A quantum version of randomization criteria

arXiv:1012.2650v7 [quant-ph] 19 May 2015

Keiji Matsumoto
November 9, 2018

Abstract
In classical statistical decision theory, comparison of experiments plays
very important role. Especially, so-called randomization criteria is most
important. In this paper, we establish two kinds of quantum analogue of
these concepts, and apply to some examples.

1 Introduction
In classical statistical decision theory, comparison of experiments plays very
important role. Especially, so-called randomization criteria is most important.
In this paper, we establish two kinds of quantum analogue these concepts, and
apply to some examples.

2 Review of classical theory

2.1 Desicion theory, frame work
A statistical experiment E = (X , X, {Pθ ; θ ∈ Θ}) consists of four parts. First,
the data space X , or the totality of all the possible data x. Second, a collection
X of subsets of X , indicating minimal unit of the events which the statistician
is concerned with.
The third element of an experiment is a parameter set Θ, which indeces all
the possible explaining theory for the outcomming data. To each θ corresponds a
probablity distribution Pθ of data, which is the fourth element of an experiment.
Pθ has to be X-measurable.
A statistician makes decision based on the data x ∈ X . The totality of
possible decisions made by the statisitcian is the decision space D, which is a
topological space and is equipped with Baire σ-field D. For example, if the
statistician is estimating the value of the parameter


θ ∈ × behind the data from
the data x ∈ X , we define (D, D) = Rl , B Rl . If the statistician is trying to


distinguish whether θ ∈ Θ0 or θ ∈ Θ1 , (D, D) = {0, 1} , 2{0,1} .
The performance is measured by a loss function lθ (t), or a function which
depends not only on the decision t ∈ D but also on the true value of the param-
eter θ ∈ Θ. It is assumed that the function t → lθ (t) is lower semicontinuous,

1
non-negative, and
−1 ≤ lθ (t) ≤ 1.
For example, in case of estimation of θ, the loss function may be

0, if kt − θk ≤ c
lθ (t) = .
1, othewise

In case of testing θ ∈ Θ0 against θ ∈ Θ1 , lθ (t) may be chosen so that lθ (1) =

1 − lθ (0) and 
0, if θ ∈ Θ0
lθ (0) = .
1 if θ ∈ Θ1
In general the statistician’s startegy (decision) is described by a bilinear map
D
D : lθ × Pθ → D (lθ , Pθ ) ∈ R,
which satisfies

|D (f, P )| ≤ kf k kP k1 ,
D (f, P ) ≥ 0, if f ≥ 0, P ≥ 0,
D (1, P ) = P (D) .

Here, P is a member of L-space of E, or a bounded singned measure such that

P ⊥ ν for all ν with Pθ ⊥ ν, ∀θ ∈ Θ. The meaning of D (lθ , Pθ ) is average of lθ
when the statistician takes the decision corresponding to D, and in many cases,
there is a Makov kernel RD such that
Z Z
D (lθ , Pθ ) = lθ (t) RD (dt, x) Pθ (dx) .
X D

D is said to be k-decision when |D| = k.

Note here that Θ can be any set, and the function θ → Pθ (B) can be any
function.

2.2 Defficiency
Let e : × → R+ be a function with 0 ≤ eθ ≤ 1, klθ k := supt∈D |lθ (t)|, and
klk := supθ∈Θ klθ k. An experiment E = (X , X, {Pθ ; θ ∈ Θ}) is said to be e-
deficient relative to another experiment F = (Y, Y, {Qθ ; θ ∈ Θ}) (denoted by
E ≥e F ), if and only if, for any loss function l with klk ≤ 1, for any finite subset
Θ0 of Θ, and any decision D on the experiment F , there is a decision D′ on the
experiment E such that

D′ (lθ , Pθ ) ≤ D (lθ , Qθ ) + eθ , ∀θ ∈ Θ0 . (1)

Defficiency δ (E, F ) is defined by

 
δ (E, F ) := inf sup eθ ; E ≥e F ,
e θ

2
E ≥0 F is denoted by E ≥ F , and when this holds, E is said to be more informa-
tive than F .
An experiment E is said to be e-deficient for k-decision problems relative to
another experiment F , if and only if, for any loss function l, any finite subset
Θ0 of Θ, and any k-decision D on the experiment F , there is a k-decision D′
on the experiment E such that (1) holds for any θ ∈ Θ0 (denoted by E ≥e,k F ).
Also, we define deficiency δk (E, F ) for k-decision problems by restricting D′ and
D to the k-desisions on E and F , respectively. E ≥0,k F is denoted by E ≥k F ,
and when this holds, E is more informative than F for k-decision problems. For
notational convienience, we define δ∞ := δ, and e-deficiency with respect to
∞-decision problems as e-deficiency.
Finally, we define

∆ (E, F ) := max {δ (E, F ) , δ (F , E)} ,

∆k (E, F ) := max {δk (E, F ) , δk (F , E)} ,
∆∞ := ∆.

When ∆ (E, F ) = 0 (, ∆k (E, F ) = 0, resp.) we say E and F are equivallent

(,equivalent for k-decision problems, resp.), and represent the situation by the
simbol E ∼ F (,E ∼k F , resp.).
Below, PΘ is the set of all probability measures over Θ whose support is
a finite set. One can prove the following necessary and sufficient condition for
E ≥e,k F [14][17]:

(i) For any loss function l with 0 ≤ lθ (t) ≤ 1, and any k-decision D on the
experiment F , there is a k-decision D′ on the experiment E such that,
for any π ∈ PΘ ,
Z Z
D′ (lθ , Pθ ) dπ ≤ {D (lθ , Qθ ) + eθ } dπ.
Θ Θ

(ii) For any loss funcetion l, and any k-decision D on the experiment F , there
is a k-decision D′ on E such that

kD′ (lθ , Pθ ) − D (lθ , Qθ )k ≤ eθ , ∀θ ∈ Θ.

Also, E ≥e F is equivalent to :

(iii) (Randomization criterion) There is an affine positive map Λ such that

kΛ (Pθ ) − Qθ k1 ≤ eθ , ∀θ ∈ Θ.

3 Notations, definitions, and basic facts

H and K are separable Hilbert spaces. B (H), B1 (H), and B0 (H) is the space of
bounded operators, trace class operators, and compact operators over Hilbert

3
 ∗
space H, respcetively. Let B0 (H) be the set of all the bounded linear functional
on B0 (H) relative to the topology induced by operator norm k·k,
kX |ψik
kXk := sup .
ψ∈H k|ψik

∗
Then, B0 (H) ≃ B1 (H) (Proposition 2.6.13 of [4]), by the duality
X ∈ B1 (H) , Y ∈ B0 (H) → tr XY.
Unless otherwise mentioned, B1 (H) ≃ B0 (H)∗ is topologized with weak* topol-
ogy.
A map Λ∗ : B (K) → B (H) is said to be completely positive if and only if
Λ ⊗ In is positive for any n, where In is the identity map form B (Cn ) to B (Cn ).
This is equivalent to
Xn
hψi | Λ (Xi∗ Xj ) |ψj i ≥ 0, (2)
i.j=1
n n
for any {|ϕi i}i=1( Xi ∈ B (K)), any {|ψi i}i=1 (|ψi i ∈ H) for any n. We
consider a dual of Λ∗ , which is Λ : B1 (H) → B1 (K). Λ is said to be trance
preserving if
tr Λ (X) = tr X, ∀X ∈ B1 (H) . (3)
We put
Ch (H, K) := {Λ; Λ linear, (2), (3)} ,
and
f (H, K) := {Λ; Λ linear, (2), tr Λ (X) ≤ tr X∀X ∈ B1 (H)} ,
Ch
f (H, K) saitisfies
Then, Λ ∈ Ch
kΛ (X)k1 ≤ 2 kXk1 . (4)
f (H, K) can be viewed as a subset of (B1 (K))B1 (H) . In this
Ch (H, K) and Ch
view,
Y
Ch (H, K) ⊂ P (H, K) := {Y ; Y ∈ B1 (K) , kY k1 ≤ 2 kXk1 } . (5)
X∈B1 (H)

Observe that {Y ; Y ∈ B1 (K) , kY k1 ≤ 2 kXk1 } is weak*-comapct by Alaoglu’s

theorem. Therefore, by Tychonoff’s theorem (Theorem 16 in Appendix A.2),
P (H, K) is compact relative to the product topology. From here, unless oth-
B (H)
erwise mentioned, (B1 (K)) 1 and Ch (H, K) are topologized by the product
topology, where B1 (K) is topologized by weak* topology.

