˚ berg1*, Renato Renner1, Oscar Dahlsten1,2,3 & Vlatko Vedral2,3,4 ´dia del Rio1*, Johan A Lı


The thermodynamic meaning of negative entropy
The heat generated by computations is not only an obstacle to circuit miniaturization but also a fundamental aspect of the relationship between information theory and thermodynamics. In principle, reversible operations may be performed at no energy cost; given that irreversible computations can always be decomposed into reversible operations followed by the erasure of data1,2, the problem of calculating their energy cost is reduced to the study of erasure. Landauer’s principle states that the erasure of data stored in a system has an inherent work cost and therefore dissipates heat3–8. However, this consideration assumes that the information about the system to be erased is classical, and does not extend to the general case where an observer may have quantum information about the system to be erased, for instance by means of a quantum memory entangled with the system. Here we show that the standard formulation and implications of Landauer’s principle are no longer valid in the presence of quantum information. Our main result is that the work cost of erasure is determined by the entropy of the system, conditioned on the quantum information an observer has about it. In other words, the more an observer knows about the system, the less it costs to erase it. This result gives a direct thermodynamic significance to conditional entropies, originally introduced in information theory. Furthermore, it provides new bounds on the heat generation of computations: because conditional entropies can become negative in the quantum case, an observer who is strongly correlated with a system may gain work while erasing it, thereby cooling the environment. ‘Erasure’ of a system is defined as taking it to a pre-defined pure state, j0æ (a familiar example is disk formatting, where a sequence of zero bits is written onto the disk). Landauer’s principle asserts that the energy dissipated to erase a system, S, using an optimal process in an environment of temperature T is given by W (S)~H (S)kT ln(2)

Our contribution is to go beyond this classical picture and consider observers who may have access to information that is represented as the state of a quantum system: a quantum memory. As an example, we could imagine a third observer, Quasimodo, who has a memory that includes n qubits, each maximally entangled with a particle of S. Our main result characterizes the work, W(SjQ), that an observer with access to a quantum memory, Q, needs to perform to erase system S. For simplicity, we formulate it here for a special case, which could be referred to as a ‘thermodynamic limit’ of erasure, where the observer erases many identical copies of S jointly. In this case, we show that there exists an erasure process whose work cost does not exceed W (SjQ)~H (SjQ)kT ln(2) ð3Þ per copy of S. Here H(SjQ) is the conditional von Neumann entropy, H(SjQ) 5 H(SQ) 2 H(Q). We show that this work cost is optimal, under the assumption that Landauer’s principle holds for a classical observer (Methods Summary). In its general form, our main result (Theorem 1 in Methods Summary) includes the more natural case where a single set of data, rather than a collection of identical copies, is to be erased. Crucially, we require that the information stored in Q be preserved by the erasure process—a non-trivial condition, given that accessing a quantum memory can in principle change it. This informationpreservation property is particularly important if we consider the erasure process to be part of a larger procedure (see, for instance, Fig. 1). For example, suppose that we erase system S and later would like to erase another system, Z. If the erasure of S removed the information about Z, erasing Z could become unnecessarily costly. More generally, we can consider a reference system, R, that models all other systems about which the memory can have information (technically, R is a purification of Q and S). The information-preservation condition can then be formulated as the requirement that the joint state of the memory and the reference system, rQR, be preserved by the erasure process (see also Fig. 2). The generalization of erasure to the quantum case exposes features not present in a classical setting. In particular, equation (3) implies that the work required for erasure may be negative for an observer with a quantum memory: the process results in a net gain of work. This happens because the uncertainty H(SjQ) can become negative for quantum observers. For instance, Quasimodo’s conditional von Neumann entropy is H(SjQ) 5 0 2 n (because the joint state of S and Q is pure and the reduced state of the memory, Q, is fully mixed). Our result provides a thermodynamic operational meaning for negative conditional entropies, which until now only had information-theoretical interpretations; for instance, they measure the entanglement needed to send a state to a receiver with side information14 (state merging), and quantify ‘violations’ of Heisenberg’s uncertainty principle15. The proof of equation (3) uses a probabilistic method to find appropriate erasure procedures. In the simple examples of our three observers, Alice, Bob and Quasimodo, we can describe them explicitly. Alice, who holds a classical description of the pure state of S, has no uncertainty about the system: H(SjA) 5 0. As expected, she does not need to


