Relating different quantum generalizations of the conditional Rényi entropy

Relating different quantum generalizations of the conditional Rényi entropy


Recently a new quantum generalization of the Rényi divergence and the corresponding conditional Rényi entropies was proposed. Here we report on a surprising relation between conditional Rényi entropies based on this new generalization and conditional Rényi entropies based on the quantum relative Rényi entropy that was used in previous literature. Our result generalizes the well-known duality relation of the conditional von Neumann entropy for tripartite pure states to Rényi entropies of two different kinds. As a direct application, we prove a collection of inequalities that relate different conditional Rényi entropies and derive a new entropic uncertainty relation.


Recently, there has been renewed interest in finding suitable quantum generalizations of Rényi’s [36] entropies and divergences. This is due to the fact that Rényi entropies and divergences have a wide range of applications in classical information theory and cryptography, see, e.g. [13].

We will review some of the recent progress here, but refer the reader to [31] for a more in-depth discussion. For our purposes, a quantum system is modeled by a finite dimensional Hilbert space. We denote by the set of positive semi-definite operators on that Hilbert space, and by the subset of density operators with unit trace.

The following natural quantum generalization of the Rényi divergence has been widely used and has found operational significance, for example, as a cut-off rate in quantum hypothesis testing [28] (see also [34]). It is usually referred to as quantum Rényi relative entropy and for all given as

for arbitrary , that satisfy . (The notation means that dominates , i.e. the kernel of lies inside the kernel of .)

While this definition has proven useful in many applications, it has a major drawback in that it does not satisfy the data-processing inequality (DPI) for . The DPI states that the quantum Rényi relative entropy is contractive under application of a quantum channel, i.e., for any completely positive trace-preserving map . Intuitively, this property is very desirable since we want to think of the divergence as a measure of how well can be distinguished from , and this can only get more difficult after a channel is applied.

Recently, an alternative quantum generalization has been investigated [30] (see also [38]). It is referred to as quantum Rényi divergence (or sandwiched Rényi relative entropy in [42]) and defined as

for all and , that satisfy . The quantum Rényi divergence has found operational significance in the converse part of quantum hypothesis testing [29]. As such, it satisfies the DPI for all as was shown by Frank and Lieb [15] and independently by Beigi [6] for . See also earlier work [30] where a different proof is given for . Furthermore, the quantum Rényi divergence has already proven an indispensable tool, for example in the study of strong converse capacities of quantum channels [42].

The definitions, and , are in general different but coincide when and commute. For , we define and as the corresponding limit. For it has been shown that [14]:

with the eigenvalue decomposition , , and the projector on the support of . In the limit both expressions converge to the quantum relative entropy [30], namely

For the limits have been evaluated in [31] and [39], respectively:

with the eigenvalue decompositions and .

It has been observed [42] that the relation

follows from the Araki-Lieb-Thirring trace inequality [1]. Furthermore, and are monotonically increasing functions. For the latter quantity, this was shown in [31] and independently in [6].

Finally, very recently Audenaert and Datta [4] defined a more general two parameter family of -z-relative Rényi entropies of the form

and explored some of its properties. We clearly have and .

2Quantum Conditional Rényi Entropies

We will in the following consider disjoint quantum systems, denoted by capital letters and . The sets and take on the expected meaning.

The conditional von Neumann entropy can be conveniently defined in terms of the quantum relative entropy as follows. For a bipartite state , we define

where is the usual von Neumann entropy. The last equality can be verified using the relation together with the fact that is positive definite.

In the case of Rényi entropies, it is not immediate which expression, , or , should be used to define the conditional Rényi entropies. It has been found in the study of the classical special case (see, e.g. [23] for an overview) that generalizations based on have severe limitations, for example they cannot be expected to satisfy a DPI. On the other hand, definitions based on the underlying divergence, as in or , have proven to be very fruitful and lead to quantities with operational significance. Together with the two proposed quantum generalizations of the Rényi divergence in and , this leads to a total of four different candidates for conditional Rényi entropies. For and , we define

The fully quantum entropy has first been studied in [39]. For the classical and classical-quantum special case this quantity gives a generalization of the leftover hashing lemma [7] for the modified mutual information to Rényi entropies with [19].

