Recoverability for Holevo’s just-as-good fidelity
Mark M. Wilde
Hearne Institute for Theoretical Physics, Department of Physics and Astronomy, Center for Computation and Technology,
Louisiana State University, Baton Rouge, Louisiana 70803, USA, Email: mwilde@lsu.edu
Abstract—Holevo’s just-as-good fidelity is a similarity measure
for quantum states that has found several applications. One of its
critical properties is that it obeys a data processing inequality: the
measure does not decrease under the action of a quantum channel
on the underlying states. In this paper, I prove a refinement of
this data processing inequality that includes an additional term
related to recoverability. That is, if the increase in the measure
is small after the action of a partial trace, then one of the states
can be nearly recovered by the Petz recovery channel, while
the other state is perfectly recovered by the same channel. The
refinement is given in terms of the trace distance of one of the
states to its recovered version and also depends on the minimum
eigenvalue of the other state. As such, the refinement is universal,
in the sense that the recovery channel depends only on one of the
states, and it is explicit, given by the Petz recovery channel. The
appendix contains a generalization of the aforementioned result
to arbitrary quantum channels.
I. I NTRODUCTION
In Holevo’s seminal 1972 work on the quasiequivalence
of locally normal states [1], he established the following
inequalities for quantum states ρ and σ:
p 1 p The function FH has also been rediscovered a number of
1 − FH (ρ, σ) ≤ kρ − σk1 ≤ 1 − FH (ρ, σ), (1)
2 times. For example, it is a particular case of Petz’s quasi-
where kρ − σk1 denotes the well known trace distance and
the function FH is Holevo’s “just-as-good fidelity,” defined as
 √ √ 2
FH (ρ, σ) ≡ Tr{ ρ σ} . (2)
After writing it down, he then remarked that “it is evident that
FH is just as good a measure of proximity of the states ρ and
σ as kρ − σk1 .” And so it is that the measure FH is known
as Holevo’s just-as-good fidelity.
Some years
after this,
Uhlmann defined the quantum fidelity trace preserving map). This inequality is a consequence of data
√ √ 2
as F (ρ, σ) ≡
ρ σ
1 [2]. It is evident that the following
relation holds
FH (ρ, σ) ≤ F (ρ, σ), (3)
due to the variational characterization of the trace norm of a
square operator X as
kXk1 = max |Tr{XU }| , (4)
U
where the optimization is with respect to a unitary operator U .
Many years after this, at the dawn of quantum computing, with
more growing interest in quantum information theory, Fuchs
and van de Graaf presented the following widely employed
inequalities [3]:
p 1 p 1/2
1 − F (ρ, σ) ≤ kρ − σk1 ≤ 1 − F (ρ, σ), (5)
2
which bear a striking similarity to (1). Indeed the lower
bound on 12 kρ − σk1 in (5) is an immediate consequence
of (3) and the lower bound in (1). The upper bound on
1
2 kρ − σk1 in (5) can be proven by first showing that it
is achieved for pure states, employing Uhlmann’s “transition
probability” characterization of F (ρ, σ) [2], and then invoking
monotonicity of trace distance with respect to partial trace.
The latter inequalities in (5) have been more widely employed
in quantum information theory than those in (1) due to
Uhlmann’s “transition probability” characterization of F (ρ, σ)
and its many implications.
Nevertheless, Holevo’s just-as-good fidelity is clearly a
useful measure of similarity for quantum states in light of (1),
and it has found several applications in quantum information
theory. For example, it serves as an upper bound on the
probability of error in discriminating ρ from σ in a hypothesis
testing experiment [4], [5], which in some sense is just a
rewriting of the lower bound in (1) (see also [6, Lemma 3.2] in
this context). In turn, this way of thinking has led to particular
decoders for quantum polar codes [7], [8].
entropies [9], [10]. It was studied under the name “quantum
affinity” in [11] and shown to be equal to the fidelity of the
canonical purifications of quantum states in [12].
One of the most important properties of FH is that it obeys
the following data processing inequality:
FH (N (ρ), N (σ)) ≥ FH (ρ, σ), (6)
where N is a quantum channel (a completely positive and
processing for Petz’s more general quasi-entropies [9], [10].
This property is one reason that FH has an interpretation as a
similarity measure: the states ρ and σ generally become more
similar under the action of a quantum channel.
The main contribution of this paper is the following refine-
ment of the data processing inequality in (6), in the case that ρ
is a bipartite density operator, σ is a positive definite bipartite
operator, and the channel is a partial trace over the B system:
p p
FH (ρA , σA ) ≥ FH (ρAB , σAB )
π 2 λmin (σAB ) 3
+ kRσA→AB (ρA ) − ρAB k1 , (7)
432 Tr{σA }
where λmin (σAB ) is the minimum eigenvalue of σAB and
h
−1/2 −1/2
i
1/2
RσA→AB (·) ≡ σAB σA (·)A σA ⊗ IB σAB (8)
is a quantum channel known as the Petz recovery channel For example, when f (x) = x ln x, then Qf reduces to the
[13], [14]. The interpretation of this inequality is the same quantum relative entropy from [27].
as that given in previous work on this topic of refining data Now consider the bipartite case and define
processing
√ inequalities
√ (see, e.g., [15]–[17]). If the difference
FH (ρA , σA ) − FH (ρAB , σAB ) is small, then one can |ΓiAÂB B̂ ≡ |ΓiAÂ ⊗ |ΓiB B̂ . (13)
approximately recover the state ρAB from its marginal ρA . We can also write this as |ΓiAB ÂB̂ with it being understood
The appendix generalizes the result in (7) to arbitrary quantum that there is a permutation of systems. Then, by the above,
channels. we have for a density operator ρAB and a positive definite
The technique that I use for proving the above data process- operator σAB that
ing refinement closely follows the elegant approach recently
put forward by Carlen and Vershynina in [18]. This technique Qf (ρAB kσAB )
appears to be different from every other approach, given in −1
= hϕσAB |AB ÂB̂ f σAB ⊗ ρTÂB̂ |ϕσAB iAB ÂB̂ . (14)


