Ke Li’s lemma for quantum hypothesis testing in

general von Neumann algebras

Yan Pautrat and Simeng Wang
arXiv:2010.02177v1 [math-ph] 5 Oct 2020

Université Paris-Saclay, CNRS, Laboratoire de Mathématiques d’Orsay

91405 Orsay, France

October 6, 2020

Abstract
A lemma stated by Ke Li in [9] has been used in e.g. [3, 8, 11, 13, 14] for
various tasks in quantum hypothesis testing, data compression with quantum
side information or quantum key distribution. This lemma was proven in finite
dimension only (with an easy extension to type I von Neumann algebras).
Here we show that the use of modular theory allows to give more transparent
meaning to the objects constructed by the lemma, and to prove it for general
von Neumann algebras.

1 Introduction
Quantum hypothesis testing is concerned with the situation where one considers a von
Neumann algebra M, equipped with a state which is either ρ or σ; this uncertainty
in the nature of the state makes sense in particular when M is viewed as modeling
the observable quantities of a quantum system, and the physical state of this system
— itself modeled by a state in the mathematical sense — is only known to be one or
the other. A natural task is then to try and determine which state is the true one
by producing, in physical terms, an experimental measurement procedure such that,
depending on the measurement outcome, one will conclude that the actual state is
ρ or σ.
In the model of orthodox quantum mechanics, a measurement procedure is de-
termined by a self-adjoint element X of the observable algebra M, and this X is
simply called an observable. Non-trivial measurements will have a random outcome;
the set of possible outcomes is exactly the spectrum of that element, and if ω is the
actual state of the system and we denote by ξX the spectral measure of X, then
the probability distribution for the measurement outcomes is ρ ◦ ξX . In the situa-
tion described above it will suffice to consider an observable with spectrum {0, 1},
that is, an orthogonal projector of M. We will take advantage of this simplification

1
and continue this preliminary discussion assuming that the observable describing the
discriminating experience, is an orthogonal projector T , which we call a test.
Suppose the decision rule is that, if the measurement outcome is 0, then the
observer concludes that the true state is ρ, and if the measurement outcome is 1, then
they conclude that the true state is σ. There are two ways in which this conclusion
can be wrong: either the actual state was ρ and the measurement outcome was 1,
or the actual state was σ and the measurement outcome was 0. Applying the rules
determining the distribution of the measurement outcomes shows that the former
occurs with probability ρ(T ), and the latter with probability σ(Id − T ). Assuming
that the possible states ρ and σ are fixed, we denote these two types or error

α(T ) := ρ(T ) β(T ) := σ(Id − T ). (1)

If, to the experimenter’s knowledge, both ρ and σ may be the state of the system,
then one typically wishes to make both α(T ) and β(T ) small. There are, however,
various ways in which this can be done, and we postpone the corresponding discussion
to section 4.
The result we are concerned with gives a test T such that α(T ) and β(T ) satisfy
a pair of upper bounds. To state it, let us assume that M is a finite-dimensional
matrix algebra: M = B(Cn ), and denote by ̺ and ς the density matrices associated
with the states ρ and σ, that is:

ρ = tr(̺ ·) σ = tr(ς ·). (2)

Assume for simplicity that ρ and σ are faithful states, or equivalently that ̺, ς are
invertible matrices. Consider then the vector space B(Cn ), on which we define the
vectors
Ωρ = ̺1/2 Ωσ = ς 1/2
and the operator
∆ρ|σ : X 7→ ̺Xς −1 .
If B(Cn ) is equipped with the scalar product hX, Y i = tr(X ∗ Y ) then ∆ρ|σ is self-
adjoint. It then holds that for any ǫ > 0 there exists a test T such that

α(T ) ≤ ǫ β(T ) ≤ hΩσ , 1(ǫ,+∞)(∆ρ|σ )Ωσ i. (3)