Lemma 1 The set Ch f (H, K), viewed as a subset of B1 (K)B1 (H) , is compact
and convex. In addition, if X ∈ B1 (H) and Y ∈ B0 (K), the linear functional
f (H, K) → tr Λ (X) Y ∈ C
Λ ∈ Ch
is continuous.

4
Proof. Since P (H, K) is compact relative to the product topology, it suffices
f (H, K) is closed. Let {Λα } a net in Ch
to show Ch f (H, K), with Λα → Λ in the
product topology. Then,

Λα (a1 X1 + a2 X2 ) − {a1 Λα (X1 ) + a2 Λα (X2 )} = 0

→ Λ (a1 X1 + a2 X2 ) − {a1 Λ (X1 ) + a2 Λ (X2 )} = 0,

and n n
X X
hψi | Λ∗α (Xi∗ Xj ) |ψj i ≥ 0 → hψi | Λ (Xi∗ Xj ) |ψj i ≥ 0,
i.j=1 i.j=1

where the convergence is in terms of weak* topology. Also, if X ≥ 0 and {|ii}

is a CONS of K,
n
X n
X
hi| Λα (X) |ii ≤ tr Λα (X) ≤ tr X → hi| Λ (X) |ii ≤ tr X.
i=1 i=1

f (H, K),
Since this holds for any n, tr Λ (X) ≤ tr X. Therefore, Λ is also in Ch
f
implying Ch (H, K) is closed and compact.
f (H, K), with Λα → Λ
To prove the second assertion, let {Λα } a net in Ch
in the product topology. Then, Λα (X) → Λ (X) in weak* topology, for any
X ∈ B1 (H). Therefore, tr Λα (X) Y → tr Λ (X) Y , for any Y ∈ B0 (H). Hence,
we have the assertion.
The set of POVM’s over a measurable space (D, D) in the Hilbert space
H is denoted by M es (D, D; H). The space of singed measures and singed
finitely additive measures over (D, D) is denoted by ca (D, D) and ba (D, D),
respectively. They are metrized by the total variation norm, which is denoted
by k·k1 . The space of bounded measurable function is denoted by Lb (D, D) .
∗
Then, ba (D, D) ≃ Lb (D, D) . We topologize ca (D, D) and ba (D, D) with
weak* topology.
The map
fM : X ∈ B1 (H) → tr XM (·) ∈ ca (D, D) (6)
is linear, positive

fM (ρ) (B) ≥ 0, ∀ρ ∈ B1 (H) , ρ ≥ 0, ∀B ∈ D,

and fM (X) (D) = tr X. Conversely, any linear, bounded, and positive map
from B1 (H) to ca (D, D) with f (X) (D) = tr X is in this form.
The space of elements f of B (B1 (H) , ba (D, D)) with positivity and

f (X) (D) = tr X, ∀X ∈ B1 (H) , (7)

is denoted by M es (D, D; H). Also, replacing the condition (7) by

f (X) (D) = tr X, ∀X ∈ B1 (H) ,

5
 ]
we deifne M ]
es (D, D; H). M es (D, D; H) and M es (D, D; H) can be viewed as
B1 (H)
a subset of ba (D, D) . In this view,
]
M es (D, D; H)
Y
⊂ P C (D, D; H) := {µ ; µ ∈ ba (D, D) , kµk1 ≤ 2 kXk1 } .
X∈B1 (H)

By Alaoglu’s theorem, {µ ; µ ∈ ba (D, D) , kµk1 ≤ 2 kXk1 } is weak*-compact.

Therefore, by Tychonoff’s theorem (Theorem 16), P C (D, D; H) is compact rela-
B (H)
tive to the product topology. From here, unless otherwise mentioned, ba (D, D) 1
is topologized with the product topology, where ba (D, D) is topologized with
weak* topology. The proof of the following lemma is almost parallel with the
one of Lemma 1, thus is omitted.
Lemma 2 Suppose D is locally compact. The set M g es (D, D; H), viewed as a
B1 (H)
subset of ba (D, D) , is compact and convex. In addition, if X ∈ B1 (H)
and l ∈ Lb (D, D), the linear functional
Z
g
f ∈ M es (D, D; H) → l (t) f (X) (d t) ∈ R
D
is continuous.

4 Quantum Theory: framework

A quantum experiment E = (H, {ρθ ; θ ∈ Θ}) consists of Hilbet space H and the
family {ρθ ; θ ∈ Θ} of regular states over H. ( More generally, though we do not
use such setting in this paper, E = (H, {ωθ ; θ ∈ Θ}), where H is a C ∗ -algebra
and ωθ is a state over H.) Also, Θ is an arbitrary set.
A quantum decision space HD is a Hilbert space, and quantum decision rule
D is a member of Ch (H, HD ). Loss function L is a non-negative function from
Θ × B1 (HD ) to R, such that

Lθ (X) ≥ 0 (X ≥ 0)
and
 
|Lθ (X) − Lθ (Y )|
sup ; X, Y ≥ 0, tr X = tr Y ≤ 1, θ ∈ Θ ≤ 1. (8)
kX − Y k1
For example:
1. Lθ (Xθ ) := kXθ − ρ0,θ k1 , with θ → ρ0,θ ∈ B1 (HD ) being continuous in
k·k1 .
2. Suppose HD < ∞ . Then, q for a continuous (in trace norm) function
1/2 1/2
θ → ρ0,θ and Lθ (X) := 1 − tr ρ0,θ Xρ0,θ satisfies (8), due to
q
1/2 1/2
1 − tr ρ0,θ Xρ0,θ ≤ kρ0,θ − Xk1 .

6
 3. Lθ (X) = tr Lθ X, where Lθ ∈ B0 (HD ) and L := {Lθ }θ∈Θ . If

sup kLθ k ≤ 1, (9)

θ∈Θ

then (8) is true.

Also, one may consider probabilistic quantum decision. If |D| < ∞, we can
consider such decision as a CPTP map D : B1 (H) → D × B1 (HD ) :
|D|
M
D:X→ Xt ,
t=1

and a proper loss function would be

1X 1
Lθ (D (Xθ )) = tr Xθ,t Xθ,t − ρ0,θ,t .
2 tr Xθ,t 1
t∈D

It is easy to see, by triangle inequality,

1X

|Lθ (D (Xθ )) − Lθ (D′ (Xθ ))| ≤ ′
Xθ,t − Xθ,t − tr Xθ,t − tr Xθ,t′
ρ0,θ,t 1
2
t∈D
1 X 
≤ pθ,t Xθ,t − p′θ,t Xθ,t
′
1
+ tr X θ,t − tr X ′
θ,t
2
t∈D
≤ kD (Xθ ) − D′ (Xθ )k1 .