where k is the Boltzmann constant . Here the von Neumann entropy, H(S) 5 2Tr[rlog2(r)], quantifies the uncertainty about system S, whose state is described by r. Different observers may have different knowledge about the same system. For instance, an observer, Alice, can prepare n quantum bits (qubits; for example n spin-1/2 particles) in a pure state of her choice, keeping a record of that state. However, from the point of view of another observer, Bob, who does not have access to Alice’s record, the state of the system is completely unknown: he describes it as a fully mixed state, of maximal entropy. Hence, rather than W(S) being defined as the ‘cost of erasure’, it may be described as the ‘cost of erasure for observer C’, which we denote by W(SjC). This leads to the following reformulation of equation (1), where H(SjC) denotes the uncertainty that observer C has about system S: W (SjC )~H (SjC )kT ln(2) ð2Þ Typically, the observer C is assumed to be classical. More explicitly, classical observers represent their information about S in a memory that consists of classical bits (as in the case of Alice and Bob above).

Institute for Theoretical Physics, ETH Zurich, 8093 Zurich, Switzerland. 2Clarendon Laboratory, University of Oxford, Oxford OX1 3PU, UK. 3Centre for Quantum Technologies, National University of Singapore, 117543 Singapore. 4Department of Physics, National University of Singapore, 117542 Singapore. *These authors contributed equally to this work. 2 J U N E 2 0 1 1 | VO L 4 7 4 | N AT U R E | 6 1