The classical version of was introduced by Arimoto for an evaluation of the guessing probability [2]. We note that he used another but equivalent expression for that we later explain in Lemma ?. Then, Gallager used (again in the form of Lemma ?) to upper bound the decoding error probability of a random coding scheme for data compression with side-information [16]. The classical and classical-quantum special cases of were, for example, also investigated in [20] and realize another type of a generalization of the leftover hashing lemma for the -distinguishability in the study of randomness extraction to Rényi entropies with .

It follows immediately from the definition and the corresponding property of that these two entropies satisfy a data-processing inequality. Namely for any quantum operation with and any , we have

while their classical-quantum versions have been obtained in [20].

The conditional entropy was proposed in [38] and investigated in [31], whereas is first considered in this paper. (Since the relative entropies and are identical for commuting operators, we note that as well as for classical distributions.) Both definitions satisfy the above data-processing inequality for .

Furthermore, it is easy to verify that all entropies considered are invariant under applications of local isometries on either the or systems. Lastly, note that the optimization over can always be restricted to for .

We use up and down arrows to express the trivial observation that and by definition. Finally, gives us the additional relations and . These relations are summarized in Figure 1. Moreover, inheriting these properties from the corresponding divergences, all entropies are monotonically decreasing functions of

For , all definitions coincide with the usual von Neumann conditional entropy . For , two quantum generalizations of the conditional min-entropy emerge, which both have been studied by Renner [35]. Namely,

(The notation and is widely used. However, we prefer our notation as it makes our exposition in this manuscript clearer.) For , we find a quantum generalization of the conditional collision entropy as introduced by Renner [35]:

For , we find the quantum conditional max-entropy first studied by König et al. [24],

where denotes the fidelity. (The alternative notation is often used.) For , we find a quantum conditional generalization of the Hartley entropy [18] that was initially considered by Renner [35],

where denotes the projector onto the support of .

Figure 1: Overview of the different conditional entropies used in this paper. Arrows indicate that one entropy is larger or equal to the other for all states \rho_{AB} \in \mathcal{S}(AB) and all \alpha \geq 0.
Figure 1: Overview of the different conditional entropies used in this paper. Arrows indicate that one entropy is larger or equal to the other for all states and all .

3Duality Relations

It is well known that, for any tripartite pure state , the relation

holds. We call this a duality relation for the conditional entropy. To see this, simply write and and note that the spectra of and as well as the spectra of and agree. The significance of this relation is manifold — for example it turns out to be useful in cryptography where the entropy of an adversarial party, let us say , can be estimated using local state tomography by two honest parties, and . In the following, we are interested to see if such relations hold more generally for conditional Rényi entropies.

It was shown in [39] that indeed satisfies a duality relation, namely

Note that the map maps the interval , where data-processing holds, onto itself. This is not surprising. Indeed, consider the Stinespring dilation of a quantum channel . Then, for pure, is also pure and the above duality relation implies that

Hence, data-processing for holds if and only if data-processing for holds.

A similar relation has recently been discovered for in [31] and independently in [6]. There, it is shown that

As expected, the map maps the interval , where data-processing holds, onto itself.

The purpose of the following is thus to show if a similar relation holds for the remaining two candidates, and . First, we find the following alternative expression for by determining the optimal in the definition .

This generalizes a result by one of the current authors [20].

Recall the definition

This can immediately be lower bounded by the expression in by substituting

for . It remains to show that this choice is optimal. We employ the following Hölder and reverse Hölder inequalities (cf. Lemma ? in Appendix A). For any , the Hölder inequality states that

Furthermore, if , we also have a reverse Hölder inequality which states that

For , we employ for , , and to find

which yields the desired upper bound since . For , we instead use . This leads us to upon the same substitutions, concluding the proof.

An alternative proof also follows rather directly from a quantum generalization of Sibson’s identity, which was introduced by Sharma and Warsi [37].

This allows us to show our main result.

