is the sub-

    ject of measurement. This can be achieved by manipu-lating the shape of the potential for the blank memory

    FIG. 2: A potential energy for a binary memory whose stateis represented by the direction of the magnetization. Whenthere is no external transverse magnetic field the memory isstable in one of the potential wells, which correspond to 0and 1 of recorded information. The transverse field lowers

    the height of the barrier at the centre. At a certain point, theprofile of the potential becomes bath-tub-like shape with aflat bottom, where the direction of the magnetization is verysensitive to the longitudinal field component. (Adapted fromRef.[19].)

    as follows. By applying a transverse external magneticfield, the peak of the central barrier becomes lower. Ata certain intensity of the field, there will be only a singlebath-tub-like flat bottom, i.e. the state of B becomesvery sensitive to a weak longitudinal component of thefield. The memory A is located so that its magnetiza-tion can cause a faint longitudinal field at the position ofthe memory B. Then because ofBs sensitivity to sucha field, the state ofA can be copied to the memory Bwith arbitrarily small (but nonzero) energy consumption.Removing the transverse field completes the correlatingprocess. The crux of the physics here is that this processcan be reversed by using the perturbation from anotherreference memory, which is in the standard state.

    Now let us focus on the erasure of information. Sincemeasurement can be done virtually without energy con-sumption, it is the dissipation due to the erasure pro-cess that compensates the entropy decrease induced byMaxwells demon in Szilards model. The physical sys-

    tem for the demons memory can be modelled as a one-molecule gas in a chamber of volumeV, which is dividedinto two parts, the left L and the right R, by a par-tition. The demon memorizes the measurement result bysetting the position of the molecule in this box. If themolecule in Szilards engine may be found in the left andthe right sides with equal probability, i.e. 1/2, then theminimum amount of work that needs to be invested anddissipated into the environment is k Tln2.

    The actual process is as follows. The molecule is in

  • 8/13/2019 0707.3400v2

    6/24

    6

    FIG. 3: Thermodynamic process to erase information. A bi-nary information is stored in a vessel as the position of the

    molecule, eitherL or R. A common procedure for both initialstates, i.e. removing the partition and halving the whole vol-ume by an isothermal compression towards the standard stateL, completes the erasure. (Adapted from Fig. 3 in Ref.[9].)

    either L or R, depending on the information it stores(Fig. 3(a)). To erase the stored information, first, weremove the partition dividing the vessel at the centre(Fig. 3(b)). Second, insert a piston at the right end(Fig. 3(c)), when the standard memory state is L, andpush it towards the left isothermally at temperature Tuntil the compressed volume becomes V /2 (Fig. 3(d)).

    The resulting state is L for both initial states and theinformation is erased. It is worth noting that the eras-ing process should not depend on the initial state of thememory. The R state in Fig. 3(a) may be transferredto L state by simply moving the region of volumeV /2to the left. However, in this case, the operator of the pis-ton needs to observe the position of the molecule and thisaction requires another memory. Thus, the erasure pro-cess should be independent of the initial memory state.The work invested to compress the volume from V toV /2is Werasure = kT ln 2 and this is dissipated as heat intothe environment, increasing its entropy byk ln 2, as Lan-dauer argued. As there is no wasted work (in the sense

    that all invested work is converted into heat to increasethe entropy of the environment),kTln 2 is the minimumamount of work to be consumed for erasure.

    If there is a biased tendency in the frequency of ap-pearance of a particular memory state, say L, how muchwould the erasure work be? The answer is simple: theerasure work is proportional to the amount of informa-tion stored, thus Werasure = kTln 2H(p), where p is theprobability for the molecule to be in the L state and

    H(p) =p logp (1 p) log(1 p) (3)

    FIG. 4: Erasure process for an unbalanced probability distri-bution. The only difference from the case of balanced distri-bution (Fig.3) is the expansion from (a) to (a), which givesusH(p) bits of work.

    is the (binary) Shannon entropy. Throughout this article,log denotes logarithms of base 2. The reason can be ex-plained by a process depicted in Fig. 4. The unbalancedtendency betweenLand Ris expressed by the numbers ofmolecules in theL and theRregions. As we consider onlyan ideal gas (with no interactions between molecules),this scenario does not change the discussion at all if weaverage the erasure work at the end. Since removing thepartition at the beginning allows the gas an undesiredirreversible adiabatic expansion/compression, we first letthe gases in both parts expand/contract isothermally by

    making the partition movable without friction (Fig. 4(a)to (a)). During this process, the gases exert work to-wards the outside. Letting pL, pR and VL denote thepressure in the left region, that in the right region, andthe volume of the region on the left of the partition, re-spectively, we can write the work done by gases as

    W =

    pVV/2

    (pLpR)dVL

    = N kT

    pVV/2

    p

    VL 1 p

    V VL

    dVL

    = N kT(ln 2 +p lnp+ (1 p) ln(1 p))

    = N kTln2(1 H(p)). (4)As the pressures in the left and the right are equal, thisis the same situation as in Fig. 3(a). Hence, at leastN kTln 2 of work needs to be consumed to set the mem-ory to the standard state (Fig. 4(c) to (d)). As a whole,we invested Werasure = kTln 2W = kTln 2H(p) ofwork per molecule.

    Maxwells demon is now exorcized. The entropy de-crease, or the equivalent work the demon could give us,should be completely consumed to make his memory

  • 8/13/2019 0707.3400v2

    7/24

    7

    state come back to its initial state. The state of thewhole system, consisting of the heat engine and the de-mon, is restored after completing a thermodynamic cycle,without violating the second law.

    IV. OTHER DERIVATIONS OF THEERASURE ENTROPY

    We have focused on the one-molecule gas model so far,however, Landauers erasure principle holds regardless ofspecific physical models. In order to see its generalitywith some concrete examples, we now briefly review twoparticularly interesting papers, one by Shizume[20] andthe other by Piechocinska[21].

    Shizume used a model of memory whose state wasrepresented by a particle having Brownian motion ina time-dependent double well potential. Assuming therandom force FR(t) to be white and Gaussian satisfy-ingFR(t1)FR(t2)= 2mT(t1 t2), the motion of thedistribution function f(x,u,t) of the particle in the po-sition (x) and velocity (u) space can be described by the

    Fokker-Planck equation. Shizume then compared Q andTdS/dt, i.e. the ensemble average of the energy givento the particle from the environment per unit time, andthe change in the entropy of the whole system per unittime multiplied by the temperature. The entropy S isthe Shannon entropy of continuous distribution, definedby

    S:=k

    dxduf(x,u,t) ln f(x,u,t). (5)

    With the help of the Fokker-Planck equation concerningf(x,u,t), one arrives at the relation,

    QTdSdt

    , (6)

    from which we obtain the lower bound of the energy dis-sipated into the environment between times ti and tf as

    Qout(ti, tf) =

    tfti

    ( Q) dt T[S(ti) S(tf)]. (7)

    Equation (7) gives us the lower bound of the heat gen-eration due to the process that erases H(p) bits of in-formation. As we expect, the lower bound is equal tokTln 2H(p). Clearly, this derivation does not use thesecond law.