recent years since [15], that has established refinements of
data processing inequalities. It builds on Petz’s approach from Now define the linear operator V by
[9], [10] for proving data processing for the quantum relative 1/2

−1/2

entropy, as well as ideas in [19]. Here, I use this same VAÂ→AB ÂB̂ ≡ σAB σA ⊗ IÂ |ΓiB B̂ . (15)
technique and establish a general lemma regarding remainder
This linear operator is an isometric extension of the Petz
terms for data processing with Petz’s quasi-entropies, and then
recovery channel, as discussed recently in [28]. One can
I specialize it to obtain the inequality in (7).
readily verify that V is an isometry and that
An interesting aspect of (7) is that the recovery channel
is explicit, given in the Petz form, and universal, having no VAÂ→AB ÂB̂ |ϕσA iAÂ = |ϕσAB iAB ÂB̂ , (16)
dependence on the state ρAB while depending only on σAB . −1
V † σAB ⊗ ρTÂB̂ V = σA

−1
⊗ ρTÂ . (17)
In the rest of the paper, I begin by giving background and
establish some notation. After that, I prove a general lemma For simple proofs of these properties, see, e.g., [25] or [28].
that refines data processing for Petz’s quasi-entropies. Then I With all these notions in place, we can recall Petz’s approach
specialize it to arrive at the inequality in (7). [9], [10], [23], [24] for establishing monotonicity of the f -
quasi-relative entropy under partial trace:
II. BACKGROUND AND N OTATION
−1
Qf (ρAB kσAB ) = hϕσAB |AB ÂB̂ f σAB ⊗ ρTÂB̂ |ϕσAB iAB ÂB̂