Substituting and employing Lemma ?, it remains to show that

is equal to

In the following we show something stronger, namely that the operators

are unitarily equivalent. This is true since both of these operators are marginals — on and  — of the same tripartite rank- operator,

To see that this is indeed true, note the first operator in can be rewritten as

The last equality can be verified using the Schmidt decomposition of with regards to the partition :. This concludes the proof.

The relation can readily be extended for all and . The limiting case is simply the duality of the conditional von Neumann entropy , whereas the case was also shown in [8]. (See [41] for a concise proof.) Again, note that the transformation maps the interval where data-processing holds for to where data-processing holds for .

We summarize these duality relations in the following theorem, where we take note that the first and second statements have been shown in [39] and [31], respectively.

4Some Inequalities Relating Conditional Entropies

Our first application yields relations between different conditional Rényi entropies for arbitrary mixed states. Recently, Mosonyi [27] used a converse of the Araki-Lieb-Thirring trace inequality due to Audenaert [3] to find a converse to the ordering relation , namely

Here we follow a different approach and show that inequalities of a similar type for the conditional entropies are a direct corollary of the duality relations in Theorem ?.

Note that the first inequality on each line follows directly from the relations depicted in Figure 1. Next, consider an arbitrary purification of . The relations of Figure 1, for any , applied to the marginal are given as

We then substitute the corresponding dual entropies according to Theorem ?, which yields the desired inequalities upon appropriate new parametrization.

We note that the fully classical (commutative) case of all these inequalities is trivial except for the second inequalities in and , which were proven before by one of authors [21]. Other special cases of these inequalities are also well known and have operational significance. For example, for states that , which relates the conditional min-entropy in to the conditional collision entropy in . To understand this inequality more operationally we rewrite the conditional min-entropy as its dual semi-definite program [24],

where is a copy of , the infimum is over all quantum channels , denotes the dimension of , and is the maximally entangled state on . Now, the above inequality becomes apparent since the conditional collision entropy can be written as [10],

where denotes the pretty good recovery map of Barnum and Knill [5]. Also, for yields , which relates the quantum conditional max-entropy in to the quantum conditional generalization of the Hartley entropy in .

We believe that the sandwich relations – for close to will prove useful in applications in quantum information theory as they allow to switch between different definitions of the conditional Rényi entropy.

5Entropic Uncertainty Relations

A series of papers [9] culminating in [12] established a general technique to derive uncertainty relations for quantum conditional entropies based on two main ingredients: (1) a duality relation, and (2) a data-processing inequality for the underlying divergence. It is evident that all our definitions of conditional Rényi entropies fit the framework of [12], which then immediately yields the following entropic uncertainty relations:

We want to point out that the first and second inequality were first shown in [12] and [31], respectively; the third inequality is novel. To verify it, we apply [12] to and note that has the required form. Furthermore, it is already pointed out in [12] that the underlying divergence, for , satisfies the required properties for the application of their theorem. As such, comparing to the corresponding duality relation , we see that in order to derive the uncertainty relation we need to restrict to to be in the regime where data-processing holds.

It is noteworthy that even for the case of classical side information (if the systems and are classical), the three relations are genuinely different. The first inequality bounds the sum of two -entropies, the second the sum of two -entropies, and the third inequality the sum of a - and an -entropy. Let us further specialize these inequalities for the case where both and are trivial. It was already noted in [31] that specializes to the well-known Maassen-Uffink relation [26]. We have

evaluated for the marginals of the states in . It is also easy to verify that and specialize to strictly weaker uncertainty relations when and are trivial.

Acknowledgments. MT is funded by the Ministry of Education (MOE) and National Research Foundation Singapore, as well as MOE Tier 3 Grant “Random numbers from quantum processes” (MOE2012-T3-1-009). MB thanks the Center for Quantum Technologies, Singapore, for hosting him while this work was done. MH is partially supported by a MEXT Grant-in-Aid for Scientific Research (A) No. 23246071 and the National Institute of Information and Communication Technology (NICT), Japan. The Centre for Quantum Technologies is funded by the Singapore Ministry of Education and the National Research Foundation as part of the Research Centres of Excellence programme.