    It is clear in Shizumes derivation that the entropy in-crease due to the erasure is independent of the secondlaw. Hence it is immune to a common criticism againstthe erasure principle that it is trivially the same as thesecond law because the second law is used in its deriva-tion. However, the above description assumes only a spe-cific physical model and thus a more general model mightbe desirable. This was done by Piechocinska, who ana-lyzed the information erasure in a quantum setting aswell as in classical settings [21].

    The key idea in her results is to make use of a quan-tity , which was introduced by Jarzynski in the contextof nonequilibrium thermodynamic processes [22]. In aclassical setting, is defined by

    (0, ) = ln[f(x, p)] + ln[i(x0, p0)]+E(x0T, p

    0T, x

    T, p

    T), (8)

    where = (x, p, xT, pT) is a set of positions and mo-menta of the degrees of freedom that describe the (mem-ory) system and the heat bath (T), respectively. Thesuperscripts, 0 and , are the initial and final times ofthe erasure process. i andf are the distribution func-tions of the particle representing the bit in a double wellpotential, and are assumed to be in the canonical distri-bution: erasing information is expressed by the form offso that it takes nonzero values only in one of the tworegions, i.e. either of those for 0 and 1, which corre-sponds to the L state in Fig. 3. Eis the change in theinternal energy of the heat bath and = (kT)1.

    The entropy increase due to erasure can be obtainedby first calculating the statistical average over all possibletrajectories . We then havee = 1, which in turnimplies 0 by the convexity of the exponentialfunction. Substituting the expressions for i and f (thecanonical distributions) to leads to an inequality

    ln 2E. (9)

    As E is the change in the internal energy of the heatbath, it includes the heat dissipated into the bath as well.Thus the conservation of energy can be written as W =

    E+ Esystem, whereWis the work done on the systemand the heat bath, and Esystem is the change in theinternal energy of the system. Due to the symmetry ofiandf, Esystem vanishes when averaged over, thereforewe now have

    kTln 2 W, (10)

    which is equivalent to Landauers erasure principle.

    Piechocinska applied a similar argument to the quan-tum case, i.e. the erasure of classical information storedin quantum states. The state of the bit can be resetafter some interaction with the heat bath, which is ini-tially in thermal equilibrium. By assuming that the bathdecoheres into one of its energy eigenstates due to theinteraction with an external environment, which may bemuch larger than the bath, we can deal with the heat dis-sipation (into the bath) quantitatively. Then, the mini-mum work consumption can be found to be kTln 2 aftercomputing a quantity corresponding to in Eq. (8). Arelated, but more general, argument in a similar spirithas been also presented in [23].

  • 8/13/2019 0707.3400v2

    8/24

    8

    FIG. 5: The one-molecule heat engine considered by Gaborto show that the light needs to behave like wave. When themolecule comes into the illuminated region, a piston is auto-matically inserted and the gas expands isothermally to ex-tract work from the heat bath. This process could in principlebe repeated infinitely, converting infinite amount of heat intomechanical work, if the light had only a particle-like nature.(Adapted from Fig. 7 in Ref. [13].)

    V. SOME INTERESTING IMPLICATIONS OFTHE SECOND LAW

    A. Wave nature of light from the second law

    Let us make a detour to another interesting and well-elaborated implication of the second law, which wasargued by Gabor. He studied Brillouins analysis ofMaxwells demon about detecting a molecule by light sig-

    nals further[13]. Gabor considered a one-molecule heatengine, a part of which is illuminated by an incandescentlight to detect the molecule wandering into this region(Fig. 5). The detection of the molecule can be done byphotosensitive elements that are placed around the lightpath, so that any scattered weak light will hit one (ormore) of them. As soon as the molecule is found by de-tecting the scattered light, a piston is inserted at the edgeof the illuminated region. Then by isothermal expansionthe gas exerts mechanical work. The same process is re-peated when the molecule wanders into the illuminatedregion again. This is a perpetuum mobile of the secondkind as it continues to convert heat from a heat bath tomechanical work. Gabor found that the second law isvulnerable if the light intensity can be concentrated in awell defined region and made arbitrarily large comparedwith the background blackbody radiation. He then de-duced that this is impossible because light behaves bothas waves and as a flux of particles. In other words, ac-cording to Gabors argument, the second law implies thewave nature of light. This is an interesting implicationof the second law in its own right, as there seems to beno direct link between thermodynamics and the natureof light.

    However, we now know that this interpretation iswrong. Even if the light behaves like particles, Gaborsengine does not violate the second law. The solution tothis apparent paradox is also the erasure principle.

    By detecting the molecule and extracting work subse-quently, the whole system stores H(p) bits of informa-tion, wherep is the probability of finding the molecule inthe illuminated region. Let us assume for simplicity that

    the sampling frequency is low enough, compared withthe time duration necessary for the molecule to travelthrough the illuminated region[111]. Then,pcan also beinterpreted as the ratio between the volume of the illumi-nated region and the whole volume of the chamber. Thisinformation is about the occurrence of the work extrac-tion and is stored in the mechanism that resets the posi-tion of the piston after the extraction. The piston forgetsthe previous action, but the resetting mechanism doesnot. Thus the whole process is not totally cyclic, thoughit should be so to work as a perpetuum mobile. BecauseGabors engine is activated with probability p, the en-gine storesH(p) bits on average. While kTln 2H(p) bits

    of work are needed to erase this information to makethe whole process cyclic, we gain onlykT p lnpof work,which is smaller than the erasure work, from this process,hence there is no violation of the second law.

    B. Gibbs paradox and quantum superpositionprinciple

    Suppose a gas chamber of volume V that is dividedinto two half regions by a removable partition. Eachhalf region is filled with a dilute ideal gas at the samepressure P. We now consider the entropy increase that

    occurs when we let gases expand into the whole volume Vby removing the partition. If the gas in one region (e.g.,the left side) is different from the gas in the other (theright side), then the entropy increase due to the mixing iskNln 2, whereNis the total number of molecules in thechamber. On the other hand, if the gases in two regionsare identical, no thermodynamic change occurs and thusthe entropy is kept constant. This discontinuous gap ofthe entropy increase with respect to the similarity of twogases is called the Gibbs paradox. Lande dealt withthis problem and derived the wave nature of physicalstate as well as the superposition principle of quantummechanics [24]. This is not only an interesting work inthe sense that it attempted to link thermodynamics andquantum mechanics, but useful to introduce the idea ofsemipermeable membranes that we will use as a tool inlater sections. Although there are a number of paperson the Gibbs paradox other than Landes, we feel thatgoing into these is out of the scope of this brief review.Interested readers may refer to, for example, Refs. [25,26,27,28].

