Abstract
The work content of nonequilibrium systems in relation to a heat bath is often analysed in terms of expectation values of an underlying random work variable. However, when optimizing the expectation value of the extracted work, the resulting extraction process is subject to intrinsic fluctuations, uniquely determined by the Hamiltonian and the initial distribution of the system. These fluctuations can be of the same order as the expected work content per se, in which case the extracted energy is unpredictable, thus intuitively more heatlike than worklike. This raises the question of the ‘truly’ worklike energy that can be extracted. Here we consider an alternative that corresponds to an essentially fluctuationfree extraction. We show that this quantity can be expressed in terms of a oneshot relative entropy measure introduced in information theory. This suggests that the relations between information theory and statistical mechanics, as illustrated by concepts like Maxwell’s demon, Szilard engines and Landauer’s principle, extends to the singleshot regime.
Introduction
The amount of useful energy that can be harvested from nonequilibrium systems not only characterizes practical energy extraction and storage, but is also a fundamental thermodynamic quantity. Intuitively, we wish to extract ordered and predictable energy, that is, ‘work’, as opposed to disordered random energy in the form of ‘heat’. The catch is that, in statistical systems, the work cost or yield of a given transformation is typically a random variable^{1}. As an example, one can think of the friction that an object experiences when forced through a viscous medium. On a microscopic level, this force resolves into chaotic molecular collisions and thus results in a random work cost each time we perform this transformation. This is further illustrated by experimental tests^{2,3,4} of ‘fluctuation theorems’, which characterize the randomness in the work cost (or ‘entropy production’) of nonequilibrium processes^{1}. These observations raise the question of a quantitative notion of work content that truly reflects the idea of work as ordered energy.
Here, we show that standard expressions for the work content^{5,6,7,8} can correspond to a very noisy and thus heatlike energy, but we also introduce an alternative that quantifies the amount of ordered energy that can be extracted. The latter can be expressed in terms of a nonequilibrium generalization of the free energy, or equivalently in terms of a oneshot informationtheoretic relative entropy, which quantifies how ‘far’ the given nonequilibrium system is from thermal equilibrium.
‘Standard’ information theory typically quantifies the resources needed to perform a given informationtheoretic task averaged over many repetitions, for example, the average number of bits needed to send many independent messages^{9}. In contrast, ‘oneshot’ (or ‘singleshot’) information theory rather focuses on single instances of such tasks (see, for example, refs 10,11). Given the strong historical links between standard information theory and the work extraction problem, via concepts like Szilard engines, Landauer’s principle and Maxwell’s demon^{12,13}, it is reasonable to ask whether also oneshot information theory has a counterpart in statistical mechanics. Together with refs 14,15,16,17,18,19, the results of this investigation suggest that this is indeed the case. A direct consequence of the present investigation is that the results of refs 14,15,19 is brought into a more physical setting, allowing, for example, systems with nontrivial Hamiltonians, proof of nearoptimality, as well as a connection to fluctuation theorems^{1}. The latter suggests that the effects we consider become relevant in the typical regimes of fluctuation theorems. Similar results as in this study have been obtained independently in ref. 16. See also recent results in ref. 17 based on ideas in ref. 18. (For further discussions on the relations to the existing literature, see Supplementary Note 1.)
Results
Work extraction
The amount of work that a system can perform while it equilibrates with respect to an environment of temperature T is often^{5,6,7,8} expressed as
Here q is the state of the system, G(h) its equilibrium state, h the system Hamiltonian and k Boltzmann’s constant. For the simple model we employ here, q is a probability distribution over a finite set of energy levels, and D(qp)=∑_{n}q_{n}log_{2}q_{n}−∑_{n}q_{n}log_{2}p_{n} is the relative Shannon entropy (Kullback–Leibler divergence)^{9}, and log_{2} denotes the base 2 logarithm.
The quantity (q,h), and the closely related cost of information erasure (Landauer’s principle), is often understood as an expectation value of an underlying random work yield (see, for example, refs 5,7,20,21). However, this tells us very little about the fluctuations, and thus the ‘quality’ of the extracted energy. Here, we show that optimizing the expected gain leads to intrinsic fluctuations. These can be of the same order as the expected work content (q, h) per se, in which case the work extraction does not act as a truly ordered energy source. As an alternative, we introduce the ɛdeterministic work content, which quantifies the maximal amount of energy that can be extracted if we demand to always get precisely this energy each single time we run the extraction process, apart from a small probability of failure ɛ. Hence, in contrast to the expected work extraction, where we do not put any restrictions on how broadly distributed the random energy gain is, we do in the ɛdeterministic work extraction demand that the probability distribution should be very peaked, that is, very predictable. In other words, the ɛdeterministic work content formalizes the idea of an almost perfectly ordered energy source.
