IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO.

9, SEPTEMBER 2023 5869

Continuity of Quantum Entropic Quantities via

Almost Convexity
Andreas Bluhm , Ángela Capel , Paul Gondolf , and Antonio Pérez-Hernández

Abstract— Based on the proofs of the continuity of the con- for some subset S0 of the quantum states and some appropriate
ditional entropy by Alicki, Fannes, and Winter, we introduce in distance measure d such as the trace distance, for example.
this work the almost locally affine (ALAFF) method. This method Already in 1973, Fannes [1] proved that the von Neu-
allows us to prove a great variety of continuity bounds for the
derived entropic quantities. First, we apply the ALAFF method mann entropy is uniformly continuous and gave a concrete
to the Umegaki relative entropy. This way, we recover known dimension-dependent bound, which was later improved to a
almost tight bounds, but also some new continuity bounds for sharp version in [2] and [3]. Similar results in the line of
the relative entropy. Subsequently, we apply our method to the almost concavity for the von Neumann entropy were provided
Belavkin-Staszewski relative entropy (BS-entropy). This yields in [4], [5], and [6] or [7], among others. Another example of a
novel explicit bounds in particular for the BS-conditional entropy,
the BS-mutual and BS-conditional mutual information. On the concrete continuity estimate is the Alicki and Fannes inequal-
way, we prove almost concavity for the Umegaki relative entropy ity for the conditional entropy [8], which was subsequently
and the BS-entropy, which might be of independent interest. improved to an almost tight version by Winter [9]. Applica-
We conclude by showing some applications of these continuity tions of this kind of continuity bounds include, but are not
bounds in various contexts within quantum information theory.
limited to, entanglement measures [10] and the capacities of
Index Terms— Umegaki relative entropy, Belavkin-Staszewski quantum channels [11], [12]. We refer the reader to textbooks
relative entropy, continuity bounds. such as [13] for a thorough discussion of continuity bounds
and their applications.
I. I NTRODUCTION
The importance of the Alicki-Fannes result [8] goes beyond

E NTROPIC quantities have proven essential in charac-

terizing the information-processing capabilities both of
classical and quantum systems. As the real world cannot
quantifying the continuity of the conditional entropy, but their
method and its improved versions [9], [14], [15] work quite
generally for entropic quantities. This has been realized by
be measured to infinite precision, such quantities must be Shirokov, who has named this approach the Alicki-Fannes-
continuous to contain meaningful information about physical Winter method [16], [17]. We continue this line of work by
systems. Often, however, we do not only want to know whether generalising the Shirokov approach further to what we call
an entropic quantity is continuous but also to quantify this the almost locally affine (ALAFF) method. The aim of this
continuity. That means we are interested in estimating for an generalization is to apply it to entropic quantities beyond
entropic quantity f those derived from the Umegaki relative entropy [18], such
sup{|f (ρ) − f (σ)| : ρ, σ ∈ S0 , d(ρ, σ) ≤ ε}. (1) as the conditional entropy. In particular, we are interested in
the Belavkin-Staszewski relative entropy (BS-entropy) [19]
Manuscript received 31 October 2022; revised 12 April 2023; and its derived entropic quantities. As the Umegaki rel-
accepted 4 May 2023. Date of publication 19 May 2023; date of current ative entropy, it generalizes the Kullback-Leibler relative
version 18 August 2023. This work was supported in part by the Deutsche
Forschungsgemeinschaft (DFG), German Research Foundation, under Project entropy of classical systems [20], but it is less well studied
470903074/TRR 352. The work of Andreas Bluhm was supported in (see [21], [22], [23], [24], [25] for some recent results).
part by the European Research Council (ERC) Grant under Agreement
81876 and in part by the VILLUM FONDEN via the Villum Centre of
The BS-entropy and the related family of geometric Rényi
Excellence for the Mathematics of Quantum Theory (QMATH) Centre of divergences have recently found an application for estimat-
Excellence under Grant 10059. The work of Antonio Pérez-Hernández was ing channel capacities [26]. Moreover, generalizations of the
supported in part by the Spanish Ministerio de Ciencia e Innovación under
Grant PID2020-113523GB-I00 and in part by the Comunidad de Madrid
mutual information and other entropic quantities based on
under Grant QUITEMAD-CMS2018/TCS-4342. (Corresponding author: the BS-entropy have been defined [27], [28]. The BS-mutual
Paul Gondolf.) information has been instrumental in proving that the mutual
Andreas Bluhm is with Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG,
38000 Grenoble, France (e-mail: [email protected]).
information in one-dimensional quantum Gibbs states of finite-
Ángela Capel and Paul Gondolf are with Fachbereich Mathematik, Eber- range, translation-invariant Hamiltonians decays exponentially
hard Karls University of Tübingen, 72076 Tübingen, Germany (e-mail: fast [29] and that Davies generators in one dimension which
[email protected]; [email protected]). converge to those Gibbs states, in the commuting case, satisfy
Antonio Pérez-Hernández is with Departamento de Matemática Aplicada,
Universidad Nacional de Educación a Distancia, 28040 Madrid, Spain (e-mail: a positive modified logarithmic Sobolev inequality at every
[email protected]). temperature, and thus rapid mixing [30], [31].
Communicated by S. Mancini, Associate Editor for Quantum. A short version of the current manuscript, with new applica-
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TIT.2023.3277892. tions in the context of quantum entropic uncertainty relations,
Digital Object Identifier 10.1109/TIT.2023.3277892 has been published in [32].
0018-9448 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5870 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

II. M AIN R ESULTS those provided for quantities derived from the relative entropy,
This section summarizes the main results of this article. The to find applications in proving the continuity of various quan-
focus of this work is not so much on the continuity bounds tities in diverse fields related to quantum information theory.
themselves, but more on the introduction of the method which In particular, we provide here a number of applications of our
allows deriving all of them in a systematic way (section IV). results in the context of quantum hypothesis testing (section
Our approach is summarized in fig. 1. For a given divergence, VII-A), to show that states that are hard to discriminate
in this paper either the Umegaki relative entropy [18] or have almost the same performance in terms of hypothesis
the BS-entropy [19], we need to prove two properties: its testing, as well as in quantum thermodynamics (section VII-
(joint) convexity and its almost (joint) concavity. Both of these B), to show continuity of the distillable athermality. We also
properties, under certain conditions on the remainder function, reprove that a state is an approximate quantum Markov chain
then directly translate into almost local affinity (definition 4.5) if and only if it is close to being recovered by the Petz recovery
of the entropic quantities derived from the divergence at hand map (section VII-C), and use our most general continuity
on a suitably defined subset S 0 of S(H). Serving as input to bounds for the relative entropy to obtain bounds for the
the ALAFF method, the remainder estimates get translated into difference between the relative entropy and the BS-entropy of
continuity bounds for said quantities. The entropic quantities two quantum states (section VII-D). Additionally, we show a
include for example, versions of the conditional entropy and new result of weak quasi-factorization for the relative entropy,
the (conditional) mutual information, as defined in fig. 1. The i.e. with an additive error term and no multiplicative error
necessity of S 0 as a restriction of S(H) becomes obvious term (section VII-E). Finally, we include continuity bounds
when trying to prove continuity bounds for the Umegaki for the relative entropy of entanglement as well as the anal-
relative entropy, for example. It is known not to be continuous ogously defined BS-entropy of entanglement (section VII-F),
on the set of all pairs of states (ρ, σ), which makes a careful and subsequently lift these results to show continuity of the
choice of S 0 inevitable. To this end, we define s-perturbed Rains information induced by the relative and the BS-entropy
∆-invariant convex subsets of S(H) (definition 4.3) for which respectively (section VII-G).
we can show that the ALAFF method works and which are
general enough to capture all situations of interest. For the III. P RELIMINARIES
formal statement of the ALAFF method, we refer the reader A. Notation and Basic Concepts
to theorem 4.6.
Thus, we are left with proving convexity and almost We denote a Hilbert space by H which, throughout this
concavity for the divergences we are interested in, paper is assumed to be finite. The dimension of such a Hilbert
namely the Umegaki relative entropy (section V) and the space will be called d and for its elements, we use |φ⟩, |ψ⟩
Belavkin-Staszewski entropy (section VI), and deriving the and |i⟩ for i ∈ N, possibly with additional indices. If we are
precise continuity estimates. For the convexity, we can rely on concerned with a bipartite or tripartite system, we will always
well-known results from the literature both for the Umegaki use capital letters in the index to identify objects associated
relative entropy [33] and the BS-entropy [23], [24]. For the with the respective subsystems. If we have, for example, the
Umegaki relative entropy, given by bipartite space H = HA ⊗ HB and consider the dimension of
HA , we write dA .
D(ρ∥σ) := Tr [ρ(log ρ − log σ)] if ker σ ⊆ ker ρ , (2) The set of (bounded) linear operators on a Hilbert space H
or + ∞ otherwise, we prove almost concavity in theorem 5.1 is B(H) and the subspace of positive semi-definite operators
and find that it is tight. The application of the ALAFF method with trace one, i.e., the quantum states or density matrices,
then allows us to recover in section V-B the almost tight results is denoted by S(H). If we want to restrict this set even
for the conditional entropy [9] and the mutual and conditional further, we indicate this with a subindex. Thus, the set of
mutual information (which can be derived from the conditional positive definite quantum states becomes S + (H), or if we
entropy [13]), but also to derive in section V-C new versions want to restrict moreover to the set of quantum states that
of what we call divergence bounds [34], [35], [36], [37], have minimal eigenvalue greater than m, we write S ≥m (H).
i.e. bounds on D(ρ||σ) in terms of 21 ∥ρ − σ∥1 . Furthermore, On the set of quantum states as well as on the set of self-
our technique produces a new result, which is the uniform adjoint operators, the relation ≤ is meant to be the partial
continuity of the relative entropy itself (in both arguments, order in the Löwner sense. That is, ρ ≥ σ if and only if ρ − σ
on a suitable set S0 ), as well as an explicit continuity bound. is positive semidefinite.
For the BS-entropy, given by We use Tr [ · ] for the usual matrix trace and ∥ · ∥1 = Tr [| · |]
h i and ∥ · ∥∞ to denote the trace norm and the spectral norm
D(ρ∥σ)
b := Tr ρ log(ρ1/2 σ −1 ρ1/2 ) if ker σ ⊆ ker ρ , on B(H), respectively. Quantum states in general are denoted
by lower Greek letters such as ρ, σ and τ , for example. For
(3)
Hermitian operators in B(H) we usually use upper Latin letters
or + ∞ otherwise, we prove the almost concavity in theo- such as X, Y . For any such X, we denote by [X]+ and [X]−
rem 6.3. The ALAFF method yields novel explicit bounds in its positive and negative parts, respectively.
particular for the BS-conditional entropy, the BS-mutual and As we later want to formally control the dependence on the
BS-conditional mutual information that we gather in section states ρ and σ that are given as arguments to the divergences,
VI-B. We expect these new continuity bounds, as well as we further introduce H × H the Cartesian product of the

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5871

Fig. 1. A flow chart demonstrating how convexity and almost concavity of a divergence can be used to obtain uniform continuity and explicit continuity
bounds on entropic quantities derived from that divergence.The subscripts of the functions fD,1/2 and fDB stand for divergence first, second argument and
divergence bound respectively.

Hilbert space H with itself. Moreover, on a bipartite system the relative entropy [23]. We further note that both can also
HAB = HA ⊗ HB , we set ρA to be the state on HA that be defined in terms of positive semi-definite operators A, B
ρ ∈ S(HAB ) is mapped to under the partial trace with (without normalisation), by just replacing ρ with A and σ
respect to the subsystem B which is a completely positive with B. We make use of this alternative definition when we
trace-preserving (CPTP) map. Furthermore, we denote by 1A define the conditional entropy and the BS-conditional entropy,
the identity matrix on A and, in a slight abuse of notation, for example. Using this notation we can write the conditional
we denote by TrA [·] both the partial trace with respect to A as entropy of ρ as
well as the complemented map on HAB by tensorizing with
1A . Hρ (A|B) := S(ρAB ) − S(ρB ) = −D(ρAB ∥ 1A ⊗ρB ) , (7)

with the last equality being a straightforward calculation. The

B. Entropies and Derived Quantities
subscript AB in ρAB = ρ just emphasises the fact that ρ stems
The von Neumann entropy of ρ ∈ S(H) is given by from S(HA ⊗ HB ) and to distinguish it from its partial trace
ρB , for example. It is noteworthy that the conditional entropy
S(ρ) := −Tr [ρ log(ρ)] . (4)
admits the following variational expression
For two quantum states ρ, σ ∈ S(H), their (Umegaki)
relative entropy [18] is defined as Hρ (A|B) = max − D(ρAB ∥ 1A ⊗σB ). (8)
σB ∈S(HB )
(
Tr [ρ log ρ − ρ log σ] if ker σ ⊆ ker ρ ,
D(ρ∥σ) := (5) Furthermore, in a similar manner as for the conditional
+∞ otherwise , entropy, one obtains the representation of the mutual informa-
and their Belavkin-Staszewski (BS) entropy [19] by tion in terms of the von Neumann entropy and the conditional
(  entropy
Tr ρ log ρ1/2 σ −1 ρ1/2 if ker σ ⊆ ker ρ ,

D(ρ∥σ) :=
b
+∞ otherwise . Iρ (A : B) := S(ρA ) + S(ρB ) − S(ρAB )
(6) = S(ρA ) − Hρ (A|B) = D(ρAB ∥ρA ⊗ ρB ) . (9)

In the event of ρ and σ commuting, the two entropies Finally, on a tripartite system H = HA ⊗ HB ⊗ HC the
coincide. Otherwise, the BS-entropy is strictly larger than conditional mutual information of a state ρ ∈ S(H) is given

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5872 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

by We have numerical results that suggest that the inequality

in the eq. (15) is strict, at least in some cases. A plot of those
Iρ (A : B|C) := S(ρAC ) + S(ρBC ) − S(ρC ) − S(ρABC )
numerics can be found in section A. Moreover, we will indeed
= Hρ (A|C) − Hρ (A|BC) formally show that both quantities are different in general in
= Iρ (A : BC) − Iρ (A : C) . (10) remark 7.17.
The last equalities are again straightforward. One easily IV. F ROM A LMOST C ONVEXITY TO C ONTINUITY B OUNDS
checks that both the mutual information and the conditional
As depicted in fig. 1, our approach is based on the convexity
mutual information are symmetric under the exchange of the
and almost concavity of a divergence. More precisely, it is
A and B system.
based on its joint convexity and almost joint concavity, but for
Let us proceed now to introduce the analogous quantities
the sake of better readability, we will just speak of convexity
from the BS instead of the relative entropy. In this framework,
and almost concavity.
we cannot construct them as sums and differences of von
It is immediately clear what is meant by convexity and this
Neumann entropies, which, for every BS-entropic quantity,
property is often even a defining property of a divergence [38]
gives rise to a zoo of different possible definitions. Some of
or a direct consequence thereof1 [39, Proposition 4.2]. The
them have already appeared before in [22], [27], and [28].
almost (joint) concavity, however, needs yet to be defined.
For a bipartite state ρ ∈ S(HA ⊗ HB ), inspired by the notion
Definition 4.1 (Almost (Joint) Concavity of a Divergence):
of conditional entropy, we define the BS-conditional entropy
A divergence D(·∥·) is called almost (jointly) concave on a
as
convex set S 0 ⊆ S(H) × S(H) if, for (ρ1 , σ1 ), (ρ2 , σ2 ) ∈ S 0 ,
Hb ρ (A|B) := −D(ρ
b AB ∥ 1A ⊗ρB ) , (11) there exists a continuous function f : [0, 1] → R with f (0) =
and building on the mutual information, we define the BS- f (1) = 0 such that, for all p ∈ [0, 1],
mutual information as D(ρ∥σ) ≥ p D(ρ1 ∥σ1 ) + (1 − p) D(ρ2 ∥σ2 ) − f (p) (16)
b AB ∥ρA ⊗ ρB ) .
Ibρ (A : B) := D(ρ (12)
holds. Here, ρ = pρ1 + (1 − p)ρ2 and σ = pσ1 + (1 − p)σ2 .
Finally, the analogue of the conditional mutual information It is important to emphasise that f in general depends on the
in this setting is a more subtle matter. Indeed, two natural ways states involved.
to construct such a quantity would be either as a difference Remark 4.2: We note that the definition of almost concavity
of BS-conditional entropies or of BS-mutual information, presented above is not itself a very strong property. For
as shown in eq. (10), which in general do not coincide. Given example, one could just choose f to be the remainders that
ρABC ∈ S(HA ⊗ HB ⊗ HC ): give equality in eq. (16). It is the behaviour of the remainder
• We define the (one-sided) BS-conditional mutual infor- functions that is pivotal, i.e., it becomes independent of ρi , σi ,
mation (os BS-CMI in short) by i = 1, 2 under certain restrictions on the states, e.g. requiring
that σi is a marginal of ρi .
Ibρos (A : B|C) := Hb ρ (A|C) − Hb ρ (A|BC)
Our approach, therefore, does not only need joint convexity
b ABC ∥ 1A ⊗ρBC ) − D(ρ
= D(ρ b AC ∥ 1A ⊗ρC ) . (13) but a well-behaved remainder function. If we find such a func-
• We define the (two-sided) BS-conditional mutual infor- tion and combine it with the boundedness of the divergence (or
mation (ts BS-CMI in short) by underlying entropic quantity), ALAFF directly gives uniform
continuity through explicit continuity bounds.
Ibts (A : B|C) := Ibρ (A : BC) − Ibρ (A : C)
ρ As we already discussed in the introduction, the predecessor
b ABC ∥ρA ⊗ ρBC ) − D(ρ
= D(ρ b AC ∥ρA ⊗ ρC ) . (14) of ALAFF was developed and used by Alicki and Fannes [8],
as well as Winter [9], to prove uniform continuity and give an
Note that both notions are clearly non-negative, as a conse-
explicit continuity bound for the conditional entropy. Shirokov
quence of the data processing inequality for the BS-entropy.
then noticed the potential beyond this specific application
In this project, we focus for simplicity on the first definition,
and moulded a method that can be applied to functions
i.e. the one-sided one. We will therefore drop the “os” notation,
defined on convex and ∆-invariant subsets of S(H) [16], [17].
as there is no possible confusion.
Independently, similar techniques were already used in [14].
Let us emphasize at this stage that the difference between
In short, ∆-invariance means that for two elements their
the aforementioned two definitions of BS-conditional mutual
normalised positive and negative part again lies in the set
information is partly related to the pathological behaviour
(see also Definition 4.3). This definition of ∆-invariance will,
of the BS-entropy with respect to continuity in general, and
however, turn out to be a limitation when trying to prove the
more specifically to the fact that the BS-conditional entropy
uniform continuity of the relative entropy, while in the case
is discontinuous on the set of positive semi-definite quantum
of the BS-entropy, it is unfitting even from the beginning,
states (cf. proposition 6.7). We suspect that as a consequence
i.e., even for the BS-conditional entropy. The problem is due
thereof, the variational definition of the BS-conditional entropy
to ∆-invariance being a rather strong property that sets like
(generalizing eq. (8)) does not agree with the one we have
given in eq. (11), namely 1 Some authors define divergences as functions on two density operators

b ρ (A|B) ≤ fulfilling a data processing inequality; however, note that convexity for a
H sup − D(ρ b AB ∥ 1A ⊗σB ) . (15) divergence implies a data processing inequality and follows from it together
σB ∈S(HB ) with additional properties, as shown in [38, Corollary 4.7].

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5873

S ≥m (H) or {(ρ, σ) : ker σ ⊆ ker ρ} do not have. Yet, those In this case, we have for ε ∈ (0, 1]
sets, or modified versions thereof, are the relevant sets for the
relative and, in particular, the BS-entropy. sup |f (ρ) − f (σ)|
ρ,σ∈S 0
1 ∥ρ−σ∥ ≤ε
In light of those problems and in an effort to make our 2 1

approach as general as possible, we propose the almost locally ε 1−s + ε max  ε 

≤ Cfs + Ef ,
affine (ALAFF) method, a generalisation of the Alicki-Fannes- 1−s 1−s 1−s + ε
Winter-Shirokov method that reduces to one implication of the (20)
former in a special case. First of all, we define a perturbed ver-
sion of the ∆-invariant subset, with the perturbation controlled with Efmax : [0, 1) → R,
by a parameter s. 
Ef (t)

Definition 4.3 (Perturbed ∆-Invariant Subset): Let s ∈ p 7→ Efmax (p) = (1 − p) max : 0 ≤ t ≤ p , (21)
1−t
[0, 1). A subset S 0 ⊆ S(H) is called s-perturbed ∆-invariant,
if for ρ, σ ∈ S 0 with ρ ̸= σ there exists τ ∈ S(H) such that where Ef = af +bf . Note that on ε ∈ (0, 1−s] Ef and Efmax
the two states coincide.
Proof: Let s ∈ [0, 1) and ε ∈ (0, 1]. Let further
∆± (ρ, σ, τ ) = sτ + (1 − s)ε−1 [ρ − σ]± (17)
ρ, σ ∈ S 0 with 12 ∥ρ − σ∥1 = ε. Then by the property of
1
lie again in S 0 . Here ε := ∥ρ − σ∥1 and [A]± denote the
2 s-perturbed ∆-invariance there exists τ ∈ S(H) such that
negative and positive part of a self-adjoint operator, respec- γ± := ∆± (ρ, σ, τ ) ∈ S 0 defined as in eq. (17). For every
tively. For s = 0, we recover the definition of ∆-invariant such γ± with a representation in terms of ρ, σ ∈ S 0 and a
subset used in [17]. τ ∈ S(H) we have that
We want to give the reader some intuition about those s- 1−s ε
perturbed ∆-invariant sets. ω∗ = ρ+ γ−
1−s + ε 1−s + ε
Remark 4.4: 1−s ε
1) Let S 0 ⊆ S(H) be s-perturbed ∆-invariant convex set. = σ+ γ+ , (22)
1−s + ε 1−s + ε
Then for t ∈ [s, 1) it is t-perturbed ∆-invariant as
well. In particular, being 0-perturbed is the strongest which can be easily checked by inserting the explicit form
condition. of γ± and using that [ρ − σ]+ − [ρ − σ]− = ρ − σ. Now
2) If S 0 ⊆ S(H) has non-empty interior with respect to ω ∗ ∈ S 0 as S 0 is convex, which allows us to evaluate f at
the 1-norm, then it is s-perturbed for some s ∈ [0, 1). ω ∗ and use eq. (18) for both of the representations we have
3) If S 0 ⊆ S(H) is s-perturbed ∆-invariant containing for the state in question. This gives us
more than one state, then there exist ρ, σ ∈ S 0 with −af (p) ≤ f (ω ∗ ) − (1 − p)f (ρ)−pf (γ− ) ≤ bf (p) ,
1
2 ∥ρ − σ∥1 = 1 − s. This follows directly from the −af (p) ≤ f (ω ∗ ) − (1 − p)f (σ)−pf (γ+ ) ≤ bf (p) , (23)
definition.
ε
In order to get well-behaved remainder functions, we define where we set p = p(ε) = 1−s+ε for better readability. Note
1
a stronger property that we call “almost local affinity”. that p ∈ (0, 2−s ] ⊆ [0, 1) as ε ∈ (0, 1] and s ∈ [0, 1) and
Definition 4.5 (Almost Locally Affine (ALAFF) Function): further that p(ε) is monotone with respect to ε. We recombine
Let f be a real-valued function on the convex set S 0 ⊆ S(H), the above to get
fulfilling
(1 − p)(f (ρ)−f (σ))
−af (p) ≤ f (pρ + (1 − p)σ)−pf (ρ) − (1 − p)f (σ) ≤ bf (p)
≤ p(f (γ+ ) − f (γ− )) + af (p) + bf (p) ,
(18)
(1 − p)(f (σ)−f (ρ))
for all p ∈ [0, 1] and ρ, σ ∈ S 0 . The functions af : [0, 1] → R ≤ p(f (γ− ) − f (γ+ )) + af (p) + bf (p) .
and bf : [0, 1] → R are required to vanish as p → 0+ , to be
(24)
non-decreasing on [0, 21 ], continuous in p and independent of
ρ, σ ∈ S 0 . We then call f an almost locally affine (ALAFF) Those two inequalities immediately give us
function.
The notion of almost locally affine functions has appeared (1 − p)|f (ρ) − f (σ)| ≤ p|f (γ+ ) − f (γ− )| + (af + bf )(p) .
previously in the literature, also under the name “approximate (25)
affinity” (see e.g. [40]). We can now formulate the following
theorem, whose proof is inspired by Shirokov [17]. If we now insert Ef = af + bf , we obtain
Theorem 4.6 (Almost Locally Affine (ALAFF) Method): Let p 1
|f (ρ) − f (σ)| ≤ |f (γ+ ) − f (γ− )| + Ef (p) .
s ∈ [0, 1) and S 0 ⊆ S(H) be an s-perturbed ∆-invariant 1−p 1−p
convex subset of S(H) containing more than one element. (26)
Let further f be an ALAFF function. We then find that f is
uniformly continuous if In the case that Cfs is finite, we can take the supremum over
all ρ, σ ∈ S 0 with 12 ∥ρ − σ∥1 = ε of the last equation and
Cfs := sup |f (ρ) − f (σ)| < +∞. (19)
ρ,σ∈S 0 even extend to 21 ∥ρ − σ∥1 ≤ ε in two steps. The first step is
1 1
1 ∥ρ−σ∥ =1−s
2 1 upper bounding 1−p Ef (p) with 1−p Efmax (p) and the second