    For convenience of the following discussion, we use theextractable work, i.e. the Helmholtz free energy, as itis equal to the entropy change (times temperature) in

  • 8/13/2019 0707.3400v2

    9/24

    9

    isothermal processes. The semipermeable membranes weintroduce here are a sort of filters that distinguish theproperty of gases, i.e. the nature of molecules, and letone (or more) particular property of gas go through it.In other words, a semipermeable membrane is transpar-ent to one type of gas, but totally opaque to other typesof gases: Each membrane thus can be characterized withthe property of the gas it lets go through. These are es-

    sentially the same as what von Neumann considered inhis discussion to define the entropy of a quantum state, orvon Neumann entropy[112], and they were scrutinized byPeres and were shown to be legitimate quantum mechan-ically [29, 30]. In Landes argument, however, quantummechanics is not assumed from the outset.

    Lande postulated the continuity of the entropy changein reality. To bridge the gap between the same and dif-ferent gases, he introduced a fractional likeness, whichis quantified as q(Ai, Bj), between two states Ai andBj . Here A (or B) represents a certain property orobservable and the indices are values of A (B) withwhich we can distinguish them completely. For simplic-ity, we assume all observables can take only discrete val-ues when measured. Two states are completely differentwhen q = 0 and they are identical when q = 1, anddifferent values of the same property are perfectly distin-guishable by some physical means, thus q(Ai, Aj) = ij.Now suppose that a semipermeable membrane that isopaque to Ai but transparent to Aj(j=i), to which wewill refer as the membraneMAi, is placed in a gas whoseproperty isBk. Then the membrane will reflect a fractionq= q(Ai, Bk) of the gas and pass the remaining fraction1 qas a result of the fractional likeness between Ai andBk. Another consequence of the membrane is that themolecules that are reflected by MAi need to change theirproperty fromBk to Aiand similarly the other molecules

    becomeAj with probabilityq(Aj , Bk) [113], in order notto change the state of molecules by a subsequent appli-cation of another MAi. In what follows, we will identifythe term property with state, although it still does notnecessarily mean a quantum state.

    This solution to the Gibbs paradox the introductionof fractional likeness between states leads, accordingto Lande, to the wave-function-like description of state.A rough sketch of his idea is as follows. First, we writedown the transition probabilities between different statesin a matrix form

    q(A1, B1) q(A1, B2) q(A2, B1) q(A2, B2)

    . (11)

    Naturally, the sum of each row or column is always unitybecause a state must take one of the possible values inany measured property. Similar matrices should be ob-tained for q(B, C), q(A, C), etc., and we expect a math-ematical relation between these matrices, for instance,

    such as q(Ai, Ck) ?=

    jq(Ai, Bj)q(Bj , Ck). A consis-tent mathematical expression can be obtained by con-sidering a matrix (A, B) whose (i, j)-th elements are

    FIG. 6: A possible configuration to confirm the continuity ofthe extractable work. The left (right) hand side of a chamberis filled with an A1 (B1)-gas. Two membranes that distin-guish A1 and A2 are used to extract work. The membraneon the left lets the A1 gas pass it through freely, but reflectsthe A2 gas completely. The other membrane works in theopposite manner. Since theB1 gas changes its state into A1with probabilityp when measured by the membrane, the right

    membrane does not reach the right end of the chamber by a(quasi-static) isothermal expansion.

    given as

    q(Ai, Bj)ei with arbitrary phase , i.e. amatrix whose rows and columns can be regarded as avector of unit norm, thus for different pairs of proper-ties are connected with an orthogonal (unitary) transfor-mation. The arbitrariness for the phase is restricted bythe condition(Ai, Aj) =

    k(Ai, Bk)(Bk, Aj) = ij.

    Identifying (Ai, Bj) =

    qei with a complex proba-

    bility amplitude for the transition from Bj to Ai in-duced by the membranes, we see a superposition rule(Ai, Ck) =

    j(Ai, Bj)(Bj , Ck). Then Lande claims

    that the introduction of complex probability amplitudes subject to the superposition rule is inseparably linkedto the admission of fractional likenessesq.

    To confirm the continuity of the extractable work (orthe entropy increase) due to the mixing of two gases,let us look at a chamber, a half of which is filled withdilute gas A1 and the other half with B1 as in Fig. 6.The number of gas molecules is N/2 each. Let us alsoassume that both A and B are a two-valued property.If we use two membranes that distinguish the state A1andA2, the work by gases will be smaller than N kTln 2because a fraction of B1 becomes A1 with a certainprobability q. The work done by the gases is given asW = (NkT/2)[2ln 2 +qln q (1 + q) ln(1 +q)] and Wdecreases smoothly from N kTln 2 (whenq= 0, for per-fectly distinct gases) to 0 (when q= 1, identical gases),therefore no discontinuous entropy change. Note thatthis choice of membranes is not optimal to maximize theamount of extractable work and we will look at this pro-cess in more detail in Section VIIB.

  • 8/13/2019 0707.3400v2

    10/24

    10

    C. Quantum state discrimination and the secondlaw

    As we have seen in the previous section, in his at-tempt to solve the Gibbs paradox, Lande deduced thata thermodynamic speculation in the form of the continu-ity principle could lead to the partial likeness (or distin-guishability) of states as a result of the wave nature of

    particles. On the other hand, starting from the distin-guishability issue of quantum states, Peres showed thatif it was possible to distinguish non-orthogonal quantumstates perfectly then the second law of thermodynamicswould necessarily be violated[29,30].

    As a background, let us consider an elementary work-extraction process using a collection of pure orthogonalstates. As shown in Fig. 7, a chamber of volume V ispartitioned by a wall into two parts, one of which hasa volume p1V and the other has p2V, where p1 + p2 =1. The vessel is filled with a gas of molecules whose(quantum) internal degree of freedom is represented by,for example, a spin. Here it suffices to consider a gas of

    spin-1/2 molecules, e.g. a gas with spin up, i.e. | , inthe left region and a spin down gas,|, in the right.Now we reintroduce semipermeable membranes, M

    and M, that distinguish the two orthogonal spins,| and|. These are essentially the same as what we haveseen in SectionV Bto consider the fractional likeness ofstates. The membraneM is completely transparent tothe| gas and completely opaque to the| gas. Theother membraneM has the opposite property.

    Suppose that the partition separating two gases is re-placed by the membranes, so that M and M face| and| , respectively, as in Fig. 7. Then, as in Sec-tion V B, gases give us some work, expanding isother-mally by contact with a heat bath of temperature T.The total work extractable can then be computed asW = p1logp1 p2 logp2 = H(pi), where H(pi) =ipilogpiis the Shannon entropy of a probability dis-tribution{pi}.

    Let us look at Peress process. The physical systemwe consider now is almost the same as the one in Fig. 7,however, we now have two non-orthogonal states. Al-though gases consist of photons of different polarizationsin the original example by Peres, we consider spin-1/2molecules to avoid the argument of the particle natureof light. The volume of the chamber here is 2V, andin the initial state, the gas of volume V is divided intotwo equal volumes V /2 and separated by an impenetra-

    ble wall (See Fig. 8(a)). The gas molecules in the leftside have a spin up|, and those in the right side havea spin|= (| + | )/2. Both parts have the samenumber of molecules, N/2, thus the same pressure. Thefirst step is to let gases expand isothermally at tempera-ture Tso that the entire chamber will now be occupiedby them (Fig.8(b)). During this expansion, gases exert 1bit of work (=N kTln 2) towards the outside, absorbingthe same amount of heat from the heat bath.