The model
Our analysis is based on a very simple model of a system interacting with a heat bath of fixed temperature T (see Fig. 1). Akin to, for example, refs 15,21,22, we model the Hamiltonian of the system as finite set of energy levels h=(h_{1},…,h_{N}). The state of the system we regard as a random variable , with a probability distribution q=(q_{1},…, q_{N}). On this system, we have two elementary operations.
The first of these two operations changes the energy levels h to a new set of energy levels h′, but leaves the state, and thus the probability distribution q, intact. We refer to this as level transformations (LTs). (For a quantum system, this would essentially correspond to adiabatic evolution with respect to some external control parameters, that is, in the limit of infinitely slow changes of the control we alter the energy levels, but not how they are occupied.) Via the LTs we define what ‘work’ is in our model. If we perform an LT that changes h to h′, and if the system is in state n, then this results in a work gain (or work cost ). As the work gain depends on the state of the system, a random state implies a random work gain.
The second elementary operation corresponds to thermalization, where one can imagine that we connect the system to the heat bath, let it thermalize and slowly deconnect it again. We model this by putting the system into the random state described by the Gibbs distribution, P(=n)=G_{n}(h), where , β=1/(kT) and is the partition function. It is furthermore assumed that the state (regarded as a random variable) after a thermalization is independent of the state before.
We construct processes by combining these two types of elementary operations into any sequence of our choice. The resulting work yield of the process is defined as the sum of the work yields of all the LTs. (For a more detailed description of the model, see Supplementary Note 2.) An example is given in Fig. 2, where we construct the analogue of isothermal reversible (ITR) processes, which serve as a building block in our analysis (see Supplementary Note 3). As opposed to other processes we will consider, the ITRs have essentially fluctuationfree work costs.
Expected work extraction
Given an initial state with distribution q, we can reproduce equation (1) within our model. A cyclic threestep process, as described in Fig. 3, gives the random work yield
By taking the expectation value, we obtain equation (1). The positivity of relative entropy, D(qp)≥0, can be used to show that no process can give a better expected work yield (Supplementary Note 4). One can in a similar fashion determine the minimal expected work cost for information erasure (see Supplementary Note 5).
Fluctuations in expected work extraction
How large are the fluctuations for a process that maximizes the expected work extraction, and thus achieves (q, h)? Equation (2) determines the noise of the specific process in Fig. 3, but it turns out that it actually specifies the fluctuations for all processes that optimize the expected work extraction. (For the exact statement, see Methods, or Supplementary Note 6.) We can conclude that to analyse the noise in the optimal expected work extraction, it is enough to consider equation (2). As we will confirm later, these fluctuations can be of the same order as (q, h) itself.
ɛdeterministic work extraction
As the optimal expected work extraction suffers from fluctuations, a natural question is how much (essentially) noisefree energy can be extracted. We say that a random variable X has the (ɛ, δ)deterministic value x, if the probability to find X in the interval [x−δ, x+δ] is larger than 1−ɛ. Hence, δ is the precision by which the value x is taken, and ɛ the largest probability by which this fails. (See Supplementary Note 7 for further properties.) We define as the highest possible (ɛ, δ)deterministic work yield among all processes that operate on the initial distribution q with initial and final energy levels h. Next, we define the ɛdeterministic work content as , that is, we take the limit of infinite precision, thus formalizing the idea of an energy extraction that is essentially free from fluctuations.
^{ɛ}(q, h) can be expressed in terms of the ɛfree energy, which is defined via restrictions to sufficiently likely subsets of energy levels. Given a subset Λ, we define . We minimize Z_{Λ}(h) among all subsets Λ such that q(Λ)=∑_{n∈Λ}q_{n}>1−ɛ. If Λ^{*} is such a minimizing set, then the ɛfree energy is defined as F^{ɛ}(q,h)=−kTlnZ_{Λ*}(h). (See Supplementary Note 8 for further explanations.) The concept of oneshot free energy has been introduced independently in ref. 16.
The distribution of fluctuations is clearly important for determining the value of ^{ɛ}(q, h). It is thus maybe not surprising that a variation (see Methods) of Crook’s fluctuation theorem^{23} can be used to prove
where ‘≈’ signifies that the equality is true up to a small error (see Methods, or Supplementary Note 9 and Supplementary Note 10). The error is small in the sense that it can be regarded as the energy of a sufficiently likely equilibrium fluctuation (see Methods and Supplementary Note 11). An example of a process that gives the righthand side of equation (3) is described in Fig. 4. In the case of completely degenerate energy levels h=(r,…,r), equation (3) reduces to the result in ref. 14. (See also Supplementary Note 12 and Supplementary Note 13 for the ɛdeterministic cost of information erasure).