©2011 Macmillan Publishers Limited. All rights reserved

however. 4). do any work to erase S: she may consult her record to check the state of S and change it to j0æ with a reversible transformation (Fig. x. If the erasure can be performed in a computationally efficient manner. He can perform a simple erasure process10. c. In addition. Bob raises all other levels in small steps. independently of the exact form of the correlations between S and Q. and a filled circle means that the system can be found in that state with certainty. it is possible to find a subsystem of S and Q that is (approximately) in a pure state. Given statistical knowledge about the properties of f. the rest of his memory. the work gained depends on the entropy of S conditioned on the information held by the quantum memory (see Supplementary Information. we show that. Bob decouples S from the bath and lowers all states back to the original level. has no access to Alice’s record and. and its reverse. one of the major obstacles to the miniaturization of circuitry. is correlated with a reference system. final ρX |0〉 |0〉 Figure 1 | Erasure in quantum computation. with the following options. (3a) Usually F is erased at the end of the algorithm. This pure subsystem furthermore has a special structure that allows us to extract work from it. we obtain a work-extraction procedure. The observer may store energy in and withdraw it from a battery. Uf. (2b) Instead. correlations between S and the memory are not as neat as for Quasimodo. for a formal proof). that swaps that state with | 0æ. creating an entangled state between two registers. By running this process in reverse. this special example contains the essence of the general case. section I.RESEARCH LETTER X (1) Uf (2a) UX ⊗ 1F Measure X (3a) Erase F Σ |x〉 |f(x)〉 x Σ |x〉 |0〉 x (2b) Erase F F a E E0 0 b E E0 0 U final ρX (3b) UX ⊗ 1 Measure X final ρX |0〉 Figure 3 | Erasure of a pure state. b. because the original state of Q1S had a fully mixed marginal in Q1—the informationpreservation condition is therefore satisfied. at the cost of increasing the work performed by an additive term (Theorem 1 in Methods Summary.11 that lets the system interact with a heat bath at temperature T (Fig. and it may not be clear why such idealizations are interesting. Using decoupling results16. Quasimodo combines two elementary procedures: the erasure process used by Bob. gains energy DE. in one single step. In its general single-shot form. it is possible to erase F while it is still partly entangled with X. Q. (3b) The information-preservation condition ensures that the rest of the algorithm is not affected by the early erasure. Q2. a ‘work-extraction’ process. we consider the period-finding algorithm for a function f. By raising or lowering an occupied level by DE. 4). used in the quantum part of Shor’s factoring algorithm. Q Reference 0 0 T Figure 2 | Erasure setting. If the Hamiltonian of S is degenerate. using the correlations (in blue) to decrease the work cost of the erasure. has maximal uncertainty: H(SjB) 5 n. Alice knows that the system is in a particular pure state. Initially. As an analogy we can think of the Carnot cycle. All rights reserved . 5). at the cost of leaving the state maximally mixed. X and F (this is done with a unitary transformation. which costs him nkTln(2) in energy. thus replacing it with a maximally mixed state. Such arguments are particularly relevant to the study of heat generation in computation. (1) Initially the algorithm evaluates f on a superposition of all inputs. leaving S and Q1 maximally mixed. this operation has no energy cost. The size of the pure state found and. and (2) raising or lowering any energy levels of their Hamiltonian (for example by tuning a magnetic field). the observer uses or. Computation can in principle be made reversible. empty levels can be changed at no energy cost. He starts by performing the elementary work extraction on the 2n-qubit state jSQ1æ to gain energy 2nkTln(2). without changing the state of the memory and the reference (Fig. The work cost of the whole process in the quasistatic limit is kTln(2) per qubit. so. S. a. Then he performs the elementary erasure process on S. here represented by a machine with a quantum memory. ©2011 Macmillan Publishers Limited. the erasure procedure we introduce has a failure probability that can be made arbitrarily small. letting S thermalize after each step. Keeping one energy level untouched. non-deterministic work extraction has been discussed before for classical observers17). In total. The erasure processes contemplated in this work require considerable control over the quantum systems involved. section V. our investigation bounds the ideal minimum work cost for the implementation of irreversible processes. whereby he transforms an initially pure state into a maximally mixed state. b. see Supplementary Information. As an example. respectively. b. it nevertheless provides the theoretical foundation for the behaviour of real heat engines. The final state of Quasimodo’s memory and the reference is rQR ~2{n 1Q1 6jQ2 RihQ2 Rj. it is possible to find highly correlated subsystems of X and F (in blue). on the two registers). (2a) The second part of the algorithm consists only of local operations on X (a unitary transformation. the bath thermalizes S. his energy gain is nkTln(2). We use a simple model for erasure. Although the ideal performance of a Carnot engine may be an unattainable limit in practice. it may be incorporated in the algorithm. Nevertheless. This is precisely the initial state of QR. In general. in state jQ2Ræ. We assume that the initial Hamiltonian of S and Q is fully degenerate (for example like that of paramagnets in a zero magnetic field). R. recall that his memory contains n qubits maximally entangled with S. 6 2 | N AT U R E | VO L 4 7 4 | 2 J U N E 2 0 1 1 Heat bath Heat bath Heat bath Figure 4 | Erasure of a fully mixed state and work extraction. The observer may couple S to a heat bath at temperature T. U. when correlations between X and F have been destroyed. Bob. which given an initial pure state yields an energy gain of kTln(2) per qubit. a. In a similar spirit. All memory contents about a reference system must be preserved. by (1) applying unitary operations to those systems. matching the prediction of equation (3). 3). a. Turning now to Quasimodo. therefore. a b nkTln(2) T a S c Battery c Heat bath b E0 E0 0 T T E0 Memory. The observer may manipulate S and Q. erases a system. the final state of X is rf X ). The work cost of this process is nkTln(2). S is in a fully mixed state. c. leaving it in a Gibbs state. where 22n1Q1 is the fully mixed state on Q1. gaining energy kTln(2) per qubit (Fig. followed by a inal measurement. The circles represent the energy eigenstates of system S. UX. We call this part of his memory Q1 and denote the entangled state jSQ1æ. For a concrete example and further discussion. As these levels are raised it is more likely to find S in the untouched level. To erase S. An observer. She performs a unitary transformation.