    In the second step, we introduce fictitious magic

    FIG. 7: The work-extracting process with semipermeablemembranes. In the initial state (a), the vessel is divided intotwo parts by an impenetrable opaque partition. The left sideof the vessel, whose volume is p1V, is occupied by the | -gas, and the right side is filled with | -gas. By replacing thepartition with two semipermeable membranes, M and M,we can extract H(p1) =p1logp1p2logp2 bits of work byisothermal expansion. The membranes reach the end of thevessel in the final state (b).

    membranes that can distinguish non-orthogonal states.We replace the partition at the centre with these mem-branes and insert an impenetrable piston at the right endof the vessel. The membraneM, which is transparent to

    the|-gas but opaque to|-gas, is fixed at the centre,while the other oneM, which has the opposite property

    toM

    , can move in the area on the left. Then, the pistoninserted at the right end andM is pushed towards theleft at the same speed so that the volume and the pressureof the|-gas in between the piston and the membraneM will be kept constant (from Fig. 8(b) to (c)). Be-cause of the property of the membranes, this step can beachieved without friction/resistance, thus needs no workconsumption or heat transfer.

    The gas in the vessel in Fig. 8(c) is a mixture of twospin states. The density matrix for this mixture is

    = 1

    2| | +1

    2| |= 1

    4

    3 11 1

    (12)

    in the{| , | }-basis. The eigenvalues of are (1 +2/2)/2 = 0.854 and (1 2/2)/2 = 0.146 with cor-

    responding eigenvectors| = cos 8 |0+ sin 8 |1 and|= cos 38 |0 + sin 38 |1, respectively.

    Now let us replace the magic membranes by ordinaryones, which discriminate two orthogonal states,| and|. The reverse process of (b)(c) with these ordinarymembranes separates | and | to reach the state (d).Then, after replacing the semipermeable membranes byan impenetrable wall, we compress the gases on the left

  • 8/13/2019 0707.3400v2

    11/24

    11

    FIG. 8: A thermodynamic cycle given by Peres to show thatdistinguishing non-orthogonal quantum states leads to a vio-lation of the second law. The arrows indicate the directionsof spin in the Bloch sphere. The use of hypothetical semiper-meable membranes, which distinguish non-orthogonal states| and | perfectly, in the step from (b) to (c) is the key toviolate the second law. (Modified from Fig. 9.2 in Ref.[30].)

    and the right parts isothermally until the total volumeand the pressure of the gases become equal to the initialones, i.e. those in state (a). This compression requires awork investment of(0.854log0.854+0.146log0.146) =0.600 bits, which is dissipated into the heat bath. In order

    to return to the initial state (a) from (e), we rotate thedirection of spins so that the left half of the gas becomes| and the right half becomes|. More specifically, weinsert an opaque wall to the vessel to halve the volume Voccupied by gases (the border between regions labelledA and B in Fig. 8(e)). Rotations| | in theregion A,| | in B and| | in C,and a trivial spatial shift restore the initial state (a).As rotations here are unitary transformations, thus anisentropic process, any energy that has to be suppliedcan be reversibly recaptured. Alternatively, we can putthe system in an environment such that all these spinpure states are degenerate energy eigenstates. Hence, wedo not have to consider the work expenditure in principlewhen the process is isentropic.

    Throughout the process depicted above and in Fig.8,the net work gained is 1 0.600 = 0.400 bits. Therefore,Peress process can complete a cycle that can withdrawheat from a heat bath and convert it into mechanicalwork without leaving any other effect in the environment.This implies that the second law sets a barrier to quan-tum state discrimination.

    D. Linearity in quantum dynamics

    Peres also showed that the second law should be vio-lated if we admit nonlinear (time) evolution of the quan-tum states[31]. His proof is concise and is summarizedbelow.

    Let the state be a mixture of two pure states, = p

    |

    |+ (1

    p)

    |

    |, with 0 < p < 1. By

    rewriting one of the state vectors, say|, as| =f| + 1 f|, where| is a vector orthogonal

    to| and f =|||2, can be written in a matrixform (in the 2-dimensional subspace that supports) as

    =

    p + f(1 p)

    f(1 f)(1 p)

    f(1 f)(1 p) (1 p)(1 f)

    . (13)

    The von Neumann entropy can be computed as S() =+log + log , where are the eigenvalues of, i.e.

    =1

    21

    4p(1

    p)(1

    f)

    1

    2

    . (14)

    We can see dS/df < 0 for all p. Therefore, in order notto make the entropy decrease in time, the change of fmust be non-positive:|(t)|(t)|2 |(0)|(0)|2.

    Now let{|k}be a complete orthogonal set spanningthe whole Hilbert space. Then, for any pure state|,

    k |k ||2 = 1. Thus, if there is some m for which|m(t)|(t)|2 |(0)|(0)|2, which meansthat the entropy of a mixture of|nn|and|| willdecrease in a closed system. Hence, f =|||2 needsto be constant for any| and| to comply with thesecond law. There are still two possibilities for the timeevolution of states|(0) |(t) to keep f constant,namely unitary and antiunitary evolutions, according toWigners theorem[32]. Nevertheless, the latter possibil-ity can be excluded due to the continuity requirement.Therefore, the evolution of quantum states is unitary,which is linear.

    E. Second law and general relativity

    The second law of thermodynamics gives an interestingimplication not only in quantum mechanics, but also inthe theory of gravity, through the impossibility of thesecond kind of perpetuum mobile. This is illustrated byBondis thought experiment. Although an assumptionin the idea presented below seems quite infeasible, let ushave a look at it because it is a nice heuristic introductionto discuss real physics later. Imagine a vertically placedconveyor belt that has a number of single-atom holderson it as in Fig.9. Let us assume that the atoms on the leftside are in an excited state and those on the right side arein a lower energy state. When an excited atom reachesthe bottom of the belt, it emits a photon, lowering its

  • 8/13/2019 0707.3400v2

    12/24

    12

    FIG. 9: Bondis thought experiment for a perpetuum mobile.Excited atoms (red balls) emit a photon at the bottom ofthe belt to lower its energy level (blue balls). The emittedphoton is reflected by the curved mirrors placed so that itwill be absorbed by an atom in the lower energy level at thetop of the belt.

    energy level, and the emitted photon will be reflected bythe curved mirrors to be directed to the atom at the topof the belt. Then this atom at the top will be excited,absorbing the photon.

    As energy is equivalent to mass, according to specialrelativity, the atoms on the right side are always heavierthan those on the left as far as the emission and ab-sorption of photon work as described above. That is,the gravitational force will keep the belt rotating for-ever. In this scenario, however, there is another assump-tion, which seems implausible, that atoms emit/absorbphotons only at the bottom/top of the belt. Such an

    assumption makes this device unlike to work. Neverthe-less, Bondis perpetuum mobile is not compatible withthe physical laws for the following reason.