Hence, this case is also satisfies (8).

A quantum experiment E is said to be q-e-deficient relative to F = (K, {σθ ; θ ∈ Θ})
for k-decision problems (denoted by E ≥qe,k F ), if and only if, for any if and only
if, for for HD with dim HD = k, any loss function L with (8), any decision D on
the experiment F ,

inf sup {Lθ (D′ (ρθ )) − Lθ (D (σθ )) − eθ } ≤ 0. (10)

D′ ∈Ch(H,HD ) θ∈Θ

When (10) is true for k = ∞, we say E is q-e-deficient relative to F , and denote

this situation by E ≥qe F . (Thus, E ≥qe,∞ F means E ≥qe F ).
Also, we can consider tasks with classical outputs. Different from purely
quantum setting, decision space is a measurable space (D, D), and loss function
t → lθ (t) is a measurable function taking values in [0, 1]. Also, decision rule is
represented by a POVM M ∈ M es (H, D, D) in H over (D, D).
E is said to be c-e-deficient relative to F (denoted by E ≥ce F ), if and only if,
for any loss function l with 0 ≤ lθ (t) ≤ 1, for any decision M on the experiment
F, Z Z
inf′ lθ (t) tr ρθ M ′ (dt) ≤ lθ (t) tr σθ M (dt) + eθ , ∀θ ∈ Θ. (11)
M D D
c-e-deficiency for k-decision problems is defined by posing the restriction |D| ≤ k
and it is denoted by E ≥ce,k F .

7
 q- and c-deficiency is defined in parallel with deficiency and denoted by
δ q (E, F ) and δ c (E, F ), respectively. Their k-decision versions δkq (E, F ) and
δkc (E, F ) are also defined analogously
.

5 Quantum randomization criterion

Theorem 3 (Fan’s minimax theorem, [3] ) Suppose that X be a compact convex
subset of vector space, and Y be a convex subset of a vector space. Assume
that f : X × Y → R satisfies following conditions: (1) x → f (x, y) is lower
semicontinuous and convex on X for every y ∈ Y: (2) y → f (x, y) is concave
on Y for every x ∈ X . Then

min sup f (x, y) = sup min f (x, y) .

x∈X y∈Y y∈Y x∈X

Theorem 4 E is q-e-deficient relative to F for k-decision problems if and only

if each of the following four holds (below, dim HD = k);

(i) For any finite subset Θ0 ⊂ Θ , for any family {Lθ }θ∈Θ0 with (8), and for
any D ∈ Ch (K, HD ) there exists a D′ ∈ Ch (H, HD ),

inf sup {Lθ (D′ (ωθ )) − Lθ (D (σθ )) − eθ } ≤ 0.

D′ ∈Ch(H,HD ) θ∈Θ0

(ii) For any finite subset Θ0 ⊂ Θ , for any L = {Lθ ; Lθ ∈ B0 } with (9), any
decision D ∈ Ch (K, HD ) on the experiment F ,

inf sup {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } ≤ 0.

D′ ∈Ch(H,HD ) θ∈Θ0

(iii) For any L = {Lθ ; Lθ ∈ B0 } with (9), any k-decision D on the experiment
F ,and any π ∈ PΘ ,
Z Z
∃Dπ′ ∈ Ch (H, HD ) , tr Lθ Dπ′ (ρθ ) dπ ≤ {tr Lθ D (σθ ) + eθ } dπ.
Θ Θ

(iv) For any D on the experiment F ,

∃D0′ ∈ Ch (H, HD ) , sup {kD0′ (ρθ ) − D (σθ )k1 − eθ } ≤ 0,

θ∈Θ

Proof. Obfiously, (10)⇒ (i)⇒(ii)⇒ (iii), and (v)⇒(10). Hence, we show

(iii)⇒(iv). R
Observe π ∈ PΘ has only finite support, and dπ is nothing but sum over
a finite subset of Θ. Thus by Lemma 1, the map
Z
′
D → {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } dπ
Θ

8
 f (H, HD ) is compact due to Lemma 1, and obviously
is continuous. Also, Ch
convex. Therefore, by Theorem 3,
Z
sup inf {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } dπ
L: Lθ ∈B0 (K),−21≤Lθ ≤0 D ∈Ch(H,HD )
′
Θ
Z
= sup inf {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } dπ
f
L: L′θ ∈B0 (K),−21≤Lθ ≤0 D′ ∈Ch(H,H D) Θ
Z
= min sup {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } dπ
f
D′ ∈Ch(H,H ′ ′
D ) L : Lθ ∈B0 (K),−21≤Lθ ≤0
′ Θ
Z

= min 2tr [D′ (ρθ ) − D (σθ )]− − eθ dπ ,
f
D′ ∈Ch(H,H D) Θ

where [X]+ ≥ 0 and [X]− ≥ 0 denotes the positive and negative the positive
part of the Hermitian operator X, or positive operators with X = [X]+ − [X]− .

Z
sup inf {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } dπ
L: Lθ ∈B0 (K),kLθ k≤1 D ∈Ch(H,HD )
′
Θ
Z
≥ sup inf {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } dπ
f
L: Lθ ∈B0 (K),kLθ k≤1 D′ ∈Ch(H,H D) Θ
Z
= min sup {tr Lθ D′ (ρθ ) − tr Lθ D (σθ ) − eθ } dπ
f
D′ ∈Ch(H,H D ) L: Lθ ∈B0 (K),kLθ k≤1 Θ
Z
= min sup {||D′ (ρθ ) − D (σθ ) ||1 − eθ } dπ
f
D′ ∈Ch(H,H D ) L: Lθ ∈B0 (K),kLθ k≤1 Θ
Z

= min 2tr [D′ (ρθ ) − D (σθ )]− − eθ dπ ,
f
D′ ∈Ch(H,H D) Θ

The map

D′ → 2tr [D′ (ρθ ) − D (σθ )]− = sup {tr Lθ D (ρθ ) − tr Lθ D (σθ )} ∈ R

Lθ ∈B0 (K),−21≤Lθ ≤0

is lower semicontinuous and convex, being pointwise supremum of bounded lin-

ear functionals. Observe also PΘ is convex. Therefore, using Theorem 3 again,
Z

sup min 2tr [D′ (ρθ ) − D (σθ )]− − eθ dπ
f
π∈PΘ D′ ∈Ch(H,H D) Θ
Z

= min sup 2tr [D′ (ρθ ) − D (σθ )]− − eθ dπ
f
D′ ∈Ch(H,H D) π∈PΘ Θ

= min sup 2tr [D′ (ρθ ) − D (σθ )]− − eθ .
f
D′ ∈Ch(H,H D ) θ∈Θ

f (H, HD ), D′′ ∈ Ch (H, HD ) defined by

For each D′ ∈ Ch

D′′ (ρ) := D (ρ) + (tr ρ − tr D (ρ)) ρ̃,

9
where ρ̃ ≥ 0, tr ρ̃ = 1, always improves D′ in the sense that

2tr [D′′ (ρθ ) − D (σθ )]−

= 2tr [D′ (ρθ ) + (tr ρθ − tr D (ρθ )) ρ̃ − D (σθ )]−
≤ 2tr [D′ (ρθ ) − D (σθ )]− .

Therefore,

min sup 2tr [D′ (ρθ ) − D (σθ )]− − eθ
f
D′ ∈Ch(H,H D) θ∈Θ

= min sup 2tr [D′ (ρθ ) − D (σθ )]− − eθ
D′ ∈Ch(H,HD ) θ∈Θ

= min sup {kD′ (ρθ ) − D (σθ )k1 − eθ } ,

D′ ∈Ch(H,HD ) θ∈Θ

and we have (iii)⇒(iv).