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5874 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

1
one using that 1−p Efmax (p) is engineered to be non-decreasing and p ∈ [0, 1]. Then, for ρ = pρ1 + (1 − p)ρ2 and σ =
ε 1 pσ1 + (1 − p)σ2 ,
on [0, 1) and thereby for the specific p = 1−s+ε ∈ [0, 2−s ]⊂
[0, 1), is non-decreasing in ε as well. Since the γ+ and
D(ρ∥σ) ≥ pD(ρ1 ∥σ1 ) + (1 − p)D(ρ2 ∥σ2 )
γ− created from ρ and σ obviously fulfill γ± ∈ S 0 and
1 1
2 ∥γ+ − γ− ∥1 = 1 − s, we immediately get the upper bound − h(p) ∥ρ1 − ρ2 ∥1 − fc1 ,c2 (p) . (30)
in eq. (20). The reduction of Efmax to Ef on ε ∈ (0, 1 − s] 2
follows immediately from Ef being non-decreasing on [0, 12 ] Here,
meaning further that Efmax inherits the vanishing property of
h(p) = −p log(p) − (1 − p) log(1 − p) ,
Ef as p → 0+ . This directly translates into Efmax (p(ε)) → 0 if
ε → 0+ , hence concluding the proof of uniform continuity. fc1 ,c2 (p) = p log(p + (1 − p)c1 )
The method presented in theorem 4.6 is named the “ALAFF + (1 − p) log((1 − p) + pc2 ) , (31)
method” to highlight the required ALAFF property necessary,
with the first one being the binary entropy. The constants in
for this technique to be applicable. We will refer to this
fc1 ,c2 are non-negative real numbers and are given by
theorem by that name in subsequent sections.
Remark 4.7: For s = 0, one recovers one implication of Z∞ h it−1 −it−1 i
the method by Shirokov, i.e., the definitions for perturbed ∆- c1 := dtβ0 (t)Tr ρ1 σ1 2 σ2 σ1 2 < ∞,
invariance and ∆-invariance coincide, Efmax reduces to Ef on −∞
the relevant domain ε ∈ [0, 1], and eq. (20) becomes Z∞ h it−1 −it−1 i
 ε  c2 := dtβ0 (t)Tr ρ2 σ2 2 σ1 σ2 2 < ∞. (32)
sup |f (ρ) − f (σ)| ≤ Cf⊥ ε + (1 + ε)Ef (27)
ρ,σ∈S 0 1+ε −∞
1 ∥ρ−σ∥ ≤ε
2 1
Here, β0 is a probability density on R (see eq. (36) for a
with concrete expression). It is noteworthy that f1,1 (·) = 0 and
fc1 ,c2 (0) = fc1 ,c2 (1) = 0.
Cf0 = sup |f (ρ) − f (σ)| Proof: It is clear that S ker is a convex set and that the
ρ,σ∈S 0
1 ∥ρ−σ∥ =1
2 1 bound holds trivially for p = 0 and p = 1. Hence let p ∈ (0, 1)
= sup |f (ρ) − f (σ)| =: Cf⊥ , (28) and (ρ1 , σ1 ), (ρ2 , σ2 ) ∈ S ker in the following. We find that
ρ,σ∈S 0
Tr[ρσ]=0
pD(ρ1 ∥σ1 ) + (1 − p)D(ρ2 ∥σ2 ) − D(ρ∥σ)
as states with maximal trace distance have orthogonal support. = −pS(ρ1 ) − (1 − p)S(ρ2 ) + S(ρ)
In the next sections, we will use theorem 4.6 together + (1 − p)Tr [ρ2 (log σ − log σ2 )]
with the almost concavity of the relative entropy and the
+ pTr [ρ1 (log σ − log σ1 )]
BS-entropy, respectively, to derive a plethora of results of
1
uniform continuity and continuity bounds for entropic quan- ≤ h(p) ∥ρ1 − ρ2 ∥1 + fc1 ,c2 (p) , (33)
tities defined through them. Depending on the case, we will 2
sometimes have to employ the whole machinery devised in where we split the relative entropies and used that the von
theorem 4.6, whereas at other times the simplification provided Neumann entropy fulfils [7, Theorem 14]
in remark 4.7 will be enough. 1
S(ρ) ≤ ∥ρ1 − ρ2 ∥1 h(p) + pS(ρ1 ) + (1 − p)S(ρ2 ) . (34)
2
V. A LMOST C ONCAVITY AND C ONTINUITY B OUNDS FOR Furthermore, we upper bound the remaining terms by
THE U MEGAKI R ELATIVE E NTROPY fc1 ,c2 (p), estimating the two separately. We will only demon-
strate the derivation for the second term, as it is completely
In this section, we apply the ALAFF method introduced
analogous to the first one. We have
in section IV for the particular case of the relative entropy,
as well as some other entropic quantities derived from it. pTr [ρ1 (log(σ) − log(σ1 ))]
All the results provided in this section are summarized in = pTr [exp(log(ρ1 ))(log(σ) − log(σ1 ))]
fig. 2.
≤ p log Tr [exp (log(ρ1 ) + log(σ) − log(σ1 ))]
Z∞ h i
1+it 1−it
A. Almost Concavity for the Relative Entropy ≤ p log dt β0 (t) Tr ρ1 (σ1−1 ) 2 σ(σ1−1 ) 2 . (35)
The (joint) convexity of the relative entropy is a −∞
well-established result with proofs found for example in [13]. The first estimate follows immediately using the well-known
In this section, we complement this result with almost con- Peierls-Bogoliubov inequality [41]. The second one involves a
cavity and further prove that the bound we obtain is tight. generalisation of the Araki-Lieb-Thirring inequality [42], [43]
Theorem 5.1 (Almost Concavity of the Relative Entropy): by Sutter et al. [44, Corollary 3.3] with
Let (ρ1 , σ1 ), (ρ2 , σ2 ) ∈ S ker with
π 1
β0 (t) = (36)
S ker := {(ρ, σ) ∈ S(H) × S(H) : ker σ ⊆ ker ρ} (29) 2 cosh(πt) + 1

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5875

Fig. 2. In this flow chart we collect the main results from this chapter, starting with the almost concavity of the relative entropy, which together with the ALAFF
method outputs a collection of continuity bounds for related entropic quantities. For the convexity and almost concavity, we are setting ρ = pρ1 +(1−p)ρ2 and
σ = pσ1 + (1 − p)σ2 , with p ∈ [0, 1]. We denote by m e σ the minimal non-zero eigenvalue of σ. The specific bounds obtained for the relative entropy fixing
the first argument and in the general case (modifying both arguments) are omitted due to their technicality.

a probability density on R. In the above steps, i.e. eq. (35), reduces to the previously known bounds for quantities derived
we relied on ρ1 , σ1 and σ to be full rank. If this is not the case from the relative entropy, e.g. the von Neumann entropy or
one obtains the same result, however, the procedure is more the conditional entropy, and it is almost tight in general.
involved. A thorough discussion can be found in section B. To illustrate that, we provide now two propositions that put
Note here that in the most general case, ·−1 in the RHS of the almost concavity of the relative entropy into perspective.
eq. (35) is the Moore-Penrose pseudoinverse. The trace in the Proposition 5.2 (Almost Concavity Estimate of the Rel-
integral can now be estimated for each t by ative Entropy Is Well Behaved): The function fc1 ,c2 +
h it−1 −it−1 i
h 12 ∥ρ1 − ρ2 ∥1 obtained in theorem 5.1 is well behaved in the
Tr ρ1 σ1 2 σ σ1 2 following sense: For j = 1, 2 and (ρj , σj ) ∈ S ker , we have
h it−1 −it−1 i the following:
= p + (1 − p)Tr ρ1 σ1 2 σ2 σ1 2 . (37)
1) If σ1 = σ2 , then c1 = c2 = 1, resulting in fc1 ,c2 +
1
Here, we just split σ and used that Tr [ρ1 ] = 1. To see that c1 < 2 ∥ρ1 − ρ2 ∥1 h ≤ h.
2) If for m e > 0 we have mρ e j ≤ σj , then fc1 ,c2 +
∞, we upper bound σ2 by 1 and σ1−1 by m e −1
σ1 1 where m e σ1 is
h 12 ∥ρ1 − ρ2 ∥1 ≤ fm e −1 + h.
e −1 ,m
the smallest non-zero eigenvalue of σ1 . This can be done, since
3) If H = HA ⊗ HB is a bipartite space and furthermore
ker σ1 ⊆ ker ρ1 . We end up with c1 ≤ m e −1
σ1 < ∞. Inserting
σj = d−1 1
A 1A ⊗ρj,B , then fc1 ,c2 + h 2 ∥ρ1 − ρ2 ∥1 ≤ h.
eq. (37) into eq. (35), we obtain the first part of fc1 ,c2 (p) and 1
4) For m1 , m2 ≥ 1 we find that both p 7→ 1−p fm1 ,m2 (p)
repeating the steps for (1 − p)Tr [ρ2 (log(σ) − log(σ2 ))] the 1
second one as well. This concludes the proof. and p 7→ 1−p h(p) are non-decreasing on [0, 1).
We remark that eq. (30) provides a result of almost con- We hence find that in the cases listed above the bound
cavity for the relative entropy in the sense of definition 4.1. becomes independent of the states and further that the remain-
Indeed, the additive “correction” term obtained for such an der functions have a desirable non-decreasing property. The
inequality to hold behaves well enough, in the sense that it proof is straightforward and can be found in section C.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5876 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

Remark 5.3: The different cases discussed in proposi- This gives us

tion 5.2 are used in the following to find almost concavity
results with a function that does not depend on the specifics pD(ρ1 ∥σ1 ) + (1 − p)D(ρ2 ∥σ2 ) − D(ρ∥σ)
of the states involved, as necessary for applying the ALAFF
 1 − t
= h(p) + p log p + (1 − p)
method. t
 1 − t
• If σ1 = σ2 , we are reducing eq. (30) to a result + (1 − p) log (1 − p) + p . (42)
of almost concavity only in the first input. This case t
was addressed in [40], where they obtained h(p) as a As [ρi , σj ] = 0 for i, j = 1, 2 and further [ρi σj , σi ] = 0 we
correction for almost concavity, a bound we are tightening find that the constants in theorem 5.1 are given by
here. Moreover, this case will yield a continuity bound
 1−t
for the relative entropy with fixed second input as shown ci = Tr ρi σi−1 σj =

, (43)
in corollary 5.9. t
• Point 3 of proposition 5.2 can be interpreted as a result of for i, j = 1, 2, i ̸= j. Since ρ1 and ρ2 orthogonal we get
almost convexity for the conditional entropy. Moreover, 1
2 ∥ρ1 − ρ2 ∥1 = 1. We hence obtain the RHS of eq. (42) from
it will yield a continuity bound for the conditional entropy the almost concavity estimate in eq. (30). This concludes the
in corollary 5.5. Since the latter result is almost tight, claim.
this shows the good behaviour of the bound obtained in
theorem 5.1. B. Reduction to Almost Tight Previously-Known Continuity
• Point 2 of proposition 5.2 is the most general setting for Bounds for the Relative Entropy
full-rank states σj , with j = 1, 2, and will be essential
for deriving the most general continuity bounds for the In this section, we will show that a number of almost tight
relative entropy in theorem 5.13. previously-known continuity bounds for quantities derived
from the relative entropy can be obtained as corollaries of
Finally, we conclude this subsection with some discussion the results of almost concavity in theorem 5.1 and proposi-
of our almost concave bound. tion 5.2 in combination with the results concerning the ALAFF
Proposition 5.4 (Almost Concavity Estimate of the Rela- method, i.e. theorem 4.6 and remark 4.7.
tive Entropy Is Tight): The bound presented in theorem 5.1
1) Uniform Continuity for the Conditional Entropy: Let us
is tight. More specifically, there are some density opera-
first consider a bipartite space and the conditional entropy of a
tors ρ1 , ρ2 , σ1 , σ2 on S(H) which saturate the inequality in
density matrix with respect to one of the subsystems. Note that,
eq. (30).
in this case, we are able to prove a result of uniform continuity
Proof: We can assume that the dimension of the under-
for any positive semidefinite matrix (with trace one), but we
lying Hilbert space is dH ≥ 2. We then find two orthonormal
do not require positive definiteness. The following coincides
states |0⟩ , |1⟩ ∈ H that we use to create
with the result of Winter [9], which he proved to be almost
tight.
ρ1 := |0⟩⟨0| , Corollary 5.5 (Uniform Continuity of the Conditional
ρ2 := |1⟩⟨1| , Entropy): The conditional entropy over the bipartite Hilbert
σ1 := t |0⟩⟨0| + (1 − t) |1⟩⟨1| , space H = HA ⊗ HB is uniformly continuous on S 0 = S(H)
σ2 := (1 − t) |0⟩⟨0| + t |1⟩⟨1| , (38) and for ρ, σ ∈ S 0 with 12 ∥ρ − σ∥1 ≤ ε ≤ 1, it holds that
 ε 
|Hρ (A|B) − Hσ (A|B)| ≤ 2ε log dA + (1 + ε)h .
for t ∈ (0, 1). We find, as of the orthonormality, that for p ∈ 1+ε
[0, 1] and (44)

ρ := pρ1 + (1 − p)ρ2 , Proof: First of all, S 0 is clearly 0-perturbed ∆-invariant.

σ := pσ1 + (1 − p)σ2 , (39) Setting f (·) = H· (A|B), we find that it is ALAFF with
aH· (A|B) = 0 as H· (A|B) is concave, and bH· (A|B) = h since
the result in theorem 5.1 becomes independent of the states as
the relative entropy between the states given by the convex
we go to H· (A|B) using point 3 of proposition 5.2. Finally,
combinations takes the value
we find that
D(ρ∥σ) = Tr [ρ log(ρ) − ρ log(σ)] Cf⊥ = sup |Hρ (A|B) − Hσ (A|B)| ≤ 2 log dA , (45)
ρ,σ∈S 0
= −h(p)−p log(pt + (1 − p)(1 − t)) Tr[ρσ]=0

− (1 − p) log((1 − p)t + p(1 − t)) , (40)

where we used − log dX ≤ H· (X|Y ) ≤ log dX shown, for
example, in [13]. Using theorem 4.6 in the form of remark 4.7,
and
we can infer the claimed continuity bound.
2) Uniform Continuity for the Mutual Information: For the
D(ρ1 ∥σ1 ) = − log(t) , mutual information, it is straightforward to derive a continuity
D(ρ2 ∥σ2 ) = − log(t) . (41) bound for such a quantity just by combining the bounds

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5877

of [2] and [9] for the von Neumann and conditional entropy, B|C) = H· (A|C) − H· (A|BC). With this representation,
respectively: we can immediately conclude that I· (A : B|C) is ALAFF
with af = h and bf = h. Finally, we have that
|Iρ (A : B) − Iσ (A : B)| ≤ 3ε log min{dA , dB }
Cf⊥ = sup |Iρ (A : B|C) − Iσ (A : B|C)|
 ε 
+ 2(1 + ε)h , (46) ρ,σ∈S 0
1+ε Tr[ρσ]=0

where ε := 12 ∥ρ − σ∥1 . For an early version, see [45, ≤ sup Iρ (A : B|C)

ρ∈S 0
Exercise 5.40]. The multiplicative factor in the first
√ term of the = sup Hρ (A|BC) − Hρ (A|C)
right-hand side was subsequently improved to 2 2 in [46] and ρ∈S 0
to 2 in [16]. Moreover, we can adapt corollary 5.5 to obtain the ≤ 2 log dA = 2 log min{dA , dB } , (50)
following bound on the mutual information, which coincides
with the tightest previously-known continuity bound for the as the conditional mutual information is non-negative and
mutual information (see e.g. [16]). again − log dX ≤ H· (X|Y ) ≤ log dX . Using theorem 4.6
Corollary 5.6 (Continuity Bound for the Mutual Informa- in the form of remark 4.7, we can conclude the claim and
tion): The mutual information on a bipartite Hilbert space obtain the given continuity bound.
H = HA ⊗ HB is uniformly continuous on S 0 = S(H) and
for ρ, σ ∈ S 0 with 21 ∥ρ − σ∥1 ≤ ε ≤ 1, we find that C. New Continuity Bounds for the Relative Entropy
|Iρ (A : B) − Iσ (A : B)| ≤ 2ε log min{dA , dB } Now, we prove some new continuity bounds for further
 ε  quantities derived from the relative entropy as a consequence
+ 2(1 + ε)h . (47)
1+ε of the results of almost concavity in theorem 5.1 and proposi-
Proof: First of all, S 0 is clearly 0-perturbed ∆-invariant. tion 5.2 in combination with the results concerning the ALAFF
With f (·) = I· (A : B) = S(·A ) − H· (A|B) one can method, i.e. theorem 4.6 and remark 4.7. All bounds in this
immediately conclude almost local affinity of I· (A : B) section can be simplified using the following lemma:
as S(·A ) is concave and fulfills eq. (34) and −H· (A|B) is Lemma 5.8: Using the notations introduced in theorem 5.1
almost locally affine with a−H· (A|B) = 0 and b−H· (A|B) = h. and remark 5.3, we have the following estimates for the error
Combined we get af = h and bf = h. We further have that bounds obtained in all results of this section:
 ε  √
Cf⊥ = sup |Iρ (A : B) − Iσ (A : B)| (1 + ε)h ≤ 2ε , (51)
ρ,σ∈S 0
1+ε
Tr[ρσ]=0

lm  ε  2 log2 me −1 √
≤ sup Iρ (A : B) ≤ 2 log min{dA , dB } , (48) e +ε
ρ∈S 0 fm
e −1 ,m
e −1 ≤ ε. (52)
lm
e lm
e +ε lm
e
where we used that 0 ≤ I· (A : B) and I· (A : B) ≤ Proof: The proofs of both inequalities follow from some
2 log min{dA , dB } [13]. Applying theorem 4.6 in the form of elementary calculus results and can be easily derived, so we
remark 4.7, we can conclude the claim and obtain the given omit them here.
continuity bound. 1) Divergence Bounds for the Relative Entropy: In this
3) Uniform Continuity for the Conditional Mutual Infor- section, we prove an upper bound on the relative entropy
mation: We can also provide a continuity bound for D(ρ∥σ) which involves the trace norm distance of ρ and σ.
the conditional mutual information of two tripartite states The literature calls these bounds upper continuity bounds [37],
ρABC , σABC ∈ S(HA ⊗ HB ⊗ HC ) from corollary 5.5, [49], [50], for which we would expect an upper bound of
by viewing it as the difference between two conditional |D(ρ1 ∥σ1 ) − D(ρ2 ∥σ2 )| in terms of the norm distance of
entropies. The following result coincides with the best ρ1 and ρ2 , and σ1 and σ2 , respectively. We hence propose
previously-known bound for the named quantity and appeared the name “divergence bound” for this kind of bound, a fit-
explicitly in [47, Lemma 4], and with a worsening of a factor ting name, since we are relating the strength of divergence
2 previously in [48] and [45, Exercise 5.41]. (between ρ and σ) to a fixed distance measure (the trace norm).
Corollary 5.7 (Uniform Continuity of the Conditional We now give the divergence bound we obtain when using
Mutual Information): The conditional mutual information with the convexity and almost concavity of D(ρ∥σ) together with
respect to H = HA ⊗ HB ⊗ HC is uniformly continuous on theorem 4.6 by going through uniform continuity of the
S 0 = S(H) and for ρ, σ ∈ S 0 with 21 ∥ρ − σ∥1 ≤ ε ≤ 1, relative entropy in its first argument.
we find that Corollary 5.9 (Uniform Continuity of the Relative Entropy
|Iρ (A : B|C) − Iσ (A : B|C)| ≤ 2ε log min{dA , dB } in the First Argument): Let σ ∈ S(H) be fixed. Then D(·∥σ) is
 ε  uniformly continuous on S 0 = {ρ ∈ S(H) : ker σ ⊆ ker ρ}
+ 2(1 + ε)h . and, for ρ1 , ρ2 ∈ S 0 with 12 ∥ρ1 − ρ2 ∥1 ≤ ε ≤ 1, it holds that
1+ε
 ε 
(49) |D(ρ1 ∥σ) − D(ρ2 ∥σ)| ≤ ε log m e −1 + (1 + ε)h ,
σ
1+ε
Proof: The procedure is now familiar. We first note that
(53)
S 0 is 0-perturbed ∆-invariant. Without loss of generality,
we can assume that dA ≤ dB and rewrite f (·) = I· (A : with m
e σ the minimal non-zero eigenvalue of σ.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5878 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

Proof: S 0 is clearly convex and 0-perturbed ∆-invariant as

for two operators A, B, ker A ∩ ker B ⊆ ker(A − B) and [A −
B]± are orthogonal. We set f (·) = D(·∥σ). Using theorem 5.1
and point 1 of proposition 5.2, we find that D(·∥σ) is ALAFF
with af = h and bf = 0. At last, we have that
Cf⊥ = sup |D(ρ1 ∥σ) − D(ρ2 ∥σ)|
ρ1 ,ρ2 ∈S 0
1 ∥ρ −ρ ∥=1
2 1 2

e −1
≤ sup D(ρ∥σ) ≤ log m σ . (54)
ρ∈S(H)

In the first inequality, we used that D(ρ∥σ) ≥ 0, and in the

second one that m e σ ρ ≤ σ hence D(ρ∥σ) ≤ log me −1
σ . Using
theorem 4.6 in the form of remark 4.7 concludes the claim.