Dyn. 1. F. Develop. T. L. Bennett. R. & Renner. Quantum discord: a measure of the quantumness of correlations. 17. & Horodecki. However. 4674–4681 (2010). asserts that the smooth Supplementary Information is linked to the online version of the paper at www. 4. we can show that the quantum bound is tight in this limit (Supplementary Information. The physics of Maxwell’s demon and information. ETH Zurich (2005).. 14. Dissipation and heat generation in the computing process. H. 21. A. P. Piani. Computing (Taylor and Francis.A. J. Theor. P. whose work cost satisfies e (SjQ)zDŠkT ln(2) W (SjQ)ƒ½Hmax pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi except with probability less than d~ 2{D=2 z12e for all D $ 0 and all e $ 0. Our main result is an upper bound for W(SjQ) that is violated only with a very small probability. de Montre Dahlsten. For instance. The work cost of erasure. U. Son. Univ.R. Author Contributions The main ideas were developed by all authors. & Preskill. Mod. and J. 19. Thermodynamics of quantum information systems — Hamiltonian description. M. Duality between smooth min. such that log2( | S1 | ) < [log2( | S | ) 2 H(S | Q)]/2. H. grant no.. Phys. A.). W. Einselection and decoherence from an information theory perspective.and maxentropies. Theory 55. Oppenheim. Int. we find that there exists an erasure  (SjQ)#H (SjQ)kT ln(2). A 223. but at the expense of keeping extra data in a memory . Piechocinska. Definition I. Phil. & Winter. A. 062314 (2000).d. Naively. 659–662 (2010). The total work cost of the process is approximately H(S | Q)kTln(2). 13. IBM J. Partial quantum information. The state of S1 6P is maximally entangled. Rev. we have used information-theoretical tools. T. E.R. Black holes as mirrors: quantum information in random subsystems.d. Renner. 012320 (2003).R. we can infer that the conditional entropy H(SjQ) cannot decrease under local operations on Q. M. M. Theor. Horodecki. Inf... Using our results. Phys. The size of S1 is limited by the correlations between S and the reference. & Horodecki. Notes on Landauer’s principle. M. (eds Vitter. The physics of forgetting: Landauer’s erasure principle and information theory. 853–862 (2000). Alicki. 25–60 (2001). We can fix d and then adjust e to minimize the work cost. at temperature T. R. R. We define the ‘work cost rate’ as the average  (SjQ)~(1=n)W (S6n jQ6n ) for an optimal erasure process. in Proc. Correspondence and requests for materials should be addressed to L. P. Open Syst.R. 219–253 (1982). Rev. of S6Q. J. and is a single-shot generalization of the von Neumann entropy26 (Supplementary Information. Plenio. 2. it may be used to obtain a single-shot generalization of discord. if we assume that Landauer’s principle holds for a classical observer (equation (2)). & Vedral. 21.d. IEEE Trans. the European Research Council (R. we expect our results to strengthen the link between information theory and statistical physics. J. J. The uncertainty principle in the presence of quantum memory.R..A. PhD thesis. A part of that memory may then be erased. hence. 080501 (2010).. High Energy Phys. 2002). 28. The Decoupling Approach to Quantum Information Theory. J. Nori. B. Int. Dupuis. and O. we pay an extra price of D < 20 in the total work cost. 12. The reduced state of P. grant no.nature. Security of Quantum Key Distribution. S. independently of the size of S. 7. ˆncia e ˜ o para a Cie 200021-119868 and the NCCR QSIT). T. 31st Annu. M. 11. H. to prove a physical result..d. W. A 61. J. & Rex. Renes. R. 3. A fully quantum asymptotic equipartition property. 6. Lett. is preserved by this process. Rev. J. V. C. 5840–5847 (2009). Information erasure. The correspondence between conditional entropy and work that we found can also be used to quantify quantum correlations18. Operational interpretations of quantum discord. Lemma I. Phys. Theory Comput. performing work log2( | S | )kTln(2). Lemma I. Then we extract work from the pure state of S1 6P. Bennett. 5 and can be found in Supplementary Information. 3495–3499 (1995). W. The thermodynamics of computation — a review. Horodecki. A 67. L. & Leighton. Finally we erase S. Rev. Readers are welcome to comment on the online version of this article at www. formulated and proved the main technical claims. 34.LETTER RESEARCH max-entropy converges to the von Neumann entropy for many identical copies of the state (Supplementary Information. 1–8 (1996). S. Shizume. 88. H. IBM J. K. 09.. 81.ch). Furthermore. & Toffoli. Quantum discord and Maxwell’s demons. Lett. originally fully mixed. 21. & Vazirani.. Aolita. J. 148–156 (1961). 13.. Rev. 673–676 (2005). Thermodynamical approach to quantifying quantum correlations. 