Letting HD = K and D = I, we obtain:

Theorem 5 E ≥qe F is equivalent to the existence of a CPTP map Λ with

kΛ (ρθ ) − σθ k1 ≤ eθ , ∀θ ∈ Θ. (12)

Corollary 6 E is q-0-deficient relative to F , if and only if

∀ (θ1 , θ2 , · · · , θn ) ∈ Θn , F (ρθ1 , ρθ2 , · · · ρθn ) ≤ F (σθ1 , σθ2 , · · · σθn ) , (13)

holds for any finite number n and any n-point functionals F such that F is
monotone increasing by CPTP map and

|F (X1 , X2 , · · · , Xk ) − F (Y1 , Y2 , · · · , Yk )|
≤ f (kX1 − Y1 k1 , kX2 − Y2 k1 , · · · , kXk − Yk k1 ) (14)

holds for any Xj , Yj ∈ S (H) (j = 1, 2, · · · , k), with f being continuous and

f (0, 0, · · · , 0) = 0. (15)

Proof. If E is q-0-deficient relative to F , by Theorem 5, we have (13). On the

other hand, if (13) holds, then, for any loss operator Lθ over K with (9),
Z Z
inf tr Lθ D (ρθ ) dπ ≤ inf tr Lθ D (σθ ) dπ.
D∈Ch(H,K) Θ D∈Ch(K,K) Θ

Therefore, by Theorem 4, we have E ≥q0 F .

10
6 Classical decision space
Lemma 7 ([14], Theorem 41.7)There is a positive linear oparator T : ba (D, D) →
ca (D, D) such that
(i) kT k = 1,
(ii) T (µ) (D) = µ (D), if µ ≥ 0.
(iii) T |ca(D,D) = id|ca(D,D) .

Theorem 8 E ≥ce F if and only if one of the following two holds:

(i) For any decision space (D, D), for any measurable loss function l with −1 ≤
lθ (t) ≤ 1, for any decision M on the experiment F , and for any π ∈ PΘ ,
there is some decision M ′ on the experiment E such that
Z Z Z Z 
′
lθ (t) tr ρθ M (dt) dπ ≤ lθ (t) tr σθ M (dt) + eθ dπ.
Θ t∈D Θ t∈D

(ii) For any decision space (D, D), any decision M on the experiment F , there
is some decision M ′ on the experiment E such that
sup {kfM ′ (ρθ ) − fM (σθ )k1 − eθ } ≤ 0,
θ∈Θ

where fM (ρ) is as of (6).

Proof. (11)⇒(i), (ii)⇒(11) is trivial. Hence, we have to show (i)⇒(ii). Sup-
pose (i) holds true. Extension of the domain of M ′ to M ] es (D, D; H) only
decreases the risk. Thus, using the argument parallel to the proof of (iii)⇒(iv)
of Theorem 4, we have, for any M ∈ M es (D, D;K),
∃f0 ∈ M es (D, D; H) , sup {kf0 (ρθ ) − fM (σθ )k1 − eθ } ≤ 0,
θ∈Θ

where Lemma 2 is used instead of Lemma 1.

Let T be as of Lemma 7,
sup {kT ◦ f0 (ρθ ) − fM (σθ )k1 − eθ }
θ∈Θ
= sup {kT ◦ f0 (ρθ ) − T ◦ fM (σθ )k1 − eθ }
θ∈Θ
≤ sup {kf0 (ρθ ) − fM (σθ )k1 − eθ } ≤ 0.
θ∈Θ

Therefore, f0′ := T ◦ f0 , which is a bounded linear map from B1 (H) to ca (D, D),
satisfies f0′ ≥ 0, and
f0′ (X) (D) = T (f0 (X)) (D) = f0 (X) (D) = tr X , ∀X ∈ B1 (H) .
Thus, there is a POVM M ′ such that fM ′ = f0′ . Thus, and (ii) is proved.
Due to (ii) of Theorem 8, we have:

11
Theorem 9 Suppose E ≥c0 F . Then, we have the following (i)-(iii).

(i) Let lθ (t) an arbitrary classical loss function which is not necessarily bounded.
Then, for any decision M on a decision space (D, D), there is decision
M ′ on (D, D) such that
Z Z
lθ (t) tr ρθ M ′ (dt) ≤ lθ (t) tr σθ M (dt) , ∀θ ∈ Θ.
t∈D t∈D

(ii) Let lθ (t) an arbitrary classical loss function which is not necessarily bounded.
Then for any decision M on an arbitrary decision space (D, D), where
× ⊂ D ⊂ Rm , and Z
t tr σθ M (dt) = θ,
t∈D

there is decision M ′ on (D, D) with

Z
t tr ρθ M ′ (dt) = θ
t∈D

such that
Z Z
lθ (t) tr ρθ M ′ (dt) ≤ lθ (t) tr σθ M (dt) , ∀θ ∈ Θ.
t∈D t∈D

(III) Let D = {0, 1}, D = 2D and Θ0 ∪ Θ1 ⊂ Θ. Then, for any decision M on

(D, D) such that
tr σ0 M ({1}) ≤ α,
there is a decision M ′ on (D, D) with

tr ρ0 M ′ ({1}) ≤ α

such that
tr ρ1 M ′ ({0}) ≤ tr σ1 M ({0}) .

Using almost parallel argument as the proof of Theorem 8, we have:

Theorem 10 E ≥ce,k F if and only if one of the following two holds:

(i) With |D| = k , for any measurable loss function l with −1 ≤ lθ (t) ≤ 1, for
any k−decision M on the experiment F , and for any π ∈ PΘ , there is
some k-decision M ′ on the experiment E such that
Z X Z (X )
′
lθ (t) tr ρθ M (t) dπ ≤ lθ (t) tr σθ M (t) + eθ dπ.
Θ t∈D Θ t∈D

12
(ii) With |D| = k , any k-decision M on the experiment F , there is some
k-decision M ′ on the experiment E such that
sup {kfM ′ (ρθ ) − fM (σθ )k1 − eθ } ≤ 0, .
θ∈Θ

In case k = 2, c-e-deficiency has more explicite expression. Since

X
lθ (t) tr ρθ M ′ (t) = {(lθ (0) − lθ (1)) tr ρθ M ′ (0) + lθ (1)} ,
t∈{0,1}

we have, letting π ∈ PΘ ,
Z (X )
′
lθ (t) (tr ρθ M (t) − tr σθ M (t)) − eθ dπ
Θ t∈D
Z
= {(lθ (0) − lθ (1)) (tr ρθ M ′ (0) − tr σθ M (0)) − eθ } dπ
Θ

Therefore, letting aθ := 21 (lθ (0) − lθ (1)),

Z (X )
′
inf′ lθ (t) (tr ρθ M (t) − tr σθ M (t)) − eθ dπ
M Θ t∈D
Z  Z  Z
= inf′ tr 2aθ ρθ dπ M ′ (0) − tr 2aθ σθ dπ M (0) − eθ dπ
M Θ
Z Z Z Θ  Z
θ

=− aθ ρθ dπ + aθ dπ − tr 2aθ σθ dπ M (0) − eθ dπ ≤ 0
Θ 1 Θ Θ θ

Since this holds for any M , we have

Z Z Z
− aθ ρθ dπ + aθ σθ dπ − eθ dπ ≤ 0,
Θ 1 Θ 1 θ
or Z Z Z
aθ ρθ dπ ≥ aθ σθ dπ − eθ dπ, (16)
Θ 1 Θ 1 θ
where θ → aθ is an arbitrary function with |aθ | < 1. Especially when Θ =
{0, 1}, this is equivalent to
kρ0 − sρ1 k1 ≥ kσ0 − sσ1 k1 − e0 − se1 , ∀s ≥ 0. (17)
In case dim H = dim K = 2, it is known that
kρ0 − sρ1 k1 ≥ kσ0 − sσ1 k1 , ∀s ≥ 0, (18)
is necessary and sufficient for E ≥q0 F [2]. In other words, E ≥q0 F is equivalent to
E ≥c0,2 F . However, in case that dim H = dim K = 3, (18) fails to be sufficinet
for E ≥q0 F [6].
In classical case, more strongly, (17), or E ≥e,2 F , is known to be equivalent
to E ≥e F [15][17]. The following theorem is found independently by [11].