We can compare eq. (53) with the findings of [51, Eq. (43) Fig. 3. The magnitude of the different bounds plotted over the relative
and (44)], based on the previous [52], where it was shown that entropy. We sampled a thousand different pairs of qubits and controlled the
  minimal eigenvalue of σ in a range from 10−4 to 10−8 . The explicit bounds
∥ρ1 − ρ2 ∥∞ can be found in table I.
|D(ρ1 ∥σ) − D(ρ2 ∥σ)| ≤ max log 1 + ,
i=1,2 mρi mσ
(55)
whenever ρi > 0 and min mρi > ∥ρ1 − ρ2 ∥∞ . Here mρi is
i=1,2
the minimal eigenvalue of ρi for i = 1, 2 and correspondingly
mσ the one of σ. This expression presents the advantage
with respect to ours of depending on the operator norm
of the difference of ρ1 and ρ2 , instead of the trace norm.
However, when ρ1 ≈ ρ2 , the upper bound in eq. (55) can
∥ρ −ρ ∥
be approximated by m1 ρ m2 σ∞ , and thus the dependence with
i
m−1
σ is linear, instead of logarithmic as in eq. (53). Further
in eq. (55) one needs ρ1 and ρ2 to be full rank and has a
condition on their minimal eigenvalues.
We can subsequently use the corollary 5.9 to prove a
divergence bound for the relative entropy.
Corollary 5.10 (Divergence Bound for the Relative Fig. 4. The difference between the bound from corollary 5.10 and the one
of Audenaert & Eisert [35, Theorem 1]. On the x-axis we plot the minimal
Entropy): Let ρ, σ ∈ S(H) with ker σ ⊆ ker ρ and eigenvalue of σ and on the y-axis ε = 12 ∥ρ − σ∥1 . The minimal eigenvalue
1
2 ∥ρ − σ∥1 ≤ ε ≤ 1, we have of ρ is set to the minimal eigenvalue of σ, thereby strengthening the bound
 ε  of Audenaert & Eisert. Both were varied between 10−20 and 12 .
D(ρ∥σ) ≤ ε log m e −1
σ + (1 + ε)h
1+ε
 log me −1 
≤ 1 + √ σ ε1/2 . (56) in the trace norm (resp. operator norm) difference between
2 the involved states, and show a dependence on the inverse
with me σ the minimal non-zero eigenvalue of σ. The √ second of the minimal eigenvalue of σ only logarithmically. This is
inequality follows from eq. (51) and the fact that ε ≤ ε for partly an advantage over the bounds in [36] and [37]. There
any ε ∈ [0, 1]. further exists a bound in [53, Proposition 5.81] which has
Proof: In the context of corollary 5.9, we just set ρ1 = ρ an explicit dependence on the dimension in addition to the
and ρ2 = σ, giving us that 21 ∥ρ1 − ρ2 ∥1 = 12 ∥ρ − σ∥1 ≤ dependence on the minimal eigenvalue of σ and therefore was
ε ≤ 1. Furthermore, D(ρ2 ∥σ) = D(σ∥σ) = 0 and |D(ρ1 ∥σ)| not investigated. The bound might have an advantage in low-
loses the absolute value, as D(·∥·) ≥ 0. The bound follows dimensional settings. In table I, fig. 3 and fig. 4 we compare
immediately. the aforementioned bounds from [35], [36], and [37]. From
There exist results on divergence bounds in the literature fig. 3 it is clear that our bound, in the majority of the cases,
which predate ours. In [34] and [37], the authors present some outperforms the bound by Vershynina and the one by Bratteli
linear bounds for the relative entropy in terms of the trace & Robinson. This is because of the logarithmic scaling with
norm difference between those states, with some multiplicative the inverse minimal eigenvalue of σ of our bound versus
factors depending on the eigenvalues of the states involved, the linear scaling with the inverse minimal eigenvalue of σ
whereas in [36] a similar bound is provided in terms of of theirs. We hence reduce the discussion to a comparison
the operator norm of the difference between the states. One between Audenaert & Eisert’s and our bound. From the first
of the bounds in [34] is further generalised in [35] and is fig. 3 and second plot fig. 4 we conclude a slight advantage of
closely related to our bound as both of them are non-linear theirs. The numerical experiments suggest, however, that the

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5879

TABLE I
A C OMPARISON OF D IFFERENT D IVERGENCE B OUNDS . H ERE ε = 12 ∥ρ − σ∥1 AND m· AND m e · A RE THE M INIMAL AND THE M INIMAL N ON -Z ERO
E IGENVALUE OF THE Q UANTUM S TATE IN THE I NDEX , R ESPECTIVELY. F URTHER λρ I S THE M AXIMAL E IGENVALUE OF ρ. T HE B OUND OF
AUDENAERT & E ISERT IN THE C ASE mρ = 0 H AS TO BE U NDERSTOOD AS THE L IMIT mρ → +0

difference between both bounds is bounded by two, hence as m̃ ≤ 1) we conclude Efmax = fm e −1 . At last, we have that
e −1 ,m
the minimal eigenvalue decreases both bounds should converge
asymptotically. Furthermore, our bound has the advantage that Cfm
e
= sup |D(ρ∥σ1 ) − D(ρ∥σ2 )|
σ1 ,σ2 ∈S 0
it does not need σ nor ρ to be full rank. This fact and its simple 1 ∥σ −σ ∥ =1−f
2 1 2 1 m

representation might give some advantages in applications. ≤ sup D(ρ∥σ)

2) Continuity Bounds for the Relative Entropy: We con- σ∈S 0
clude our section on continuity bounds with the most involved e −1 ) ,
≤ log(m (60)
continuity bound until now. It concerns the relative entropy
and regards it in all its power as a function of two variables, where we used that D(ρ∥·) ≥ 0 and for the last inequality
i.e., it constitutes a continuity bound both for the first and the e ≤ σ for all σ ∈ S 0 . Employing now theorem 4.6 we
that mρ
second input simultaneously. This presents some challenges obtain uniform continuity and the claimed continuity bound.
that need to be dealt with, as the relative entropy exhibits
discontinuity whenever the kernel of the second input is not As in the case of the continuity bound for the relative
contained in that of the first one. To overcome these issues, entropy in the first input, we can compare eq. (58) with
we need to employ the ALAFF method in its full generality. [51, Eq.(39) and (40)], as an extension of the previous [52],
In the first step, we fix the first input of the relative entropy in which it was shown that
 
and provide a continuity bound for the relative entropy in the ∥σ1 − σ2 ∥∞
|D(ρ∥σ1 ) − D(ρ∥σ2 )| ≤ max − log 1 − ,
second argument. i=1,2 mσi
Corollary 5.11 (Uniform Continuity of the Relative Entropy (61)
in the Second Argument): Let ρ ∈ S(H) be fixed and 1 > m e >
0. Then, D(ρ∥·) is uniformly continuous on whenever min mσi > ∥σ1 − σ2 ∥∞ , where mσi is the minimal
i=1,2
eigenvalue of σi for i = 1, 2. In the low ε regime the bound
S 0 := {σ ∈ S(H) : mρ
e ≤ σ} . (57) in eq. (61) as well as the bound in eq. (58) scales linearly in
m−1σ .
1
We further get that, for σ1 , σ2 ∈ S 0 with 2 ∥σ1 − σ2 ∥1 ≤ ε, In the above corollary, two choices need some more jus-
tification. The first choice is 1 > me and the second one is
ε
|D(ρ∥σ1 ) − D(ρ∥σ2 )| ≤ log(me −1 ) s = m. e We want to put them into context by the following
lm
e lemma, demonstrating that these assumptions are necessary to
lm
e +ε
 ε 
obtain a non-trivial S0 .
+ fm −1 ,m −1
lm
e
e e
lm
e +ε Lemma 5.12 Let ρ ∈ S(H) and s ∈ [0, 1) with rank ρ ≥ 2,
2 −1 √
3 log m further me ∈ (0, ∞) and
≤
e
ε, (58)
1−m e
S 0 := {σ ∈ S(H) : ker σ ⊆ ker ρ, mρ
e ≤ σ} . (62)
where lme = 1 − m.e The second
√ inequality follows from (52),
and by noticing that ε ≤ ε for any ε ∈ [0, 1]. Then, the following is true:
Proof: We have that S 0 is clearly convex as, for σ1 , σ2 ∈ e then S 0 is s-perturbed ∆-invariant if and only
1) If 1 > m,
S 0 and λ ∈ [0, 1], if s ≥ m.
e
e then S 0 = {ρ}.
2) If 1 = m,
λσ1 + (1 − λ)σ2 ≥ λmρ
e + (1 − λ)mρ
e = mρ
e , (59) e S 0 = ∅.
3) If 1 < m,
We will only give proof for the first one in section D and
giving the kernel inclusion as well as the condition for the leave the last two for the reader. Next, we proceed to state
smallest eigenvalue on the support of ρ. Furthermore, S 0 is and prove the main result of this subsection on continuity
s-perturbed ∆-invariant with s = m. e This is because one can bounds, namely the uniform continuity bound for the relative
perturb with τ = ρ and get subminorization by mρ. e Employing entropy in both arguments on a suitable subspace. Since we
point 2 of proposition 5.2 we further find that f (·) = D(ρ∥·) have already explored the cases in which we either fix the
satisfies eq. (18) with bf = 0 and af = fm e −1 , hence
e −1 ,m second (corollary 5.9) or first (corollary 5.11) density operator,
Ef = fm e −1 ,m
e −1 . Using again proposition 5.2 (point 4, since we now combine both results in the proof of the next theorem.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5880 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

Theorem 5.13 (Uniform Continuity of the Relative We therefore obtain

Entropy): Let 1 > 2m
e > 0 and δ
|D(ρ1 ∥σ1 ) − D(ρ1 ∥σ)| ≤ e −1 )
log(m
2lm
S 0 = {(ρ, σ) : ρ, σ ∈ S(H), ker σ ⊆ ker ρ, 2m
e ≤m
e σ} ,
e
−1  2−1 δ 
lme +2 δ
(63) + fm
e −1 ,m
e −1 −1 δ
,
lm
e lm
e +2
δ
with m e σ the minimal non-zero eigenvalue of σ. Then, |D(ρ2 ∥σ) − D(ρ2 ∥σ2 )| ≤ log(me −1 )
2lm
D(·∥·) is uniformly continuous on S 0 and we find that for
e
−1  2−1 δ 
lme +2 δ
(ρ1 , σ1 ), (ρ2 , σ2 ) ∈ S 0 with 12 ∥ρ1 − ρ2 ∥ ≤ ε ≤ 1 and + fm
e −1 ,m
e −1 −1 δ
.
1 lm lm
e +2
2 ∥σ1 − σ2 ∥1 ≤ δ ≤ 1
e
(70)
 δ 
|D(ρ1 ∥σ1 ) − D(ρ2 ∥σ2 )| ≤ ε + log(m e −1 ) Combining the bounds and further using that
lm
−1  2−1 δ 
 ε  e
lm
e +2 δ
+ (1 + ε)h fm
e −1 ,m
e −1 −1 δ
1+ε lm
e lme +2
lm
e +δ
 δ 
lm
e +δ
 δ 
+2 fm −1 ,m −1 ≤ fm −1 ,m −1 , (71)
lm
e
e e
lm
e +δ lm
e
e e
lm
e +δ
−1  2
 log m 5 log e −1 1/2
m we obtain the claimed bound, and thereby also uniform
≤ 1+ √ ε1/2 + √
e
δ , (64)
2 2(1 − m)e continuity.
Let us conclude this section by emphasizing that there might
with lm e √ = 1 − m. e The
√ second inequality follows from (52) be some room for improvement in the previous result. For
and ε ≤ ε and δ ≤ δ for ε, δ ∈ [0, 1]. instance, it should be possible to improve the interpolation
Proof: We will prove the uniform continuity by between σ1 and σ2 considered in eq. (65) by optimizing
proving that the bound eq. (64) holds. Therefore, let over the interpolation parameter instead of setting it to 1/2.
(ρ1 , σ1 ), (ρ2 , σ2 ) ∈ S 0 with 12 ∥ρ1 − ρ2 ∥ ≤ ε ≤ 1 and However, we believe this would not change the appearance of
1
2 ∥σ1 − σ2 ∥ ≤ δ ≤ 1. We define the bound drastically and thus the reason for not performing
this optimization.
1 1
σ= σ1 + σ2 , (65)
2 2
VI. A LMOST C ONCAVITY AND C ONTINUITY B OUNDS FOR
and obtain THE B ELAVKIN -S TASZEWSKI E NTROPY

1 1 δ Following the same lines as in the previous section, we now

∥σ − σ1 ∥1 = ∥σ1 − σ2 ∥1 ≤ ≤ 1 , apply the ALAFF method introduced in section IV for the
2 4 2
1 1 δ particular case of the BS-entropy. For that, we need to prove
∥σ − σ2 ∥1 = ∥σ1 − σ2 ∥1 ≤ ≤ 1 . (66) a result of almost concavity for the BS-entropy, which we do in
2 4 2
section VI-A. However, in contrast to the case of the relative
Using this, we get entropy, our result for the BS-entropy is not tight. We leave the
discussion on the almost concavity bound and the difficulties
|D(ρ1 ∥σ1 ) − D(ρ2 ∥σ2 )| ≤ |D(ρ1 ∥σ1 ) − D(ρ1 ∥σ)| that appear in the BS-entropy case to the next subsection.
+ |D(ρ1 ∥σ) − D(ρ2 ∥σ)| Subsequently, we combine our result of almost concav-
ity for the BS-entropy with the ALAFF method to provide
+ |D(ρ2 ∥σ) − D(ρ2 ∥σ2 )| .
certain results of uniform continuity and explicit continuity
(67) bounds for entropic quantities derived from the BS-entropy
in section VI-B. All the results provided in this section are
The middle term can be bounded using corollary 5.9 and the summarized in fig. 5.
fact that

e σ−1 ≤ log(2 max{m

e −1 e −1 e −1 . A. Almost Concavity for the BS-Entropy
log m σ1 , m σ2 }) ≤ log m (68)
In this section, we prove the almost concavity of the
One obtains BS-entropy and thereby complement the established result of
 ε  convexity [24, Theorem 4.4], [38, Corollary 4.7]. We first want
e −1 + (1 + ε)h
|D(ρ1 ∥σ) − D(ρ2 ∥σ)| ≤ ε log m . to give some auxiliary results that will be needed later. The first
1+ε of those concerns an operator inequality for the term inside
(69) the trace in the definition of the BS-entropy.
Lemma 6.1: Let A1 , A2 ∈ B(H) positive semi-definite, p ∈
The other two terms are bound using corollary 5.11 and [0, 1] and
e 1 ≤ 12 σ1 ≤ σ and mρ
the fact that mρ e 2 ≤ 12 σ2 ≤ σ
by construction of S 0 and the definition of σ, respectively. A = pA1 + (1 − p)A2 . (72)

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5881

Fig. 5. In this flow chart we collect the main results from this section, starting with the almost concavity for the BS-entropy, which together with the
ALAFF method outputs a plethora of continuity bounds for related entropic quantities. For the convexity and almost concavity of the BS-entropy we are
setting ρ = pρ1 + (1 − p)ρ2 and σ = pσ1 + (1 − p)σ2 , with p ∈ [0, 1]. We denote by mσ the minimal eigenvalue of σ. In the almost concavity bound,
ĉ0 is the maximum of σ1−1 and σ2−1 . Additionally, we assume in all the continuity bounds that m ≤ η −1 ∞ , for η = σ, ρ.
∞ ∞

Then We now use the operator monotonicity of the logarithm to find

−A log(A) ≤ −pA1 log(A1 ) − (1 − p)A2 log(A2 )

+ hA1 ,A2 (p) 1 (73) −Tr [A log(A)]
= −pTr [A1 log(A)] − (1 − p)Tr [A2 log(A)]
with hA1 ,A2 (p) = −p log(p)Tr [A1 ] − (1 − p) log(1 − p)Tr [A2 ]
≤ −pTr [A1 log(pA1 )] − (1 − p)Tr [A2 log((1 − p)A2 )]
a distorted binary entropy.
Proof: It holds that = −pTr [A1 log(A1 )] − (1 − p)Tr [A2 log(A2 )]
+ hA1 ,A2 (p). (78)
−A log(A) + pA1 log(A1 ) + (1 − p)A2 log(A2 )
≤ ∥−A log(A) + pA1 log(A1 )
Inserting this into eq. (77) and then into eq. (74) yields the
+(1 − p)A2 log(A2 )∥1 1 . (74) claimed result.
Now, since x 7→ −x log(x) is operator concave [54, Theorem The next auxiliary result concerns an equivalent formulation
2.6], we have for the BS-entropy constructed from the function x 7→ x log x
and has already appeared in the literature (see e.g. [41, Eq.
−A log(A) ≥ −pA1 log(A1 ) − (1 − p)A2 log(A2 ) , (75) (7.35)]). We include here a short proof of this result for
completeness.
giving us that Lemma 6.2: Let ρ ∈ S(H) and σ ∈ S + (H), then
−A log(A) + pA1 log(A1 ) + (1 − p)A2 log(A2 ) ≥ 0 , (76) h i
D(ρ∥σ)
b = Tr σ(σ −1/2 ρσ −1/2 ) log(σ −1/2 ρσ −1/2 ) . (79)
and hence

∥−A log(A) + pA1 log(A1 ) + (1 − p)A2 log(A2 )∥1

Proof: Slightly misusing notation, we can replace the
= Tr [−A log(A) + pA1 log(A1 ) + (1 − p)A2 log(A2 )] . regular log with a log that evaluates to 0 at 0 and thereby
(77) artificially add 0 to the domain. This changes neither the RHS

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5882 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

nor the LHS but allows us to derive prove

−D(ρ∥σ)
h i
= Tr ρ log(ρ1/2 σ −1 ρ1/2 )
b
D(ρ∥σ)
b
h  i
h i = Tr σ −σ −1/2 ρσ −1/2 log(σ −1/2 ρσ −1/2 )
= Tr log(ρ1/2 σ −1/2 σ −1/2 ρ1/2 )ρ1/2 σ −1/2 σ 1/2 ρ1/2 h  i
≤ pTr σ −σ −1/2 ρ1 σ −1/2 log(σ −1/2 ρ1 σ −1/2 )
h i
= Tr ρ1/2 σ −1/2 log(σ −1/2 ρσ −1/2 )σ 1/2 ρ1/2 h  i
+ (1 − p)Tr σ −σ −1/2 ρ2 σ −1/2 log(σ −1/2 ρ2 σ −1/2 )
h i
= Tr σ(σ −1/2 ρσ −1/2 ) log(σ −1/2 ρσ −1/2 ) . (80)
+ hA1 ,A2 (p)
We used the cyclicity of the trace several times, and the = −pD(ρ
b 1 ∥σ) − (1 − p)D(ρ b 2 ∥σ) + hA ,A (p) . (86)
1 2
well-known fact that we have f (L∗ L)L∗ = L∗ f (LL∗ ) in
At last we can estimate Tr [Aj ] = Tr σ ρj ≤ σ −1 ∞ ≤
 −1 
case the spectrum of L∗ L and LL∗ lie in the domain of f
[55, Lemma 61.]. ĉ0 for j = 1, 2 using Hölder’s inequality, giving us
Building on the previous results from this section, we pro- hA1 ,A2 (p) ≤ ĉ0 h(p).
ceed to prove now the main result, namely the almost con- We now have to estimate terms of the form D(ρ b j ∥σj ) −
cavity for the BS-entropy. This falls in the line of results of b j ∥σ) for j = 1, 2. This is done using the
D(ρ
almost concavity discussed in definition 4.1. Peierls-Bogoliubov inequality [41] and the multivariate trace
Theorem 6.3 (Almost Concavity of the BS-Entropy): Let inequalities of Sutter et al. [44]:
(ρ1 , σ1 ), (ρ2 , σ2 ) ∈ S ker,+ with b j ∥σj ) − D(ρ
b j ∥σ)
D(ρ
h  i
1/2 1/2 1/2 1/2
S ker,+ := {(ρ, σ) ∈ S(H) × S(H) : σ ∈ S + (H)} (81) = Tr ρj log(ρj σj−1 ρj ) − log(ρj σ −1 ρj )
h 
and p ∈ [0, 1]. Then, for ρ = pρ1 + (1 − p)ρ2 , σ = pσ1 + (1 − 1/2 1/2
≤ Tr exp log(ρj ) + log(ρj σj−1 ρj )
p)σ2 , we have i
1/2 1/2
− log(ρj σ −1 ρj )
D(ρ∥σ)
b ≥ pD(ρ
b 1 ∥σ1 ) + (1 − p)D(ρ
b 2 ∥σ2 ) h 
1/2 1/2
− ĉ0 (1 − δρ1 ρ2 )h(p) − fĉ1 ,ĉ2 (p) , (82) ≤ Tr exp log(ρj ) + log(ρj σj−1 ρj )
i
−1/2 −1/2
+ log(ρj σρj )
with
1/2 1/2 −1/2 −1/2
≤ log α(ρj , ρj σj−1 ρj , ρj σρj )
h(p) = −p log(p) − (1 − p) log(1 − p) , (
log(p + (1 − p)ĉ1 ) j = 1
fĉ1 ,ĉ2 (p) = p log(p + ĉ1 (1 − p)) = . (87)
log((1 − p) + pĉ2 ) j = 2
+ (1 − p) log((1 − p) + ĉ2 p) ,
( In the third line, we use that
1 if ρ1 = ρ2 1/2 1/2 −1/2 −1/2
δρ1 ρ2 = , (83) − log(ρj σ −1 ρj ) ≤ log(ρj σρj ) (88)
0 otherwise
which is true since for Pρj the projection on the support of
and the constants ρj , we have