AHölder Inequalities

We prove the following Hölder and reverse Hölder inequalities for traces of operators.

The first statement also immediately follows from a Hölder inequality for unitarily invariant norms (the trace norm in this case), e.g. in [11]. However, we believe that the following reduction of the proof to the commutative case is noteworthy.

For commuting and , the above result immediately follows from the corresponding classical Hölder and reverse Hölder inequalities. Now, let be a pinching in the eigenbasis of . Since commutes with , we have

under the respective constraints. Now, note that for , we have by the pinching inequality for the Schatten -norm [11] and follows. On the other hand, for , we use [11], which implies that . This yields and concludes the proof.


  1. On an inequality of Lieb and Thirring.
    H. Araki. Letters in Mathematical Physics, 19(2):167–170, Feb. 1990.
  2. Information Measures and Capacity of Order alpha for Discrete Memoryless Channels.
    S. Arimoto. Colloquia Mathematica Societatis János Bolya, 16:41–52, 1975.
  3. On the Araki-Lieb-Thirring inequality.
    K. M. R. Audenaert. Int. J. of Inf. and Syst. Sci., 4(1):78–83, Jan. 2008.
  4. -z-relative Renyi entropies.
    K. M. R. Audenaert and N. Datta. Oct. 2013.
  5. Reversing Quantum Dynamics with Near-Optimal Quantum and Classical Fidelity.
    H. Barnum and E. Knill. J. Math. Phys., 43(5):2097, 2002.
  6. Sandwiched Rényi Divergence Satisfies Data Processing Inequality.
    S. Beigi. J. Math. Phys., 54(12):122202, June 2013.
  7. Generalized Privacy Amplification.
    C. H. Bennett, G. Brassard, C. Crepeau, and U. M. Maurer. IEEE Trans. on Inf. Theory, 41(6):1915–1923, 1995.
  8. Single-Shot Quantum State Merging.
    M. Berta. Master’s thesis, ETH Zurich, 2008.
  9. The Uncertainty Principle in the Presence of Quantum Memory.
    M. Berta, M. Christandl, R. Colbeck, J. M. Renes, and R. Renner. Nat. Phys., 6(9):659–662, July 2010.
  10. An equality between entanglement and uncertainty.
    M. Berta, P. J. Coles, and S. Wehner. page 5, Feb. 2013.
  11. Matrix Analysis.
    R. Bhatia. Graduate Texts in Mathematics. Springer, 1997.
  12. Uncertainty Relations from Simple Entropic Properties.
    P. J. Coles, R. Colbeck, L. Yu, and M. Zwolak. Phys. Rev. Lett., 108(21):210405, May 2012.
  13. Generalized Cutoff Rates and Renyi’s Information Measures.
    I. Csiszár. IEEE Trans. on Inf. Theory, 41(1):26–34, 1995.
  14. A Limit of the Quantum Renyi Divergence.
    N. Datta and F. Leditzky. Aug. 2013.
  15. Monotonicity of a Relative Rényi Entropy.
    R. L. Frank and E. H. Lieb. J. Math. Phys., 54(12):122201, June 2013.
  16. Source Coding with Side Information and Universal Coding.
    R. G. Gallager. In Proc. IEEE ISIT, volume 21, Ronneby, Sweden,, June 1979. IEEE.
  17. Multiplicativity of Completely Bounded p-Norms Implies a Strong Converse for Entanglement-Assisted Capacity.
    M. K. Gupta and M. M. Wilde. Oct. 2013.
  18. Transmission of Information.
    R. V. L. Hartley. Bell Syst. Tech. J., 7(3):535–563, July 1928.
  19. Exponential decreasing rate of leaked information in universal random privacy amplification.
    M. Hayashi. Information Theory, IEEE Transactions on, 57(6):3989–4001, 2011.
  20. Large Deviation Analysis for Quantum Security via Smoothing of Renyi Entropy of Order 2.
    M. Hayashi. Feb. 2012.
  21. Security analysis of epsilon-almost dual universal2 hash functions.
    M. Hayashi. Sep 2013.