I begin by reviewing some background and establish no-
−1
= hϕσA |AÂ V † f σAB ⊗ ρTÂB̂ V |ϕσA iAÂ


tation. Basic concepts of quantum information theory can
be found in [6], [20], [21]. Let f be an operator convex  −1
≥ hϕσA |AÂ f V † σAB ⊗ ρTÂB̂ V |ϕσA iAÂ


function defined on [0, ∞). Examples include f (x) = x ln x, −1
= hϕσA |AÂ f σA ⊗ ρTÂ |ϕσA iAÂ = Qf (ρA kσA )


f (x) = −xα for α ∈ (0, 1), f (x) = xα for α ∈ (1, 2]. (18)
According to [22, Section 8], such a function has the following where we made use of everything above and the operator
integral representation: Jensen inequality [29].
f (x) = f (0) + ax + bx2 III. G ENERAL STATEMENT FOR QUASI - ENTROPIES
Z ∞  
x t I now modify the approach from [18] for lower bounds for
+ dµ(t) −1+ , (9)
0 1+t x+t relative entropy differences to use an arbitrary operator convex
where a ∈ R, b R≥ 0, and µ is a non-negative measure on function f instead. So we are considering the following f -
∞ 2
(0, ∞) satisfying 0 dµ(t)/ (1 + t) < ∞. quasi-relative entropy difference:
Define the maximally entangled vector as
Qf (ρAB kσAB ) − Qf (ρA kσA ). (19)
|S|−1
Recall the integral representation of f from (9). Let
X
|ΓiS S̃ ≡ |iiS |iiS̃ , (10)
−1 −1
i=0
∆AB ÂB̂ ≡ σAB ⊗ ρT , ∆AÂ ≡ σA ⊗ ρTÂ , (20)
for orthonormal bases {|iiS }i and {|iiS̃ }i , and for a positive  ÂB̂ 
1/2 −1/2
VAÂ→AB ÂB̂ ≡ σAB σA ⊗ IÂ |ΓiB B̂ (21)
semi-definite operator σ, define its canonical purification by
and recall from (17) that V † ∆AB ÂB̂ V = ∆AÂ . This implies
 
1/2
|ϕσ iS S̃ ≡ σS ⊗ IS̃ |ΓiS S̃ . (11)

∆
=
V † ∆ † †








AB ÂB̂ V ∞ = V V ∆AB ÂB̂ V V



Then, following Petz [9], [10], [23], [24], as well as what was AÂ ∞


∞
discussed later in [25], [26], we define the f -quasi-relative ≤
∆

AB ÂB̂ ∞ (22)
entropy Qf (ρkσ) of a density operator ρ and a positive definite
with the last equality following from isometric invariance of
operator σ as
the operator norm and the inequality from submultiplicativity
Qf (ρkσ) ≡ hϕσ |S S̃ f σS−1 ⊗ ρTS̃ |ϕσ iS S̃ . of the operator norm and the fact that V V † is a projection.


(12)
 Lemma 1: Let µ be a measure. For an operator X, define So we set
Z ∞  
1 1 t 1/2 −1/2
−1 1/2
ν(X) = dµ(t) t − , (23) wAB ÂB̂
≡ σAB σA ∆AÂ + t σA
0 t t+X
−1 1/2
RT − ∆AB ÂB̂ + t σAB , (32)
and for T >R 0, define µ([0, T ]) ≡ 0 dµ(t). For c > 0, define
∞ 1
g(c, T ) ≡ T dµ(t) 1+t/c . Let ρAB be a density operator so that
t
and σAB a positive definite operator. Then for all T > 0, the |ϕw iAB ÂB̂ = wAB
t
|ΓiAÂB B̂ . (33)
ÂB̂
following inequality holds

h i
Now invoking the definition in (23) we find that

1/2 −1/2 1/2 1/2

σAB σA ν(∆AÂ )σA − ν(∆AB ÂB̂ )σAB |ΓiAÂB B̂



1/2 −1/2 1/2 1/2
2 σAB σA ν(∆AÂ )σA − ν(∆AB ÂB̂ )σAB
1/2 1/2
≤ µ([0, T ]) [Qf (ρAB kσAB ) − Qf (ρA kσA )] Z ∞
1/2 −1/2

1 1

1/2



+ 2g(
∆AB ÂB̂
∞ , T ) Tr{σA }. (24) = dµ(t) t σAB σA − σA
0 t t + ∆ AÂ
Z ∞  
Proof. The proof follows [18] quite closely at times but also 1 1 1/2
− dµ(t) t − σAB (34)
features some departures. Since V is an isometry satisfying 0 t t + ∆ AB ÂB̂
V † V = IAÂ , it follows that V V † is a projection, so that Z ∞ "  
1/2 −1/2 1 1 1/2
V V † ≤ IAB ÂB̂ . Using the integral representation in (9), we = dµ(t) t σAB σA − σA
0 t t + ∆AÂ
arrive at the chain of inequalities in (25), where we made use #
of (17) and the fact that V V † is a projection so that V V † ≤
 