ĉ0 := max{ σ1−1 , σ2−1 ∞ } , Pρj (Pρ σPρj )−1 Pρj ≤ Pρj σ −1 Pρj , (89)
∞
1/2 1/2 −1/2 −1/2
ĉ1 := α(ρ1 , ρ1 σ1−1 ρ1 , ρ1 σ2 ρ1 ) , as x → x−1 is operator convex and hence fulfills the
1/2 1/2 −1/2 −1/2 Sherman-Davis inequality [54, Theorem 4.19]. Note that σ
ĉ2 := α(ρ2 , ρ2 σ2−1 ρ2 , ρ2 σ1 ρ2 ) , (84)
is invertible and that by (Pρj σPρj )−1 we mean the Moore-
R +∞ h 1+it 1−it
i Penrose pseudoinverse. We find
with α(O, P, Q) := −∞ dtβ0 (t) tr OP 2 QP 2 and
1/2 1/2 1/2 1/2
β0 the probability density defined in eq. (36). − log(ρj σ −1 ρj ) = − log(ρj Pρj σ −1 Pρj ρj )
Proof: The formula for p = 0, 1 is trivial, hence let p ∈ 1/2
≤ − log(ρj Pρj (Pρj σPρj )−1 Pρj ρj )
1/2
(0, 1). We find that −1/2 −1/2
= log(ρj Pρj σPρj ρj )
b 1 ∥σ1 ) + (1 − p)D(ρ
pD(ρ b 2 ∥σ2 ) − D(ρ∥σ) −1/2 −1/2
= log(ρj σρj ). (90)
b
≤ p(D(ρ
b 1 ∥σ1 ) − D(ρ
b 1 ∥σ))
The argument why the inequalities in eq. (87) hold in the
+ (1 − p)(D(ρ
b 2 ∥σ2 ) − D(ρ
b 2 ∥σ)) + ĉ0 h(p) . case of ρj not being full rank is simpler than in the case
(85) of the corresponding inequality for the relative entropy (cf.
theorem 5.1 and section B). For the BS-entropy, we can
Indeed, as of lemma 6.2 and then lemma 6.1 with A1 = already restrict eq. (87) to the support of ρj as all operators
1/2 1/2 1/2 1/2
σ −1/2 ρ1 σ −1/2 , A2 = σ −1/2 ρ2 σ −1/2 respectively, we can involved, ρj , ρj σj−1 ρj and ρj σ −1 ρj , commute with

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5883

the projection onto this support. In the last step, we split σ Proposition 6.6 (Bounds on BS-Entropic Quantities): For
and evaluated the first term to p in case j = 1 or the second ρ ∈ S(HA ⊗ HB ), we find:
term in case j = 2 to (1−p) and left the other one untouched, 1) For the BS-conditional entropy:
respectively. This concludes the proof.
− log min{dA , dB } ≤ Hb ρ (A|B) ≤ log dA . (91)
Remark 6.4: We strongly suspect that theorem 6.3 can be
improved because of two reasons. The first one is that we 2) For the BS-mutual information:
would expect the results of almost concavity of the relative and
0 ≤ Ibρ (A : B) ≤ log min{dA , dB }
the BS-entropy to coincide in the case that the involved states
commute. The reason is that in this case, both quantities reduce + log min{ ρ−1
A ∞
, ρ−1
B ∞
},
to the classical relative entropy. A straightforward calculation (92)
shows that then ĉ1 = c1 and ĉ2 = c2 , hence fc1 ,c2 = fĉ1 ,ĉ2 , −1
with · the Moore-Penrose pseudoinverse.
but h ≤ ĉ0 h with equality if, and only if, σ1 and σ2 are pure,
3) For ρ ∈ S(HA ⊗ HB ⊗ HC ), we find that the
which in addition to σ1 and σ2 being full rank means H = C.
BS-conditional mutual information satisfies:
The other reason is given by the continuity bound we obtain
for the BS-conditional entropy in corollary 6.8. Numerics 0 ≤ Ibρ (A : B|C) ≤ log min{d2A , dABC }. (93)
suggest an almost convex bound that is independent of the The first two bounds are shown to be tight. For the third one,
minimal eigenvalue (cf. fig. 6) if the inputs are full rank2 . we expect that similar reasoning should also show its tightness.
Hence we would also suspect that an optimal almost concave The proof can be found in section F. We further want to
remainder of the BS-entropy reduces to an eigenvalue inde- remark that the scaling of the bound with respect to ρ−1 A ∞
pendent bound in the case of the BS-conditional entropy.
or ρ−1B ∞ is justified. The reasoning can be found in section
Analogous to the case of the relative entropy we provide an
F as well.
additional proposition to give context to the above result, i.e. 1) Uniform Continuity for the BS-Conditional Entropy:
to provide simpler expressions if the involved states satisfy We encounter a slight complication when it comes to the
specific conditions. uniform continuity of the BS-conditional entropy compared
Proposition 6.5 (Almost Concavity Estimate of the to the uniform continuity of the conditional entropy that we
BS-Entropy Is Well Behaved): The function ĉ0 h + fĉ1 ,ĉ2 have covered in corollary 5.5. This is because the almost
obtained in theorem 6.3 is well behaved in the following concave bound of the BS-entropy depends on the minimal
sense: Let j = 1, 2 and (ρj , σj ) ∈ S ker,+ . We have the eigenvalue of the second argument (see eq. (84)), i.e. it has
following: to be full rank. This means the input to the BS-conditional
1) If σ1 = σ2 , then ĉj = 1, resulting in fĉ1 ,ĉ2 + ĉ0 h = ĉ0 h. entropy has to be full rank as well. Although we think that the
2) If the σj have a minimal eigenvalue that is bounded result of almost concavity for the BS-entropy can be improved,
from below by m > 0 respectively, then fĉ1 ,ĉ2 + ĉ0 h ≤ we know that there is no extension of uniform continuity nor
fm−1 ,m−1 + m−1 h. continuity for the BS-conditional entropy to positive semi-
3) If H = HA ⊗ HB is a bipartite space, ρj has a minimal definite states, as this quantity is not continuous on those.
eigenvalue bounded from below by m > 0, and further This is the content of the next proposition. We also refer the
σj = d−1 A 1A ⊗ρj,B , then fĉ1 ,ĉ2 + ĉ0 h ≤ fm−1 ,m−1 + reader to [56, Remark 3.3] for a similar behaviour of the sharp
m−1 h. quantum Rényi divergences.
1
4) We find for m1 , m2 ≥ 1, p 7→ 1−p fm1 ,m2 (p) and p 7→ Proposition 6.7 (Discontinuity of the BS-Conditional
1
1−p ĉ0 h(p) are non-decreasing on [0, 1). Entropy): The BS-conditional entropy is discontinuous on
This result should be compared to proposition 5.2, its the set of positive semi-definite operators over HA ⊗ HB if
analogue for the relative entropy. The proof can be found in dA , dB ≥ 2.
section E. We will use the reductions from proposition 6.5 to Proof: Since dA ≥ 2 as well as dB ≥ 2, we find
simplify the terms in theorem 6.3 for the various applications orthogonal |iA ⟩ ∈ HA , |iB ⟩ ∈ HB , i = 0, 1. For ε ∈ (0, 1)
presented in the subsequent section. we then define
√ √
|εB ⟩ = 1 − ε |0B ⟩ + ε |1B ⟩ , (94)
B. Continuity Bounds for the BS-Entropy which is clearly normalised. Furthermore,
In this section, we will use the almost concavity for 1
ρ0 := (|0A ⟩⟨0A | + |1A ⟩⟨1A |) ⊗ |0B ⟩⟨0B | ,
the BS-entropy from theorem 6.3 together with the ALAFF 2
method in its full generality. 1 1
ρε := |0A ⟩⟨0A | ⊗ |0B ⟩⟨0B | + |1A ⟩⟨1A | ⊗ |εB ⟩⟨εB | ,
Before we dive into the continuity and divergence bounds, 2 2
we want to collect some lower and upper estimates of entropic (95)
quantities derived from the BS-entropy (see section III-A for The above are states and fulfil
the specific definitions). 1
∥ρ0 − ρε ∥1 = ∥|1A ⟩⟨1A | ⊗ (|0B ⟩⟨0B | − |εB ⟩⟨εB |)∥1
2 The full rank requirement is necessary, as we will show in proposition 6.7 2
1 √
that the BS-conditional entropy is discontinuous in the presence of vanishing = ∥|0B ⟩⟨0B | − |εB ⟩⟨εB |∥1 = ε . (96)
eigenvalues 2

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5884 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

To see the last equality, we can identify the subspace spanned

by |0B ⟩ and |1B ⟩ with C2 and then get that
 
1 0
|0B ⟩⟨0B | →
0 0
 √ √ 
1−ε ε 1−ε
and |εB ⟩⟨εB | → √ √ . (97)
ε 1−ε ε
Calculating the eigenvalues of the difference
√ and taking the
sum of their absolute value gives 2 ε and thereby eq. (96).
Since clearly [ρ0 , 1 ⊗TrA [ρ0 ]] = 0, the BS and conditional
entropy coincide and we find

Ĥρ0 (A|B) = Tr [|0B ⟩⟨0B | log |0B ⟩⟨0B |]


1
− Tr (|0A ⟩⟨0A | + |1A ⟩⟨1A |) ⊗ |0B ⟩⟨0B |
2 Fig. 6. We investigate the dependence of the almost convex remainder term
 of the BS-conditional entropy on the minimal eigenvalue of the involved
1 states. For the minimal eigenvalues 10−4 , 10−8 , 10−16 , 10−32 we sampled
· log (|0A ⟩⟨0A | + |1A ⟩⟨1A |) ⊗ |0B ⟩⟨0B |
2 five hundred pairs of qubits (ρ, σ) both of them with controlled eigenvalues.
We then sampled for every state pair ten values of p, the convex interpolation
1
= 0 − log = log 2 . (98) parameter, and plotted the remainder. As can be seen from the plot, the
2 remainder appears to be independent of the minimal eigenvalue and the shape
suggests a binary entropy or Gini impurity. The result shows a similar pattern
The result for ρε cannot be calculated so easily. We have that if the dimension is increased.

b ρ (A|B) = − 1 Tr [|0B ⟩⟨0B |

H ε
2 1
∥ρ − σ∥1 ≤ ε ≤ 1 it holds that
1/2 1/2 2
· log(|0B ⟩⟨0B | (|εB ⟩⟨εB | + |0B ⟩⟨0B |)−1 |0B ⟩⟨0B | )]
|H b σ (A|B)| ≤ 2 l−1 ε log dA
b ρ (A|B) − H
1 m
− Tr [|εB ⟩⟨εB | lm + ε  ε 
2 + (fm−1 ,m−1 + m−1 h) ,
1/2 1/2
· log(|εB ⟩⟨εB | (|εB ⟩⟨εB | + |0B ⟩⟨0B |)−1 |εB ⟩⟨εB | )] lm lm + ε
1 (102)
= − log Tr |0B ⟩⟨0B | (|εB ⟩⟨εB | + |0B ⟩⟨0B |)−1
 
2 with lm = 1 − dH m.
1 Proof: We find that S 0 is s-perturbed ∆-invariant with
− log Tr |εB ⟩⟨εB | (|εB ⟩⟨εB | + |0B ⟩⟨0B |)−1 ,
 
2 s = mdH . The justification of this choice is completely
(99) analogous to the reasoning in lemma 5.12 with ρ = d−1 H 1,
i.e. the maximally mixed state. Furthermore, f (·) = H b · (A|B)
where in the first equality we used that |0B ⟩⟨0B | |1B ⟩⟨1B | =
is ALAFF with af = 0 as H b · (A|B) is concave, and bf =
|1B ⟩⟨1B | |0B ⟩⟨0B | = 0 and in the second equality that |εB ⟩⟨εB |
m−1 h + fm−1 ,m−1 since the result in section VI-A becomes
and |0B ⟩⟨0B | are rank-one projections. We find, using again
the matrix representation in eq. (97), that independent of the states as we restrict to H
b · (A|B) using point
! 3 of proposition 6.5. We further find that
1 √ ε−1
√
(|εB ⟩⟨εB | + |0B ⟩⟨0B |)−1 → ε 1−ε
. (100) Cfs ≤ sup |H
b ρ (A|B) − H
1
b ρ (A|B)| ≤ 2 log dA ,
2
√ ε−1 2
ε −1
√ ρ1 ,ρ2 ∈S(H)
ε 1−ε
(103)
By forming matrix products and calculating the trace, we can
immediately conclude that using proposition 6.6. This allows us to apply theorem 4.6
where Efmax coincides with Ef as of point 4 in proposition 6.5.
Tr |εB ⟩⟨εB | (|εB ⟩⟨εB | + |0B ⟩⟨0B |)−1 = 1 ,
 
This concludes the claim.
Tr |0B ⟩⟨0B | (|εB ⟩⟨εB | + |0B ⟩⟨0B |)−1 = 1 .
 
(101) Even though a continuity bound for the BS-conditional
entropy can only be proven for positive definite states, numer-
If we insert this into eq. (99), we get H
b ρ (A|B) = 0.
ε ical simulations show us that we could expect a tighter
This previous result shows in particular that we could bound on the previous proposition coinciding with that of
only expect continuity and uniform continuity for the corollary 5.5, i.e., without the dependence on the minimal
BS-conditional entropy on the set of full-rank states. The pres- eigenvalues of the states involved. One can find a visualisation
ence of the minimal eigenvalue of the states in the continuity of those numeric simulations that underlie the conjecture in
bound provided below is thus not surprising. fig. 6. The possibility of obtaining such a tighter bound is left
Corollary 6.8 (Uniform Continuity of the BS-Conditional for future work.
Entropy): The BS-conditional entropy over the bipartite 2) Uniform Continuity for the BS-Mutual Information:
Hilbert space H = HA ⊗ HB is for d−1 H > m > 0 uniformly Let us address now the case of the BS-mutual information.
continuous on S 0 = S ≥m (H) and for ρ, σ ∈ S 0 with Since the BS-conditional entropy is a particular case of the

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5885

latter (by assuming that one of the reduced states of ρAB and analogously for ρB , we find
is maximally mixed), the discontinuity issues presented in
the previous subsection are expected to arise in the current Ibpρ1 +(1−p)ρ2 (A : B)
one as well. More specifically, the example of discontinuity ≥ pD(ρ
b 1 ∥ρ1,A ⊗ (pρ1,B + (1 − p)ρ2,B ))
of the BS-conditional entropy presented in proposition 6.7 + (1 − p)D(ρ
b 2 ∥ρ2,A ⊗ (pρ1,B + (1 − p)ρ2,B ))
also constitutes an example of discontinuity of the BS-mutual
information. Thus, we can only expect to prove uniform − m−1 h(p) − fm−1 ,m−1 (p)
continuity for the BS-mutual information for full-rank states ≥ pIbρ1 (A : B) + (1 − p)Ibρ2 (A : B)
However, there is a subtle difference between the settings − m−1 h(p) − 2 fm−1 ,m−1 (p) . (107)
of the BS-conditional entropy and the BS-mutual information.
As shown in proposition 6.6, the former is bounded between In the last step we used again that D(·∥·)b is monotone
the same values as the (usual) conditional entropy, whereas the decreasing in its second argument and that pρ1,B + (1 −
latter presents some pathological behaviour. Pathological in p)ρ2,B ≤ (p + (1 − p)m−1 )ρ1,B and pρ1,B + (1 − p)ρ2,B ≤
the sense that its (tight) upper bound depends on the minimal (m−1 p + (1 − p))ρ2,B , giving us another fm−1 ,m−1 (p). Hence
eigenvalues of the reduced state, as shown in eq. (92). For this af = m−1 h+2 fm−1 ,m−1 . We conclude the proof by noticing
reason, a continuity bound for the BS-mutual information will again that ρ−1 −1
A ∞ ≤ ρAB ∞ ≤ m
−1
, yielding the upper
necessarily depend on the minimal eigenvalues of the states bound
involved.
Cfs ≤ sup Ibρ (A : B) ≤ log min{dA , dB } + log m−1 .
Corollary 6.9 (Uniform Continuity for the BS-Mutual Infor- ρ∈S 0
mation): The BS-mutual information on a bipartite Hilbert (108)
space H = HA ⊗ HB is for d−1 H > m > 0 uniformly continu-
ous on S 0 = S ≥m and for ρ, σ ∈ S 0 with 21 ∥ρ − σ∥1 ≤ ε ≤ Finally, we apply theorem 4.6 and get the claimed bounds as
1 we find that Ef coincides with Efmax , due to point 4 in proposition 6.5.
3) Uniform Continuity for the BS-Conditional Mutual Infor-
mation: Next, we provide a result of uniform continuity for
|Ibρ (A : B) − Ibσ (A : B)| the BS-conditional mutual information, defined in eq. (13).
−1 lm + ε  
As a difference between two BS-conditional entropies, it will
≤ 2 lm ε(log min{dA , dB } + log m−1 ) + zm lmε+ε
lm not present the pathological behaviour from the BS-mutual
√
2 log min{dA , dB } + 6 log2 m−1 + 2(m−1 + 1) √ information, as the BS-conditional entropies are bounded
≤ ε, between the same limits as the (usual) conditional entropies.
1−mdH
(104) See proposition 6.6 for the specific bounds on all these BS-
entropic quantities.
Nevertheless, the continuity bound we obtain below for
with lm = 1−mdH and the BS-conditional mutual information also depends on the
minimal eigenvalues of the states involved, as happened in
the case of the BS-conditional entropies.
zm (p) = 2 fm−1 ,m−1 (p) + (m−1 + 1)h(p) . (105) Corollary 6.10 (Uniform Continuity of the BS-Conditional
Mutual Information): The BS-conditional mutual information
√ over H = HA ⊗ HB ⊗ HC is for d−1
For the second inequality, we used lemma 5.8 and ε ≤ ε for H > m > 0 uniformly
ε ∈ [0, 1]. continuous on S 0 = S ≥m (H) and for ρ, σ ∈ S 0 with
1
Proof: As in the case of the BS-conditional entropy, 2 ∥ρ − σ∥1 ≤ ε ≤ 1 we find that
we find that S 0 is s-perturbed ∆-invariant with s = mdH . |Ibρ (A : B|C) − Ibσ (A : B|C)|
To conclude that Ib· (A : B) is ALAFF we first note that −1
p
because of the convexity of D(·∥·), ≤ 2 ε lm log min{dA , dABC } + 2gm (ε)
√ √
b
2 log min{dA , dABC } + 4 log2 m−1 + 2 2m−1 √
≤ ε,
1−mdH
Ibpρ1 +(1−p)ρ2 (A : B)
(109)
≤ pD(ρ
b 1 ∥ρ1,A ⊗ (pρ1,B + (1 − p)ρ2,B ))
with lm = 1−mdH and
+ (1 − p)D(ρ
b 2 ∥ρ2,A ⊗ (pρ1,B + (1 − p)ρ2,B ))
lm + ε  ε 
≤ pIbρ1 (A : B) + (1 − p)Ibρ2 (A : B) + h(p) . (106) gm (ε) = (fm−1 ,m−1 + m−1 h) . (110)
lm lm + ε
√
For the second inequality, we used lemma 5.8 and ε ≤ ε for
In the last step, we used that D(·∥·)
b is monotone decreasing ε ≤ 1.
in its second argument, and pρ1,B ≤ pρ1,B + (1 − p)ρ2,B , Proof: We have that S 0 is s-perturbed ∆-invariant using
(1−p)ρ2,B ≤ pρ1,B +(1−p)ρ2,B , respectively. Hence bf = h. the same reasoning as in the proof of corollary 6.8. Because of
We follow similar lines to obtain af . Starting with theorem 6.3 the representation Ib· (A : B|C) = Hb · (A|C) − Hb · (A|BC) we
and point 2 in proposition 6.5 using that ρ−1 −1
A ∞ ≤ ρAB ∞ , can immediately conclude that Ib· (A : B|C) is ALAFF with

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5886 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

af = fm−1 ,m−1 + m−1 h and bf = fm−1 ,m−1 + m−1 h arguing the relative entropy would give analogous continuity bounds
along the same lines as in corollary 6.8. Using proposition 6.6 also in this setting.
we can conclude
p VII. A PPLICATIONS
Cfs ≤ sup Ibρ (A : B|C) ≤ 2 log min{dA , dABC } .
ρ∈S(H) In this section, we use some of the previously derived
(111) bounds to provide applications in various contexts within the
field of quantum information.
Applying theorem 4.6 and using point 4 of proposition 6.5 we
get that Ef = Efmax and thereby conclude the assertion. A. Quantum Hypothesis Testing
4) Divergence Bound for the BS-Entropy: We conclude this In this section, we interpret our bounds in terms of hypothe-
section by following the same lines as in the case of the relative sis testing. Quantum state discrimination and quantum hypoth-
entropy to provide a divergence bound for the BS-entropy. esis testing are both well-studied tasks in quantum information
Firstly, we will prove the uniform continuity of the BS-entropy theory.
in the first argument and subsequently derive from that result In quantum state discrimination, you are given a source
the divergence bound. These results should be compared to which prepares quantum states ρ1 and ρ2 with equal probabil-
their relative entropy analogues, namely corollary 5.9 and ity. The task is to perform a measurement in order to identify
corollary 5.10, respectively. whether the state prepared by the source is ρ1 or ρ2 . In this
Corollary 6.11 (Uniform Continuity of the BS-Entropy in setting, the optimal probability of successfully identifying the
the First Argument): Let σ ∈ S + (H) be fixed. Then D(·∥σ)
b is state is given in terms of the trace distance as
uniformly continuous on S 0 = S(H), and for ρ1 , ρ2 ∈ S 0 with  
1 1 1
2 ∥ρ1 − ρ2 ∥ ≤ ε ≤ 1 we find that psucc = 1 + ∥ρ1 − ρ2 ∥1 (115)
2 2
b 2 ∥σ)| ≤ ε log(m−1 −1
 ε 
|D(ρ
b 1 ∥σ) − D(ρ σ ) + (1 + ε)mσ h ,
1+ε using the Helstrom measurement (see textbooks such as [57]).
(112) In quantum hypothesis testing, we consider an asymmetric
setting with n copies and we are interested in the asymptotic
with mσ the minimal eigenvalue of σ.
performance. Again, the task is to discriminate between ρ and
Proof: The procedure is familiar. First, S 0 is 0-perturbed
σ, using a measurement {E, 1 −E} where 0 ≤ E ≤ 1. Upon
∆-invariant. Second f (·) = D(·∥σ)
b is ALAFF with af =
−1 the first outcome, the guess is ρ, and upon the second σ.
mσ h and bf = 0 employing theorem 6.3 and point 1 of
Therefore, we define the errors of the first and second kind as
proposition 6.5. Further
α(E)n = Tr[ρ⊗n (1 − E)] (116)
Cf⊥ ≤ sup D(ρ∥σ)
b ≤ log m−1
σ (113)
ρ∈S(H) and
1/2 −1 1/2
since ρ σ ρ ≤ 1 m−1
σ .Applying now remark 4.7 gives β(E)n = Tr[σ ⊗n E] . (117)
the claimed result.
Utilizing the above result we obtain a divergence bound We now want to fix the error of the first kind to be at most ε
for the BS-entropy which constitutes the analogue to the one and define
of the relative entropy in corollary 5.10. Note that even the βε (ρ⊗n ||σ ⊗n ) := min{β(E)n : α(E)n ≤ ε}, (118)
divergence bounds obtained in both cases are similar, except
where the minimum runs over 0 ≤ E ≤ 1. Then, the quantum
for the presence of a factor m−1σ in the second term of the
Stein’s lemma [3], [58] states that
bound.
Corollary 6.12 (Divergence Bound for the BS-Entropy): Let 1
lim log[βε (ρ⊗n ||σ ⊗n )] = −D(ρ||σ). (119)
ρ ∈ S(H) and σ ∈ S + (H), then for 21 ∥ρ − σ∥1 ≤ ε ≤ 1, n→∞ n

we have Therefore, we can interpret the continuity bound in the way

 ε  that two states that are hard to discriminate have almost the
D(ρ∥σ)
b ≤ ε log m−1
σ + (1 + ε)mσ h
−1
, (114) same performance in terms of hypothesis testing. We can
1+ε
illustrate this with corollary 5.9, just by taking 1 + ε there
with mσ the minimal eigenvalue of σ.
to be 2psucc following eq. (115).
Proof: In the context of corollary 6.11, we just set ρ1 = ρ
Corollary 7.1: Let σ ∈ S(H) be fixed, 0 < ε < 1 and let us
and ρ2 = σ, giving us that 12 ∥ρ1 − ρ2 ∥1 = 12 ∥ρ − σ∥1 ≤ ε ≤
b 2 ∥σ) = D(σ∥σ) consider a source which produces ρ1 , ρ2 with equal probabil-
1. Further D(ρ b = 0 and |D(ρ
b 1 ∥σ)| loses the
ity. Moreover, let p be an upper bound on the probability psucc
absolute value, as D(·∥·)
b ≥ 0. The bound follows immediately.
of successfully identifying the state. Then, the difference in the
asymptotic error exponent in hypothesis testing is bounded by
With this, we conclude our section on continuity bounds
1 1
for entropic quantities derived from the BS-entropy. We have lim log[βε (ρ⊗n 1 ||σ ⊗n )] − lim log[βε (ρ⊗n
2 ||σ
⊗n
)]
n→∞ n n→∞ n
deliberately omitted the analogues of corollary 5.11 and the-  2p − 1 
orem 5.13 for the BS-entropy, due to their high technicality ≤ (2p − 1) log m e −1
σ + 2p h , (120)
and the complexity of the continuity bounds that we would 2p
obtain with our method. However, the same procedure as for e −1
with m σ the minimal non-zero eigenvalue of σ.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5887