This result was first proven in [9], and it is generally quoted as “Ke Li’s lemma” for
quantum hypothesis testing, even though it is not the only result of Ke Li relevant
to this field. The statement as written above, however, does not appear in [9] and
in particular, there is no mention of the operator ∆ρ|σ in that article. The result
was later reformulated in [3] to involve the operator ∆ρ|σ , which the reader will
already have recognized to be the finite-dimensional instance of a relative modular
operator. From then one it was expected that a proof written solely in terms of
modular theory on B(Cn ) should be possible, in which case, according to a standard
rule of thumb discussed in [7] it should extend with minimal effort to general von
Neumann algebras. It turns out that such an extension holds for any two faithful

2
normal states ρ, σ on a σ-finite von Neumann algebra. To state it we need, however,
to recall how the relative modular operator ∆ρ|σ is defined in the general case, and
we therefore postpone its statement to Section 3.
Even though our proof is written in the language of modular theory, it did not
proceed from a simple “modular translation” of Ke Li’s finite dimensional proof, as we
could not find such a thing. Note that until now, no extension of inequalities (3) was
found, except for the simple extension to separable type I von Neumann algebras,
in which case ρ and σ were still of the form (2) and one could simply proceed by
finite-dimensional approximations of ̺ and ς.
Let us recall shortly the definition of T in the proof of (3) as in [3, 9] in order
to underline its connection with modular theory, and motivate our definition of T
below. Consider spectral decompositions of ̺, ς:
n
X n
X
̺= λx |ax ihax | ς= µy |by ihby |
x=1 y=1

where (λx )x and (µy )y are labeled in nondecreasing order, and both families (ax )x
and (by )y are orthonormal bases of Cn . Define for any y = 1, . . . , n:
n
Qy = 1[0,ǫµy ] (̺) = 1λx ≤ǫµy |axihax |,
X

x=1

so that (Qy )y is a nondecreasing family of projectors, and

n
1λx ≤ǫµy hax , by iax .
X
ξy = Qy by =
x=1

Ke Li then defines his test as the orthogonal projector TKL onto the vector space
spanned by the ξy , y = 1, . . . , n.
For X an operator on Cn denote now by LX the operator on B(Cn ) acting by
Y 7→ XY , and let J be the antilinear involution X 7→ X ∗ on B(Cn ). Notice then
that (JLX J)Y = Y X ∗ and let M, M′ be the sets of operators of the form LX or
JLX J respectively. We then have:
n
X
∆ρ|σ = λx µy −1 L|ax ihax | JL|by ihby | J
x,y=1

so that
n n
1(0,ǫ](∆ρ|σ ) = 1λx ≤ǫµy L|ax ihax | JL|by ihby | J =
X X
LQy JL|by ihby | J
x,y=1 y=1

and n
1(0,ǫ] (∆ρ|σ ) Ωσ = 1λx ≤ǫµy µ1/2
X
y hax , by i |ax ihby |. (4)
x,y=1

3
It is then immediate to remark that 1(0,ǫ] (∆ρ|σ ) Ωσ by = µy ξy for any y = 1, . . . , n, so
that

Ran 1(0,ǫ] (∆ρ|σ ) Ωσ = span ξy , y = 1, . . . , n



where the last equation is due to the faithfulness of σ. Last, remark that the range
Ran 1(0,ǫ] (∆ρ|σ ) Ωσ is the same as M′ 1(0,ǫ] (∆ρ|σ ) Ωσ . Therefore, TKL is equivalently

defined by the fact that LTKL is the orthogonal projection on M′ 1(0,ǫ] (∆ρ|σ ) Ωσ .
Since all quantities appearing in this last sentence are well-defined in the standard
representation of any von Neumann algebra (see section 2), this gives a satisfactory
starting point for our proof of the extension of Ke Li’s lemma.
The structure of the paper is as follows. In section 2 we recall the elements of the
modular theory of von Neumann algebras required by the statement of our result.
In section 3 we give our result and its proof. In section 4 we discuss the merits of
our result and its possible applications.