13
Theorem 11 Suppose Θ = {0, 1}, and [ρ0 , ρ1 ] = 0. Then, E ≥ce F if and only
if (17) holds, or E ≥ce,2 F .

Proof. Let F M be a classical experiment consisted with QM θ respecitively,

c
where QM
θ (dx) = tr σθ M (dx). Then, by Theorem 8, E ≥e F if and only if

E ≥e F M , ∀M.

As noted above, this equivalent to [15]

kρ0 − sρ1 k1 ≥ QM M
0 − sQ1 1
− e0 − se1 , ∀M, ∀s ≥ 0.

Therefore, since
max QM M
0 − sQ1 1
= kσ0 − sσ1 k1 ,
M

we have the assertion.

[5] introduced the notion of statistical morphism, which we use here with
some non-essential modifications. A map Γ from {ρθ }θ∈Θ ⊂ B1 (H) into B1 (K)
is said to be k-statistical morphism if and only if for k-decision M over H, there
exists a k-decision M ′ over K with

tr Γ (ρθ ) M (t) = tr ρθ M ′ (t) , ∀θ ∈ Θ. (19)

E ≥c0,k F is equivalent to the existence of k-statistical morphism Γ on {ρθ }θ∈Θ

with Γ (ρθ ) = σθ .
Obviously, any positive linear, and trace preserving map Γ with Γ (ρθ ) = σθ
, ∀θ ∈ Θ, is k-statistical morphism, for any k. The following lemma has some
implications on its converse statement.

Lemma 12 Suppose dim H < ∞. Then, any k-statistical morphism Γ on

{ρθ }θ∈Θ can be extended to a linear, trace preserving, and positive map Γ′ to
span {ρθ }θ∈Θ .
n
Proof. Let {ρθi }i=1 be linear independent elements of {ρθ }θ∈Θ and define Γ′
by linear combination of {Γ (ρθi )}i :
n
! n
X X
′
Γ a i ρθ i = ai Γ (ρθi ) .
i=1 i=1

Obviously, Γ is linear and trace preserving. First, we prove Γ (ρθ ) = Γ′ (ρθ );

′
′
By
Pn definiton, for any M and for any ε > 0, there is M with (19). Let ρθ =
i=1 ai ρθi . Then,

tr Γ (ρθ ) M (t) = tr ρθ M ′ (t)

Xn
= ai tr ρθi M ′ (t)
i=1
= tr Γ′ (ρθ ) M (t) .

14
Since M is arbitrary k-valued measurement, we have Γ (ρθ ) = Γ′ (ρθ ), and Γ′
is a linear extention of Γ.
Finally, we prove that Γ′ is positive on span {ρθ }θ∈Θ . For any positive matrix
P
M ≤ 1 and any ρ = i ai ρθi ≥ 0,
n
X n
X
tr Γ′ (ρ) M = tr ai Γ (ρθi ) M ≥ tr ai ρθi M ′ − nε ≥ −nε.
i=1 i=1

Since ε > 0 and M ≥ 0 are arbitrary, we have positivity of Γ′ on span {ρθ }θ∈Θ .

Theorem 13 Suppose dim H < ∞ and span {ρθ }θ∈Θ is the totality of Her-
mitian matrices. Then, E ≥c0,k F holds if and only if there is a positive trace
preserving map Γ with Γ (ρθ ) = σθ , ∀θ ∈ Θ . Namely, E ≥c0,2 F , E ≥c0,3 F , · · · ,
E ≥c0,k F are all euqivalent to E ≥c0 F .

Proof. The first statement follows directly from Lemma 12. As for the second
statement, it is obvious that E ≥c0 F implies E ≥c0,2 F , E ≥c0,3 F , · · · , E ≥c0,k F .
Conversely, suppose E ≥c0,2 F . Then by Lemma 12, there is a positive linear,
and trace preserving map Γ with Γ (ρθ ) = σθ , ∀θ ∈ Θ, which implies E ≥c0 F .
Classically, it is known that E ≥0,2 F , E ≥0,3 F , · · · , E ≥0,k F are all equiv-
alent to E ≥0 F , provided Θ is a finite set [15][17]. The above theorem is a
quantum version of this statement.

7 Compact covariant experiments

Let dim H < ∞, dim K < ∞. Let G be a compact group, and g → Ug ∈ SU (H)
and g → Vg ∈ SU (K) be representations of G. Suppose that there is a natural
action θ → gθ of g ∈ G on θ ∈ Θ. Moreover, we suppose that for any θ, there is
g ∈ G with g0 = θ. Then we consider the covariant experiments, which satisfy

ρgθ = Ug ρθ Ug† , σgθ = Vg σθ Vg† ,

or
ρg0 = Ug ρ0 Ug† , σg0 = Vg σ0 Vg† .
We further suppose that the assumptions of 5, which are conditions (A’) and
(B), hold true. Then, Due to Theorem 5, we have

δ q (E, F ) = inf sup kΦ (ρθ ) − σθ k1

Φ θ∈Θ


= inf sup Φ Ug ρ0 Ug† − Vg σ0 Vg† 1
Φ g∈G


= inf sup Vg† Φ Ug ρ0 Ug Vg − σ0 †
1
.
Φ g∈G

15
 Denote by M the average with respect to Haar measure of G, and define


Φ∗ (ρ) := MVg† Φ Ug ρ0 Ug† Vg .

Then, Φ∗ is covariant,


Φ∗ Ug ρUg† = Vg Φ∗ (ρ) Vg† , (20)

and, by convexity of the norm k·k1 ,

sup Vg† Φ Ug ρ0 Ug† Vg − σ0 1
g∈G

≥ kΦ∗ (ρ0 ) − σ0 k1


= Vg† Φ∗ Ug ρ0 Ug† Vg − σ0 1
, ∀g ∈ G.

Therefore,
δ q (E, F ) = inf kΦ∗ (ρ0 ) − σ0 k1 ,
Φ∗

where Φ∗ runs over all the CPTP maps with (20).