B. Free Energy C. Approximate Quantum Markov Chains

In section VII-A, we already saw one interpretation of In this section, we consider a tripartite Hilbert space
our results in terms of hypothesis testing. This section gives HABC = HA ⊗ HB ⊗ HC and ρABC ∈ S + (HABC ). We fur-
another interpretation using the language of quantum thermo- ther consider the conditional mutual information of ρABC
dynamics. between A and C conditioned on B. The well-known property
A ubiquitous quantity in quantum thermodynamics is free of strong subadditivity of the von Neumann entropy [60] is
energy. To define it, we need to fix a Hamiltonian H ∈ B(H), equivalent to the non-negativity of the conditional mutual
H = H ∗ , and some inverse temperature β > 0. The Gibbs information, which is furthermore known [61], [62] to vanish
state of this system, describing a quantum system in thermal if, and only if,
equilibrium, is 1/2 −1/2 −1/2 1/2
ρABC = ρAB ρB ρBC ρB ρAB , (126)
−βH
e i.e., whenever ρABC is a quantum Markov chain. In particular,
ρβ (H) = . (121)
tr[e−βH ] 1/2 −1/2
if we denote PB→AB (ρBC ) = ρAB ρB ρBC ρB ρAB ,
−1/2 1/2

Now, we can define the free energy as we have

IPB→AB (ρBC ) (A : C|B) = 0 . (127)
F (ρ) = tr[Hρ] − β −1 S(ρ) . (122)
Moreover, by the decomposition of the CMI of ρABC in terms
It can be related to the relative entropy as of a difference of conditional entropies, as well as the data
processing inequality, we have
D(ρ||ρβ (H)) = β(F (ρ) − F (ρβ (H))) , (123)
Iρ (A : C|B) = Hρ (C|B) − Hρ (C|AB)
which can easily be verified by direct computation. ≤ HPB→AB (ρBC ) (C|AB) − Hρ (C|AB) .
Inspired by quantum information theory, in particular entan- (128)
glement theory, during the last years various descriptions of
quantum thermodynamics as a resource theory have emerged. Here we w.l.o.g. assumed that dA ≤ dC using the symmetry of
Resource theories are described in terms of free states and the CMI in A and C. Therefore, we can apply our continuity
free operations. In quantum thermodynamics, the free state bound for the CE from corollary 5.5 (which provides, in this
is ρβ (H), whereas the choices of free operations can differ. case, a tighter result than corollary 5.7), cf. also [9], to obtain
Possible choices include the thermal operations (TO), their an upper bound on the CMI of ρABC in terms of how far
closure (CTO), and the Gibbs preserving covariant operations it is from being recovered with the Petz recovery map, i.e.,
(GPC). Instead of giving a formal definition here, we refer the in terms of
reader to [59, Section II.C]. In entanglement theory, we are 1/2 −1/2 −1/2 1/2
ρABC − ρAB ρB ρBC ρB ρAB . (129)
interested in the distillation of EPR pairs from other states, 1
possibly taking many copies. In the same spirit, in quantum A similar direction was previously explored in [48, Eq. (26)].
thermodynamics, the corresponding task is the distillation of Note that, as a direct consequence of corollary 5.5, we get the
athermality. The asymptotic distillable athermality is quan- following bound for any state ρABC ∈ S(HABC ):
tified by the free energy difference in eq. (123) [40], [59].  ε 
Theorem V.1 of [59] states that for the asymptotic distillation Iρ (A : C|B) ≤ 2ε log min{dA , dC } + (1 + ε)h ,
1+ε
rate of athermality
(130)
DistillF (ρ, ρβ (H)) = D(ρ||ρβ (H)) , (124) with
1 1/2 −1/2 −1/2 1/2
where F ∈ {TO, CTO, GPC}. Again, we refer the reader ε := ρABC − ρAB ρB ρBC ρB ρAB . (131)
to [59] for the formal definitions. Thus, we can interpret corol- 2 1

lary 5.9 as quantifying the continuity of distillable athermality. Moreover, we can use the following inequality
Corollary 7.2: Let H be a fixed Hamiltonian with maximal 
x

√
eigenvalue λmax , minimal eigenvalue λmin , and β > 0 an (1 + x)h ≤ 2x , (132)
1+x
inverse temperature. Then, for ρ1 , ρ2 ∈ S(H) such that 12 ∥ρ1 −
ρ2 ∥1 ≤ ϵ ≤ 1, it holds that x ∈ [0, 1], as well as the fact that, since ε ∈ [0, 1],
for every √
then ε ≤ ε, to upper bound the CMI of ρABC by
|DistillF (ρ1 , ρβ (H)) − DistillF (ρ2 , ρβ (H))| √ 
 ε  Iρ (A : C|B) ≤ 2 log min{dA , dC } + 1
≤ ε βλmax + log(Tr e−βH ) + (1 + ε)h
 

 1 ε+ ε
1/2
1/2 −1/2 −1/2 1/2
· ρABC − ρAB ρB ρBC ρB ρAB .
≤ εβ(λmax − λmin ) + ε log(d) + (1 + ε)h , 1
1+ε (133)
(125)
This bound should be compared to lower bounds for the
where F ∈ {TO, CTO, GPC}. conditional mutual information. On the one hand, Fawzi and

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5888 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

Renner proved in [63] the following lower bound for such a Proof: We can upper bound the difference between the
√ √ 2
quantity in terms of the fidelity F (ρ, σ) = ρ σ 1: entropies by

Iρ (A : C|B) ≥ − log F (ρABC , RB→AB (ρBC )) , (134) D(ρ∥σ)

b − D(ρ∥σ)
h  i
where RB→AB is another recovery map, the so-called rotated = Tr ρ log(ρ1/2 σ −1 ρ1/2 ) − log ρ + log σ
Petz recovery map, which was explicitly constructed in [64].
 n o
= −D ρ exp log σ + log(ρ1/2 σ −1 ρ1/2 )
Several results have been provided in this line in the past h n oi
decade. Here we specifically focus on [65], in which Carlen ≤ log Tr exp log σ + log(ρ1/2 σ −1 ρ1/2 )
and Vershynina proved: h i
 π 4 ≤ log Tr σρ1/2 σ −1 ρ1/2 , (139)
−2 −2
Iρ (A : C|B) ≥ ρ−1
B ∞ ρABC ∞
−1
8 where we have used the non-negativity for the relative entropy
4
1/2 −1/2 −1/2 1/2 of density matrices and Golden-Thompson inequality [70],
· ρABC − ρAB ρB ρBC ρB ρAB ,
1 [71]. Next, we can write
(135) h i h i
Tr σρ1/2 σ −1 ρ1/2 = Tr σρ1/2 (σ −1 − ρ−1 )ρ1/2 + 1 .
Therefore, by combining eq. (133) with eq. (135) we obtain the
following “sandwich” for the conditional mutual information (140)
of a tripartite density matrix ρABC in terms of its trace distance Therefore, using log(x + 1) ≤ x, we have
to its Petz recovery map: h i
 π 4 D(ρ∥σ)
b − D(ρ∥σ) ≤ Tr σρ1/2 (σ −1 − ρ−1 )ρ1/2 . (141)
−2 −2
ρ−1
B ∞ ρABC ∞
−1
8 Now, we can use the following expression for invertible
4
1/2 −1/2 −1/2 1/2 matrices X and Y :
· ρABC − ρAB ρB ρBC ρB ρAB
1
≤ Iρ (A : C|B) ≤ X −1 − Y −1 = Y −1 (Y − X)X −1 (142)
2 (log min{dA , dC } + 1) Then,
1/2 h i h i
1/2 −1/2 −1/2 1/2
· ρABC − ρAB ρB ρBC ρB ρAB . (136) Tr σρ1/2 (σ −1 − ρ−1 )ρ1/2 = Tr σρ−1/2 (ρ − σ)σ −1 ρ1/2
1

In particular, this implies that a state ρABC ∈ ≤ ∥ρ−1/2 (ρ − σ)σ −1 ρ1/2 ∥∞

S(HA ⊗ HB ⊗ HC ) is an approximate quantum Markov ≤ σ −1 ∞
∥ρ − σ∥∞ , (143)
chain [66] (i.e. Iρ (A : C|B) < ϵ) if, and only if, it is
close to its reconstructed state under the Petz recovery map. by [72, Proposition IX.1.1] and Hölder’s inequality.
This idea was used in [67] to prove that a Gibbs state The previous proposition provides a general upper bound
of a one-dimensional local Hamiltonian is an approximate for the distance between both entropies in terms of the
quantum Markov chain, and subsequently, in [68] to provide spectral norm and the minimal eigenvalue of the second input.
an estimate on the time it takes for a Markovian evolution This is valid for any pair of states but does not yield any
of a density matrix to become an approximate quantum further information on specific pairs with better conditions.
Markov chain. Moreover, a similar inequality has recently Alternatively, we can prove the following bound, from which
been employed in [69] to study the decay of the CMI for it is obvious that D(ρ||σ) = D(ρ||σ)
b if ρ and σ commute.
purely generated finitely correlated states. Proposition 7.4: Consider two positive definite states
ρ, σ ∈ S+ (H). Then, the following inequality holds:

D. Difference Between Relative Entropy and BS-Entropy D(ρ∥σ)

b ≤ D(ρ∥σ) + f ([ρ1/2 , σ −1/2 ]) (144)
It is well-known that the BS-entropy is an upper bound on where the last term is given by
the Umegaki relative entropy [23], [24], [41], i.e., that 2
f (X) := ∥X∥∞ + 2 ∥X∥∞ . (145)
D(ρ||σ) ≤ D(ρ||σ)
b , (137)
In particular, whenever ρ and σ commute, f vanishes.
and they coincide if and only if ρ and σ commute (see, Proof: The proof proceeds in the same way as for
e.g., [58] and [23, Proposition 4.7]). In this section, our proposition 7.3 until
aim is to quantify how large the difference between the two h i
divergences can become. We start with two upper bounds on D(ρ∥σ)
b − D(ρ∥σ) ≤ Tr σρ1/2 (σ −1 − ρ−1 )ρ1/2 . (146)
D(ρ||σ)
b in terms of D(ρ||σ). Let us define
Proposition 7.3: Consider two positive definite states
ρ, σ ∈ S+ (H). Then, the following inequality holds: η := σ 1/2 ρ1/2 σ −1 ρ1/2 σ 1/2 . (147)

D(ρ∥σ)
b ≤ D(ρ∥σ) + m−1 ∥ρ − σ∥∞ , (138) Then,
h i
where m is the minimal eigenvalue of σ. Tr σρ1/2 (σ −1 − ρ−1 )ρ1/2 = Tr [η − σ] . (148)

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5889

Introducing ρ gives since we can verify with pi = λi tr[σPi ], qi = tr[σPi ] that

k   
Tr [η − σ] = Tr [η − σ + ρ − ρ]
X Pi pi qi
D(p∥q) = Tr pi log − log
= Tr [η − ρ] + Tr [ρ − σ] i=1
Tr [Pi ] Tr [Pi ] Tr [Pi ]
k
= Tr [η − ρ] ≤ ∥η − ρ∥1 . (149) X
= pi (log pi − log qi )
Moreover, as appears in [22, Remark 2.2], the right-hand side i=1
k
above can be estimated by X
= λi tr[σPi ] log λi
h i 2 h i i=1
∥η − ρ∥1 ≤ ρ1/2 , σ −1/2 +2 1/2
ρ ,σ −1/2
. h 1 1 1 1
i
∞ ∞ = Tr σσ − 2 ρσ − 2 log(σ − 2 ρσ − 2 )
(150) h 1 1
i
= Tr ρ log(ρ 2 σ −1 ρ 2 ) . (156)
This concludes the proof of the proposition. qi
Finally, we want to compare our previous bounds, proven Obviously, if m is the minimal eigenvalue of σ, then Tr[P i]
≥
using inequalities such as Golden-Thompson or Hölder, with m for all i ∈ {1, . . . , k}. Thus, the assertion follows from
those we could obtain by means of our continuity bounds, theorem 5.13. Moreover, it is clear that if [ρ, σ] = 0 there is
as the BS-entropy can, in particular, be regarded as a relative a unitary U which diagonalizes ρ and σ simultaneously such
entropy. We can also apply the continuity bound we derived that ρ = p and σ = q.
in theorem 5.13. Finally, the last simplification from eq. (154) is a direct
Corollary 7.5: Let ρ ∈ S(H), σ ∈ S+ (H) and m̃ such that consequence of theorem 5.13 and lemma 5.8.
d−1
H > 2m e > 0 and the minimal eigenvalue of σ is lower
bounded by 2m. e Let E. Weak Quasi-Factorization of the Relative Entropy
k
Results of quasi-factorization for a divergence allow us to
− 21 − 12
X split such a divergence in a bipartite space in terms of the
σ ρσ = λi Pi (151)
sum of two “conditional” divergences on subsystems and a
i=1
multiplicative error term related to the correlations between
be the spectral decomposition with eigenvalues λi and projec- both subsystems on the second input of the divergences.
tions Pi . Define density matrices A weak version of such a result presents instead an additive
error term.
k k
X Pi X Pi More specifically, it was proven in [73] that, given a
p= λi tr[σPi ] , q= tr[σPi ] . (152) bipartite space HAB = HA ⊗ HB and ρAB , σAB ∈ S(HAB ),
i=1
tr[Pi ] i=1
tr[Pi ]
the following inequality holds:
1 1
Then, for 2 ∥ρ − p∥ ≤ ε ≤ 1 and 2 ∥σ − q∥1 ≤ δ ≤ 1, it holds D(ρAB ∥σAB )
that 1
≤ [DA (ρAB ∥σAB ) + DB (ρAB ∥σAB )] ,
 δ  1 − 2 ∥h(σAB )∥∞
|D(ρ∥σ)
b − D(ρ∥σ)| ≤ ε + log(me −1 )
lme  (157)
ε 
+ (1 + ε)h with
1+ε
lm −1/2 −1/2 −1/2 −1/2
e +δ
 δ 
+2 fm , h(σAB ) := σA ⊗ σB σAB σA ⊗ σB − 1AB ,
e −1 ,m
e −1
lme lm
e +δ (158)
(153)
and for X = A, B
e = 1 − m.
with lm e In particular, if [ρ, σ] = 0, ε and δ can be DX (ρAB ∥σAB ) := D(ρAB ∥σAB )−D(ρX c ∥σX c ) , (159)
taken as 0 such that the RHS of eq. (153) is zero.
Moreover, we can further simplify the previous bound to whenever ∥h(σAB )∥∞ < 1/2. Note that the term ∥h(σAB )∥∞
provides a measure of how far σAB is from being a tensor
√ √ e√
5 log2 m product between A and B. This result, and subsequent exten-
|D(ρ∥σ)
b − D(ρ∥σ)| ≤ ( 2 − log m)
e ε+ δ.
1−m e sions with additional conditions on σAB , are expected to find
(154) applications on various tasks in quantum information theory,
and in particular, have proven to be essential for some recent
Proof: Our argument is a slight variation of Matsumoto’s proofs of positivity of modified logarithmic Sobolev inequal-
minimal reverse test [24] (see also [23]). We can write the ities (MLSIs) for quantum Markov semigroups modelling
BS-entropy as the relative entropy of two commuting density thermal dissipative evolutions on quantum spin systems [30],
matrices [31], [74], [75]. It is important to remark that eq. (157) is
equivalent to a generalization of the property of superadditivity
D(ρ∥σ)
b = D(p∥q), (155) of the relative entropy, as shown in [76].

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5890 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

In [22], some authors of the current manuscript tried to using lemma 5.8. We then have
extend the previous result for the Umegaki relative entropy
to the BS-entropy framework. However, we showed that the |D(ρAB ∥σAB ) − D(ρA ⊗ ρB ∥σA ⊗ σB )|
BS-entropy cannot satisfy a property of superadditivity, which  √ √ 5 log2 me −1 √
e −1 + 2
≤ log m ε+ δ,
makes it impossible to obtain a quasi-factorization for the BS- lm
e
entropy in the spirit of eq. (157) without an additive error (165)
term. Instead, we proved a result of weak quasi-factorization,
from which we recovered eq. (157) if the marginals of ρAB concluding thus the proof.
and σAB commute. Here, we can prove another result along Note that, even though there is a caveat in this result in
these lines as a consequence of our continuity bound for the the form of an additive error term, which prevents it from
relative entropy. Indeed, as a consequence of theorem 5.13, being useful to prove the positivity of MLSIs, it presents
we obtain the following result of quasi-factorization for the the advantage with respect to eq. (157) that there is no
relative entropy with an additive error term. multiplicative error term in this case, which might be of
Corollary 7.6 (Weak Quasi-Factorization for the Relative more interest for some other contexts, such as for entropy
Entropy): Given ρAB , σAB ∈ S(HA ⊗ HB ) such that accumulation [77], [78] or in the line of the applications given
ker(σX ) ⊂ ker(ρX ) for X = A, B, AB, we have: by the Brascamp-Lieb dualities [79].

D(ρAB ∥σAB ) ≤ DA (ρAB ∥σAB ) + DB (ρAB ∥σAB )

F. Minimal Distance to Separable States
+ ξRE (ρAB , σAB ) ,
In this section, we show how to reprove the continuity
(160) bounds for the relative entropy of entanglement in [9] from
the ALAFF method and how this strategy generalizes if we
with quantify the minimal distance to the set of separable states in

e −1
log m
 terms of the BS-entropy instead.
1/2
ξRE (ρAB , σAB ) := √ + 1 ∥ρAB − ρA ⊗ ρB ∥1 Let C ⊂ S(H) be a compact convex subset of the set
2
of quantum states with at least one positive definite state.
5 log2 m
e −1 1/2 We can define the minimal distance to C in terms of the relative
+√ ∥σAB − σA ⊗ σB ∥1 ,
2(1 − m)
e entropy as
(161)
DC (ρ) := inf D(ρ∥γ). (166)
γ∈C
−1 −1 −1 −1
n o
−1
where me = 12 max ⊗
σA σB ,∞
.σAB ∞ As explained in [9], the fact that C contains a positive definite
Proof: The difference between the relative and the two state guarantees that DC (ρ) < ∞ for all ρ ∈ S(H). Moreover,
conditional entropies can be written as the infimum is attained, as follows from the fact that the
relative entropy is lower semi-continuous [41] and Weierstrass’
D(ρAB ∥σAB ) − DA (ρAB ∥σAB ) − DB (ρAB ∥σAB ) theorem on extreme values of such functions [80, Theorem
= −D(ρAB ∥σAB ) + D(ρA ⊗ ρB ∥σA ⊗ σB ) . 2.43]. Examples of C include SEPAB , the set of separable
(162) states for systems A, B, and

Therefore, we can apply theorem 5.13 to obtain a continu- {d−1

A 1A ⊗ σB : σB ∈ S(HB )}, (167)
ity bound for the difference between the last two relative which yields DC (ρAB ) = −Hρ (A|B) + log dA . The quantity
entropies, obtaining DSEPAB is known as the relative entropy of entanglement [81],
[82]. It constitutes a tight upper bound on the distillable
|D(ρAB ∥σAB ) − D(ρA ⊗ ρB ∥σA ⊗ σB )| entanglement [82], [83]. This is the quantity we focus on for
 δ   ε 
now.
≤ ε+ log(me −1 ) + (1 + ε)h
lm
e 1+ε Lemma 7.7: Let C ⊂ S(H) be a compact convex set
lm + δ  δ 
containing at least one positive definite state. Then, DC is
+2 fm ,
e
e −1 ,m
e −1
lme lm
e +δ convex on S(H).
(163) Proof: This follows directly from the joint convexity of
the relative entropy.
with In order to apply the ALAFF method, we need to prove
almost concavity next.
1 1
ε := ∥ρAB − ρA ⊗ ρB ∥1 , δ := ∥σAB − σA ⊗ σB ∥1 , Lemma 7.8: Let C ⊂ S(H) be a compact convex set
2 2 containing at least one positive definite state. Moreover, let
(164)
ρ1 , ρ2 ∈ S(H) and p ∈ [0, 1]. Then,
−1 −1 −1 −1
n o
−1 DC (pρ1 + (1 − p)ρ2 ) ≥ pDC (ρ1 ) + (1 − p)DC (ρ2 ) − h(p).
e = 12 max σA
and for m ⊗ σB ∞
, σ AB ∞ , lm e = 1−
m,.
e Moreover, we can apply the simplification of theorem 5.13 (168)

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5891

Proof: We can use the almost concavity of the relative Almost concavity requires more work in this case.
entropy. Let τ the state that achieves the infimum in DC (pρ1 + Lemma 7.12: Let C ⊂ S(H) be a compact convex set
(1 − p)ρ2 ). By theorem 5.1 and point 1 of proposition 5.2, containing the maximally mixed state. Moreover, let ρ1 , ρ2 ∈
we obtain that S(H), p ∈ [0, 1), and d ∈ N, d ≥ 2 the dimension of H.
Then,
DC (pρ1 + (1 − p)ρ2 ) ≥ pD(ρ1 ∥τ ) + (1 − p)D(ρ2 ∥τ ) − h(p)
b C (pρ1 + (1 − p)ρ2 ) ≥ pD
D b C (ρ1 ) + (1 − p)D
b C (ρ2 ) − gd (p).
≥ pDC (ρ1 ) + (1 − p)DC (ρ2 ) − h(p) ,
(169) (175)
d
which is the assertion. Here, gd (p) := p1/d h(p) − log(1 − p1/d ) for p ∈ (0, 1) and
Finally, we need the following estimate: gd (0) := 0.
Lemma 7.9: Let H = HA ⊗ HB . It holds that Proof: In order to apply the almost concavity of the BS-
entropy, we need to control the minimal eigenvalue of τ , the
sup |DSEPAB (ρ) − DSEPAB (σ)| ≤ log min{dA , dB }. best approximation of ρ = pρ1 + (1 − p)ρ2 in C. To this end,
ρ,σ∈S(H)
1 we will use a strategy inspired by [84]. Let τs be the state
2 ∥ρ−σ∥1 =1
achieving the infimum in
(170)
b ρ sτ + (1 − s) 1
 
inf D (176)
PddAA ≤ dB . For a pure
Proof: Without loss of generality, let τ ∈C d
state |ψ⟩ with Schmidt decomposition i=1 λi |iA ⟩ ⊗ |iB ⟩, let
for some s ∈ (0, 1) which we will specify later. Clearly,
dA
1 X b ρ sτs + (1 − s) 1 .
 