    What prevents this machine from perpetual motion isactually the distorted spacetime, i.e. a change of the met-ric, which is seen as gravity by us. The theory of generalrelativity tells that the space is more stretched if one goesfarther from the horizon: the length of the geodesic lineis longer near the horizon for a given length in the normalsense, which is defined as the circumference of the spherearound the massive object divided by 2. Because lighttravels along a geodesic, stationary observers see that thewavelength becomes longer when light leaves away fromthe object: the light becomes red-shifted.

    An experiment to confirm this red-shift was carriedout in 1960 by Pound and Rebka [33]. They made useof the Mossbauer effect of nuclear resonance, which canbe used to detect extremely small changes in frequency.Their results demonstrated that the photons do changethe frequency by a few parts of 1015 when they travel for22.5m vertically, which agreed Einsteins prediction witha high accuracy (only one percent error in the end [34]).The existence of the gravitational red-shift directly rulesout Bondis perpetuum mobile.

    F. Einstein equation from thermodynamics

    Einsteins equation, which describes the effect ofenergy-mass on the geometrical structure of the four di-mensional spacetime, can be derived from a fundamentalthermodynamic relation. In thermodynamics, knowingthe entropy of a system as a function of energy and vol-ume is enough to get the equation of state from the funda-

    mental relation,Q = T dS. Jacobson tried to obtain thefield equation as an equation of state, starting from ther-modynamic properties of black holes [35]. It had beenknown by then that there was a strong analogy betweenthe laws of black hole mechanics and thermodynamics[36]. That is, the horizon area of a black hole does notdecrease with time, just as the entropy in thermodynam-ics. Bekenstein then argued that the black hole entropyshould be proportional to its horizon area after introduc-ing the entropy as a measure of information about theblack hole interior which is inaccessible to an exteriorobserver [37,38].

    Bekensteins idea suggests that it is natural to regard

    the (causal) horizon as a diathermic wall that preventsan observer from obtaining information about the otherside of it. On the other hand, a uniformly accelerated ob-server sees a black body radiation of temperature T fromvacuum (the Unruh effect) [39,40]. The origin of the Un-ruh effect lies in the quantum fluctuation of the vacuum,which is also the origin of the entropy of the horizon, i.e.the correlation between both sides of the horizon. Thus,in order to start from the above relation,Q= T dS, Ja-cobson associatedQ and Twith the energy flow acrossthe causal horizon and the Unruh temperature seen bythe observer inside the horizon, respectively. Then, Ein-steins field equation can be obtained by expressing theenergy flow in terms of the energy-momentum tensorTand the (horizon) area variation in terms of the expan-sion of the horizon generators. Another essential elementin the field equation, the Ricci tensor R, appears in theform of the expansion through the Raychaudhuri equa-tion (for example, see Ref.[41]), which describes the rateof the volume (area) change of an object in a Rieman-nian manifold. The resulting equation is thus (by settingc= 1)

    R12

    Rg+ g=2k

    T, (15)

    where is the proportionality constant between the en-tropy and horizon area, viz. dS = d

    A, and R and

    are the scalar curvature and the cosmological constant.Comparing with the standard expression of the equation,in which the coefficient for T is 8G, we identify tobe k/(4G) =k/(4l2P) with the Planck length lP, whichagrees with the derivation of in Ref.[42].

    Einsteins field equation can indeed be seen as a ther-modynamic equation of state. An important assump-tion for the above derivation is, however, the existenceof a local equilibrium condition, for which the relationQ = T dS is valid. This means that it would not be

  • 8/13/2019 0707.3400v2

    13/24

    13

    appropriate to quantize the field equation as it is not ap-propriate to quantize the wave equation for sound propa-gation in air. Further, Jacobson speculated that the Ein-stein equation might not describe the gravitational fieldwith sufficiently high frequency or large amplitude distur-bances, because the local equilibrium conditions wouldbreak down in such situations as in the case of soundwaves.

    Bekensteins conjecture about the black hole entropy isnow widely accepted as a real physical property, partic-ularly after the discovery of the Hawking radiation [42]that showed that black holes do radiate particles in athermal distribution at finite temperature. The thermo-dynamics of black holes is still an extensive and activefield, whose famous problems include the black hole in-formation paradox. Covering these topics in detail gobeyond the scope of this brief overview, thus here wesimply list a few references [43,44,45,46,47,48].

    It is now clear that this example illustrates a close con-nection between information, thermodynamics, and thegeneral relativity, which might look unrelated with each

    other at first sight. This strongly re-suggests the dual-ity of entropy, which we have mentioned at the end ofSectionIIC,and the universality of the thermodynamicrelations in generic physics. We shall attempt to explorethis duality in the paradigm of quantum information the-ory later on in SectionVII B.

    VI. ERASURE OF CLASSICAL INFORMATIONENCODED IN QUANTUM STATES

    In classical information theory, an alphabet i {1,...,n}, which appears with probability pi in a mes-sage [114] generated by a source, is represented by oneof the n different classical states. On the other hand, inquantum information theory[49], information is encodedin quantum states, so that each alphabeti is representedby one of the n different quantum states whose densitymatrices are denoted by i. We will refer to each statecarrying an alphabet as a message state. We will alsocall a set of quantum states used in a message, in whichthe state i appears with probability pi, an ensemble ofquantum states{pi, i}, i {1,...,n}.

    A smart way to erase classical information encodedin quantum states was first considered by Lubkin [2],who introduced erasure by thermal randomization, andby Vedral [50, 51] in a more general setting. Thermalrandomization makes use of the randomness of states ina heat bath that is in thermal equilibrium. If we put amessage state in contact with a heat bath at temperatureT, the state will approach thermal equilibrium with theheat bath. More precisely, a message state i changesgradually after colliding (interacting) with the heat bathand sufficiently many collisions make the state to becomeindistinguishable with that of the heat bath. We assumethat the baths state as a whole will not change muchsince its size is very large. Due to the uncertainty stem-

    FIG. 10: The erasure of classical information carried by quan-tum states. Each message state interacts with a heat bath attemperatureTand reaches thermal equilibrium. The infor-mation originally encoded in a state is lost and all states endup in , which is the thermal state at the temperature T.

    ming from thermal fluctuations, we irreversibly lose theinformation that was carried by the statei.

    Because of the generic nature of this erasure processby thermalization, entropy of the whole system, consist-ing of the message state and the heat bath, necessar-ily increases. How much would this increase be? Letus first simplify the discussion by considering that eachmessage state is a pure state as in Fig. 10 [9, 50]. Be-fore erasing, the whole message is an ensemble{pi, |i},thus its average state is described by a density operator=

    ipi|ii|. The thermalization process brings all

    states|i to the same state , which is in thermal equi-librium at temperatureT. The density matrix is givenby

    = eH

    Z =j

    qj |ejej |, (16)

    where H =

    i ei |eiei| is the Hamiltonian of the mes-sage state with energy eigenstates|ei, Z= Tr(eH) isthe partition function, and= (kT)1.