  22. Tight exponential analysis of universally composable privacy amplification and its applications.
    M. Hayashi. Information Theory, IEEE Transactions on, 59(11):7728–7746, 2013.
  23. Information Theoretic Security for Encryption Based on Conditional Renyi Entropies.
    M. Iwamoto and J. Shikata. 2013.
  24. The Operational Meaning of Min- and Max-Entropy.
    R. König, R. Renner, and C. Schaffner. IEEE Trans. on Inf. Theory, 55(9):4337–4347, Sept. 2009.
  25. Inequalities for the Moments of the Eigenvalues of the Schrödinger Hamiltonian and Their Relation to Sobolev Inequalities.
    E. H. Lieb and W. E. Thirring. In The Stability of Matter: From Atoms to Stars, chapter III, pages 205–239. Springer Berlin Heidelberg, 2005.
  26. Generalized Entropic Uncertainty Relations.
    H. Maassen and J. Uffink. Phys. Rev. Lett., 60(12):1103–1106, Mar. 1988.
  27. Rényi Divergences and the Classical Capacity of Finite Compound Channels.
    M. Mosonyi. 2013.
  28. On the Quantum Rényi Relative Entropies and Related Capacity Formulas.
    M. Mosonyi and F. Hiai. IEEE Trans. on Inf. Theory, 57(4):2474–2487, Apr. 2011.
  29. Quantum Hypothesis Testing and the Operational Interpretation of the Quantum Renyi Relative Entropies.
    M. Mosonyi and T. Ogawa. Sept. 2013.
  30. Quantum Relative Rényi Entropies.
    M. Müller-Lennert. Master thesis, ETH Zurich, Apr. 2013.
  31. On Quantum Rényi Entropies: A New Generalization and Some Properties.
    M. Müller-Lennert, F. Dupuis, O. Szehr, S. Fehr, and M. Tomamichel. J. Math. Phys., 54(12):122203, June 2013.
  32. The Converse Part of The Theorem for Quantum Hoeffding Bound.
    H. Nagaoka. Nov. 2006.
  33. We use the convention that and .
  34. On Error Exponents in Quantum Hypothesis Testing.
    T. Ogawa and M. Hayashi. IEEE Trans. on Inf. Theory, 50(6):1368–1372, June 2004.
  35. Security of Quantum Key Distribution.
    R. Renner. PhD thesis, ETH Zurich, Dec. 2005.
  36. On Measures of Information and Entropy.
    A. Rényi. In Proc. Symp. on Math., Stat. and Probability, pages 547–561, Berkeley, 1961. University of California Press.
  37. Fundamental Bound on the Reliability of Quantum Information Transmission.
    N. Sharma and N. A. Warsi. Phys. Rev. Lett., 110(8):080501, Feb. 2013.
  38. Smooth entropies—A tutorial: With focus on applications in cryptography, Sept. 2012.
    M. Tomamichel. Available online:
  39. A Fully Quantum Asymptotic Equipartition Property.
    M. Tomamichel, R. Colbeck, and R. Renner. IEEE Trans. on Inf. Theory, 55(12):5840–5847, Dec. 2009.
  40. Uncertainty Relation for Smooth Entropies.
    M. Tomamichel and R. Renner. Phys. Rev. Lett., 106(11):110506, Mar. 2011.
  41. Leftover Hashing Against Quantum Side Information.
    M. Tomamichel, C. Schaffner, A. Smith, and R. Renner. IEEE Trans. on Inf. Theory, 57(8):5524–5535, Aug. 2011.
  42. Strong Converse for the Classical Capacity of Entanglement-Breaking and Hadamard Channels.
    M. M. Wilde, A. Winter, and D. Yang. June 2013.
  43. Finite Blocklength Bounds for Multiple Access Channels with Correlated Sources.
    H. Yagi. In Proc. IEEE ISITA, pages 377–381, 2012.
This is a comment super asjknd jkasnjk adsnkj
The feedback cannot be empty
Comments 0
The feedback cannot be empty
Add comment

You’re adding your first comment!
How to quickly get a good reply:
  • Offer a constructive comment on the author work.
  • Add helpful links to code implementation or project page.