τψ = |iA ⟩⟨iA | ⊗ |iB ⟩⟨iB | . (171) b C (ρ) ≤ D
D (177)
dA i=1 d
This state is manifestly separable. Then, Furthermore, with τ̂ a state such that Db C (ρ) = D(ρ∥τ̂
b ),
b ρ sτs + (1 − s) 1 ≤ D b ρ sτ̂ + (1 − s) 1
   
sup |DSEPAB (ρ) − DSEPAB (σ)| D
ρ,σ∈S(H) d d
1
2 ∥ρ−σ∥1 =1 ≤ sup D(|ψ⟩⟨ψ| ∥τψ ) ≤D b C (ρ) − log s , (178)
|ψ⟩⟨ψ|∈S(H)

= log dA . (172) as sτ̂ +(1−s) 1d ≥ sτ̂ and the logarithm is operator monotone.
Since D b C (ρ) < ∞ we have ker τ̂ ⊆ ker ρ, thus, we can restrict
In the first inequality, we have used that DSEPAB is positive τ̂ to the support of ρ, where τ̂ is positive definite. Combining
and convex. this bound with theorem 6.3, we infer
This allows us to prove via the ALAFF method a continuity
b C (pρ1 + (1 − p)ρ2 )
D
bound for the relative entropy of entanglement:
Theorem 7.10: For ε ∈ [0, 1] and H = HA ⊗ HB , it holds b pρ1 + (1 − p)ρ2 sτs + (1 − s) 1 + log s
 
≥D
that for ρ, σ ∈ S(H) with 12 ∥ρ − σ∥1 ≤ ε d
d
|DSEPAB (ρ) − DSEPAB (σ)| ≤ ε log min{dA , dB } ≥ pDb C (ρ1 ) + (1 − p)D
b C (ρ2 ) − h(p) + log s.
  1−s
ε (179)
+ (1 + ε)h .
1+ε
Here, we have used point 1 of proposition 6.5. Finally,
(173) d
we have to choose s such that 1−s h(p) − log s goes to zero
+
Proof: This follows from remark 4.7, using lemma 7.7, for p → 0 and is non-decreasing on p ∈ [0, 1/2]. It turns out
lemma 7.8, point 4 of proposition 5.2, and lemma 7.9. that s = 1 − p1/d is a convenient choice, see lemma G.1 and
theorem 7.10 recovers the bound [9, Corollary 8], proven lemma G.2.
with very similar methods, which improved over the earlier Remark 7.13: Note that we could have substituted gd in
bound in [84]. The interest of executing the proof here is that lemma 7.12 by a symmetrized version
a similar strategy will give us bounds on a BS-entropy version (
gd (p) p ∈ [0, 1/2]
of the relative entropy of entanglement, as we will show now. g̃d (p) := (180)
We define gd (1 − p) p ∈ [1/2, 1]
in order to obtain
D
b C (ρ) = inf D(ρ∥γ)
b , (174)
γ∈C
b C (pρ1 + (1 − p)ρ2 ) ≥ pD
D b C (ρ1 ) + (1 − p)D
b C (ρ2 ) − g̃d (p)
which measures how far ρ is from C in terms of the BS-
(181)
entropy. The infimum is attained as the BS-entropy is also
lower semi-continuous [25, Section 10]. Convexity follows for all p ∈ [0, 1] and g̃d (0) = g̃d (1) = 0. For the ALAFF
again from the joint convexity of the BS-entropy. method with s = 0, however, it is only relevant what happens
Lemma 7.11: Let C ⊂ S(H) be a compact convex set on [0, 1/2].
containing at least one positive definite state. Then, D
b C is The final estimate we need in order to apply the ALAFF
convex on S(H). method is proven in a very similar way as lemma 7.9.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5892 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

d
Lemma 7.14: Let H = HA ⊗ HB . It holds that Here, gd (p) := p1/d h(p) − log(1 − p1/d ) for p ∈ (0, 1) and
sup |Db SEP (ρ) − D
b SEP (σ)| ≤ log min{dA , dB }. gd (0) = 0.
ρ,σ∈S(H)
AB AB
Proof: It holds that for ρ, σ ∈ S(H) with 12 ∥ρ − σ∥1 ≤ ε
1
2 ∥ρ−σ∥1 =1
b var (A|B) − H
|H b var (A|B)| = |D
b C (ρ) − D
b C (σ)| , (189)
(182) ρ σ 0 0

since the normalization does not matter. Thus to apply ALAFF,

PddAA ≤ dB . For a pure
Proof: Without loss of generality, let
state |ψ⟩ with Schmidt decomposition i=1 λi |iA ⟩ ⊗ |iB ⟩, let we need to bound
again sup |D
b C (ρ) − D
b C (σ)| . (190)
0 0
dA ρ,σ∈S(H)
1 X 1
τψ = |iA ⟩⟨iA | ⊗ |iB ⟩⟨iB | , (183) 2 ∥ρ−σ∥1 =1
dA i=1 b C (ρ) ≥ 0 for all states ρ,
Using eq. (91) and the fact that D 0
which is a separable state. Then, we obtain
sup |D
b SEP (ρ) − D b SEP (σ)|
ρ,σ∈S(H)
AB AB
sup |D
b C (ρ) − D
0
b C (σ)|
0
1 ρ,σ∈S(H)
2 ∥ρ−σ∥1 =1 1
2 ∥ρ−σ∥1 =1
b ρvar (A|B) + log dA
≤ sup −H
≤ sup D(|ψ⟩⟨ψ|
b ∥τψ ) ρ∈S(H)
|ψ⟩⟨ψ|∈S(H)
≤ 2 log dA . (191)
= log dA . (184)
The assertion follows from combining the above with
In the above inequality, we have used that D
b SEP is positive
AB lemma 7.11, lemma 7.12 with lemma G.1 and lemma G.2,
and convex. Note that |ψ⟩ is in the support of τψ . and lemma G.3 to apply remark 4.7.
Theorem 7.15: For ε ∈ [0, 1], H = HA ⊗ HB , and dAB ∈ Remark 7.17: Note that the findings of corollary 7.16 and
N, dAB ≥ 2, it holds that for ρ, σ ∈ S(H) with 12 ∥ρ − σ∥1 ≤ proposition 6.7 provide a formal proof that H b ρ (A|B) and
ε var
Hρ (A|B) are different in general. Indeed, while we have
b
|Db SEP (ρ) − D
AB
b SEP (σ)| ≤ ε log min{dA , dB }
AB just shown that the latter quantity is continuous on S(H) as a
consequence of the results of this section, in proposition 6.7
 
ε
+ (1 + ε)gdAB . we showed that the former quantity is in general discontinuous
1+ε
on S(H).
(185)
d
Here, gd (p) := p1/d h(p) − log(1 − p1/d ) for p ∈ (0, 1) and G. Rains Information
gd (0) = 0.
Proof: As shown in lemma G.3, it holds that gd (p)/(1−p) Inspired by the Rains bound from entanglement theory [85],
is non-decreasing on [0, 1] for all d ∈ N, d ≥ 2. Thus, for any divergence D, the generalized Rains bound of a
the assertion follows from remark 4.7 using lemma 7.11, quantum state ρAB ∈ S(HA ⊗ HB ) was defined in [86] by
lemma 7.12 with lemma G.1 and lemma G.2, and lemma 7.14. R(ρAB ) := min D(ρAB ∥σAB ) , (192)
σAB ∈PPT′ (A:B)
To end this section, let us investigate the choice where the minimization is taken over the Rains set
C0 := {d−1
A 1A ⊗ σB : σB ∈ S(HB )}. (186)
n o
PPT′ (A : B) := σAB : σAB ≥ 0, σAB TB
≤ 1 . (193)
1
From the discussion after eq. (15), we know that
TB
b ρ (A|B) ≤ b ρvar (A|B) , Where · denotes the partial transpose in the B-system. This
H sup − D(ρ b AB ∥ 1A ⊗σB ) =: H
σB ∈S(HB )
definition can be easily extended to channels in the following
way. For a quantum channel TA′ →B : S(HA ⊗ HA′ ) →
(187)
S(HA ⊗ HB ), we define
but equality does not hold in general. This is different from the
Umegaki relative entropy, where the conditional entropy coin- R(T ) := max R(TA′ →B (ϕAA′ )) , (194)
ρA ∈S(HA )
cides with its variational expression. Nonetheless, we obtain
a continuity bound for H b ρvar (A|B) from the approach in this for ϕAA′ a purification of ρA . In particular, for the Umegaki
section. relative entropy, we introduce the Rains information as
Corollary 7.16: Let H = HA ⊗ HB . For ε ∈ [0, 1] and R(T ) := max min D(TA′ →B (ϕAA′ )∥σAB ) ,
dAB ∈ N, dAB ≥ 2, it holds that for ρ, σ ∈ S(H) with ρA ∈S(HA ) σAB ∈PPT′ (A:B)
1
2 ∥ρ − σ∥1 ≤ ε (195)
b var (A|B) − H
|H b var (A|B)| ≤ 2ε log dA as well as the BS-Rains information by
ρ σ
 
ε
+ (1 + ε)gdAB . R(T
b ) := max min D(T
b A′ →B (ϕAA′ )∥σAB ) .
1+ε ρA ∈S(HA ) σAB ∈PPT′ (A:B)

(188) (196)

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5893

In the rest of the subsection, we will drop the subindex from Proof: Let us drop the subscripts from the channels for
the channels whenever it is clear in which systems they act. ease of notation. Firstly, note that SEPAB ⊂ PPT′ (A : B).
In [26], it was proven that the latter two quantities constitute Therefore,
upper bounds to the quantum capacity of a quantum channel.
Indeed, the following inequality holds for any channel T : R(T ) = max DPPT′ (A:B) (T (ϕAA′ ))
ρA ∈S(HA )

Q(T ) ≤ R(T ) ≤ R(T

b ). (197) ≤ max DSEPAB (T (ϕAA′ )) . (202)
ρA ∈S(HA )
Moreover, the BS-Rains information is a limit of Rains infor-
Hence, in general
mations induced by α-geometric Rényi divergences, which
can be written as single-letter formulas and computed via a max |DPPT′ (A:B) (ρAB ) − DPPT′ (A:B) (σAB )|
semidefinite program (SDP), as shown in [26]. The study of ρAB , σAB ∈S(HAB )
1
2 ∥ρAB −σAB ∥1 =1
these quantities is therefore of great interest for application
in the context of strong converses of quantum capacities of ≤ max DPPT′ (A:B) (ρAB )
ρAB ∈S(HAB )
channels.
≤ max DSEPAB (ρAB )
Here, as a consequence of corollary 5.9 and corollary 6.11, ρAB ∈S(HAB )
respectively, we can provide continuity results for both the
Rains information and the BS-Rains information, respectively, ≤ log min{dA , dB } , (203)
following the lines of theorem 7.10. Beforehand, we need to where in the last inequality we have used lemma 7.9. Fol-
justify that both quantities are well-defined, i.e., that each of lowing the lines of theorem 7.10, we have for ρAB , σAB ∈
these quantities is attained at a certain ρA ∈ S(HA ) and S(HA ⊗ HB ) with 12 ∥ρAB − σAB ∥1 ≤ ε the following
σAB ∈ PPT′ (A : B), and thus the minimum and maximum continuity bound:
in their definitions are properly written. For that, note that
we are first taking an infimum on the second input over the |DPPT′ (A:B) (ρAB ) − DPPT′ (A:B) (σAB )|
compact set PPT′ (A : B). Then, the infimum is attained and  ε 
the expression obtained is a continuous function, as we will ≤ ε log min{dA , dB } + (1 + ε)h .
1+ε
show below in eq. (203). Next, we perform an optimization (204)
problem on the first input over another compact set, namely
S(HA ). Thus, that supremum is also attained and both Rains Note that since PPT′ (A : B) does not only contain states
informations are well defined. but also subnormalized states, lemma 7.7 and lemma 7.8
From now on, for simplicity and for similarity with the are not directly applicable. One can however verify that the
quantities introduced in the previous section, given ρAB ∈ corresponding statements for PPT′ (A : B) still hold using the
S(HA ⊗ HB ), let us define same arguments. For simplicity, let us denote
 ε 
DPPT′ (A:B) (ρAB ) := min D(ρAB ∥σAB ) . (198) b(ε) := ε log min{dA , dB } + (1 + ε)h . (205)
σAB ∈PPT′ (A:B) 1+ε
Then, it is clear that we can rewrite, for a quantum channel To estimate an upper bound on the difference that
T : S(HA ⊗ HA′ ) → S(HA ⊗ HB ), appears in eq. (201), first note that, given T 1 , T 2 :
R(T ) := max DPPT′ (A:B) (T (ϕAA′ )) , (199) S(HA ⊗ HA′ ) → S(HA ⊗ HB ) two quantum channels with
1 1 2
2 T −T ≤ ε, and ρA ∈ S(HA ) with ϕAA′ a
ρA ∈S(HA )
1→1
for ϕAA′ a purification of ρA . The next step before applying purification of it, we have
the ALAFF method is bounding the difference between two 1 1
Rains informations of two quantum channels. For that, we will T 1 (ϕAA′ ) − T 2 (ϕAA′ ) 1
≤ T1 − T2 1→1
≤ ε.
2 2
use the 1 → 1 norm of the difference between channels. Let us (206)
recall that for T : S(HA ⊗ HA′ ) → S(HA ⊗ HB ) a quantum
channel, its 1 → 1 norm is given by Consider now ρ1 , ρ2 ∈ S(HA ) with respective purifications
ϕ1AA′ , ϕ2AA′ , the states in which the respective maxima of
∥T ∥1→1 := max ∥T (η)∥1 . (200)
η:∥η∥1 ≤1 R(T 1 ) and R(T 2 ) are attained. Then, we clearly have, for
For TA′ →B , the 1 → 1 norm coincides with the diamond norm. i, j = 1, 2 and i ̸= j,
Now, as a consequence of lemma 7.9 and theorem 7.10 from |R(T j ) − DPPT′ (A:B) (T i (ϕjAA′ ))|
the previous section, we can derive the following continuity
bound for the Rains information. = |DPPT′ (A:B) (T j (ϕjAA′ )) − DPPT′ (A:B) (T i (ϕjAA′ ))|
Theorem 7.18: For ε ∈ [0, 1] and TA1 ′ →B , TA2 ′ →B : ≤ b(ε) , (207)
S(HA ⊗ HA′ ) → S(HA ⊗ HB ) two quantum channels with
1 1 2 and thus,
2 ∥TA′ →B − TA′ →B ∥1→1 ≤ ε, we have:

|R(TA1 ′ →B ) − R(TA2 ′ →B )| ≤ ε log min{dA , dB } R(T i ) ≥ DPPT′ (A:B) (T i (ϕjAA′ )) ≥ R(T j ) − b(ε) . (208)
 ε 
+ (1 + ε)h . Therefore, we can conclude
1+ε
(201) |R(T 1 ) − R(T 2 )| ≤ b(ε) , (209)
Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5894 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

and consequently A natural question arises from the findings of this paper: Is
 ε  our method applicable to any other family of divergences? We
|R(T 1 ) − R(T 2 )| ≤ ε log min{dA , dB } + (1 + ε)h . expect this to be the case, since, as shown in section II, our
1+ε
(210) method only requires almost concavity and convexity (already
known for divergences) in order to work. Therefore, a result
of almost concavity with a “well-behaved” correction factor
In a similar way, we can also prove uniform continuity would be enough for the ALAFF method and is expected to
and provide explicit continuity bounds for the BS-Rains infor- exist, for families such as the α-sandwiched Rényi divergences
mation. Analogously to what we have done above for the or the α-geometric Rényi divergences, as they converge to
Rains information, we can define for ρAB ∈ S(HA ⊗ HB ) the quantities studied in this paper. This possibility will be
the following quantity: explored in a future manuscript.
Let us conclude this section, and our paper, with some
D
b PPT′ (A:B) (ρAB ) := min b AB ∥σAB ) , (211)
D(ρ
σAB ∈PPT′ (A:B) analysis of the results obtained here. For both the Umegaki
and the BS-entropies, we have presented results of almost
and thus, we can rewrite, for a quantum channel T : concavity in order to provide some continuity bounds. How-
S(HA ⊗ HA′ ) → S(HA ⊗ HB ), ever, while for the former (cf. theorem 5.1) we have shown
that the result is tight, for the latter (cf. theorem 6.3) we
R(T
b ) := max D
b PPT′ (A:B) (T (ϕAA′ )) , (212)
ρA ∈S(HA ) are certain that there is room for improvement. Indeed, our
almost concavity bound for the BS-entropy depends on the
for ϕAA′ a purification of ρA . We can finally use lemma 7.14
minimal eigenvalues of some of the states involved even in
and theorem 7.15 from the previous section, for the BS-
the simplified case of the BS-conditional entropy. In such a
entropy, to obtain a continuity bound for the BS-Rains infor-
case, numerical simulations, as well as analytical proof, have
mation. However, the bound obtained, as well as the procedure
shown us that there is a universal bound for the BS-conditional
employed to derive it, are a straightforward combination of the
entropy of a state which is independent of the state involved.
strategies of the continuity bound for the Rains information
Therefore, we would expect an almost convexity result for
theorem 7.18 and the continuity bound for the BS-entropy
the BS-conditional entropy being independent of the states
of entanglement from theorem 7.15. Therefore, we omit it,
involved, and this is clearly not the case at the moment. Nev-
to avoid unnecessary repetitions.
ertheless, there is no doubt that the BS-entropy, and quantities
Theorem 7.19: For ε ∈ [0, 1] and TA1 ′ →B , TA2 ′ →B :
derived from it, are “pathological” in some sense. First of
S(HA ⊗ HA′ ) → S(HA ⊗ HB ) two quantum channels with
1 1 2 all, we have shown that the BS-conditional entropy exhibits
2 ∥T − T ∥1→1 ≤ ε, we have: discontinuities in the presence of vanishing eigenvalues (cf.
b 1 ) − R(T
b 2 )| ≤ ε log min{dA , dB } + (1 + ε)gd
 ε  proposition 6.7), as opposed to the conditional entropy, which
|R(T AB ,
1+ε behaves well in that setting. This motivates the idea that the
(213) minimal eigenvalue of the involved states should appear in
where gd (t) := d
h(t) − log(1 − t1/d ). the most general bounds of almost concavity and continuity.
t1/d
Additionally, we can compare some upper bounds of some
entropic quantities derived from the Umegaki and the BS-
VIII. O UTLOOK entropy:
In this paper, we have introduced a generalisation of the • For the relative entropy, we have the following 3 bounds:
Alicki-Fannes-Winter method by Shirokov and applied it to
derive results of uniform continuity and explicit continuity −Hρ (A|B) ≤ log dA ,
bounds for divergences. We gave this generalisation the name Iρ (A : B) ≤ 2 log min{dA , dB } ,
ALAFF (cf. theorem 4.6) after the functions to which it applies e −1
D(ρ∥σ) ≤ log m σ . (214)
(almost locally affine functions). The method allows deriving
various continuity bounds for entropic quantities, by simply • For the BS-entropy, we have the following 3 bounds (cf.
proofing (joint) convexity and almost (joint) concavity of the proposition 6.6):
underlying divergence. −H b ρ (A|B) ≤ log dA ,
In particular, in the current paper, we have applied our
e −1
Ibρ (A : B) ≤ log dA m (ρA ) ,
ALAFF method to the specific cases of the Umegaki and
the Belavkin-Staszewski relative entropies. For both of them, D(ρ∥σ)
b ≤ log m−1
σ . (215)
we have proven results of almost concavity (for the Umegaki In the above m· , m
e · denote the minimal respectively minimal
case, our result is shown to be tight), and these, together with non-zero eigenvalue of the state in the index. It is remarkable
the well-known results of convexity for these quantities, have that for the conditional and BS-conditional entropy and the
yielded a plethora of results of continuity bounds for both mutual information, there appears no dependence on the
the Umegaki and BS-entropies, as well as for many other minimal eigenvalue of the argument, whilst for the BS-mutual
quantities derived from them. In particular, our results recover information this is the case.
the previously known almost tight continuity bounds for the Moreover, let us recall that, from the discussion in
conditional entropy and the (conditional) mutual information. remark 7.17, we know that the conditional BS-entropy and