25. Nature 436.B). ´ al (2009). The erasure proceeds in three steps.com/reprints. that is decoupled from the reference. 23. O. we consider a ‘thermodynamic limit’ of erasure. our analysis shows that erasure can be optimized if information stored in other parts of the memory is used (Fig. our work is related to discord20–23. 9. D. S1. Contemp. Maxwell’s Demon: Entropy. the Portuguese Fundaça Tecnologia (L.. See Supplementary Information for details. gaining energy 2log2( | S1 | )kTln(2).. ACM Symp. 15. Leff. 1. V. Horodecki. Lett. Phys. 22. reversible computation and Maxwell’s demon. Classical and Quantum Information. 2 J U N E 2 0 1 1 | VO L 4 7 4 | N AT U R E | 6 3 ©2011 Macmillan Publishers Limited. & Horodecki.nature. Res. & Vitelli. Paterek. C. Colbeck. M. Inf.. R. Acknowledgements We thank R. 89. Zurek. 24. Theory 56.com/nature. 5. 905–940 (1982). Rev. This we state as Theorem 1: there exists a process to erase a system S. P. 1999). Phys.com/nature. E. V. Colbeck. Horodecki. The authors declare no competing financial interests. Colbeck. 1–23 (2009). J. the quantum asymptotic equipartition property27. the reduced state of S1 is approximately fully mixed16. 120 (2007).. Phys. F.. keeping the rest intact.V. SFRH/BD/43263/2008). Tomamichel. In general. Similarly. 180402 (2002).. Phys. Ann. & Vedral. Phys. More precisely. M. Author Information Reprints and permissions information is available at www. Received 24 November 2010. V. Modi. All rights reserved . 6. R. R. and the corresponding heat generation. R. M. where a large collection of independent and identical copies of the system is erased.R.2 METHODS SUMMARY Here we characterize a single instance of erasure.16. Nature Phys. wrote the manuscript. These are weak if S is highly correlated with Q (because the global state is pure28). the work cost of that operation... V. Colbeck for discussions. W(SjQ). Combining Theorem 1 with the asymptotic equipartition property. the average work cost never process such that w exceeds W(SjQ) 5 H(SjQ)kTln(2).6).R. Our results suggest that discord can quantify the difference between the respective work costs of erasure using quantum and classical memories24. M. 8. Zurek. 11. B. is a random variable and may fluctuate. Fredkin.. Lett. F. & Renner. H. & Zurek. Logical reversibility of computation.. & Winter. and R. Inadequacy of von Neumann entropy for characterising extractable work. 032324 (2011).5). 20. Computing (Taylor and Francis. Phys. e The quantity Hmax (SjQ) denotes the smooth max-entropy of system S conditioned on the quantum memory. A. N..d. R. (delrio@phys. Hist. 104.nature. & Rex. Q. Phys. S. Rev. 42. Horodecki. Berta. if we allow a probability of failure of 3%. Leff. M. Horodecki. Res. Christandl. A 83. 258932) and Singapore’s National Research Foundation and Ministry of Education (V. 1990). M.3). Mod. First we find a subsystem. Reference S S1 P Q Figure 5 | General erasure procedure. 501–510 (2003). C. R. L. Schulman. 525–532 (1973). R. Phys. in the limit of large w n (Supplementary Information. The proof of Theorem 1 is sketched in Fig. Phys. 017901 (2000). PhD thesis. the work required to erase a system S cannot be reduced by locally processing information about S (see Supplementary Information. conditioned on memory Q. section II. Ollivier. J. J. section II).A. J. 1). Cavalcanti.R. 9. S. To understand the exact meaning of equation (3). & Renner. M. would be given solely by the entropy of the part to be erased. H. Rieper. 17.ethz. 18. Phys. Oppenheim. Landauer. Hayden. Renner. Inf. & Boixo. Maruyama. R. K. 205–217 (2004). Bennett. H. 26. R. Heat generation required by information erasure.25. Horodecki. & Williamson. Separability of mixed states: necessary and sufficient conditions. K. K... Phys.) 322–329 (Association for Computing Machinery. For instance.. E 52. Maxwell’s Demon 2: Entropy. We acknowledge support from the Swiss National Science Foundation (L. 27. accepted 18 April 2011. Rev. which is a fundamental result in information theory known as the data processing inequality. Because our relation is valid for a single instance of erasure. 16. 5. 053015 (2011). IEEE Trans. F. because an energy gain in erasure relies on entanglement between the system and the memory. L. Tomamichel. an erasure process with negative work cost can serve as an entanglement witness19. H. Information.D. P. it is purified by a subsystem. Vedral. Develop. Stud. W. Modi. Larmore. grant no. Conservative logic. R. A result from information theory.. 10. such as decoupling14. It is also possible to translate thermodynamic statements into information-theoretical ones. L. Because S1 is decoupled from the reference. Unified view of quantum and classical correlations. H. section I. In this work. Phys.