Acknowledgements YP wishes to thank Nilanjana Datta and Cambyse Rouzé for

initiating the pleasant collaboration that led to [3] and the present endeavors, Ke Li
and Magdalena Musat for instructive discussions and most of all Yoshiko Ogata who
helped chase the modular in Ke Li’s original proof. YP was supported by ANR grant
“NonStops” ANR-17-CE40-0006. SW was partially supported by a public grant as
part of the FMJH and ANR grant “ANCG” ANR-19-CE40-0002.

2 Modular theory and Haagerup Lp spaces

We recall first a few notions on von Neumann algebras and relative modular opera-
tors. Suggested references are [2], and for a pedagogical introduction we recommend
section 2 (and in particular sections 2.11 and 2.12) of [7]. We then move on to
describe a few properties of Haagerup’s Lp spaces following [5] and [12].
Let M be a von Neumann algebra, i.e. a ∗-algebra of bounded linear operators
on some Hilbert space, that is closed in the weak operator topology and contains
the identity. The set of operators commuting with all elements of M is called the
commutant of M, and denoted M′ ; von Neumann’s bicommutant theorem then
states that a ∗-algebra of bounded linear operators containing the identity is a von
Neumann algebra if and only if M = M′′ , the latter being the bicommutant (M′ )′
of M.
The notion of nonnegative operators induces a partial order on M: by defi-
nition, X ≤ Y if Y − X is a nonnegative operator. A map on M such that
the null operator is the only nonnegative operator mapped to the zero element is
called faithful. Linear forms ω on M with the additional regularity property that
ω(lim supn Xn ) = lim supn ω(Xn ) if Xn is an increasing sequence of self-adjoint op-
erators are called normal. A linear form which maps nonnegative operators to non-
negative scalars is called positive. A positive linear form ω from M to [0, +∞] will
be called a weight. A weight ω such that the set of nonnegative elements A with

4
ω(A) < +∞ is weakly dense is called semifinite. A weight ω with the property that
ω(Id) = 1 is called a state (and necessarily takes values in [0, +∞)). A weight or
state with the property that ω(XY ) = ω(Y X) for all X, Y in M is called tracial, or
a trace.
A standard representation of some von Neumann algebra M is a quadruple
(π, H, H+ , J) where H is a Hilbert space (with scalar product denoted h·, ·i), π is
a faithful morphism of ∗-algebras from M to B(H), H+ is a self-dual cone of H (i.e.
the set of φ in H such that hφ, ψi ≥ 0 for all ψ in H+ is H+ itself), J is an anti-unitary
involution of H, and these different objects satisfy the following properties:

• JMJ = M′ ,
• JXJ = X ∗ for X in M ∩ M′ ,
• Jψ = ψ for ψ in H+ ,
• JXJX H+ ⊂ H+ for X ∈ H.

A standard representation of a von Neumann algebra M always exists (we will

describe one of them below), and if (π1 , H1 , H1+ , J1 ) and (π2 , H2 , H2+ , J2 ) are two
standard representations, then there exists a unitary operator U : H1 → H2 such
that Uπ1 (X)U ∗ = π2 (X) for all X ∈ M, UJ1 U ∗ = J2 and UH1+ = H2+ .
We will now describe the modular structure associated with a von Neumann
algebra M: fix a standard representation (π, H, H+ , J) of M. Then for any normal
state ω on M there exists a unique Ωω in H+ such that

ω(X) = hΩω , π(X)Ωω i for any X ∈ M. (5)

This Ωω is cyclic in the sense that the closure of π(M)H is H itself. If in addition ω
is faithful, then Ωω is separating in π(M), that is, π(X)Ωω = 0 if and only if X = 0.
We continue by fixing two normal faithful states ρ and σ. We can then define a
densely defined operator Sρ|σ by

Sρ|σ XΩσ = X ∗ Ωρ .