Let CΦ∗ be the Choi’s representation of a channel Φ∗ ,
 
dim
XH
CΦ∗ := Φ∗ ⊗ I  |ii |ii hj| hj| ,
i,j=1

where {|ii} is a CONS of H. Then, (20) can be written as

 
Ug ⊗ Vg , CΦ∗ = 0 , (g ∈ G) , (21)

Example 14 H = K =Cd , G = SU (d), Ug = g, and Vg = V gV † Then, Φ∗ has

to be depolarizaing channel,

(1 − λ) (tr X)
Φ∗ (X) := 1 + λV † XV, (0 ≤ λ ≤ 1) .
d
Hence,
λ
E ≥q0 F ⇔ σ0 = 1 + (1 − λ) V † ρ0 V .
d
Especially,
 suppose ρ0 and σ0 have
 the same spectrum.q Then, although the set
U ρ0 U † U∈SU(d) equals the set U σ0 U † U∈SU(d) , E 0 F unless V † ρ0 V = σ0 .
Now, let H = K = C2 , and
 
1 1+u 0
V † ρ0 V = (u ≥ 0) ,
2 0 1−u
 √ 
1 1√+z x − −1y
σ0 = .
2 x + −1y 1−z

16
Then,
δ q (E, F ) = inf kΦ∗ (ρ0 ) − σ0 k1
Φ∗
1p
= inf (z − λu) + x2 + y 2
λ:0≤λ≤1 2
 q
 2
 12 (z − u) + x2 + y 2 , (z ≥ u) ,
p
= 1
x2 + y2, (0 ≤ z ≤ u) ,

 p2
1 2 2 2
2 z +x +y , (z ≤ 0) .
When z ≤ 0, the optimal Φ∗ (ρ0 ) = 12 1. Thus, best approximate experiment

E to F with E ≥q0 E ′ consists of 21 1 only. Put differently, E ′ = C2 , {ρ′θ ; θ ∈ Θ} ,
′

where ρ′θ = 12 1 for all θ ∈ Θ.

Example 15 H = K =Cd , d is prime power, and G = {Xds Zdt }s,t∈{0,1,··· ,d−1} ,
where
 
0 ··· ··· 0 1
 .. .. 
 1 0 . . 0 
 
 .. .. ..  ,
Xd =  0 1 . . . 
 
 . . . . 
 .. .. . . 0 .. 
0 ··· 0 1 0
 
1 0 ··· ··· 0
 √
.. .. 
 0 exp −12π/d 0 . . 
 
 .. √
. .. .
.. 
Zd =  . 0 exp −14π/d .
 
 . .. .. .. 
 .. . . . 0 
√

0 ··· ··· 0 exp −12π (d − 1) /d
Also, Ug = Vg = g. Note that
X ′ ′
CΦ∗ = at,s,t′ s′ Xdt Zds ⊗ Xdt Zds ,
t,s,t′ ,s′ ∈{0,1,··· ,d−1}

where at,s,t′ ,s′ are complex numbers. Since

 ′′ ′′ ′′ ′′
 ′ ′

Xdt Zd−s ⊗ Xdt Zds Xdt Zds ⊗ Xdt Zds
s′′ (t′ −t)−t′′ (s+s′ )
 ′ ′
  ′′ ′′ ′′ ′′

= ωd Xdt Zds ⊗ Xdt Zds Xdt Zd−s ⊗ Xdt Zds ,

(21) implies that at,s,t′ s′ takes non-zero value for t,s, t′ ,s′ with t′ = t and
s′ = d − s. Therefore, considering that ChΦ∗ is Hermitian, and that Φ∗ is trace
preserving, the space of channels satisfying (20) is (as a real vector space) d2 − 1
dimensional. On the other hand, a channel
X

†
Φ∗ (ρ) = pt,s Xdt Zds ρ Xdt Zds (22)
t,s∈{0,1,··· ,d−1}

17
satisfies (20), and the space of channels with (22) is d2 − 1. Hence, (20) is
equivalent to (22).
Therefore,
δ q (E, F ) = min kρ′ − σ0 k1
′ ρ

where ρ moves all over the convex hull of the set {(Xdt Zds ) ρ0 (Xdt Zds ) ; t, s ∈ {0, 1, · · · , d − 1}}.
′

Especially, when d = 2, letting ~x0 = (x01 , x02 , x03 ) and ~y0 be the Bloch
representation of ρ0 and σ0 , respectively, we have

δ q (E, F ) = min k~x − ~y0 k ,

~
x

where and ~x moves all over the convex hull of (x01 , x02 , x03 ), (−x01 , −x02 , x03 ),
(x01 , −x02 , −x03 ), and (−x01 , x02 , −x03 ).

8 Translation experiments
8.1 Models and questions
Let dim H = dim K = ∞ (countable), and define
† †
ρθ := WAθ ρWAθ , σθ := WBθ σWBθ ,

where
√
−1(θ 1 P −θ 2 Q)
Wθ := e . θ ∈ R2 ,
is a Weyl operator, and A and B are real invertible 2 × 2 matrices. Appling
Theorem 5, we have

δ q (E, F ) = inf sup kΦ (ρθ ) − σθ k1 .

Φ θ∈Θ

8.2 Restriction to covariant maps

The argument of this section draws upon [12]. For any Φ, define
 
† †
Φθ (X) := WBθ Φ WAθ XWAθ WBθ .

Then,

δ q (E, F ) = inf sup kΦθ (ρ) − σk1

Φ θ∈Θ

= inf sup sup tr (Φθ (ρ) − σ) X

Φ θ∈Θ kXk≤1

= inf sup sup tr (Φθ (ρ) − σ) X

Φ kXk≤1 θ∈Θ

≥ inf sup Mθ tr (Φθ (ρ) − σ) X,

Φ kXk≤1

18
where Mθ is the invariant mean of the translation group in R2 . Note, if ρ is a
density operator, the map

Φ∗ (ρ) : X → Mθ tr Φθ (ρ) X

is linear and bounded, and maps 1 to 1. Also, the mapping Φ∗ : ρ → Φ∗ (ρ) is

linear, and covariant :
  h i
† †
Φ∗ WAθ ρWAθ [X] = Φ∗ (ρ) WBθ XWBθ . (23)

Thus,

sup sup (Φ∗ (ρθ ) [X] − tr σθ X)

θ∈Θ kXk≤1
 h i 
† †
= sup sup Φ∗ (ρ) WBθ XWBθ − tr σWBθ XWBθ
θ∈Θ kXk≤1

= sup (Φ∗ (ρ) [X] − tr σX) .

kXk≤1

Hence, in optimizing Φ, we just have to consider Φ∗ with covariant property

(23).
Φ∗ is seemingly difficult to handle, since its output state may not be normal,
i.e., may not have the density. However, it turns out that Φ∗ with non-normal
output is not optimal.
Since B1 (H) is the dual of the space of compact operators B0 (H), there is
a positive Yρ ∈ B1 (H) with

Φ∗ (ρ) [X] = tr Yρ X, ∀X ∈ B0 (H) .

Consider the map

ρ→ sup Φ∗ (ρ) [P ] = tr Yρ .
P :finite rank projector

Since this is linear in ρ, positive and bounded, there is a positive bounded

operator T with
tr Yρ = tr T ρ.
Due to covariant property of Φ∗ (23), we have
†
tr T (WAθ ρWAθ ) = tr T ρ

for any ρ. Therefore, T commutes WAθ for all θ ∈ R2 . Therefore, T = c1.

Thus, c = tr Yρ is independent of the input ρ. Therefore, ρ∗ := 1c Yρ is a density
operator. We denote by Φ′∗ the CPTP map which sends ρ to ρ∗ .

19
 Letting {Xn } be a sequence of compact operators such that limn→∞ tr (cρ∗ − σ) Xn =
tr [cρ∗ − σ]− (0 ≤ c ≤ 1),

sup (Φ∗ (ρ) [X] − tr σX)

kXk≤1

=2 sup (Φ∗ (ρ) [X] − tr σX)

X≤0,kXk≤1

≥ 2 lim tr (cρ∗ − σ) Xn
n→∞
= 2tr [cρ∗ − σ]−
≥ 2tr [ρ∗ − σ]− = kΦ′∗ (ρ) − σk1 . (24)

Therefore, Φ′∗ is at least as good as Φ∗ .