    The total entropy change Serasure is the sum of theentropy change of the message system and that of theheat bath: Serasure = Ssys+ Sbath. Since the mes-sage state before the erasure is pure and its state afterthe erasure is the same as the heat bath, the minimumentropy change in the message state is given by[115]

    Ssys

    = k ln 2S(), (17)

    where S() = Tr( log ) is the von Neumann en-tropy of the state . Von Neumann introduced this en-tropy by contemplating the disorder of quantum statesso that it has the same meaning as the entropy in (phe-nomenological) thermodynamics in a setting where gasesof molecules with quantum properties were considered[52]. The factor k ln 2 is just a conversion factor to makeit consistent with the previous discussion of Landauersprinciple.

  • 8/13/2019 0707.3400v2

    14/24

    14

    The entropy change in the heat bath is equal to the av-erage heat transfer from the bath to the message systemdivided by the temperature: Sbath = Qbath/T. Theheat change in the heat bath is the same as that in thesystem with an opposite sign, i.e. Qbath =Qsys.The heat transfer can be done quasistatically so that themechanical work required for the state change is arbitrar-ily close to 0. Therefore, due to the energy conservation,

    Qsys must be equal to the change of internal energyof the message system Usys, which can be computed asthe change of average values of the Hamiltonian Hbeforeand after the erasure process. Hence,

    Sbath = QsysT

    =UsysT

    = Tr(H) Tr(H)T

    = Tr[( )H]T

    . (18)

    By using Eq. (16), the Hamiltonian Hcan be expressedin terms of the partition function Zas

    kTln(Z). Now

    we have

    Sbath = kTr [( )ln(Z)]= kTr [( ) ln ]= k ln2[S() + Tr( log )]. (19)

    Combining Eq. (17) and Eq. (19) gives the total entropychange after the erasure:

    Serasure = Ssys+ Sbath =Tr( log ), (20)where the unimportant conversion factor k ln 2 is set to beunity as a unit of entropy. The minimum of the entropy

    change Serasure can be obtained as:

    Serasure =Tr( log )S(). (21)The inequality follows from the property of the quan-tum relative entropy,S(||) :=S()Tr( log )0.This minimum can be achieved by choosing the temper-ature of the heat bath and the set {pi, |i} such that =

    pi|ii| is the same as the thermal equilibrium

    state . Consequently, the minimum entropy increaserequired for the erasure of classical information encodedin quantum states is given by the von Neumann entropyS(), where is the average state of the system, insteadof the Shannon entropyH(p) in the case of erasing infor-

    mation in classical states.

    VII. THERMODYNAMIC DERIVATION OFTHE HOLEVO BOUND

    A. from the erasure principle

    Landauers erasure principle, together with itsLubkins version for quantum states, is simple in form;

    however, it implies some significant results in the theoryof quantum information. For example, it can be used toderive the efficiency of the compression of data carried byquantum states[9] and also the upper bound on the ef-ficiency of the entanglement distillation process[50,53].Here, we look at the derivation of the Holevo bound fromLandauers principle, which was first discussed by Plenio[9,54], as we will examine the same problem from a dif-

    ferent perspective in the next section.To give a precise form of the Holevo bound let us con-sider two parties, Alice and Bob. Suppose Alice has aclassical information source preparing symbols i= 1,...,nwith probabilities p1,...,pn. The aim for Bob is to de-termine the actual preparation i as best as he can. Toachieve this goal, Alice prepares a state i with prob-ability pi and gives the state to Bob, who makes ageneral quantum measurement (Positive Operator Val-ued Measure or POVM) with elements Ej =E1,...,Em,m

    j=1 Ej =l, on that state. On the basis of the measure-ment result he makes the best guess of Alices prepara-tion. The Holevo bound [55] is an upper bound on theaccessible information, i.e.

    I(A: B)S() i

    piS(i), (22)

    where I(A : B) is the mutual information between theset of Alices preparationsi and Bobs measurement out-comesj, and =

    ni pii. The equality is achieved when

    all density matrices commute, namely [i, j ] = 0.Let us first consider a simple case in which all i

    are pure: i = |ii|. The average state will be =

    ipi|ii|. Then, the Shannon entropy of the

    message is always greater than or equal to the von Neu-mann entropy of the encoded quantum state (Theorem

    11.10 in Ref. [49]), that is, H(pi) S(), with equalityif and only ifi|j= ij for all i and j .

    How much information can Bob retrieve from the state? The above analysis of erasure of information in quan-tum states tells us that we have to invest at least S() bitsof entropy to destroy all available information. This im-plies that the amount of information that Bob can accessis bounded by S(), because if he could obtain S() + bits of information then the minimum entropy increaseby the erasure should be at least S() + , by Landauersprinciple. In other words, one cannot obtain more infor-mation than it could be erasable.

    Therefore, the accessible information I(A : B) is

    smaller than or equal to the minimum entropy increaseby the erasure and it is this relation that corresponds tothe inequality (22). If Alice encodes her information iinto a pure state i, the relation reads

    I(A: B)S(), (23)which is the same inequality as Eq. (22) becauseS(i) =0 for all pure states i.

    If Alice uses mixed statesito encodei, Eq. (21) needsto be modified. Instead of Eq. (17), the entropy increase

  • 8/13/2019 0707.3400v2

    15/24

    15

    for each state is now [116] Sjsys = S()S(j) aftercontact with a heat bath whose state is given by Eq. (16).The average entropy change of the heat bath is the sameas Eq. (19): Sbath =S() log . The total entropychange by the thermalization will be

    Serasure =j

    pjSjsys+ Sbath

    =j

    pj(S() S(j)) S() Tr( log )

    = j

    pjS(j) Tr( log )

    S() j

    pjS(j), (24)

    which, together with the above argument, implies theHolevo bound in the form of Eq. (22). This analysisof erasure of information encoded with mixed states ismore straightforward and less ambiguous than that inRefs.[9,54].

    The above analysis thus justifies the Holevo bound.However, it does not give the precise condition for theequality in Eq. (22), which is [i, j] = 0. The condi-tion we can derive here is that all density matrices{i}support orthogonal subspaces, i.e. Tr(ij) = 0. This ismore restrictive than the commutativity of density ma-trices mentioned above. In the next section, we will see ifthe second law implies the Holevo bound more directly.

    B. from the second law

    Since the work-informationdualityin the erasure prin-

    ciple supports Brillouins hypothesis, which we have seenin SectionII C, on the equivalence between informationtheoretic and thermodynamic entropies, it might be nat-ural to expect that the second law may put a certainbound on the quality of information or the performanceof information processing. In this section, we derive thegeneral bound on storage of quantum information, theHolevo bound[55] derived from the second law of ther-modynamics. As the second law is the most fundamen-tal physical law that governs the behavior of entropy, thisproblem is interesting in terms of the spirit of the physicsof information and deserves to be investigated in its ownright.