This operator turns out to be closable, and its closure S ρ|σ has polar decomposition

S ρ|σ = J∆ρ|σ 1/2

where ∆ρ|σ is a positive-definite operator on H, which is in general unbounded.

We are now in a position to state our main result, but will first introduce ad-
ditional elements required for our proof. The reader may wish to take a peek at
Theorem 1 in section 3.
To prove the result we consider the Haagerup Lp -spaces associated with M. We
will not need the detailed construction of these spaces and only recall their properties
(see chapter 2 of [12] for a complete presentation; in the rest of this section all
numbered references point to those notes, with e.g. Theorem 2.7 meaning Theorem
7 of chapter 2 of [12]). An important role will be played by a distinguished von
Neumann algebra N , equipped with a normal semifinite faithful trace τ . This algebra

5
N then acts on the Hilbert space L2 (R, H) which we simply denote K, for its nature
will not matter. For a definition of this algebra N , see the beginning of chapter 2
of [12]: between saying too much and saying too little, it seems more convenient to
say too little. This N is such that there exists a faithful normal representation of
M as a sub-von Neumann algebra of N ; in order to spare ourselves an additional
notation for this representation, we simply assume that M is realized as a subalgebra
of N , and therefore acts on K = L2 (R, H). We say that a densely defined unbounded
operator on K is measurable with respect to (N , τ ) if for every δ > 0 there exists
a projection P in N such that the range of P is included in the domain of N , and
τ (Id − P ) < δ. We say that an unbounded operator on K is affiliated with N if any
Borel bounded functional of either 21 (K + K ∗ ) or 2i1 (K − K ∗ ) is an element of N .
Then for all 1 ≤ p ≤ ∞, there exists a space Lp (M) which is a subspace of the set of
those measurable operators with respect to (N , τ ) which are affiliated with N , which
satisfies the following properties. For all 1 ≤ p < ∞, an operator X on K is in Lp (M)
if and only its polar decomposition X = U|X| satisfies U ∈ M and |X|p ∈ L1 (M).
Moreover there is a faithful positive linear functional tr : L1 (M) → C such that

tr(XY ) = tr(Y X) for all X ∈ Lp (M), Y ∈ Lq (M) with 1/p + 1/q = 1.

If we define kXkp = tr(|X|p )1/p then the closed product of measurable operators on
these spaces satisfies the Hölder inequality, and in particular we have Lp (M)Lq (M) ⊂
Lr (M) for any p, q, r ∈ [1, +∞] with 1/r = 1/p + 1/q. This induces a natural Hilbert
space structure on L2 (M) by taking p = q = 2. In addition (Theorem 2.7 and
Proposition 2.10), M = L∞ (M) and the normal states ρ, σ on M are of the form

ρ = tr(̺·) σ = tr(ς·) (6)

where ̺, ς are positive elements of L1 (M) with tr(σ) = 1. This of course echoes (2)
and will allow us to mimic some of the “spatial” arguments in Ke Li’s proof. Taking
the values of p, q, r appropriately, we obtain a faithful representation of M on L2 (M)
as
π(X)Y = XY, X ∈ M, Y ∈ L2 (M).
Denote J : X 7→ X ∗ and L2 (M)+ the 2

subset of L (M) made of nonnegative oper-
2 2
ators. Then π, L (M), L (M)+ , J is a standard form of M (see Theorem 2.36).
Last, we will use the crucial fact that Lp (M) is isometrically isomorphic to a sub-
space of the tracial weak Lp -space on (R, τ ), and more precisely we have (Lemma
2.5):
tr(X) = τ 1(1,∞) (X) for any X ∈ L1 (M)+ .


(7)

3 Our result and its proof

We first state the general form of Ke Li’s lemma. As discussed in section 2, the
vector Ωσ and the operator ∆ρ|σ are meant in any standard representation of M.