After all, we have

inf sup kΦ (ρθ ) − σθ k1 = inf kΦ (ρ) − σk1 ,

Φ θ∈Θ Φ

where Φ runs over all the CPTP maps with

 
† †
Φ WAθ ρWAθ = WBθ Φ (ρ) WBθ , (25)

or  
† †
Φ∗ WBθ XWBθ = WAθ Φ∗ (X) WAθ . (26)

8.3 Characterization of covariant maps

Inserting X = Wξ to (26), one has
√
−1ξ T JBθ
e− WAθ Φ∗ (Wξ ) = Φ∗ (Wξ ) WAθ ,

where  
0 1
J= .
−1 0
Since this holds for any θ and ξ, we have

Φ∗ (Wξ ) = c (ξ) WCξ ,

where C satisfies
C T JA = JB.
Using the identity
AT JA = (det A) J,
or
JA = (det A) AT −1 J,
we have
det B
C= AB −1 .
det A

20
 Suppose det A = det B, then

det C = 1.

Hence, according to Lemma 18, for Φ∗ to be identity preserving and completely

positive, c (ξ) has to be a characteristic function of a classical probability dis-
tribution F over R2 ,
Z √
2 1 dF ( x)
c (ξ) = e −1(ξ x −ξ x )
1 2
.
2π
Letting Pρ denote the P -function of ρ, we have
Z Z √
dz dz
Pρ (z) e −1(ξ z −ξ z ) Wz Wξ |0i hz|
1 2 2 1
tr ρWξ = tr Pρ (z) Wξ Wz |0i hz| = tr
2π 2π
Z √ dz
= h0| ξi Pρ (z) e −1(ξ z −ξ z )
1 2 2 1
.
2π
and thus,
Z √ dz
−1(ξ 1 z 2 −ξ 2 z 1 )
h0| ξi PΦ(ρ) (z) e
2π
= tr Φ (ρ) Wξ = tr ρΦ∗ (Wξ ) = c (ξ) tr ρWCξ
Z √ dz
= c (ξ) h0| ξi Pρ (z) e −1((Cξ) z −(Cξ) z )
1 2 2 1

2π
Z √ dz
= c (ξ) h0| ξi Pρ (z) e −1 det C (ξ z −ξ z )
1 2 2 1
.
2π
Therefore,
Z √
Z √
dz dz
PΦ(ρ) (z) e −1(ξ z −ξ z ) = c (ξ) Pρ (z) e −1(ξ z −ξ z )
1 2 2 1 1 2 2 1
.
2π 2π
Taking inverse Fourer transform of both sides, we have
Z
PΦ(ρ) (x) = Pρ (x − y) dF ( y) .

Therefore, (25) is equivalent to

Z Z
dx
Φ (ρ) = Pρ (x − y) |xi hx| dF ( y)
2π
Z Z
dx
= Pρ (x) |x + yi hx + y| dF ( y)
2π
Z Z 
dx
= Wy Pρ (x) |xi hx| Wy† dF ( y)
2π
Z
= Wy ρ Wy† dF ( y) , (27)

21
which is analogous to its classical version [16].
On the other hand, by Lemma 18, if A 6= B, c (ξ) has to be a non-commutative
characteristic function
c (ξ) = ω (WΩξ ) ,
where ω is a state. Ω is an operator satisfying

J − C T JC = ΩT JΩ,

or  1/2
det B
Ω = (1 − det C)1/2 S = 1− S,
det A
with
det S = 1.
Hence,
tr Φ (ρ) Wξ = tr ρΦ∗ (Wξ ) = ω (WΩξ ) tr ρWCξ . (28)

8.4 Gaussian shift models

When ρ is gaussian state with mean value zero, ρ satisfies
T
tr ρWξ = e−ξ Σρ ξ/4
, (29)

where  
1
1 tr ρQ2 2 tr ρ (P Q + QP )
Σρ = 1 .
2 2 tr ρ (P Q + QP ) tr ρP 2
Suppose det A = det B. Then, due to (27),

E ≥q0 F ⇔ Σρ ≤ Σσ .

In classical case, it had been shown that the same condition on the variances is
necessary and sufficient for E ≥0 F [16][17].
Suppose det A 6= det B. Then, if E ≥q0 F , due to (28), ω (WΩξ ) is also Gaus-
sian,
T T
ω (WΩξ ) = e−ξ Ω Σω Ωξ/4 .
Also, by (28), we have

ΣΦ(ρ) = ΩT Σω Ω + C T Σρ C,

or, with A′ = AB −1
1  
−2
S T Σω S = −1 ΣΦ(ρ) − (det A′ ) A′T Σρ A′ . (30)
1− (det A′ )

22
Therefore, by Lemma 19, for ω with (28) to exist, the following is necessary and
sufficient:
√
Σω + −1J ≥ 0
√
√
⇔ S T Σω + −1J S = S T Σω S + −1J ≥ 0
⇔ tr S T Σω S ≥ 0, det S T Σω S ≥ 1.
By (30), these are equivalent to
 
−2 −2
tr ΣΦ(ρ) − (det A′ ) A′T Σρ A′ = trΣΦ(ρ) − (det A′ ) tr A′ A′T Σρ ≥ 0, (31)
  h i
−1 −2 −2
1 − (det A′ ) det ΣΦ(ρ) − (det A′ ) A′T Σρ A′ ≥ 1.
(32)
Further, we suppose ΣΦ(ρ) = Σρ = a2 1. Then, these conditions can be
written as
2
2 (det A′ ) ≥ tr A′T A′
h i  
−2 −1 2
det 1 − (det A′ ) A′T A′ ≥ a−4 1 − (det A′ ) .
Without loss of generality, let
 
′ α 0
A′T A = .
0 β
where α ≥ 0, β ≥ 0. Then, it follows that
α (β − 1) + β (α − 1) ≥ 0, (33)
p 2
(α − 1) (β − 1) ≥ a−4 αβ − 1 . (34)

By (34), α − 1 and β − 1 have to have the same sign. Therefore, by (33),

α ≥ 1, β ≥ 1. (35)
2
In classical case, with ΣΦ(ρ) = Σρ = a 1, E ≥0 F is equivalent to (35)
[10][17]. Even in quantum case, when a ≫ 1, the system is almost classical.
Therefore, it is expected that (35) is sufficient for E ≥q0 F . In fact, this is easily
verified by noticing the right hand side of (34) is almost 0.
However, when a is not very large, quantum case is very much different from
classical case. For example, suppose a = 1. Then, (34) is written as
p
α + β ≤ 2 αβ.
Therefore, we have to have
α = β.
Hence, E ≥q0 F is equivalent to
√
AB −1 = A′ = αO,
where α ≥ 1 and O is an orthogonl matrix. This is very much stronger than
the classical condition (35).