    In order to see the genuine thermodynamic bound, weneed to minimize the axiomatic assumptions that stemfrom quantum mechanics. Assumptions we make hereare, (a) Entropy: the von Neumann entropy is equivalentto the thermodynamic entropy, (b) Statics and measure-ment: a physical state is described by a density matrix,and the state after a measurement is a new state thatcorresponds to the outcome (projection postulate), (c)Dynamics: there exist isentropic transformations.

    Employing the density matrix-based description meansthat we presume the existence of superpositions of states.

    Allowing superpositions might sound rather abrupt; how-ever, we can assume that we are taking a similar standas Gabors picture on the possibility of superpositionspurely from thermodynamic considerations. Thus, thiscould be stated in the other way around: assuming thesuperpositions of states, we can describe a state by adensity matrix, which can be defined as a convex sumof outer products of normalized state vectors. The

    nonzero components of state vectors represent super-posed state elements. Probability distributions in clas-sical phase space can also be described consistently: alldiagonal elements of a classical density matrix are real,representing probabilities, and all off-diagonal elementsare zero. When a measurement is performed, one of thediagonal elements becomes 1, replacing all others with 0.

    We will use an arrow to denote a state vector, such as ,to make it clear that we do not use the full machineryof the Hilbert space (such as the notion of inner prod-uct) and we never use the Born trace rule for calculatingprobabilities.

    Consider a chamber of volume V, which is divided into

    two regions of volumes p1V and p2V, respectively. Theleft-side region (L) is filled with p1Nmolecules in state1, and p2N molecules in state 2 are located in the

    right-side region (R). The two states, 1 and 2, can bethought of a representation of an internal degree of free-dom. Generalizations to arbitrary numbers of general(mixed) states and general measurements are straight-forward.

    We can now have a thermodynamic loop formed by twodifferent paths between the above initial thermodynamicstate to the final state (Fig.11). In the final state, both

    constituents, 1 and 2, are distributed uniformly overthe whole volume. Hence, each molecule in the final state

    can be described by =

    ipi i i , regardless of theposition in the chamber. One of the paths converts heatinto work, while the other path, consisting of a quasi-static reversible process and isentropic transformations,requires some work consumption.

    In the work-extracting process, we make use of twosemipermeable membranes, M1 and M2, which separatetwo perfectly distinguishable (orthogonal) states e1 ande2 (=e1). The membraneMi (i = 1, 2) acts as a com-pletely opaque wall to molecules inei, but it is transpar-

    ent to molecules inej=i. Thus, for example, a state i isreflected byM1to become e1with (conditional) probabil-ityp(e1

    |i) and goes through with probability p(e2

    |i),

    being projected onto e2. This corresponds to the quan-tum (projective) measurement on molecules in the basis{e1, e2}, however, we do not compute these probabilitiesspecifically as stated above.

    By replacing the impenetrable partition with the twomembranes, we can convert heat from the heat bath intomechanical work Wext, which can be as large as the ac-cessible information I(A : B), i.e. the amount of in-formation Bob can obtain about Alices preparation bymeasurement in the basis{e1, e2}[117]. The transforma-

  • 8/13/2019 0707.3400v2

    16/24

    16

    FIG. 11: The thermodynamic cycle, which we use to discussthe second law. The cycle proceeds from the initial state (a)to the final state (c) via the post-work-extraction state (b), and returns to the initial state with a reversible process.

    tion from the post-work-extraction state, which we call

    hereafter, to the final state can be done by a processshown in Fig.12 and the minimum work needed is givenby S= S() S().

    Another path, which is reversible, from the initial state

    to the final state is as follows. Let{1, 2} be an or-thonormal basis which diagonalizes the density matrix,such that

    =i

    pi i i =

    k

    kk k, (25)

    wherek are eigenvalues of. We can extract S() bits of

    work by first transforming

    {1,2

    }to

    {1, 2

    }unitarily,

    and second using a new set of semipermeable membranesthat perfectly distinguish 1 and 2.

    If the initial state is a combination of mixed states withcorresponding weights given by{pi, i}, the extractablework during the transformation to =

    ipii becomes

    S()ipiS(i). This can be seen by considering aprocess

    {pi, i} (i) {piij , ij} (ii) {k, k} (iii),

    where{ij, ij} and{k, k} are the sets of eigenvaluesand eigenvectors ofi and, respectively[56].

    The function of the semipermeable membrane can al-ternatively be understood as a Maxwells demon whocontrols small doors on a partition depending on the re-sult of his measurement of each molecule. Then do weneed to consume some work to reset his memory? Unlikethe previous discussions (such as that of Szilards engine),it turns out that the demons memory can be erased isen-tropically due to the remaining (perfect) correlation be-tween the state of each molecule and his memory regis-ters. This can be sketched as follows. Once the demonobserves a molecule, a correlation between the state of

    FIG. 12: The thermodynamic process to transform the inter-

    mediate state into the final state . Firstly, after attachingan empty vessel of the same volume to that containing thegas , the membranes Mj are used to separate two orthog-onal states e1 and e2 ((a) to (c)). As the distance betweenthe movable opaque wall and the membrane M2 is kept con-stant, this process entails no work consumption/extraction.

    As =P

    cjejej , compressing each ej-gas into the volume

    ofcjVas in (d) makes the pressures of gases equal and thiscompression requiresS() =

    Pcjlog2 cj bits of work. Sec-

    ond, quantum states of gases are isentropically transformed,thus without consuming work, so that the resulting state (e)

    will have jNmolecules in j, where =P

    j j j is the

    eigendecomposition of. To reach (f), S() bits of work can

    be extracted by using membranes that distinguish j. As

    a result, the work needed for the transformation isS() S() bits.

    the molecule and his memory is created. Since he can inprinciple keep track of all molecules, a perfect correlationbetween the state of the n-th molecule and that of the n-th register of demons memory will be maintained. Thena controlled-NOT-like isentropic operation between themolecules and the corresponding memory registers (withmolecules as a control bit) can reset the demons memoryto a standard initial state without consuming work.

    The second law states (in Kelvins form) that the network extractable from a heat bath cannot be positiveafter completing a cycle, i.e. Wext Winv 0. For thecycle described in Fig.11, it can be expressed as

    I(A: B)S() i

    piS(i) + S, (26)

    where S= S() S(). As is identical to the result-ing state of a projective measurement on in the basis

    {e1, e2}, =

    jPjPj withPj = ejej and consequently

  • 8/13/2019 0707.3400v2

    17/24

    17

    Sis always non-negative (See Ref. [52]). The inequality(26) holds even if the measurement by membranes was ageneralized (POVM) measurement[56].

    The form of Eq. (26) is identical with that of Eq. (22)for the Holevo bound, except an extra non-negative term,S. This illustrates that there is a difference betweenthe bound imposed by quantum mechanics (the Holevobound) and the one imposed by the second law of ther-

    modynamics. Namely, there is a region in which wecould violate quantum mechanics while complying withthe thermodynamical law. In the classical limit, themeasurement is performed in the joint eigenbasis of mu-tually commuting is, consequently S = 0, and, inaddition, the Holevo bound is saturated: I(A : B) =S()ipiS(i). Thus, the classical limit and the ther-modynamic treatment give the same bound.