6
Theorem 1. Let M be a von Neumann algebra, and ρ, σ be two faithful normal
states on M. For any ǫ > 0 there exists a test T ∈ M such that

α(T ) ≤ ǫ β(T ) ≤ hΩσ , 1(ǫ,+∞)(∆ρ|σ )Ωσ i.

Remark 2. Our assumption that ρ and σ are faithful is unessential, simplifies the
proof and the description of ∆ρ|σ in section 2, and can be dropped by considering an
approximation of general ρ, σ by faithful states.
The rest of this section is dedicated to the proof

of Theorem 1. We consider
2 2
the standard representation π, L (M), L (M)+ , J of M that we discussed in the
second part of section 2. Once again we simply write M for π(M), so that M will
act by left-multiplication on L2 (M).
Let P be the following orthogonal projection operator on L2 (M):

P = 1(0,ǫ] (∆ρ|σ ).

We recall that if X is an element of L2 (M) then its support in M, which we denote

by ℓ(X) (with an ℓ for “left support”), is defined as the smallest orthogonal projector
P in M such that P X = X. It is immediate to show that this projector has range
M′ X. Similarly we define the support of X ∈ L2 (M) in M′ , which we denote by
r(X) (with an r for “right support”) as the smallest orthogonal projector P ′ in M′
such that P ′ X = X.
First define our test T as the support in M of P(Ωσ ). By definition P(Ωσ ) is
an element of L2 (M), and therefore a densely defined closed operator on K which is
affiliated with N ; the polar decomposition of P(Ωσ ) as an operator on K has partial
isometry U which from the properties of L2 (M) belongs to M. Then UU ∗ is by
definition the smallest orthogonal projector P in B(K) such that P P(Ωσ ) = P(Ωσ ),
and belongs to M. Therefore T is equivalently defined as the orthogonal projector
onto M′ P(Ωσ ), as UU ∗ , or as left-multiplication by the orthogonal projector onto
the closed range of P(Ωσ ) when the latter is viewed as an operator on K. Note that
by definition, T ∈ M.
From the first expression for T we have

σ(Id − T ) = kΩσ − T Ωσ k2 ≤ kΩσ − P(Ωσ )k2 = hΩσ , 1(ǫ,+∞)(∆ρ|σ )Ωσ i

where:

• the first equality follows from (5) and the fact that Id − T is a self-adjoint
projection,
• the inequality follows from the inequality kΨ − T Ψk ≤ kΨ − Φk, valid for any
Ψ in L2 (M) and Φ ∈ Ran T by definition of an orthogonal projection, and the
fact that P(Ωσ ) is an element of M′ P(Ωσ ) ⊂ Ran T ,
• the second equality follows from (5) and the fact that Id − P is a self-adjoint
projection.

7
This proves the second bound in Theorem 1, and we now move on to prove the first
bound.
R1 R1
Let ̺ = 0 λ deρ (λ) and ς = 0 µ deσ (µ) be spectral decompositions of ̺, ς in K.
Then from the definition in the second part of section 2,

∆ρ|σ = ̺ Jς −1J on MΩσ

so that, denoting ẽσ := Jeσ J,

Z
∆ρ|σ = λµ−1 deρ (λ) dẽσ (µ)
2
R+

at least on MΩσ , and since the latter integral expression defines a closed operator,
equality holds everywhere. Approximating 1(0,ǫ] by polynomials weakly, it is easy to
see that Z
P=
2
1(0,ǫ] (λµ−1 ) deρ (λ)dẽσ (µ).
R+

Applying Fubini’s theorem twice, we have

Z
P(Ωσ ) =
2
1(0,ǫ] (λµ−1 )µ1/2 deρ (λ)dẽσ (µ)
R+
Z ∞
= µ1/2 1(0,ǫµ] (̺) dẽσ (µ). (8)
0