23
References
[1] P. Alberti, ”On the simultaneous transformation of density operators by
means of a completely positive”, Unity Preserving Linear Map. Publ. RIMS
(Kyoto), 21:617–644, 1985.
[2] P. Alberti and A. Uhlmann, ”A problem relating to positive linear maps on
matrix algebras”, Rep. Math. Phys. 18, 163-176 (1980).
[3] J. M. Borwein and D. Zhuang, On Fan’s minimax theorem, Math. Pro-
gramming, Vol. 34, 232–234 (1986).
[4] O. Bratteli, D. W. Robinson, ”Operator Algebras and Quantum Statistical
Mechanics 1”, Springer-Verlag (1979).
[5] F. Buscemi, Comparison of quantum statistical models: a ”Quantum Black-
well theorem”, http://arxiv.org/abs/1004.3794 (2010).
[6] A. Chefles, R. Jozsa, and A. Winter, ”On the existence of physical trans-
formations between sets of quantum states ”arXiv:quant-ph/0307227.
[7] J. B. Conway, ”A Cource in Functional Analysis”, (2nd ed.) Springer-Verlag
(1997).
[8] B. Demoen, P.Vanheuverzwijn and A.Verbeure, ”Completely positive maps
on the CCR-algebra,” Lett. Math. Phys. 2, 161 (1977).
[9] N. Dunford, J. Schwartz, ”Linear Operators, Part I”, Interscience (1952).
[10] O. Hansen, E. Torgersen, ”Comparison of linear normal experiments”, The
Annals of Statistics, Vol. 2, No. 2, 367-373 (1974).
[11] A. Jencova, private communication (2011).
[12] O. Kruger, ”Quantum Information Theory with Gaussian Systems”, doctral
dissertation (2006).
[13] J. L. Kelley, ”General Topology”, Graduate Texts in Mathematics,
Springer-Verlag (1975).
[14] H. Strasser, ”Mathematical Theory of Statistics”, de Gruyter (1985).
[15] E. Torgersen, ”Comparison of statistical experiments when the parame-
ter space is finite”, Z. Wahrscheinlichkeitstheorie verw. Geb. 16, 219-249
(1970).
[16] E. Torgersen, ”Comparison of translation experiments”, The Annals of
Mathematical Statistics, Vol. 43, No. 5 (1972).
[17] E. Torgersen, ”Comparison of Statistical Experiments”, Cambridge Univer-
sity Press (1991).

24
A Backgrounds from analysis
A.1 Weak and weak* topology
For the detail of the following statements, see [7], for example. Let E and E ′
be a normed Banach space and the totality of continuous linear functional on
E, respectively. If we take as the norm of f ∈ E ′ the operator norm kf k as a
functional, then E ′ become a normed linear space called conjugate space. E ′ is
complete, and thus is a Banach space. The topology introduced by kf k is called
strong topology.
The weak* topology σ (E ′ , E) in E ′ is indtroduced as follows. For every
α > 0 and every finite number of elements xi (i = 1, · · · , n), we denote by
W (x1 , · · · xn , α) the set of all f such that |f (xi )| ≤ α. The topology for which
the sets W (x1 , · · · xn , α) form the fundamental system of the neighbours of zero
is called weak* topology. In other words, an open set containig 0 is a union of
the sets in the form of W (x1 , · · · xn , α). The weak topology σ (E, E ′ ) in E is
defined by exchanging the role of E and E ′ above.
The weak and weak* topologies are locally convex topologies since sets
W (x1 , · · · xn , α) are convex.
n
The sequence {fi }i=1 in E ′ is called weakly convergent to the functional f0
∞
if it converges to f0 in the weak* topology. In order for {fi }i=1 to be weakly
convergent to f0 , it is necessary and sufficient that limn→∞ fn (x) = f0 (x) for
every x ∈ E.
A convex set in a normed linear space E ′ has the same closure both in
the initial topology and in the weak* topology σ (E ′ , E). In particular, if the
n
sequence {fi }i=1Pis weakly convergent to f0 , there exists a sequence of linear
m
combinatitons { i=1 λm i fi } converging in the norm to f0 .
Every closed sphere in E ′ is compact in the weak* topology σ (E ′ , E) (Alaoglu’s
theorem ).

A.2 Product topology and Tychonoff’s theorem

Q
Given a set Y and topological spaces (Xy , xy ), and we furnish y∈Y Xy with
the product topology, or the coarsest topology which makes projection P :
Q
y∈Y Xy → Xy , P (x) = xy continuous. A local base of this topology is a
family of sets in the form of
\
P −1 (Uy ) ,
y∈F

whereTeach Uy is an open set in Xy and F is a finite subset of Y [13]. (Note

that y∈Y P −1 (Uy ) is not necessarily open.)
Weak* topology σ (E ′ , E) can be viewed as a product topology, where X = R
and Y = E.

Theorem 16 (Tychonoff ’s theorem) The cartesian product of a collection of

compact topological spaces is compact relative to the product topology.

25
 .

B Proof of (4)
Let
X +X X −X
XR := , XI := √ .
2 2 −1
Define XR+ , XR− , XI+ and XI− so that XR+ , XR− , XI+ and XI− are positive
self-adjoint, and satisfy

XR+ − XR− = XR , XI+ − XI− = XI .

Then,

kΛ (X)k1
√
= Λ (XR+ ) − Λ (XR− ) + −1 (Λ (XI+ ) − Λ (XI− )) 1
≤ kΛ (XR+ )k1 + kΛ (XR− )k1 + kΛ (XI+ )k1 + kΛ (XI− )k1
= tr Λ (XR+ ) + tr Λ (XR− ) + tr Λ (XI+ ) + tr Λ (XI− )
= tr XR+ + tr XR− + tr XI+ + tr XI−
= kXR k1 + kXI k1
X +X X −X
= + √ ≤ kXk1 + X 1
2 1 2 −1 1
= 2 kXk1 .

C CP maps on CCR algebra

Let J be a bilinear antisymmetiric form on R2n induced by the matrics [Ji,j ].
(We will not distinguish bilinear forms and their implimenting matrices.) Define
Weyl operators Wξ (ξ ∈ R2n ) which are unitary operators with

W0 = 1,
 √ 
−1 ′
Wξ Wξ′ = exp − J (ξ, ξ ) Wξ+ξ′ .
2
√

= exp − −1J (ξ, ξ ′ ) Wξ′ Wξ .

The algebra generated by Weyl operators is called CCR algebra and denoted
by CCR(J).
Given a state ω, ω (Wξ ) is called characteristic function of the state ω.
Define array of self-adjoint operators


R~ = Q1 , P 1 , Q2 , P 2 , · · · , Qn , P n

26
by
n
√ X

Wξ = exp −1 ξ 2j Qj − ξ 2i−1 P j .
j=1

Then, they satisfy √

[Rj , Rk ] = −1Jj,k .

Lemma 17 (Lemma 2.2 of [12])Let X ∈ CCR(J). If

√

Wξ XWξ† = exp − −1J (ξ, ξ ′ ) X

for any ξ ∈ R2n , then X = const. × Wξ′ .

Lemma 18 (Theorem 2.3 of [8]) Let c (ξ) be a functional over R2n with c (0) =
1. Then, the linear map Wξ → c (ξ) WAξ is liner and completely positive only
˜ where
if c (ξ) is a characteristic function of a state ω over CCR(J),

J˜ (ξ, ξ ′ ) := J (ξ, ξ ′ ) − J (Aξ, Aξ ′ ) .

The state ω is called a Gaussian state if its characteristic function is Gaus-

sian,  T 
ξ Σξ √
ω (Wξ ) = exp − + −1ξ · η ,
4
where Σ is a real positive symmetric matrix, and η ∈ R2n . It holds that
 
η=ω R ~ ,
1 √

Σ + −1J = ω ((Rj − ηj ) (Rk − ηk )) .
2
 T √ 
Lemma 19 (p. 18 of [12]) exp − ξ 4Σ ξ + −1ξ · η is a characteristic func-
tion of a state over CCR(J) if and only if
√
γ + −1J ≥ 0.

Below, let n = 1 and  

0 1
J := ,
−1 0
and define the vacume state |0i by the equation
√
Q1 + −1P1
√ |0i = 0.
2
Also, define coherent state |zi with z ∈ R2 by

|zi := Wz |0i ,

27
Then, any density matrix ρ can be written
Z
dz
ρ = Pρ (z) |zi hz| ,
2π

where Pρ (z) is called P-function of ρ.

28