1 1 1/2
IAB ÂB̂ . Similarly, we find that − − σAB (35)
t t + ∆AB ÂB̂
Qf (ρA kσA ) = hϕσA |AÂ f ∆AÂ |ϕσA iAÂ

 Z ∞ "
1/2 −1/2 1 1/2
= dµ(t) t − σAB σA σA
= f (0) + hϕσA |AÂ a∆AÂ + b∆2AÂ |ϕσA iAÂ
 
0 t + ∆ AÂ
Z ∞  
σA ∆AÂ t #
+ dµ(t) hϕ |AÂ −1+ |ϕσA iAÂ . 1 1/2
0 1+t ∆AÂ + t + σ (36)
t + ∆AB ÂB̂ AB
(26) Z ∞
t
Thus, we find that =− dµ(t) twAB ÂB̂
. (37)
0
Qf (ρAB kσAB ) − Qf (ρA kσA ) ≥ Thus, for any T > 0, we have that
Z ∞ "

−1
h i

dµ(t) t hϕ |AÂ V † ∆AB ÂB̂ + t
σA
V
1/2 −1/2 1/2 1/2

σAB σA ν(∆AÂ )σA − ν(∆AB ÂB̂ )σAB |ΓiAÂB B̂



0
Z ∞
2
#
t




−1 =
dµ(t) twAB |ΓiAÂB B̂
(38)
− ∆AÂ + t |ϕσA iAÂ . (27)

0
ÂB̂

2
Z T
t

Now consider that for t > 0 ≤ dµ(t) t
wAB ÂB̂
|ΓiAÂB B̂
2
0


∞
h
−1 i σA Z

t hϕσA |AÂ V † (∆AB ÂB̂ + t)−1 V − ∆AÂ + t |ϕ iAÂ +


t


dµ(t) t wAB ÂB̂ |ΓiAÂB B̂
(39)


t t T 2
= t hϕw |AB ÂB̂ ∆AB ÂB̂ + t |ϕw iAB ÂB̂



t

2 Let us study the two terms separately. For the first term, from
≥ t2
|ϕw iAB ÂB̂
, (28)



2
Cauchy–Schwarz
2 " Z
where
Z # "Z #
 T  T T
dµ(t)f (t)g(t) ≤ dµ(t)f 2 (t) dµ(t)g 2 (t) ,
 
t
−1 
|ϕw iAB ÂB̂ ≡ V ∆AÂ + t |ϕσA iAÂ  0  0 0

−1 σAB
− ∆AB ÂB̂ +t |ϕ iAB ÂB̂ . (29) we have that
#2
Consider that
"Z
T
t

t 1/2 −1/2
−1 1/2 dµ(t) t
wAB ÂB̂ |ΓiAÂB B̂
2
|ϕw iAB ÂB̂ = σAB σA ∆AÂ + t σA |ΓiAÂ |ΓiB B̂ 0

−1 1/2 T
− ∆AB ÂB̂ + t σAB |ΓiAÂ |ΓiB B̂ (30)
Z

2
dµ(t) t2
wAB

t
 ≤ µ([0, T ]) ÂB̂
|ΓiAÂB B̂
2 (40)
1/2 −1/2
−1 1/2 0
= σAB σA ∆AÂ + t σA Z ∞
2
dµ(t) t2
wAB

t
 ≤ µ([0, T ]) ÂB̂
|ΓiAÂB B̂
2 (41)

−1 1/2 0
− ∆AB ÂB̂ + t σAB |ΓiAÂ |ΓiB B̂ . (31)
≤ µ([0, T ]) [Qf (ρAB kσAB ) − Qf (ρA kσA )] . (42)
 Qf (ρAB kσAB ) = hϕσA |AÂ V † f ∆AB ÂB̂ V |ϕσA iAÂ