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5895

A PPENDIX B
S UPPLEMENTS TO THE P ROOF OF T HEOREM 5.1
We will now show that the result of the inequality in eq. (35)
is still true, even if ρ1 , ρ2 , σ1 , σ2 are not full rank. We have
that
ker σ ⊆ ker σ1 ⊆ ker ρ1 . (217)
If ker σ ⊊ ker ρ1 we set
−1
Π
e ρ := Pker ρ ∩(ker σ)⊥ ,
1 1
Πρ1 := Π
eρ
1
Π
e ρ , (218)
1
1

and if ker σ ⊊ ker σ1 ,

−1
e σ := Pker σ ∩(ker σ)⊥ ,
Π 1 1
Πσ1 := Π
eσ
1 Π
e σ , (219)
1
1
Fig. 7. The red line is the BS-conditional entropy defined via the partial
trace evaluated at ρAB . The dots are the BS-entropy between the state ρAB normalised projections on the spaces in the index. Both of the
and 1A ⊗σB with σB ∈ S(HB ) sampled at random. The orange dots are latter are quantum states and fulfil
the cases when the −D(ρ b AB ∥ 1A ⊗σB ) exceeds H(A|B)b ρ . We sampled
a total of 100.000 pairs of ρAB and σB and evaluated both H(A|B)
b ρ and Πρ1 ρ1 = ρ1 Πρ1 = 0, Πσ1 σ1 = σ1 Πσ1 = 0,
−D(ρ
b AB ∥ 1A ⊗σB ). Only a tenth of all samples were kept in addition to (220)
the ones that violated the bound. Those were then plotted in ascending order
Πσ1 ρ1 = ρ1 Πσ1 = 0 .
w.r.t the magnitude of their BS-conditional entropy. We further controlled the
minimal eigenvalue and set HA ⊗ HB = C2 ⊗ C2 to reduce the risk of For 1 > ε > 0 and 1 > δ > 0, let
numerical flaws. (
εΠρ1 + (1 − ε)ρ1 if ker σ ⊊ ker ρ1
ρ1,ε = ,
ρ1 if ker σ = ker ρ1
its variational counterpart are different because the latter (
is continuous on S(H) and the former is not. One could δΠσ1 + (1 − δ)σ1 if ker σ ⊊ ker σ1
σ1,δ = . (221)
wonder whether the same difference appears for the BS- σ1 if ker σ = ker σ1
mutual information. Analogously to the case of the (Umegaki)
We then have that ker ρ1,ε = ker σ1,δ = ker σ. This means,
mutual information, we could define four possible versions
however, considering Tr [ρ1,ε (log σ − log σ1,δ )] we can reduce
of such a notion by optimizing over one marginal, both or
to the subspace where they are all full rank. We then apply the
none. Remarkably, we find that, when optimizing over both
Peierls-Bogoliubov inequality [41] and the multivariant trace
marginals, we have, assuming w.l.o.g. dA ≤ dB ,
inequality by Sutter et al. [44, Corollary 3.3]
Ibρvar (A : B) := inf D(ρ
b AB ∥σA ⊗ σB )
σA ,σB Tr [ρ1,ε (log σ − log σ1,δ )]
≤ log dA − b ρvar (A|B)
H ≤ 2 log min{dA , dB } . ≤ log Tr [exp (log(ρ1,ε ) + log(σ) − log(σ1,δ ))]
(216) Z∞ h it−1 −it−1 i
≤ log dt β0 (t) Tr ρ1,ε σ1,δ2 σσ1,δ2 .
Comparing this bound to the one shown above for Ibρ (A : B), −∞
which we prove to be tight in proposition 6.6, we realize that
(222)
the BS-mutual information and its variational counterpart (with
optimization over both marginals) are also different in general. Both of the traces on the LHS and RHS of eq. (222) can
To conclude, the literature concerning continuity bounds for without change be extended to the full Hilbert space again.
entropic quantities is much broader than the results collected Next, we take limits on both sides of the inequality and in
here. For Rényi and Tsallis entropies, many results concerning doing so recover the claim. We first note that the limit ε →
their continuity can be derived from other techniques, such as 0 requires no more argument as both sides are linear in ε.
majorization flows, and can be found in texts such as [87], Hence, we get
[88], and [53]. Additionally, some of these results for the von
Neumann entropy, Rényi and Tsallis entropies, as well as their Tr [ρ1 (log σ − log σ1,δ )]
classical counterparts, can be extended to energy-constrained Z∞ h it−1 −it−1 i
systems in infinite dimensions, as shown in [89] and [9] (see ≤ log dt β0 (t) Tr ρ1 σ1,δ2 σσ1,δ2 . (223)
also the recent [90]). We leave for future work the possibility −∞
of extending the results presented here to a similar framework.
The limit δ → 0 on the other hand is, in the case of
ker σ ⊊ ker σ1 , a little more involved. Due to the orthogonality
A PPENDIX A in eq. (220) we cannot only split up the logarithm but also
N UMERICAL I NVESTIGATION OF THE VARIATIONAL eliminate terms. More specifically, we have
D EFINITION OF THE BS-C ONDITIONAL E NTROPY
See Fig. 7. log σ1,δ = log(δΠσ1 ) + log((1 − δ)σ1 ) , (224)

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5896 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

where the logarithms in the RHS have to be understood 2) With j, k = 1, 2, j ̸= k and mρ

e j ≤ σj , we find
as living in the support of the respective argument (and
it−1 −it−1 it−1 −it−1
complemented with zeros in the rest). Hence, we obtain for σj 2
ρj σj 2
≤ σj 2
e −1 σj σj
m 2

the LHS of eq. (223)

e −1 Pσj ≤ m
≤m e −1 1 (230)
Tr [ρ1 (log σ − log σ1,δ )]
where Pσj is the projection onto the support of σj .
= Tr [ρ1 (log σ − log(δΠσ1 + (1 − δ)σ1 ))]
We therefore find
= Tr [ρ1 (log σ − log((1 − δ)σ1 )]
Z∞
+ Tr [ρ1 log(δΠσ1 )]
cj ≤ e −1 Tr [σk ] = m
dtβ0 (t)m e −1 . (231)
= Tr [ρ1 (log σ − log((1 − δ)σ1 )]
−∞
= Tr [ρ1 (log σ − log σ1 ] + log(1 − δ) . (225)
By the monotonicity of the logarithm, we obtain fc1 ,c2 ≤
Moreover, for the RHS of eq. (223) we use that fm e −1 and hence fc1 ,c2 + h ≤ fm
e −1 ,m e −1 + h.
e −1 ,m
z
σ1,δ = δ z Πzσ1 + (1 − δ)z σ1z , (226) 3) For j, k = 1, 2, j ̸= k we have

for any z ∈ C, where the last exponential has to be under- Z∞ h it−1 −it−1 i

stood again in the support of the respective argument. Thus, cj = dtβ0 (t)Tr ρj,AB 1A ⊗(ρj,B
2
ρk,B ρj,B2 )
we obtain −∞
h it−1 −it−1 i Z∞ h it−1 −it−1 i
Tr ρ1 σ1,δ2 σ σ1,δ2 = dtβ0 (t)Tr ρj,B (ρj,B
2
ρk,B ρj,B2 )
h it−1 −it−1 i
= (1 − δ)−1 Tr ρ1 σ1 2 σσ1 2 −∞
h it−1 −it−1 i
Z∞
it−1 −it−1
+ (1 − δ) 2 δ 2 Tr ρ1 σ1 2 σΠσ1 2 = dtβ0 (t)Tr [ρk,B ] = 1. (232)
it−1 −it−1
h it−1 −it−1 i −∞
+ δ 2 (1 − δ) 2 Tr ρ1 Πσ12 σσ1 2
h it−1 −it−1 i We used that the functional calculus has the property
+ δ −1 Tr ρ1 Πσ12 σΠσ1 2 that f (A ⊗ B) = f (A) ⊗ f (B) for A, B self-adjoint,
h it−1 −it−1 i as can easily be verified by direct computation, and that
= (1 − δ)−1 Tr ρ1 σ1 2 σσ1 2 . (227) the trace is cyclic. This gives us fc1 ,c2 = f1,1 = 0 which
Taking the limit δ → 0 now directly follows from the concludes the claim.
1
continuity of the logarithm. We thereby conclude 4) The derivative of p 7→ 1−p h(p) at p ∈ (0, 1) is
log(p)
− (1−p)2 ≥ 0, which proves the second assertion. For
pTr [ρ1 (log(σ) − log(σ1 ))] p 7→ 1−p 1
fm1 ,m2 (p) = 1−p p
log(p + m1 (1 − p)) +
Z∞ h it−1 −it−1 i log(1−p + m2 p) we use similar reasoning. First we use
≤ p log dt β0 (t) Tr ρ1 σ1 2 σσ1 2 , that m2 ≥ 1 hence log(1−p+m2 p) = log(1+(m2 −1)p)
−∞ is monotone in p, i.e. in particular non-decreasing.
p
(228) Second we note that p 7→ 1−p log(p + m1 (1 − p))
is monotone in p, because forming the derivative at
for σ1 , σ2 , ρ1 not full rank. p ∈ (0, 1), we get

A PPENDIX C 1  p 
+ log(p + m 1 (1 − p)) − p
P ROOF OF P ROPOSITION 5.2 (1 − p)2 p + (1 − p)m1
1  p
We first of all note that for all ρ1 , ρ2 ∈ S(H) we have ≥ 2
1
∥ρ1 − ρ2 ∥1 ≤ 1, hence as a direct consequence fc1 ,c2 + (1 − p) p + (1 − p)m1
2
1
∥ρ1 − ρ2 ∥1 h ≤ fc1 ,c2 + h. We therefore will drop the p + (1 − p)m1 − 1 
2 + −p
1
2 ∥ρ1 − ρ2 ∥ in front of the h here already. p + m1 (1 − p)
1  m (1 − p) + 2p − 1 
1) If σ1 = σ2 =: σ, we find for j = 1, 2 that =
1
− p
Z∞ (1 − p)2 p + (1 − p)m1
(m1 − 1)(p − 1)2 
h i
cj =
it−1 −it−1
dtβ0 (t)Tr ρj σ 2 σσ 2 1 
= 2
(1 − p) p + (1 − p)m1
−∞
Z∞ ≥ 0. (233)
= dtβ0 (t)Tr [ρj ] = 1 . (229)
We used that for x ≥ 1, log(x) ≥ x−1
x (this can be seen
−∞
by taking the derivative and realizing that both sides
The reduction of fc1 ,c2 + h to h then happens because coincide for x = 1) and m1 ≥ 1. This concludes the
log(p + (1 − p)) = log(1) = 0 gives fc1 ,c2 = 0. claim.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5897

A PPENDIX D A PPENDIX E
P ROOF OF L EMMA 5.12 P ROOF OF P ROPOSITION 6.5
We first show that for s ≥ m,
e S 0 is s-perturbed ∆-invariant. 1) If σ1 = σ2 = σ, then for j = 1, 2 with α(O, P, Q) from
For that purpose let σ1 , σ2 ∈ S 0 , then we find theorem 6.3
∆± (σ1 , σ2 , ρ) = sρ + (1 − s)[σ1 − σ2 ]± ≥ mρ, (234) 1/2 1/2 −1/2 −1/2
ĉj = α(ρj , ρj σ −1 ρj , ρj σρj )
e
which immediately gives the kernel inclusion as well as Z∞ Z∞
the condition to be lower bounded by mρ. e Therefore, = dtβ0 (t)Tr [ρj ] = dtβ0 (t) = 1 (241)
∆± (σ1 , σ2 , τ ) ∈ S 0 which makes S 0 an s-perturbed ∆- −∞ −∞
invariant set. We show the other direction by contrapositive.
Let s < m.
e Since m e < 1 and rank ρ ≥ 2 we find an ε > 0 and which gives us immediately fĉ1 ,ĉ2 + ĉ0 h = ĉ0 h.
e < ρ− 2ε |i⟩⟨i|
two orthonormal |0⟩ , |1⟩ ∈ supp ρ, such that mρ 2) For j, k = 1, 2 with j ̸= k we first have σk ≤ m−1 σj
for i = 0, 1. We then have that giving us
ε ε 1/2 1/2 −1/2 −1/2
σ1 = ρ + |0⟩⟨0| − |1⟩⟨1| ĉj ≤ α(ρj , ρj σj−1 ρj , ρj m−1 σj ρj )
2 2
ε ε Z∞
σ2 = ρ − |0⟩⟨0| + |1⟩⟨1| (235) −1
2 2 =m dtβ0 (t)Tr [ρj ] = m−1 . (242)
1
manifestly are contained in S 0 . Furthermore, 2 ∥σ1 − σ2 ∥1 = −∞
ε and
using α(O, P, Q) as in theorem 6.3 defined. Since
ε −1
[σ1 − σ2 ]+ = |0⟩⟨0| ĉ0 ≤ m−1 and because the logarithm is monotone this
ε−1 [σ1 − σ2 ]− = |1⟩⟨1| . (236) immediately gives fĉ1 ,ĉ2 + ĉ0 h ≤ fm−1 ,m−1 + m−1 h.
3) The proof is along the same lines as the one for 2.,
We will now show that there exists no τ ∈ S(H) such that however with σj = d−1 A 1A ⊗ρj,B . We just have to
∆± (σ1 , σ2 , τ ) ∈ S 0 again, meaning S 0 is not s-perturbed show that the minimal eigenvalue of σj is bounded from
∆-invariant. Assume there is an operator τ ≥ 0 such that below by m. We use that TA : τ 7→ d−1 A 1A ⊗τB is
∆± (σ1 , σ2 , τ ) ∈ S 0 we then would have a conditional expectation and that d−1 A 1 A ⊗τB is full
⊥ ⊥ ⊥ ⊥ rank if τ was full rank [54, Theorem 4.13]. This means,
|0⟩⟨0| ∆+ (σ1 , σ2 , τ ) |0⟩⟨0| = |0⟩⟨0| sτ |0⟩⟨0| however,
⊥ ⊥
≥m
e |0⟩⟨0| ρ |0⟩⟨0|
⊥ ⊥ ⊥ ⊥ (d−1
A 1A ⊗ρj,B )
−1
= TA (ρj )−1 ≤ TA (ρ−1
j ), (243)
|1⟩⟨1| ∆− (σ1 , σ2 , τ ) |1⟩⟨1| = |1⟩⟨1| sτ |1⟩⟨1|
⊥ ⊥ where we used [54, Theorem 4.16]. This gives us
≥m
e |1⟩⟨1| ρ |1⟩⟨1| (237)
⊥
where |i⟩⟨i| := Pρ − |i⟩⟨i| for i = 0, 1. Here Pρ is the projec- (d−1
A 1A ⊗ρj,B )
−1
∞
≤ TA (ρ−1 ) ∞
tion on the support of ρ. We further used ∆± (σ1 , σ2 , τ ) ≥ mρ
e ≤ ρ−1 ≤ m−1 . (244)
∞
as ∆± (σ1 , σ2 , τ ) are in S 0 by assumption. To fulfil eq. (237)
we clearly need to choose s > 0 and since s < m e we directly −1
Hence, we have that (d−1 A ⊗ ρj,B )
−1
∞
the minimal
obtain the conditions −1
eigenvalue of dA ⊗ ρj,B is bounded from below by
⊥ ⊥ ⊥ ⊥ m. From here on the proof is analogous to the one in 2.
|0⟩⟨0| τ |0⟩⟨0| ≩ |0⟩⟨0| ρ |0⟩⟨0| (238)
We obtain fĉ1 ,ĉ2 + ĉ0 h ≤ fm−1 ,m−1 + ĉ0 h and again use
and eq. (244) to get fm−1 ,m−1 + ĉ0 h ≤ fm−1 ,m−1 + m−1 h.
⊥ ⊥
|1⟩⟨1| τ |1⟩⟨1| ≩ |1⟩⟨1| ρ |1⟩⟨1|
⊥ ⊥
. (239) 4) The proof is completely analogous to the one in 4.
of section C.
This gives us,
h i
⊥ ⊥ A PPENDIX F
Tr [τ ] ≥ Tr |0⟩⟨0| τ |0⟩⟨0| + |0⟩⟨0| τ |0⟩⟨0|
h i P ROOF OF P ROPOSITION 6.6
⊥ ⊥ ⊥ ⊥
= Tr |0⟩⟨0| τ |0⟩⟨0| + |0⟩⟨0| |1⟩⟨1| τ |1⟩⟨1| |0⟩⟨0|
h i 1) We begin with the BS-conditional information. The
⊥ ⊥ ⊥ ⊥ upper bound on H
b · (A|B) can be obtained by
> Tr |0⟩⟨0| ρ |0⟩⟨0| + |0⟩⟨0| |1⟩⟨1| ρ |1⟩⟨1| |0⟩⟨0|
h i
⊥ ⊥
= Tr |0⟩⟨0| ρ |0⟩⟨0| + |0⟩⟨0| ρ |0⟩⟨0| b AB ∥d−1 1A ⊗ρB ) + log dA
b ρ (A|B) = −D(ρ
H A

= Tr [Pρ ρ] = Tr [ρ] = 1, (240) ≤ log dA . (245)

where we used that |0⟩ and |1⟩ are orthogonal, hence where we used the non-negativity of D(·∥·)
b on quantum
⊥ ⊥ 2
|0⟩⟨0| |1⟩⟨1| = |1⟩⟨1| |0⟩⟨0| = |0⟩⟨0| and |0⟩⟨0| = states. The bound is attained if one inserts the maximally
⊥ 2 ⊥
|0⟩⟨0| , (|0⟩⟨0| ) = |0⟩⟨0| . We thus conclude τ ̸∈ S(H) mixed state, i.e., ρAB = d−1
AB 1AB . For the lower bound,
proving the claim. we use that −D(·∥·)
b is jointly concave and TrA [·] linear

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5898 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

which means without loss of generality one can assume bipartite space HA ⊗ HB with HA having dimension
ρ to be pure, i.e., a rank one projection. Then dA and HB dimension dB = dA + 1. Furthermore, let
ε ∈ (0, 1). We then consider sets of orthonormal vectors
H
b |ψ⟩⟨ψ| (A|B)
{|iA ⟩}di=1
A
⊂ HA , {|iB ⟩}di=1
A
⊂ HB and define
= −D(|ψ⟩⟨ψ|
b ∥ 1A ⊗PB )
A −1 r
dX
h
1/2 1/2
i ε
= −Tr |ψ⟩⟨ψ| log |ψ⟩⟨ψ| (1A ⊗PB−1 ) |ψ⟩⟨ψ| |ψ⟩ := |iA ⟩ ⊗ |iB ⟩
i=1
dA − 1
= − log Tr |ψ⟩⟨ψ| (1A ⊗PB−1 ) √
 
+ 1 − ε |(dA )A ⟩ ⊗ |(dA )B ⟩
= − log Tr PB PB−1 ,
 
(246) dA p
X
with PB = TrA [|ψ⟩⟨ψ|]. Employing the Schmidt decom- = λi |iA ⟩ ⊗ |iB ⟩ . (252)
position to |ψ⟩⟨ψ| we find that i=1

d
X with the λi defined accordingly. We find that
PB = λ2i Pi (247) dA
X
i=1
ρA := TrB [|ψ⟩⟨ψ|] = λi |iA ⟩⟨iA | ,
with Pi orthogonal rank one projections on HB , λ2i > i=1
d dA
λ2i = 1. Further d ≤ min{dA , dB } the Schmidt
P
0 and
X
i=1
ρB := TrA [|ψ⟩⟨ψ|] = λi |iB ⟩⟨iB | , (253)
rank. This gives us that i=1

 Xd and the Moore-Penrose pseudoinverse (in the case of ρA

Tr PB PB−1 = λ2i λ−2

i = d ≤ min{dA , dB }. it is an inverse)
i=1
dA
(248) X
ρ−1
A = λ−1
i |iA ⟩⟨iA | ,
Through monotonicity of the logarithm, we obtain the i=1
lower bound, i.e., dA
X
b ρ (A|B) ≥ − log min{dA , dB }. ρ−1
B = λ−1
i |iB ⟩⟨iB | . (254)
H (249)
i=1
This bound is attained for ρ a pure state with full We find
Schmidt rank, which can directly be seen from the above
Tr |ψ⟩⟨ψ| ρ−1 −1
 
calculations. A ⊗ ρB
2) We now tackle the BS-mutual information. The lower √ p
X λi λj
bound, i.e. Ibρ (A : B) ≥ 0, is a direct consequence of = ⟨iA |kA ⟩ ⟨kA |jA ⟩ ⟨iB |lB ⟩ ⟨lB |jB ⟩
λk λl
the data processing inequality [38]. Applying TrA [ · ], i,j,k,l
√ p
we find X λi λj
= δik δkj δil δlj
b AB ∥ρA ⊗ ρB ) ≥ D(ρ
Ibρ (A : B) = D(ρ b B ∥ρB ) = 0. λk λl
i,j,k,l
(250) X 1 (dA − 1)2 1
= = + , (255)
λi ε 1−ε
To proof the upper bound, we w.l.o.g assume that i
ρ−1
A ∞ ≤ ρ−1
B ∞ . We then use that ρA ⊗ ρB ≥ with which, as |ψ⟩⟨ψ| is a rank one projection
−1 −1
ρA ∞ PρA ⊗ ρB , where PρA is the projection to the
support of ρA . This gives us Ib|ψ⟩⟨ψ| (A : B)
h i
1/2 1/2
b AB ∥ρA ⊗ ρB )
Ibρ (A : B) = D(ρ = Tr |ψ⟩⟨ψ| log(|ψ⟩⟨ψ| ρ−1 A ⊗ ρ −1
B |ψ⟩⟨ψ| )
b AB ∥Pρ ⊗ ρB ) + log ρ−1 −1 −1
 
≤ D(ρ A A ∞
= log Tr |ψ⟩⟨ψ| ρA ⊗ ρB
= D(ρAB ∥ 1A ⊗ρB ) + log ρ
b −1 (dA − 1)2 1 (dA − 1)2
A ∞ = log( + ) ≥ log( ).
b ρ (A|B) + log ρ−1 ε 1−ε ε
= −H A ∞ (256)
≤ log min{dA , dB } + log ρ−1
A ∞
≤ log min{dA , dB } We directly obtain ρ−1
A ∞ = ρB ∞ =
−1 dA −1
ε and
by construction dA < dB , hence the bound in eq. (92)
+ log min{ ρ−1
A ∞
, ρ−1
B ∞
} (251)
gives
In the second equality we used that (ker ρA ) ⊗ HB ⊆
dA (dA − 1)
ker ρAB , so extending PρA to 1A has no effect. With Ib|ψ⟩⟨ψ| (A : B) ≤ log( ). (257)
ε
the next example, we will see that the bound is tight
and scales with log max{ ρ−1 −1
A ∞ , ρB ∞ } in some We first note that for ε = 1 − d1A we get equality in
cases. For that purpose let dA ∈ N, dA ≥ 2 and a eq. (257). What is, however, more interesting is the fact

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5899

that We will now show monotonicity in d of gd′ (p) for all p ∈

2 (0, 1/2). This will allow us to show the non-negativity of
(dA − 1) dA (dA − 1)
log( ) ≤ Ib|ψ⟩⟨ψ| (A : B) ≤ log( ), eq. (262) on (0, 1/2) only for d = 2 and conclude it for all
ε ε
(258) d ≥ 2. We have


with ∂ ′ p2/d−1 d(p1/d − 1) + (p1/d − 2) log(p)
g (p) =
dA (dA − 1) (dA − 1)2 dA ∂d d d3 (p1/d − 1)2
log( ) − log( ) = log( ). + log(1 − p) − log(p) . (263)
ε ε dA − 1
(259) The above is non-negative for p ∈ (0, 1/2), if
I.e., the error of the bound is of order log( dAdA−1 )
inde- 1
(2 − p1/d ) log ≥ d(1 − p1/d ) (264)
pendent of the ε. This means, that the scaling behaviour p
of the bound, in terms of the minimal non-zero eigen- equivalent to
value of ρA and ρB respectively is the best one can do.
1
1
3) The lower bound of the BS-CMI is again a consequence 1+ 1/d
log ≥ d (265)
of the data processing inequality. The upper bound is 1−p p
a direct consequence of the bounds obtained for the One obtains the last inequality by substitution of p = edt with
BS-conditional information due to the definition of the t ∈ (−∞, − log(2)
d ) giving us
conditional mutual information in eq. (13) 1
1