We now prove a regularity lemma that will allow us to approximate T thanks to

Riemann sums approximating the expression (8) for P(Ωσ ). Assume therefore that
(Xn )n is a sequence of bounded operators on K that converges to X strongly, and
let pn (respectively p) be the support of Xn (respectively X) as an operator on K.
For any ξ in K we have using pn Xn ξ = Xn ξ and p Xξ = Xξ:

kpn pXξ − pXξk ≤ kpn Xξ − pn Xn ξk + kpn Xn ξ − Xξk ≤ 2kXn ξ − Xξk

and the upper bound goes to zero as n → ∞. In particular, pn p converges to p

strongly, which further implies that (1 − p)pn p and ppn (1 − p) converge to 0 weakly.
Now consider Pn (respectively P ) the operator of left-multiplication by pn (respec-
tively p) on L2 (M). By the normality of ρ, we get

ρ(P ) = lim ρ(P Pn P )

n
= lim ρ(P Pn P ) + lim ρ((1 − P )Pn P ) + lim ρ(P Pn (1 − P ))
n n

n
= lim ρ(Pn ) − ρ((1 − P )Pn (1 − P ))
n
≤ lim sup ρ(Pn ). (9)
n

8
 Recall now that our test T is defined as the support ℓ(X) of X := P(Ωσ ). Let
(Xn )n be a sequence of Riemann sums of the form

µi 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 )

X 1/2

Xn =
i

that converges strongly to X, and let Tn be the support ℓ(Xn ) of Xn . From (9), we
have ρ(T ) ≤ lim supn ρ(Tn ) and we therefore consider ρ(Tn ). Obviously

ℓ 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 ) =: Tn′ ,

X


Tn ≤
i

1/2
where we omitted the scalar coefficients λi since X and its multiple λX must have
the same range for any X ∈ L2 (M). Note that

1[0,ǫµi ) (̺) is a projection whose range
contains that of ℓ 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 ) , so we write

ℓ 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 ) = 1[0,ǫµi ) (̺) ℓ 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 ) .

Note that (see e.g. [4])

τ 1(1,∞) (X) ≤ τ 1(1,∞) (Y ) if 0 ≤ X ≤ Y.

Thus together with (7),

tr ̺1[0,ǫµi ) (̺) ℓ 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 )

X
tr(̺Tn′ ) =




i
X  


= τ 1(1,∞) ̺1[0,ǫµi ) (̺) ℓ 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 )
i
X  


≤ τ 1(1,∞) ǫµi ℓ 1[0,ǫµi ) (̺) ẽσ (µi ) − ẽσ (µi−1 )
i



1(1,∞) (ǫµi ) τ ℓ 1[0,ǫµi ) (̺) ẽσ (µi) − ẽσ (µi−1 )
X
≤ .
i

Note that pi,1 := 1[0,ǫµi ) (̺) and pi,2 := ẽσ (µi ) − ẽσ (µi−1 ) are projections. In this case
it is folklore that ℓ(pi,1 pi,2 ) is equivalent to r(pi,1 pi,2 ), i.e. that there exists a unitary
U in M such that ℓ(pi,1 pi,2 ) = UU ∗ and r(pi,1 pi,2 ) = U ∗ U, and

r(pi,1 pi,2 ) = pi,1 ∨ p⊥ ⊥ ⊥

i,2 − pi,2 ≤ 1 − pi,2 = pi,2 .


In other words, ℓ(pi,1 pi,2 ) is equivalent to a subprojection of pi,2 , whence τ ℓ(pi,1 pi,2 ) ≤
τ (pi,2 ). So the previous inequality reads
X 
tr(̺Tn′ ) ≤ τ 1(1,∞) (ǫµi ) ẽσ (µi ) − ẽσ (µi−1 )
i

Taking the limit of Riemann sums and using (7) again, we get
Z 
1(1,∞) (ǫµ) dẽσ (µ) = τ 1(1,∞) (ǫς) = tr(ǫς) = ǫ.


tr(̺T ) ≤ τ

The proof is complete.