The above result can be reformulated in terms of an ɛsmoothed relative Rényi 0entropy, defined as This relative entropy was (up to some technical differences) introduced in ref. 24 in the context of oneshot information theory. (See refs 25,26 for quantum versions.) One can see that
Comparisons
An immediate question is how (q, h) compares with ^{ɛ}(q, h), and with the fluctuations in the optimal expected work extraction. The latter we measure by the s.d. of W_{yield} in equation (2), . We compare how these three quantities scale with increasing system size (for example, in number of spins, or other units).
Our first example is a collection of m systems whose state distributions are independent and identical, q^{m}(n_{1},…, n_{m})=q(n_{1})…q(n_{m}), and which have noninteracting identical Hamiltonians, corresponding to energy levels h^{m}(n_{1},…,n_{m})=h(n_{1})+…+h(n_{m}). In this case (q^{m}, h^{m})=m(q, h), and , where σ(qr)^{2}=∑_{n}q_{n}[log_{2}(q_{n}/r_{n})]^{2}−D(qr)^{2}. By Berry–Esseen’s theorem^{27,28} (see Methods and Supplementary Note 14), one can show that ^{ɛ}(q^{m}, h^{m}) is equal to m(q, h) to the leading order in m (see ref. 29 for a similar result in a resource theory framework). The difference only appears at the next to leading order. Hence, in these systems the fluctuations are comparably small, and the dominant contribution to ^{ɛ}(q^{m}, h^{m}) is (q^{m}, h^{m}). It appears reasonable to expect similar results for nonequilibrium systems with sufficiently fast spatial decay of both correlations and interactions, which may explain why issues concerning as a measure of work content appear to have gone largely unnoticed.
A simple modification of the state distribution in the previous example results in a system with large fluctuations. With probability 1−ɛ (independent of m), the system is prepared in the joint ground state 0,…, 0, and with probability ɛ in the Gibbs distribution. This results in and yields (q^{m},h^{m})∼−mkTln(2)(1−ɛ)log_{2}G_{0}(h), and ^{ɛ}(q^{m}, h^{m})∼−mkT ln(2)log_{2}G_{0}(h). Hence, all three quantities grow proportionally to m.
For a second case of large fluctuations, we choose the distribution , for a collection of dlevel systems. For large m, we assume that the energy levels are dense enough that they can be replaced by a spectral density. One example is Wigner’s semicircle law, where for x≤R(m). With , this is the asymptotic energy level distribution of large random matrices from the Gaussian unitary ensemble^{30}. For the semicircle distribution (q^{m}, h^{m})∼R(m), , and ^{ɛ}(q^{m},h^{m})∼c(ɛ)R(m), where c(ɛ) is independent of m.
Discussion
We have here employed what one could refer to as a discrete classical model. Relevant extensions include a classical phasespace picture, as well as a quantum setting that allows superpositions between different energy eigenstates (for example, in the spirit of refs 16,29,31) and where the workextractor can possess quantum information about the system^{15}. An operational approach, based on what ‘work’ is supposed to achieve, rather than ad hoc definitions, may yield deeper insights to the question of the truly worklike energy content.
It is certainly justified to ask for the relevance of the effects we have considered here. The evident role of fluctuations suggests that the noise in the expected work extraction should become noticeable in the same nanoregimes as where fluctuation theorems are relevant. The considerable experimental progress on the latter (see, for example, refs 2,3,4) should reasonably be applicable also to the former. Also the theoretical aspects of the link to fluctuation theorems merits further investigations.
We have seen that the ɛdeterministic work content to the leading order becomes equal to the expected work content for systems with identical noninteracting Hamiltonians and identical uncorrelated state distributions. However, we have also demonstrated by simple examples that the expected work extraction can become very noisy when we deviate from this simple setup. In these cases, the expected work extraction thus fails to capture our intuitive notion of work as ordered energy, while the ɛdeterministic work extraction is predictable by construction. One might object that many realistic systems are approximately noninteracting and approximately uncorrelated, and thus presumably show no significant difference between the ɛdeterministic and the expected work extraction. However, as we here consider a general nonequilibrium setting, there is no particular reason to assume, for example, weak correlations. It is maybe also worth to point out that the fluctuations in the expected work extraction can be large also outside the microscopic regime, as this only requires a sufficiently ‘violent’ relation between the nonequilibrium state and the Hamiltonian of the system. As opposed to the expected work content, the ɛdeterministic work content retains its interpretation as the ordered energy. It is no coincidence that this is much analogous to how singleshot information theory generalizes standard information theory^{10,11}. In this spirit, the present study, along with refs 14,15,16,17,18,19, can be viewed as the first glimpse of a ‘singleshot statistical mechanics’.
