Quantum information theory (MAT4430) Spring 2021

Lecture 7: Measuring distances between quantum states

Lecturer: Alexander Müller-Hermes

To analyze protocols in quantum information theory, we need to introduce some quanti-

ties measuring the distance between quantum states and quantum channels. To measure the
distance between quantum states we will often use the trace distance based on the Schatten
1-norm and the fidelity. Another distance measure is the quantum relative entropy, which
we will introduce later.

1 Positivity of block operators and operator inequalities

Recall, that we discussed the following lemma in the exercises:

Lemma 1.1 (Positivity of a block matrix). Let H denote a complex Euclidean space. For
X ∈ B(H), we have
1H X †
 
∈ B(H ⊕ H)+ ,
X 1H
if and only if kXk∞ 6 1.

Proof. By the singular value decomposition there are unitaries U, V ∈ U (H) and an operator
S ∈ B(H), that is diagonal in the computational basis with positive diagonal entries, such
that X = U SV . We can then verify that

1H X †
   †   
V 0 1H S V 0
= .
X 1H 0 U S 1H 0 U†

This operator is positive if and only if

 
1H S
∈ B(H ⊕ H)+ .
S 1H

Finally, we note that there exists a unitary W ∈ U (H ⊕ H) such that

       
1H S † 1 s1 1 s2 1 sdim(H)
W W = ⊕ ⊕ ··· ⊕ ,
S 1H s1 1 s2 1 sdim(H) 1

R
where si ∈ + are the diagonal entries of S. This operator is positive if and only if maxi si =
kXk∞ 6 1, and the proof is finished.

A consequence of this lemma is the following lemma:

Lemma 1.2 (Positivity of a block matrix). Let H denote a complex Euclidean space. For
X, Y ∈ B(H)+ and Z ∈ B(H), we have

X Z†
 
∈ B(H ⊕ H)+ ,
Z Y
1 1
if and only if there exists a K ∈ B(H) satisfying kKk∞ 6 1 and Z = Y 2 KX 2 .

1
Proof. Note that the cone B(H ⊕ H)+ is closed, and therefore we have

X Z†
 
∈ B(H ⊕ H)+ ,
Z Y

if and only if
X + 1C Z†
 
∈ B(H ⊕ H)+ ,
Z Y + 1C
for all  > 0. Thus, it is sufficient to consider the case where X and Y are invertible. In this
case, we have
1
! 1
! 1 1
!
X− 2 X Z† X− 2 X − 2 Z †Y − 2

0 0 1H
1 1 = 1 1 .
0 Y −2 Z Y 0 Y −2 Y − 2 ZX − 2 1H
1 1
By Lemma 1.1, this matrix is positive if and only if K = Y − 2 ZX − 2 satisfies kKk∞ 6 1.
Since the completely positive map AdM for
1
!
X− 2 0
M= 1 ,
0 Y −2

is invertible with completely positive (and in particular positive) inverse, the statement of
the lemma follows.

We will also need the following operator inequality, which we will prove in the exercises:

Theorem 1.3 (Choi’s inequality). For any
positive and
invertible operator X ∈ B( Cd A )+
and any positive and unital map P : B C
dA → B C
dB we have

P (X)−1 6 P (X −1 ),

where we used the Moore-Penrose pseudoinverse.

2 Schatten p-norms
The most important norms in quantum information theory are the Schatten norms, i.e., the
non-commutative analogues of the lp -norms. In particular, the Schatten 1-norm, also known
as the trace norm, and the Schatten ∞-norm, which you know as the operator norm, are
ubiquitous.

2.1 Basic definition and properties

For a complex Euclidean space H, an operator X ∈ B (H) and p ∈ [1, ∞) the Schatten
p-norm is given by
kXkp = Tr [|X|p ]1/p ,
√
where |X| = X ∗ X denotes the unique positive square root of the positive operator X ∗ X.
It can be shown that
kXk∞ := lim kXkp ,
p→∞

coincides with the usual operator norm kXk. For p = 2, we recover the Hilbert-Schmidt norm
kXk2 = kXkHS induced by the Hilbert-Schmidt inner product on B(H). A useful alternative

2
expression of the Schatten norms is given in terms of the singular values s1 (X), . . . , sR (X)
of the operator X ∈ B(H), and we have

R
! p1
X
kXkp = si (X)p ,
i=1

which coincides with the lp -norms of the vector of singular values. The Schatten norms
behave mostly like the lp -norms, and you might want to verify the following standard facts:

• We have the ordering k · k∞ 6 k · kp 6 k · k1 for any p ∈ [1, ∞].

• The norms k · kp and k · kq for 1/p + 1/q = 1 are dual norms with respect to the
Hilbert-Schmidt inner product.

• We have the Hölder’s inequality

kXY k1 6 kXkp kY kq ,

for any operators X, Y ∈ B(H) and any p, q ∈ [1, ∞) satisfying 1/p + 1/q = 1.

2.2 Fun fact: Extreme points of the unit balls B∞ and B1

We will start with the following easy theorem:

Theorem 2.1. Let H denote a complex Euclidean space and consider the unit ball

B1 = {X ∈ B(H) : kXk1 6 1}.

Then, we have
Ext (B1 ) = {|vihw| ∈ B(H) : hv|vi = hw|wi = 1}.