    The same saturation occurs when an appropriatecollective measurement is performed on blocks of mmolecules, each of which is taken from an ensemble{pi, i}. When m tends to infinity 2m(S()

    Pi piS(i))

    typical sequences (the sequences in which i appears

    aboutpim times) become mutually orthogonal and canbe distinguished by square-root or pretty good mea-surements [57,58]. This situation is thus essentially clas-sical, hence, S0 and the Holevo bound will be sat-urated.

    VIII. ENTANGLEMENT DETECTION BYMAXWELLS DEMON(S)

    Now let us move on to see how we can deal with entan-glement from the point of view of Maxwells demon. Thereason we pick up the topic of entanglement in particu-

    lar is that it is not only crucially important in quantuminformation theoretic tasks, such as quantum cryptogra-phy [59, 60], dense coding [61], quantum teleportation[62] and quantum computation[63,64], but it is also di-rectly linked to the foundations of quantum mechanics.However, the theory of entanglement is too broad anddeep to explore comprehensively in this article. In addi-tion, the most of it appears irrelevant or at least unclearin the context of Maxwells demon at any rate. Thereforewe focus on only some recent work that have discussedthe quantumness of correlation and/or the problem ofentanglement detection, which is one of the most impor-tant topics in its own right.

    When we discuss entanglement in this article we areprimarily interested in bipartite entanglement, unlessotherwise stated. Formally, entanglement is defined asa form of quantum correlation that is not present in anyseparable states. LetHA andHB be the Hilbert spacesfor two spatially separated (non-interacting) subsystemsA and B, which are typically referred to as Alice and Bob,andHAB =HA HB the whole (joint) Hilbert space ofthe two. We also letS(H) denote the state space, whichis a set of density operators acting onH.

    A state of a bipartite system is said to be separable or

    classically correlated [65] if its density operator can bewritten as a convex sum of products of density operators

    =n

    i=1

    piAi Bi , (27)

    where allpi are nonnegative andn

    i=1pi= 1. Any statethat cannot be written in the form of Eq. (27) is called

    entangled. We will letSsep denote the subspace that con-tains all separable states.

    It is natural to ask whether or not a given state S(HAB) is separable, considering the importanceof entanglement in many quantum information theoretictasks [118]. Quite a few separability criteria, i.e. a con-dition that is satisfied by all separable states, but notnecessarily by entangled states, have been proposed sofar to answer this simple question. Separability crite-ria are typically expressed in terms of an operator ora function, such as an entanglement witness [66] or thecorrelation function in Bells inequality[67]. Despite thesimplicity of the question, it is generally very hard to

    find good separability criteria. By a good separabilitycriterion, we mean an efficient separability criterion thatsingles out as many entangled states as possible. Thehardness of the problem is primarily related to the con-vexity of the separable subspace, which is formed by allthe separable states: Because of the bulgy surface ofthe separable subspace, there does not exist any oper-ator/function that is linear with respect to the matrixelements of density operator, e.g. eigenvalues of, anddistinguishes separable and entangled states perfectly.

    Another simple question is about the amountof entan-glement a pair (or a set) of quantum objects contains. Itplays a major role when it comes to the characterizationor manipulation of entanglement. Since entanglementcan be regarded as a valuable resource in quantum in-formation processing, the quantification of entanglementis a problem of great interest and importance. Despiteits profoundness, we will not go into details on the quan-tification issue here: Instead, we refer interested readersto some references, such as Refs.[68,69,70,71,72,73].Also, Ref. [51] contains not only a review on entangle-ment measures, but also some discussions on quantuminformation processing from the thermodynamic point ofview.

    A. Work deficit

    In this subsection, we will review the concept ofworkdeficit, which was introduced by Oppenheim et al. [74].An apparent goal of this work was to quantify entan-glement via a thermodynamic quantity; this idea shed anew light on the quantumness of correlations by takinga thermodynamic approach.

    As we have emphasized, information is always storedin a physical system with physical states that are distin-guishable by measurement so that stored information can

  • 8/13/2019 0707.3400v2

    18/24

    18

    be extracted. No generality is lost when we think of agas in a chamber, such as the one considered by Szilard,as a general information-storage apparatus. Even if wehad a different type of physical system for informationstorage, the information can be perfectly transferred forfree to the memory of the type of Szilards engine, if theinitial state of Szilards engine is provided in a standardstate. Since the measurement can be done with negligible

    energy consumption (as we have seen in SectionIII), theinformation transfer can be completed by converting thestate from the initial standard state to the state corre-sponding to the stored (measured) information; the finalconversion requires no energy.

    Now that we have identified a memory with the onemolecule gas of Szilards engine we can present a gen-eral statement: from an ensemble of memories, each ofwhich stores the value of an n-bit random variable X,one can extract mechanical work whose average amountper single memory register is (by taking units such thatk ln2 = 1)

    WC=n H(X), (28)where H(X) is the Shannon entropy of X. The ex-tractable work is the work done by the gas for memory,thus it is nothing but Eq. (4) whenn = 2. Equation (28)can be easily understood in the following way. Supposethere areNmemory registers. If we measure all N regis-ters the remaining uncertainty in the memory is zero; wecan obtain N n bits of work. Nevertheless, we still keepthe information due to the measurement on memory andthis needs to be erased to discuss solely the amount ofextractable work. The minimum energy consumption toerase the information is, according to the erasure prin-ciple, equal to N H(X) bits. Thus the maximum totalamount of extractable work is given by N(nH(X))bits. Alternatively, one can use the first law of thermo-dynamics to arrive at the same expression as Eq. (28).The work done by the gas in an isothermal process isequal to the entropy change multiplied by the tempera-ture.

    The same argument is applicable to work extractionfrom quantum bits (qubits). Let be the density opera-tor describing the state in a given ensemble. Qubits after(non-collective) measurements are in a known pure state,which is essentially a classical system in terms of infor-mation. Thus the information stored in this set of purestates can be copied to the Szilard-type memory and each

    register can give us 1 bit of work. Then, after erasing theinformation acquired by the measurement, the net max-imum amount of work we get becomes 1 S() bits ofwork.

    The work deficit is a difference between the globallyand the locally extractable work within the frameworkof LOCC, i.e. local operations and classical communi-cation, when is a system with spatially separated sub-systems. Suppose that we have an n qubit state AB,which is shared by Alice and Bob, then the optimal work

    extractable is

    Wglobal= n S(AB), (29)if one can access the entire system globally. On the otherhand, we let Wlocal be the largest amount of work thatAlice and Bob can locally extract from the same systemunder LOCC. The deficit is defined as

    =Wglobal Wlocal (30)In order to grasp this picture, let us compute the

    deficits for a classically correlated state

    ABcl =1

    2(|0000| + |1111|) (31)

    and a maximally entangled state

    |AB= 12

    (|00 + |11). (32)

    The globally extractable workWclglobalfromABcl is simply

    1 bit. The locally extractable workWcllocalturns out to bealso 1 bit. The protocol is as follows. Alice can mea