 Z ∞  
∆AB ÂB̂ t
= hϕσA |AÂ V † f (0) + a∆AB ÂB̂ + b∆2AB ÂB̂ + dµ(t) −1+ V |ϕσA iAÂ
0 1 + t ∆ AB ÂB̂ + t
= f (0) + hϕσA |AÂ aV † ∆AB ÂB̂ V + bV † ∆2AB ÂB̂ V |ϕσA iAÂ
 
Z ∞  † 
σA V ∆AB ÂB̂ V † t
+ hϕ |AÂ dµ(t) −1+V V |ϕσA iAÂ
0 1+t ∆AB ÂB̂ + t
≥ f (0) + hϕσA |AÂ aV † ∆AB ÂB̂ V + bV † ∆AB ÂB̂ V V † ∆AB ÂB̂ V |ϕσA iAÂ
 
Z ∞  † 
σA V ∆AB ÂB̂ V † t
+ hϕ |AÂ dµ(t) −1+V V |ϕσA iAÂ
0 1+t ∆AB ÂB̂ + t
 Z ∞    
∆AÂ t
= f (0) + hϕσA |AÂ a∆AÂ + b∆2AÂ + dµ(t) −1+V† V |ϕσA iAÂ , (25)
0 1 + t ∆ AB ÂB̂ + t

Moving to the second term, from the reasoning in the proof Substituting into (24), we find that
of [18, Theorem 1.7], we find that for any positive operator X h i
1/2 −1/2 α 1/2 1/2
σAB σA ∆AÂ σA − ∆α σ |ΓiAÂB B̂ ≤
   
1 1 1 1 AB ÂB̂ AB 2
t − ≤t − I (43) 1/2
t t+X t t + kXk∞

sin(απ) α 1/2
1 T [Qα (ρAB kσAB ) − Qα (ρA kσA )]
= I (44) απ
1 + t/ kXk∞

2 sin(απ) ∆AB ÂB̂ ∞
+ Tr{σA }. (53)
πT 1−α (1 − α)
R∞  
so that T dµ(t) t 1t − t+X 1
≤ g(kXk , T ) I. This leads to
the development in (45), and after putting everything together, We can consider this for an arbitrary α ∈ (0, 1), but the
we get (24). most interesting and physically relevant case seems to occur
when α = 1/2. So I now prove the claim in (7).
IV. A PPLICATION TO H OLEVO ’ S JUST- AS - GOOD FIDELITY For α = 1/2, the lower bound in (53) simplifies to
I now specialize the above analysis to the case of the h i
operator convex function −xα for α ∈ (0, 1), and I abbreviate 1/2 −1/2 1/2 1/2
σAB σA ρA − ρAB |ΓiAÂB B̂

2
the corresponding quasi-entropy as Qα . For this case, from
1/2 −1/2 1/2

1/2
[22, Section 8], we have that dµ(t) = sin(απ)
π tα−1 dt. Plugging = σAB σA ρA − ρAB , (54)
2
into the quantities in Lemma 1, we find that
while the upper bound in (53) becomes
sin(απ) T α−1
Z
sin(απ) α
µ([0, T ]) = t dt = T . (46) 1/2 1/2
T 1/4 Q1/2 (ρAB kσAB ) − Q1/2 (ρA kσA )

π 0 απ [2/π]

We also find that 4 ∆AB ÂB̂ ∞
+ Tr{σA }. (55)
πT 1/2
g( ∆AB ÂB̂ ∞ , T )
Z ∞ Now minimizing over T > 0 gives the choice
1
= dµ(t) (47) " #4/3
T 1 + t/ ∆AB ÂB̂ ∞ 8 ∆AB ÂB̂ ∞ Tr{σA }
T = ,
sin(απ) ∞
Z
1 [2π]
1/2 
Q1/2 (ρAB kσAB ) − Q1/2 (ρA kσA )
1/2
= dt tα−1 (48)
π T 1 + t/ ∆
AB ÂB̂ ∞ (56)
Z ∞
sin(απ) 1 leading to the upper bound
≤ dt tα−1 (49)
π T t/ ∆
AB ÂB̂ ∞
h
(3/[22/3 ]) (8/π 2 )[Q1/2 (ρAB kσAB )
sin(απ) ∆AB ÂB̂ ∞
= . (50) i1/3
πT 1−α (1 − α)