9
4 Comparison with the Neyman–Pearson test
Let us now return to the practical task of interest, which is of discriminating between
the states ρ and σ. As we mentioned in the introduction, one will typically try to
make both error probabilities α(T ) and β(T ) small (recall that α(T ) and β(T ) are
defined by (1)). However, it is expected that there is a tradeoff between one error
probability and the other. One must therefore make more precise in what sense
we want to “make both α(T ) and β(T )” small. One possible sense is natural if we
assume that prior probabilities p and q := 1 − p can be assigned to the states ρ and
σ respectively. One may then decide to minimize the quantity p α(T ) + q β(T ). This
is the realm of symmetric hypothesis testing, as opposed to assymetric hypothesis
testing where one wishes e.g. to minimize α(T ) under a specific constraint on β(T ).
To state the relevant result for symmetric hypothesis testing we define a specific
test TNP called the Neyman–Pearson test. In the finite-dimensional test it is defined
as
TNP := ℓ 1R+ (p̺ − qς) .

In the general case the definition is formally similar:

TNP := ℓ (pρ − qσ)+

but here one needs to rely on the (unique) Jordan decomposition

pρ − qσ = (pρ − qσ)+ − (pρ − qσ)−

with (pρ − qσ)± two positive normal forms on M. We will freely denote TNP as
TNP (p, q) when we need to emphasize the dependency on p, q.
The following result is known as the quantum Neyman–Pearson lemma and was
proved in [1] in the finite-dimensional case and in [6] in the general case:


p α(TNP ) + q β(TNP ) = inf p α(T ) + q β(T ) (10)
T

where the infimum is indifferently over the set of orthogonal projections of M, or

over the set of elements T of M satisfying 0 ≤ T ≤ Id. The associated relevant
upper bound was also proven in [1, 6] and is called the Chernoff bound :

p α(TNP ) + q β(TNP ) ≤ inf ps q 1−s hΩσ , ∆sρ|σ Ωσ i. (11)

s∈[0,1]

Note that the Chernoff bound (with an additional lower bound, see [6,10]) is sufficient
to prove the standard Stein’s lemma and Hoeffding bounds for asymmetric hypothesis
testing, see [6].
Denote by TKL (or, again, by TKL (ǫ) when we need to emphasize the dependence
on ǫ) the test constructed in section 3 (we confess a slight inconsistency in our choice
of notation, as the present TKL is the LTKL of section 1). A first interesting observation
is that, in the case where M is commutative, or when ̺ and ς commute, then
1 ǫ

TKL (ǫ) = TNP , .
1+ǫ 1+ǫ

10
This can be easily seen in the finite-dimensional case, i.e. when M = {LX , X ∈
B(Cn )} as described in section 1 (and using the same notation). Using the fact that
the families (ax )x and (by )y are the same (up to permutation) when ̺ and ς commute,
expression (4) takes the form
n
1(ǫ,+∞) (∆ρ|σ ) Ωσ = 1λx >ǫµx µ1/2
X
x |ax ihax |. (12)
x=1
1
On the other hand, when p = 1+ǫ
,
n
1R+ (p̺ − qς) = 1λx >ǫµx |ax ihax |.
X
(13)
x=1

Therefore, the supports of the operators in (12) and (13) are the same, so that
TKL = TNP .
Now let us compare the merits of Theorem 1, in comparison to the Chernoff
1
bound (11). From now on we always consider p = 1+ǫ so that p satisfies the relation
q/p = ǫ. The Chernoff bound (11) implies
ρ(TNP ) ≤ ǫ σ(TNP ) ≤ ǫ−s hΩσ , ∆sρ|σ Ωσ i (14)
whereas Theorem 1 gives
ρ(TKL ) ≤ ǫ σ(TKL ) ≤ hΩσ , 1(ǫ,+∞) (∆ρ|σ )Ωσ i. (15)
If the upper bound for σ(TKL ) in (15) is bounded by an application of Markov’s ex-
ponential inequality, then (14) and (15) yield the same pair of bounds. It therefore
turns out that the estimates (15) given by Theorem 1 are no better than those (14)
given by the Chernoff bound, unless one has an estimate for hΩσ , 1(ǫ,+∞)(∆ρ|σ )Ωσ i
more precise than that given by Markov’s exponential inequality. It therefore seems
that our result is better suited to situations where specific information on the tails
of the distribution of ∆ρ|σ with respect to the state Ωσ is available. This was the
case, in particular, in [9] or in [11], which used specific information on the struc-
ture of the states ρ and σ (namely, multiplicativity or submultiplicativity) to derive
concentration inequalities on the latter distribution. This, however, was done in a
finite-dimensional setting for which Theorem 1 follows from Ke Li’s original result
and the present general extension is not needed.