−dt 1 + ≥ d ⇔ −t 1 + ≥ 1 (266)
b ρ (A|C) − H
Ibρ (A : B|C) = H b ρ (A|BC) 1 − et 1 − et
≤ log dA + log min{dA , dBC } which is true for t ∈ (−∞, 0) hence in particular on
= log min{d2A , dABC } . (260) (−∞, − log(2)d ). We thereby have that for d ≥ 2 p ∈ (0, 1/2)
gd′ (p) ≥ g2′ (p). It is straightforward to see that g2′ (p) ≥ 0 on
We expect that the tightness of such a bound can be p ∈ (0, 1/2). This finally lets us conclude the claim that gd (p)
proven in a similar way to the one for the BS-mutual is non-decreasing on p ∈ [0, 1/2] as gd (p) is continuous on
information. [0, 1/2] by lemma G.1.
Lemma G.3: Let d ∈ N, d ≥ 2. Then, the function p 7→
A PPENDIX G
gd (p)/(1 − p) is non-decreasing on [0, 1).
B EHAVIOR OF gd
d
Proof: The argument follows similar lines as the one in
In this section, we study the function gd (p) := p1/d h(p) − lemma G.2. We first note that p 7→ 1−p 1
is non decreasing on
1/d
log(1 − p ) for p ∈ (0, 1) and a fixed d ∈ N, d ≥ 2. This [0, 1/2) and p 7→ gd (p) is as well, as proven in lemma G.2.
function appears in some of the continuity bounds in section Hence p 7→ 1−p 1
gd (p) is non decreasing on [0, 1/2]. What
VII-F. now remains to show is that it is non-decreasing on [1/2, 1).
Lemma G.1: Let d ∈ N, d ≥ 2. Then, limp→0+ gd (p) = 0. We can differentiate the function on the interval [1/2, 1) and
In particular, gd is continuous on p ∈ [0, 1). obtain
Proof: Since limp→0+ log(1 − p1/d ) = 0, we can focus ∂ gd (x) 1  dp−1/d h(p) − log(1 − p1/d )
d
on p1/d h(p). The assertion follows from applying L’Hospital’s =
∂x 1 − x x=p 1−p 1−p
rule twice. Indeed,
p1/d−1
d d(log(1 − p) − log(p)) + − p−1/d p−1 h(p)
lim h(p) = lim+ d(1 − p1/d )
p→0+ p1/d p→0 p1/d−1 /d 
d(−(1 − p)−1 − p−1 ) + dp−1/d (log(1 − p) − log(p))
= lim+
(1 − d)p1/d−2 /d2
  
p→0 1 −1/d 1 1
≥ p h(p) −
d3
 2−1/d 
p 1−p 1−p p
= lim+ + p1−1/d  
p→0 d − 1 1−p −1/d h(p)
+ (d − 1)p + log(1 − p)
= 0. (261) 1−p
log(1 − p1/d )

Continuity, therefore, follows from the definition of the func- +p −1/d
log(1 − p) −
tion. 1−p
Lemma G.2: Let d ∈ N, d ≥ 2. Then, the function gd is ≥ 0. (267)
non-decreasing on [0, 1/2]. 1
The last inequality holds since p ≥ 2 and d ≥ 2 hence
Proof: We can differentiate gd (p) on (0, 1/2). This yields
1 1
∂ 1  p2/d−1 − ≥ 0,
gd (p) = 1/d + (d − 1 + p−1 ) log(1 − p) 1−p p
∂p p d(1 − p1/d )
h(p)
+ log(1 − p) ≥ 0 ,

− (d − 1) log(p) 1−p
1 ′ log(1 − p1/d )
=: 1/d
gd (p) . (262) p−1/d log(1 − p) − ≥ 0. (268)
p 1−p

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
5900 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023

To see the last inequality, one can verify that p−1/d ≤ 1/(1−p) [23] F. Hiai and M. Mosonyi, “Different quantum f -divergencies and the
in this regime and that log(1 − p) ≥ log(1 − p1/d ). Thus reversibility of quantum operations,” Rev. Math. Phys., vol. 29, no. 7,
2017, Art. no. 1750023.
p 7→ g1−p
d (p)
is non-decreasing on [1/2, 1), which concludes the [24] K. Matsumoto, “Reverse test and characterization of quantum relative
proof. entropy,” 2010, arXiv:1010.1030.
[25] K. Matsumoto, “A new quantum version of f -divergence,” in Reality
ACKNOWLEDGMENT and Measurement in Algebraic Quantum Theory. Singapore: Springer,
2018, pp. 229–273.
The authors are grateful to Li Gao for the interesting [26] K. Fang and H. Fawzi, “Geometric Rényi divergence and its applications
discussions, to Peter Brown for the example used in propo- in quantum channel capacities,” Commun. Math. Phys., vol. 384, no. 3,
pp. 1615–1677, Jun. 2021.
sition 6.7 and to Mark Wilde for the discussion on the [27] S. O. Scalet, Á. M. Alhambra, G. Styliaris, and J. I. Cirac, “Computable
difference between BS-mutual information and its variational Rényi mutual information: Area laws and correlations,” Quantum, vol. 5,
counterpart. Moreover, the authors would like to thank Álvaro p. 541, Sep. 2021.
Alhambra for pointing them to [59]. [28] Y. Zhai, B. Yang, and Z. Xi, “Belavkin–Staszewski relative entropy,
conditional entropy, and mutual information,” Entropy, vol. 24, no. 6,
p. 837, Jun. 2022.
R EFERENCES [29] A. Bluhm, Á. Capel, and A. Pérez-Hernández, “Exponential decay of
[1] M. Fannes, “A continuity property of the entropy density for spin lattice mutual information for Gibbs states of local Hamiltonians,” Quantum,
systems,” Commun. Math. Phys., vol. 31, no. 4, pp. 291–294, Dec. 1973. vol. 6, p. 650, Feb. 2022.
[2] K. M. R. Audenaert, “A sharp continuity estimate for the von Neumann [30] I. Bardet, Á. Capel, L. Gao, A. Lucia, D. Pérez-García, and C. Rouzé,
entropy,” J. Phys. A, Math. Theor., vol. 40, no. 28, pp. 8127–8136, “Entropy decay for davies semigroups of a one dimensional quantum
Jul. 2007. lattice,” 2021, arXiv:2112.00601.
[3] D. Petz, Quantum Information Theory and Quantum Statistics. Berlin, [31] I. Bardet, Á. Capel, L. Gao, A. Lucia, D. Pérez-García, and C. Rouzé,
Germany: Springer, 2008. “Rapid thermalization of spin chain commuting Hamiltonians,” Phys.
[4] I. H. Kim, “Modulus of convexity for operator convex functions,” Rev. Lett., vol. 130, no. 6, Feb. 2023, Art. no. 060401.
J. Math. Phys., vol. 55, no. 8, Aug. 2014, Art. no. 082201. [32] A. Bluhm, P. Capel, A. Gondolf, and A. Pérez-Hernández, “General
[5] E. A. Carlen and E. H. Lieb, “Remainder terms for some quan- continuity bounds for quantum relative entropies,” in Proc. IEEE Int.
tum entropy inequalities,” J. Math. Phys., vol. 55, no. 4, Apr. 2014, Symp. Inf. Theory (ISIT), May 2023, pp. 1–9.
Art. no. 042201. [33] G. Lindblad, “Expectations and entropy inequalities for finite quantum
[6] I. Kim and M. B. Ruskai, “Bounds on the concavity of quantum entropy,” systems,” Commun. Math. Phys., vol. 39, no. 2, pp. 111–119, Jun. 1974.
J. Math. Phys., vol. 55, no. 9, Sep. 2014, Art. no. 092201. [34] K. M. R. Audenaert and J. Eisert, “Continuity bounds on the quan-
[7] K. M. R. Audenaert, “Quantum skew divergence,” J. Math. Phys., tum relative entropy,” J. Math. Phys., vol. 46, no. 10, Oct. 2005,
vol. 55, no. 11, Nov. 2014, Art. no. 112202. Art. no. 102104.
[8] R. Alicki and M. Fannes, “Continuity of quantum conditional informa- [35] K. M. R. Audenaert and J. Eisert, “Continuity bounds on the quantum
tion,” J. Phys. A, Math. Gen., vol. 37, no. 5, pp. L55–L57, Feb. 2004. relative entropy—II,” J. Math. Phys., vol. 52, no. 11, Nov. 2011,
[9] A. Winter, “Tight uniform continuity bounds for quantum entropies: Art. no. 112201.
Conditional entropy, relative entropy distance and energy constraints,” [36] O. Bratteli and D. W. Robinson, Operator Algebras and Quantumsta-
Commun. Math. Phys., vol. 347, no. 1, pp. 291–313, Oct. 2016. tistical Mechanics II. Equilibrium States. Models in Quantum Statisti-
[10] M. A. Nielsen, “Continuity bounds for entanglement,” Phys. Rev. A, cal Mechanics (Texts and Monographs in Physics). Berlin, Germany:
Gen. Phys., vol. 61, no. 6, Apr. 2000, Art. no. 064301. Springer, 1981.
[11] D. Leung and G. Smith, “Continuity of quantum channel capacities,” [37] A. Vershynina, “Upper continuity bound on the quantum quasi-relative
Commun. Math. Phys., vol. 292, no. 1, pp. 201–215, Nov. 2009. entropy,” J. Math. Phys., vol. 60, no. 10, Oct. 2019, Art. no. 102201.
[12] M. E. Shirokov, “Tight uniform continuity bounds for the quantum [38] F. Hiai, M. Mosonyi, D. Petz, and C. Bény, “Quantum f -divergences
conditional mutual information, for the Holevo quantity, and for capac- and error correction,” Rev. Mat. Phys., vol. 23, no. 7, pp. 691–747, 2011.
ities of quantum channels,” J. Math. Phys., vol. 58, no. 10, Oct. 2017, [39] M. Tomamichel, Quantum Information Processing With Finite
Art. no. 102202. Resources. Cham, Switzerland: Springer, 2016.
[13] M. M. Wilde, Quantum Information Theory. Cambridge, U.K.: [40] F. G. S. L. Brandão, M. Horodecki, J. Oppenheim, J. M. Renes,
Cambridge Univ. Press, 2016. and R. W. Spekkens, “Resource theory of quantum states out of ther-
[14] M. Mosonyi and F. Hiai, “On the quantum Rényi relative entropies mal equilibrium,” Phys. Rev. Lett., vol. 111, no. 25, Dec. 2013,
and related capacity formulas,” IEEE Trans. Inf. Theory, vol. 57, no. 4, Art. no. 250404.
pp. 2474–2487, Apr. 2011. [41] M. Ohya and D. Petz, Quantum Entropy and Its Use (Texts and
[15] B. Synak-Radtke and M. Horodecki, “On asymptotic continuity of Monographs in Physics). Berlin, Germany: Springer, 1993.
functions of quantum states,” J. Phys. A, Math. Gen., vol. 39, no. 26, [42] H. Araki, “On an inequality of Lieb and Thirring,” Lett. Math. Phys.,
pp. L423–L437, Jun. 2006. vol. 19, no. 2, pp. 167–170, Feb. 1990.
[16] M. E. Shirokov, “Advanced Alicki–Fannes–Winter method for energy- [43] E. H. Lieb and W. E. Thirring, “Inequalities for the moments of
constrained quantum systems and its use,” Quantum Inf. Process., the eigenvalues of the Schrödinger Hamiltonian and their relation to
vol. 19, no. 5, p. 164, May 2020. Sobolev inequalities,” in Studies in Mathematical Physics, E. H. Lieb,
[17] M. E. Shirokov, “Quantifying continuity of characteristics of com- Ed. Princeton, NJ, USA: Princeton Univ. Press, 1976, pp. 269–304.
posite quantum systems,” Phys. Scripta, vol. 98, no. 4, Mar. 2023, [44] D. Sutter, M. Berta, and M. Tomamichel, “Multivariate trace inequali-
Art. no. 042002. ties,” Commun. Math. Phys., vol. 352, no. 1, pp. 37–58, May 2017.
[18] H. Umegaki, “Conditional expectation in an operator algebra. IV. [45] M. Hayashi, Quantum Information: An Introduction. Berlin, Germany:
Entropy and information,” Kodai Math. J., vol. 14, no. 2, pp. 59–85, Springer, 2006.
Jan. 1962. [46] M. E. Shirokov, “Adaptation of the Alicki–Fannes–Winter method for
[19] V. P. Belavkin and P. Staszewski, “C ∗ -algebraic generalization of the set of states with bounded energy and its use,” Rep. Math. Phys.,
relative entropy and entropy,” Ann. Inst. Henri Poincaré, Sect. A, vol. 37, vol. 81, no. 1, pp. 81–104, Feb. 2018.
no. 1, pp. 51–58, 1982. [47] M. E. Shirokov, “Uniform continuity bounds for information charac-
[20] S. Kullback and R. A. Leibler, “On information and sufficiency,” Ann. teristics of quantum channels depending on input dimension and on
Math. Statist., vol. 22, no. 1, pp. 79–86, 1951. input energy,” J. Phys. A, Math. Theor., vol. 52, no. 1, Jan. 2019,
[21] A. Bluhm and Á. Capel, “A strengthened data processing inequality for Art. no. 014001.
the Belavkin–Staszewski relative entropy,” Rev. Math. Phys., vol. 32, [48] D. Sutter and R. Renner, “Necessary criterion for approximate recov-
no. 2, Mar. 2020, Art. no. 2050005. erability,” Annales Henri Poincaré, vol. 19, no. 10, pp. 3007–3029,
[22] A. Bluhm, Á. Capel, and A. Pérez-Hernández, “Weak quasi-factorization Oct. 2018.
for the Belavkin–Staszewski relative entropy,” in Proc. IEEE Int. Symp. [49] K. M. R. Audenaert and N. Datta, “α-z-Rényi relative entropies,”
Inf. Theory (ISIT), Jul. 2021, pp. 118–123. J. Math. Phys., vol. 56, Oct. 2015, Art. no. 022202.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.
BLUHM et al.: CONTINUITY OF QUANTUM ENTROPIC QUANTITIES VIA ALMOST CONVEXITY 5901

[50] A. E. Rastegin, “Upper continuity bounds on the relative q-entropy for [78] T. Metger, O. Fawzi, D. Sutter, and R. Renner, “Generalised entropy
q > 1,” J. Math. Phys., vol. 52, no. 6, Jun. 2011, Art. no. 062203. accumulation,” in Proc. IEEE 63rd Annu. Symp. Found. Comput. Sci.
[51] G. Gour and M. Tomamichel, “Optimal extensions of resource measures (FOCS), Oct. 2022, pp. 844–850.
and their applications,” Phys. Rev. A, Gen. Phys., vol. 102, no. 6, [79] M. Berta, D. Sutter, and M. Walter, “Quantum Brascamp–Lieb dualities,”
Dec. 2020, Art. no. 062401. Commun. Math. Phys., vol. 2023, pp. 1–24, Mar. 2023.
[52] G. Gour and M. Tomamichel, “Entropy and relative entropy from [80] C. D. Aliprantis and K. C. Border, Infinite Dimensional Analysis: A
information-theoretic principles,” IEEE Trans. Inf. Theory, vol. 67, Hitchhiker’s Guide, 3rd ed. Berlin, Germany: Springer, 2006.
no. 10, pp. 6313–6327, Oct. 2021. [81] V. Vedral and M. B. Plenio, “Entanglement measures and purification
[53] E. P. Hanson, “Entropic continuity bounds & eventually entanglement- procedures,” Phys. Rev. A, Gen. Phys., vol. 57, no. 3, pp. 1619–1633,
breaking channels,” 2020, arXiv:2010.02408. Mar. 1998.
[54] E. A. Carlen, “Trace inequalities and quantum entropy: An introductory [82] V. Vedral, M. B. Plenio, M. A. Rippin, and P. L. Knight, “Quantifying
course,” in Entropy and the Quantum (Contemporary Mathematics), entanglement,” Phys. Rev. Lett., vol. 78, no. 12, pp. 2275–2279, 1997.
vol. 529, R. Sims and D. Ueltschi, Eds. Providence, RI, USA: American [83] E. M. Rains, “Bound on distillable entanglement,” Phys. Rev. A, Gen.
Mathematical Society, 2009, pp. 73–140. Phys., vol. 60, no. 1, pp. 179–184, Jul. 1999.
[55] V. Katariya and M. M. Wilde, “Geometric distinguishability measures [84] M. J. Donald and M. Horodecki, “Continuity of relative entropy of
limit quantum channel estimation and discrimination,” Quantum Inf. entanglement,” Phys. Lett. A, vol. 264, no. 4, pp. 257–260, Dec. 1999.
Process., vol. 20, no. 2, p. 78, Feb. 2021. [85] E. M. Rains, “A semidefinite program for distillable entanglement,”
[56] H. Fawzi and O. Fawzi, “Defining quantum divergences via convex IEEE Trans. Inf. Theory, vol. 47, no. 7, pp. 2921–2933, Nov. 2001.
optimization,” Quantum, vol. 5, p. 387, Jan. 2021. [86] M. Tomamichel, M. M. Wilde, and A. Winter, “Strong converse rates
[57] M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum for quantum communication,” IEEE Trans. Inf. Theory, vol. 63, no. 1,
Information, 10th ed. Cambridge, U.K.: Cambridge Univ. Press, 2010. pp. 715–727, Jan. 2017.
[58] F. Hiai and D. Petz, “The proper formula for relative entropy and its [87] E. P. Hanson and N. Datta, “Maximum and minimum entropy states
asymptotics in quantum probability,” Commun. Math. Phys., vol. 143, yielding local continuity bounds,” J. Math. Phys., vol. 59, no. 4,
no. 1, pp. 99–114, Dec. 1991. Apr. 2018, Art. no. 042204.
[59] G. Gour, “Role of quantum coherence in thermodynamics,” PRX Quan- [88] E. P. Hanson and N. Datta, “Universal proofs of entropic continuity
tum, vol. 3, no. 4, Nov. 2022, Art. no. 040323. bounds via majorization flow,” 2019, arXiv:1909.06981.
[60] E. H. Lieb and M. B. Ruskai, “Proof of the strong subadditivity of [89] S. Becker, N. Datta, and M. G. Jabbour, “From classical to quan-
quantum mechanical entropy,” J. Math. Phys., vol. 14, pp. 1938–1941, tum: Uniform continuity bounds on entropies in infinite dimen-
1973. sions,” IEEE Trans. Inf. Theory, early access, Feb. 23, 2023, doi:
[61] D. Petz, “Monotonicity of quantum relative entropy revisited,” Rev. 10.1109/TIT.2023.3248228.
Math. Phys., vol. 15, no. 1, pp. 79–91, Mar. 2003. [90] M. E. Shirokov, “Quantum relative entropy: General convergence crite-
rion and preservation of convergence under completely positive linear
[62] P. Hayden, R. Jozsa, D. Petz, and A. Winter, “Structure of states which
maps,” 2022, arXiv:2205.10341.
satisfy strong subadditivity of quantum entropy with equality,” Commun.
Math. Phys., vol. 246, no. 2, pp. 359–374, Apr. 2004.
[63] O. Fawzi and R. Renner, “Quantum conditional mutual information and
approximate Markov chains,” Commun. Math. Phys., vol. 340, no. 2, Andreas Bluhm received the Ph.D. degree in mathematics from the Technical
pp. 575–611, Dec. 2015. University of Munich, Germany, in 2019. From 2019 to 2022, he was a
[64] M. Junge, R. Renner, D. Sutter, M. M. Wilde, and A. Winter, “Universal Post-Doctoral Researcher with the Centre for the Mathematics of Quantum
recovery from a decrease of quantum relative entropy,” Ann. Henri Theory (QMATH) hosted by the Department of Mathematical Sciences,
Poincaré, vol. 19, p. 2955, Aug. 2018. University of Copenhagen, Denmark. Since 2022, he has been a Researcher
[65] E. A. Carlen and A. Vershynina, “Recovery map stability for the data with the French National Centre for Scientific Research (CNRS), Grenoble
processing inequality,” J. Phys. A, Math. Theor., vol. 53, no. 3, Jan. 2020, Computer Science Laboratory (LIG), Grenoble, France. His research interests
Art. no. 035204. include quantum information theory and quantum spin systems.
[66] D. Sutter, Approximate Quantum Markov Chains (Springer Briefs in
Mathematical Physics). Cham, Switzerland: Springer, 2018.
[67] K. Kato and F. G. S. L. Brandão, “Quantum approximate Markov Ángela Capel received the Ph.D. degree in mathematics from Universidad
chains are thermal,” Commun. Math. Phys., vol. 370, no. 1, pp. 117–149, Autónoma de Madrid, Spain, in 2019. From 2020 to 2021, she was a
Aug. 2019. Post-Doctoral Researcher with the Technical University of Munich, where
[68] E. P. Hanson, C. Rouzé, and D. Stilck França, “Eventually entanglement she has been an MCQST Distinguished Post-Doctoral Fellow since October
breaking Markovian dynamics: Structure and characteristic times,” Ann. 2020 hosted by the Chair of Mathematical Physics. Since October 2021,
Henri Poincaré, vol. 21, pp. 1517–1571, Mar. 2020. she has been a Junior Professor with the Department of Mathematics, Eber-
[69] P. Svetlichnyy and T. A. B. Kennedy, “Decay of quantum conditional hard Karls University of Tübingen, Germany. Her research interests include
mutual information for purely generated finitely correlated states,” quantum information theory, mathematical physics, and quantum many-body
J. Math. Phys., vol. 63, no. 7, Jul. 2022, Art. no. 072201. systems.
[70] S. Golden, “Lower bounds for the Helmholtz function,” Phys. Rev.,
vol. 137, no. 4B, pp. B1127–B1128, Feb. 1965.
[71] C. J. Thompson, “Inequality with applications in statistical mechanics,” Paul Gondolf received the bachelor’s degree in physics from Heidelberg
J. Math. Phys., vol. 6, no. 11, pp. 1812–1813, 1965. University, Germany, in 2020, and the master’s degree in mathematical physics
[72] R. Bhatia, Matrix Analysis (Graduate Texts in Mathematics), vol. 169. from the Eberhard Karls University of Tübingen, Germany, in 2022, where he
New York, NY, USA: Springer, 1997. is currently pursuing the Ph.D. degree in mathematics under the supervision
[73] Á. Capel, A. Lucia, and D. Pérez-García, “Quantum conditional relative of Ángela Capel, Marius Lemm, and Cambyse Rouzé. His research interests
entropy and quasi-factorization of the relative entropy,” J. Phys. A, Math. include matrix analysis, convex analysis, and bosonic open quantum systems.
Theor., vol. 51, no. 48, Nov. 2018, Art. no. 484001.
[74] I. Bardet, Á. Capel, A. Lucia, D. Pérez-García, and C. Rouzé, “On the
modified logarithmic Sobolev inequality for the heat-bath dynamics for Antonio Pérez-Hernández received the Ph.D. degree in mathematics from
1D systems,” J. Math. Phys., vol. 62, no. 6, Jun. 2021, Art. no. 061901. the University of Murcia, Spain, in 2017. From 2018 to 2019, he was a
[75] Á. Capel, C. Rouzé, and D. S. França, “The modified logarithmic Post-Doctoral Researcher with Instituto de Ciencias Matemáticas (ICMAT),
Sobolev inequality for quantum spin systems: Classical and commuting Madrid. In 2020, he held a post-doctoral position with Universidad Com-
nearest neighbour interactions,” 2021, arXiv:2009.11817. plutense de Madrid (UCM) supported by a Juan de la Cierva Fellowship.
[76] Á. Capel, A. Lucia, and D. Pérez-García, “Superadditivity of quantum In November 2020, he joined Universidad Nacional de Educación a Distancia
relative entropy for general states,” IEEE Trans. Inf. Theory, vol. 64, (UNED) as an Assistant Professor and was promoted to an Associate Professor
no. 7, pp. 4758–4765, Jul. 2018. in December 2022. His research interests include functional analysis and its
[77] T. Metger and R. Renner, “Security of quantum key distribution from applications to other fields, with a recent focus on quantum information theory
generalised entropy accumulation,” 2022, arXiv:2203.04989. and quantum many-body systems.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:46:59 UTC from IEEE Xplore. Restrictions apply.