− Q1/2 (ρA kσA )] ∆AB ÂB̂ ∞ Tr{σA } . (57)
Furthermore, we have that
Z ∞   Thus, the final inequality is
1 1
ν(X) = dµ(t) t − (51) π2
1/2 −1/2 1/2
3
0 t t+X σ ρ − ρ
1/2
AB A A AB

54 ∆AB ÂB̂ ∞ Tr{σA }
Z ∞ σ
2
 
sin(απ) 1 1
= dt tα−1 t − = X α . (52)
π 0 t t + X ≤ Q1/2 (ρAB kσAB ) − Q1/2 (ρA kσA ). (58)
 Z ∞
t


dµ(t) t wAB ÂB̂
|ΓiAÂB B̂

T 2
Z ∞ h
−1

−1  1/2 1/2 −1/2

−1

−1  1/2
i
≤
dµ(t) t t − ∆AB ÂB̂ + t σAB − t − ∆AÂ + t
σAB σA σA |ΓiAÂB B̂

ZT∞ 2
∞
 Z
−1

−1  1/2 1/2 −1/2

−1

−1  1/2
≤
dµ(t) t t − ∆AB ÂB̂ + t σAB |ΓiAÂB B̂ +
dµ(t) t σAB σA t − ∆AÂ + t σA |ΓiAÂB B̂

ZT∞ 2 T 2
∞
 Z
−1

−1  1/2 
−1

−1  1/2
=
dµ(t) t t − ∆AB ÂB̂ + t σAB |ΓiAÂB B̂
+
dµ(t) t t − ∆AÂ + t σA |ΓiAÂ

T 2 T 2

1/2 1/2
≤ g( ∆AB ÂB̂ ∞ , T ) σAB |ΓiAÂB B̂ + g( ∆AÂ ∞ , T ) σA |ΓiAÂ

 2  2
= g( ∆AB ÂB̂ ∞ , T ) + g( ∆AÂ ∞ , T ) Tr{σA } ≤ 2g( ∆AB ÂB̂ ∞ , T ) Tr{σA }. (45)

Using definitions, this is then equivalent to where λmin (τ ) now denotes the minimum non-zero eigenvalue
of τ and P denotes the Petz recovery map for τ and N ,
π2
1/2 −1/2 1/2
3
1/2
AB σA ρ A − ρ AB
defined as
54 ∆AB ÂB̂ ∞ Tr{σA }
σ
2
1/2 1/2 1/2 1/2 P(·) = τ 1/2 N † [(N (τ ))−1/2 (·)(N (τ ))−1/2 ]τ 1/2 . (64)
≤ Tr{ρA σA } − Tr{ρAB σAB }. (59)
The estimate from [18, Lemma 2.2] then gives Proof. Let us start by returning to (7) and reflecting on its
statement as well as its proof. If supp(ρAB ) ⊆ supp(σAB ),
π 2 /432
1/2 −1/2 −1/2 1/2
3
then without loss of generality, we can restrict the whole space
σ ρ σ σ − ρ


∆

Tr{σA } σ AB A A A AB AB
1 of systems A and B to the support of σAB and the inequality
AB ÂB̂ ∞
1/2 1/2
≤ Tr{ρA σA } − Tr{ρAB σAB }. (60)
1/2 1/2 in (7) holds with λmin (σAB ) equal to the minimum non-zero
eigenvalue of σAB . Now we can apply this result, as well
−1 1
Observe that ∆AB ÂB̂ ∞ = σAB ⊗ ρTÂB̂ ≤ λmin (σ as the well known Stinespring dilation theorem, in order to