Methods
Randomness in optimal expected work extraction
In the main text, we briefly mentioned the fact that processes that optimize the expected work extraction converge to the random variable in equation (2). We can phrase this result more precisely as follows. For a process that operates on an initial state with distribution q, we let W_{yield}(, ) denote the corresponding random work yield. We here consider cyclic processes that starts and ends in the energy levels h. If is a family of processes such that lim_{m→∞}〈W_{yield}(_{m}, )〉=(q, h), then in probability. (For a proof see Supplementary Note 6).
Bounds on the ɛdeterministic work content
The exact statement of equation (3) is
In Supplementary Note 11, it is shown that −kTln(1−ɛ) is an upper bound to the ɛdeterministic work content of equilibrium systems. Equation (5) thus determines the value of ^{ɛ}(q, h) up to an error with the size of a sufficiently probable equilibrium fluctuation. We obtain the lower bound in equation (5) by the process described in Fig. 4. The upper bound is obtained by a combination of a variation (Supplementary Note 10, Supplementary Equation (S73)) on Crook’s fluctuation theorem^{23} and a work bound for LTs (Supplementary Note 10). For a discussion on an alternative singleshot work extraction quantity, and its relation to ^{ɛ}, see Supplementary Note 15.
Expansion of ^{ɛ} in the multicopy case
In the case of a state distribution q^{m}(n_{1},…,n_{m})=q(n_{1})…q(n_{m}), and energy levels h^{m}(n_{1},…,n_{m})=h(n_{1})+…+h(n_{m}), the ɛdeterministic work content has the expansion
where is a correction term that grows slower than , and Φ^{−1} is the inverse of the cumulative distribution function of the standard normal distribution. The smaller our error tolerance ɛ, the more the correction term lowers the value of ^{ɛ}(q^{m}, h^{m}) as compared with (q^{m}, h^{m}). This expansion is proved via Berry–Esseen’s theorem^{27,28}, which determines the convergence rate in the central limit theorem (see Supplementary Note 14).
Additional information
How to cite this article: Åberg J. Truly worklike work extraction. Nat. Commun. 4:1925 doi: 10.1038/ncomms2712 (2013).
References
 1.
Jarzynski, C.. Equalities and inequalities: irreversibility and the second law of thermodynamics at the nanoscale. Annu. Rev. Condens. Matter Phys. 2, 329–351 (2011).
 2.
Liphardt, J., Dumont, S., Smith, S. B., Tinoco, I. Jr & Bustamante, C.. Equilibrium information from nonequilibrium measurements in an experimental test of Jarzynski’s equality. Science 296, 1832–1835 (2002).
 3.
Collin, D., Ritort, F., Jarzynski, C., Smith, S. B., Tinoco, I. Jr & Bustamante, C.. Verification of the Crooks fluctuation theorem and recovery of RNA folding free energies. Nature 437, 231–234 (2005).
 4.
Toyabe, S., Sagawa, T., Ueda, M., Muneyuki, E. & Sano, M.. Experimental demonstration of informationtoenergy conversion and validation of the generalized Jarzynski equality. Nat. Phys. 6, 988–992 (2010).
 5.
Procaccia, I. & Levine, R. D.. Potential work: a statisticalmechanical approach for systems in disequilibrium. J. Chem. Phys. 65, 3357–3364 (1967).
 6.
Lindblad, G.. NonEquilibrium Entropy and Irreversibility Reidel: Lancaster, (1983).
 7.
Takara, K., Hasegawa, H.H. & Driebe, D. J.. Generalization of the second law for a transition between nonequilibrium states. Phys. Lett. A 375, 88–92 (2010).
 8.
Esposito, M. & Van den Broeck, C.. Second law and Landauer principle far from equilibrium. Euro. Phys. Lett. 95, 40004 (2011).
 9.
Cover, T. M. & Thomas, J. A.. Elements of Information Theory 2nd edn Wiley, Hoboken (2006).
 10.
Holenstein, T. & Renner, R.. On the randomness of independent experiments. IEEE Trans. Inf. Theor. 57, 1865–1871 (2011).
 11.
Tomamichel, M., Colbeck, R. & Renner, R.. A fully quantum asymptotic equipartition property. IEEE Trans. Inf. Theor. 55, 5840–5847 (2009).
 12.