References
[1] Konrad Audenaert, Michael Nussbaum, and Arleta Szkoła. Asymptotic error
rates in quantum hypothesis testing. Comm. Math. Phys., 279, 2008.
[2] Ola Bratteli and Derek W. Robinson. Operator algebras and quantum statistical
mechanics. 1. Texts and Monographs in Physics. Springer-Verlag, New York,
second edition, 1987. C ∗ - and W ∗ -algebras, symmetry groups, decomposition of
states.

11
[3] Nilanjana Datta, Yan Pautrat, and Cambyse Rouzé. Second-order asymptotics
for quantum hypothesis testing in settings beyond i.i.d.—quantum lattice systems
and more. J. Math. Phys., 57(6):062207, 26, 2016.

[4] Thierry Fack and Hideki Kosaki. Generalized s-numbers of τ -measurable opera-
tors. Pacific J. Math., 123(2):269–300, 1986.

[5] Uffe Haagerup. Lp -spaces associated with an arbitrary von Neumann algebra.
In Algèbres d’opérateurs et leurs applications en physique mathématique (Proc.
Colloq., Marseille, 1977), volume 274 of Colloq. Internat. CNRS, pages 175–184.
CNRS, Paris, 1979.

[6] Vojkan Jakšić, Yoshiko Ogata, Claude-Alain Pillet, and Robert Seiringer. Quan-
tum hypothesis testing and non-equilibrium statistical mechanics. Rev. Math.
Phys., 24(6):1230002, 67, 2012.

[7] Vojkan Jakšić, Yoshiko Ogata, Yan Pautrat, and Claude-Alain Pillet. Entropic
fluctuations in quantum statistical mechanics. an introduction. Quantum Theory
from Small to Large Scales, Lecture Notes of the Les Houches Summer School,
95(978-0-19-965249-5):pp.213–410, 2012.

[8] Eneet Kaur and Mark M. Wilde. Upper bounds on secret-key agreement over
lossy thermal bosonic channels. Phys. Rev. A, 96:062318, Dec 2017.

[9] Ke Li. Second-order asymptotics for quantum hypothesis testing. Ann. Statist.,
42(1):171–189, 02 2014.

[10] Michael Nussbaum and Arleta Szkoła. The Chernoff lower bound for symmetric
quantum hypothesis testing. Ann. Statist., 37(2):1040–1057, 2009.

[11] Cambyse Rouzé and Nilanjana Datta. Finite blocklength and moderate devia-
tion analysis of hypothesis testing of correlated quantum states and application
to classical-quantum channels with memory. IEEE Transactions on Information
Theory, 64(1):593–612, 2018.

[12] Marianne Terp. Lp spaces associated with von neumann algebras. Notes, Math.
Institute, Copenhagen University, 1981.

[13] Marco Tomamichel and Vincent Y. F. Tan. Second-order asymptotics for the
classical capacity of image-additive quantum channels. Comm. Math. Phys.,
338(1):103–137, 2015.

[14] Mark M. Wilde, Marco Tomamichel, and Mario Berta. Converse bounds for pri-
vate communication over quantum channels. IEEE Transactions on Information
Theory, 63(3):1792–1817, 2017.

12