AB )
∞
because ρAB is a density operator. So we then get arrive at the statement of the theorem. Stinespring’s theorem
states that for a quantum channel N acting on a state ω of a
π 2 λmin (σAB )
1/2 −1/2 1/2
3
1/2 system S, there exists an isometry US→AB such that
σAB σA ρA − ρAB
54 Tr{σA } 2
1/2 1/2
≤ Tr{ρA σA } − Tr{ρAB σAB }, (61)
1/2 1/2 N (ω) = TrB {US→AB ω(US→AB )† }. (65)
π 2 λmin (σAB )
1/2 −1/2 −1/2 1/2
3
So we pick
σAB σA ρA σA σAB − ρAB

432 Tr{σA } 1
1/2 1/2 1/2 1/2
≤ Tr{ρA σA } − Tr{ρAB σAB }, (62) ρAB = US→AB ω(US→AB )† , (66)
†
σAB = US→AB τ (US→AB ) , (67)
the latter of which being what was claimed in (7).
Acknowledgements. I thank Marco Piani and Anna Ver- so that ρA = N (ω), σA = N (τ ), and then find that
shynina for discussions related to the topic of this paper, and
I acknowledge support from the NSF under grant no. 1714215.
p
FH (N (ω), N (τ )) ≥
Note: The results and proofs in the main text of this p
paper were developed after [18] but independently of FH (US→AB ω(US→AB )† , US→AB τ (US→AB )† )
arXiv:1710.08080 and were communicated privately by email π 2 λmin (US→AB τ (US→AB )† )
+ ×
in mid-October 2017. 432 Tr{N (τ )}
RA→AB (N (ω)) − US→AB ω(US→AB )† 3 . (68)
σ
A PPENDIX
1
This appendix contains a generalization of the result in (7)
to arbitrary quantum channels. Due to isometric invariance of Holevo’s just-as-good fidelity
Theorem 2: Let ω be a density operator and τ a positive and the minimum non-zero eigenvalue, and the fact that N is
semi-definite operator such that supp(ω) ⊆ supp(τ ). Let N trace preserving, we find that
be a quantum channel. Then
FH (US→AB ω(US→AB )† , US→AB ω(US→AB )† ) = FH (ω, τ ),
p p
FH (N (ω), N (τ )) ≥ FH (ω, τ ) λmin (US→AB τ (US→AB )† ) = λmin (τ ),
π 2 λmin (τ ) 3 Tr{N (τ )} = Tr{τ }.
+ kP(N (ω)) − ωk1 , (63)
432 Tr{τ } (69)
Also, the Petz map RσA→AB simplifies for our choices as [22] F. Hiai, M. Mosonyi, D. Petz, and C. Beny, “Quantum f -divergences
and error correction,” Rev. Math. Phys., vol. 23, pp. 691–747, 2011.
RσA→AB (·) [23] D. Petz, “From quasi-entropy,” September 2010, arXiv:1009.2679.
1/2 −1/2 −1/2 1/2
[24] ——, “From f -divergence to quantum quasi-entropies and their use,”
= σAB [σA (·)σA ⊗ IB ]σAB Entropy, vol. 12, no. 3, pp. 304–325, 2010.
[25] M. Tomamichel, R. Colbeck, and R. Renner, “A fully quantum asymp-
= [US→AB τ (US→AB )† ]1/2 [N (τ )−1/2 (·)N (τ )−1/2 ⊗ IB ] totic equipartition property,” IEEE Transactions on Information Theory,
vol. 55, no. 12, pp. 5840–5847, December 2009.
× [US→AB τ (US→AB )† ]1/2 [26] N. Sharma, “Equality conditions for the quantum f -relative entropy and
= US→AB τ 1/2 (US→AB )† [N (τ )−1/2 (·)N (τ )−1/2 ⊗ IB ] generalized data processing inequalities,” in 2010 IEEE International
Symposium on Information Theory, June 2010, pp. 2698–2702.
× US→AB τ 1/2 (US→AB )† [27] H. Umegaki, “Conditional expectations in an operator algebra IV,” Kodai
Math. Sem. Rep., vol. 14, no. 2, pp. 59–85, 1962.
= US→AB τ 1/2 N † [N (τ )−1/2 (·)N (τ )−1/2 ]τ 1/2 (US→AB )† . [28] M. M. Wilde, “Optimized quantum f-divergences and data processing,”
(70) October 2017, arXiv:1710.10252.
[29] F. Hansen and G. K. Pedersen, “Jensen’s operator inequality,” Bulletin
Isometric invariance of the trace norm and the above then gives London Math. Soc., vol. 35, no. 4, pp. 553–564, July 2003.