Leff, H. S. & Rex, A. F.. Maxwell's Demon: Entropy, Information, Computing Taylor and Francis (1990).
 13.
Leff, H. S. & Rex, A. F.. Maxwell’s Demon 2: Entropy, Classical and Quantum Information, Computing Taylor and Francis (2002).
 14.
Dahlsten, O., Renner, R., Rieper, E. & Vedral, V.. Inadequacy of von Neumann entropy for characterizing extractable work. New J. Phys. 13, 053015 (2011).
 15.
del Rio, L., Åberg, J., Renner, R., Dahlsten, O. & Vedral, V.. The thermodynamic meaning of negative entropy. Nature 474, 61–63 (2011).
 16.
Horodecki, M. & Oppenheim, J.. Fundamental limitations for quantum and nano thermodynamics. Nat. Commun. doi:10.1038/ncomms3059 (2013).
 17.
Egloff, D., Dahlsten, O., Renner, R. & Vedral, V.. Laws of thermodynamics beyond the von Neumann regime. Preprint at http://arxiv.org/abs/1207.0434 (2012).
 18.
Egloff, D.. Work Value of Mixed States in Single Instance Work Extraction Games Master's thesis, ETH Zurich (2010).
 19.
Faist, P., Dupuis, F., Oppenheim, J. & Renner, R.. A quantitative Landauer's principle. Preprint at http://arxiv.org/abs/1211.1037 (2012).
 20.
Shizume, K.. Heat generation required by information erasure. Phys. Rev. E 52, 3495–3499 (1995).
 21.
Piechocinska, B.. Information erasure. Phys. Rev. A 61, 062314 (2000).
 22.
Crooks, G. E.. Nonequilibrium measurements of free energy differences for microscopically reversible markovian systems. J. Stat. Phys. 90, 1481–1487 (1998).
 23.
Crooks, G. E.. Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences. Phys. Rev. E 60, 2721–2726 (1999).
 24.
Wang, L., Colbeck, R. & Renner, R.. Simple channel coding bounds. IEEE Int Symp Inf Theory 2009, 1804–1808 (2009).
 25.
Datta, N.. Min and maxrelative entropies and a new entanglement monotone. IEEE Trans. Inf. Theor. 55, 2816–2826 (2009).
 26.
Wang, L. & Renner, R.. Oneshot classicalquantum capacity and hypothesis testing. Phys. Rev. Lett. 108, 200501 (2012).
 27.
Berry, A. C.. The accuracy of the gaussian approximation to the sum of independent variates. Trans. Amer. Math. Soc. 49, 122–136 (1941).
 28.
Esseen, C.G.. On the Liapunoff limit of error in the theory of probability. Ark. Mat. Astr. o. Fys. 28A, 1–19 ((1942).
 29.
Brandão, F. G. S. L., Horodecki, M., Oppenheim, J., Renes, J. M. & Spekkens, R. W.. The resource theory of quantum states out of thermal equilibrium. Preprint at http://arxiv.org/abs/1111.3882 (2011).
 30.
Mehta, M. L.. Random Matrices Elsevier (2004).
 31.
Janzing, D., Wocjan, P., Zeier, R., Geiss, R. & Beth, Th.. Thermodynamic cost of reliability and low temperatures: tightening Landauer’s principle and the second law. Int. J. Theor. Phys. 39, 2717–2753 (2000).
Acknowledgements
I thank Lídia del Rio, Renato Renner and Paul Skrzypczyk for useful comments. This research was supported by the Swiss National Science Foundation through the National Centre of Competence in Research ‘Quantum Science and Technology’, by the European Research Council, grant no. 258932, and the Excellence Initiative of the German Federal and State Governments (grant ZUK 43).
Author information
Affiliations
Institute for Physics, University of Freiburg, HermannHerderStrasse 3, Freiburg D79104, Germany
 Johan Åberg
Institute for Theoretical Physics, ETH Zurich, Zurich 8093, Switzerland
 Johan Åberg
Authors
Search for Johan Åberg in:
Contributions
J.Å. is responsible for the whole content of this paper.
Competing interests
The author declares no competing financial interests.
Corresponding author
Correspondence to Johan Åberg.
Supplementary information
PDF files
 1.
Supplementary Information
Supplementary Notes 115 and Supplementary References
Rights and permissions
To obtain permission to reuse content from this article visit RightsLink.
About this article
Further reading

Correlations as a resource in quantum thermodynamics
Nature Communications (2019)

Quantum majorization and a complete set of entropic conditions for quantum thermodynamics
Nature Communications (2018)

Holographic second laws of black hole thermodynamics
Journal of High Energy Physics (2018)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.