905–940 (1982). 21. In Encyclopaedia of Complexity and Systems Science (ed. the cost of erasure is always positive. which says that if there is no information available about the data being erased. they do not violate the second law of thermodynamics.com/nature. We thank Charles H. 2. with work cost arbitrarily close to zero1.1038/nature10395 The thermodynamic meaning of negative entropy ˚ berg. Renato Renner. 7174–7201 (Springer. J. the overall work cost of such a computation is always positive or zero (even though temporary quantum correlations may be created and exploited during the course of the computation). Bennett for remarks on reversible computation. standard techniques of reversible information processing allow any deterministic classical algorithm to be performed. Watrous. J. Johan A Lı & Vlatko Vedral Nature 474. because in a computation with deterministic classical output the joint entropy of all registers conditioned on the output cannot be negative. because they are not cyclic processes. Theor. H. Similarly. All rights reserved . although the erasure processes we considered in our Letter can have negative work cost (that is. they can yield work). which can only be restored by doing work. C.2. Phys. These clarifications are developed in more detail in the Supplementary Information to this Addendum. A negative work cost is associated with the consumption of entanglement. on a classical or quantum computer. Oscar Dahlsten ´dia del Rio. Quantum computational complexity. In fact. we note that. 61–63 (2011) To clarify the implications of our result. Supplementary Information is linked to the online version of the Addendum at www. Our results are also consistent with the original unconditional form of Landauer’s principle. Bennett. in a thermodynamically reversible fashion. 1. Int. The thermodynamics of computation—a review.CORRECTIONS & AMENDMENTS ADDENDUM doi:10. 2009). 4 7 6 | N AT U R E | VO L 4 7 6 | 2 5 AU G U S T 2 0 1 1 ©2011 Macmillan Publishers Limited.nature. R. A.) Part 17. Meyers.

Sign up to vote on this title
UsefulNot useful

Master Your Semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master Your Semester with a Special Offer from Scribd & The New York Times

Cancel anytime.