RA→AB (N (ω)) − US→AB ω(US→AB )†

σ
1
= kP(N (ω)) − ωk1 , (71)
concluding the proof.
R EFERENCES
[1] A. S. Holevo, “On quasiequivalence of locally normal states,” Theor.
Math. Phys., vol. 13, no. 2, pp. 1071–1082, November 1972.
[2] A. Uhlmann, “The “transition probability” in the state space of a *-
algebra,” Rep. Math. Phys., vol. 9, no. 2, pp. 273–279, 1976.
[3] C. A. Fuchs and J. van de Graaf, “Cryptographic distinguishability mea-
sures for quantum mechanical states,” IEEE Transactions on Information
Theory, vol. 45, no. 4, pp. 1216–1227, May 1998.
[4] Audenaert, K. M. R. et al., “Discriminating states: The quantum
Chernoff bound,” Phys. Rev. Lett., vol. 98, p. 160501, April 2007.
[5] Calsamiglia, J. et al., “Quantum Chernoff bound as a measure of
distinguishability between density matrices: Application to qubit and
Gaussian states,” Physical Review A, vol. 77, p. 032311, March 2008.
[6] M. Hayashi, Quantum Information: An Introduction. Springer, 2006.
[7] M. M. Wilde and S. Guha, “Polar codes for classical-quantum channels,”
IEEE Transactions on Information Theory, vol. 59, no. 2, pp. 1175–
1187, February 2013, arXiv:1109.2591.
[8] S. Guha and M. M. Wilde, “Polar coding to achieve the Holevo capacity
of a pure-loss optical channel,” in Proceedings of the 2012 Int. Symp.
Inf. Theory, Boston, MA, USA, 2012, pp. 546–550, arXiv:1202.0533.
[9] D. Petz, “Quasi-entropies for states of a von Neumann algebra,” Publ.
RIMS, Kyoto University, vol. 21, pp. 787–800, 1985.
[10] ——, “Quasi-entropies for finite quantum systems,” Reports in Mathe-
matical Physics, vol. 23, pp. 57–65, 1986.
[11] S. Luo and Q. Zhang, “Informational distance on quantum-state space,”
Physical Review A, vol. 69, no. 3, p. 032106, March 2004.
[12] A. Winter, ““Extrinsic” and “intrinsic” data in quantum measurements:
asymptotic convex decomposition of positive operator valued measures,”
Comm. Math. Phys., vol. 244, no. 1, pp. 157–185, January 2004.
[13] D. Petz, “Sufficient subalgebras and the relative entropy of states of a
von Neumann algebra,” Communications in Mathematical Physics, vol.
105, no. 1, pp. 123–131, March 1986.
[14] ——, “Sufficiency of channels over von Neumann algebras,” Quarterly
Journal of Mathematics, vol. 39, no. 1, pp. 97–108, 1988.
[15] O. Fawzi and R. Renner, “Quantum conditional mutual information and
approximate Markov chains,” Communications in Mathematical Physics,
vol. 340, no. 2, pp. 575–611, December 2015, arXiv:1410.0664.
[16] M. M. Wilde, “Recoverability in quantum information theory,” Proc.
Roy. Soc. A, vol. 471, no. 2182, p. 20150338, October 2015.
[17] M. Junge et al., “Universal recovery from a decrease of quantum relative
entropy,” September 2015, arXiv:1509.07127.
[18] E. A. Carlen and A. Vershynina, “Recovery map stability for the data
processing inequality,” October 2017, arXiv:1710.02409.
[19] D. Petz, “Monotonicity of quantum relative entropy revisited,” Rev.
Math. Phys., vol. 15, no. 1, p. 79, March 2003.
[20] A. S. Holevo, Quantum Systems, Channels, Information, ser. de Gruyter
Studies Math. Phys. (Book 16). de Gruyter, November 2012.
[21] M. M. Wilde, Quantum Information Theory, 2nd ed. Cambridge
University Press, 2017, available as arXiv:1106.1445.