Proof. By the singular value decomposition any operator X ∈ B1 can be written as

dim(H)
X
X= si |vi ihwi |,
i=1

Pdim(H)
such that si > 0 for all i ∈ {1, . . . , dim (H)} and i=1 si = 1. This shows that the
extreme points of B1 are a subset of the set of rank-1 operators |vihw| ∈ B(H) satisfying
hv|vi = hw|wi = 1. Finally, it is clear that any such rank-1 operators is extremal since

|vihw| = (1 − λ)X1 + λX2

for X1 , X2 ∈ B1 and λ ∈ (0, 1) implies that X1 = X2 = |vihw|.

The next theorem determines the extreme points of the k · k∞ -unit ball.

Theorem 2.2. Let H denote a complex Euclidean space and consider the unit ball

B∞ = {X ∈ B(H) : kXk∞ 6 1}.

Then, we have
Ext (B∞ ) = U(H),
i.e., the extreme points are the unitary operators U ∈ U(H).

3
Proof. Consider a X ∈ B(H) satisfying kXk∞ 6 1. By the singular value decomposition, we
have X = U DV for unitary operators U, V ∈ U (H) and an operator D ∈ B(H) diagonal in
the computational basis such that Dii = si ∈ [0, 1] for each i ∈ {1, . . . , dim (H)}. Observe,
that every s ∈ [0, 1] can be written as
1 it
e + e−it


s=
2
for some t ∈ R. Using this decomposition for all singular values shows that
1 1
X = U DV = U (D1 + D2 ) V = (U1 + U2 ) ,
2 2
for unitary operators D1 , D2 ∈ U (H) diagonal in the computational basis, and unitary
operators U1 , U2 ∈ U (H) obtained as U1 = U D1 V and U2 = U D2 V . We conclude that each
contraction can be written as a convex combination of two unitary operators!
Clearly, kU k∞ = 1 for any unitary operator U ∈ U(H). Consider now a unitary operator
U ∈ U(H), and assume that there are X1 , X2 ∈ B∞ \ {0} satisfying

U = pX1 + (1 − p)X2 , (1)

for some p ∈ (0, 1). By the singular value decomposition, we have X1 = V SW for V, W ∈
U (H) and an operator S ∈ B∞ diagonal in the computational basis and containing the
singular values s1 , . . . , sd of X1 on its diagonal. We define U 0 = V † U W † ∈ U (H) and
X20 = V † X2 W † ∈ B∞ . Then, we have

U 0 = pD + (1 − p)X20 .

Consider now a normalized eigenvector |vi ∈ H of the unitary operator U 0 . We find that
d
X
1 = |hv|U 0 |vi| 6 p|hv|D|vi| + (1 − p)|hv|X 0 |vi| 6 p si |vi |2 + (1 − p) 6 1,
i=1

where d = dim (H) and vi ∈ C are the entries of |vi in the computational basis. This implies
that
d
X d
X
si |vi |2 = |vi |2 = 1,
i=1 i=1

and therefore si = 1 for each i ∈ {1, . . . , d}. By repeating the same argument as above, we
conclude that X1 and X2 in (1) are unitary operators. In this case, we can compute
 
1H = U † U = p2 1H + (1 − p)2 1cH + p(1 − p) X1† X2 + X2† X1 ,

and using that p(1 − p) 6= 0 we find that

1 † 
X1 X2 + X2† X1 = 1H .
2
But an operator of the form 21H − W for a unitary W ∈ U(H) is only itself a unitary
operator if W = 1H , since the eigenvalues of 21H − W have modulus strictly larger than
1 when W ∈ U(H) \ {1cH }. We conclude that X1† X2 = 1H and thus X1 = X2 = U . This
means that U is an extreme point of B∞ .

4
2.3 Induced norms and the Russo-Dye Theorem
Inspired by the operator norm, we can use the Schatten p-norms to define norms on the
space of linear maps L : B(HA ) → B(HB ). We will write
kL(x)kβ
kLkα→β = sup ,
x∈B(HA ) kxkα

for any linear map L : B(HA ) → B(HB ) and α, β ∈ [1, ∞]. These norms have many nice
properties which they inherent from the Schatten p-norms, or get by specializing general
properties of operator norms. For example we have the following properties, which you
might want to prove yourself:
• For any linear map L : B(HA ) → B(HB ) and all α, β ∈ [1, ∞] we have
kLkα→β = kL∗ kβ 0 →α0 ,
where L∗ : B(HB ) → B(HA ) is the adjoint operator with respect to the Hilbert-
Schmidt inner product, and 1/α + 1/α0 = 1 = 1/β + 1/β 0 .
• For linear maps L1 : B(HA ) → B(HB ) and L2 : B(HB ) → B(HC ) we have
kL2 ◦ L1 kα→γ 6 kL2 kβ→γ kL1 kα→β ,
for any α, β, γ ∈ [1, ∞].
In quantum information theory, we will mostly use the two special cases of α = β = 1 and
α = β = ∞. These norms are closely related to the trace norm and the operator norm, and
they behave particularly nicely when applied to positive maps:
Theorem 2.3 (Russo-Dye). For any positive map P : B(HA ) → B(HB ), we have
kP k1→1 = kP ∗ k∞→∞ = kP ∗ (1HB ) k∞ .
In particular, kP k1→1 = 1 if P is positive and trace-preserving.
Proof of Theorem 2.3. Using the duality of the norms k·k1→1 and k·k∞→∞ , it will be enough
to show that
kP k∞→∞ := sup kP (X)k∞ = kP (1HA ) k∞ .
X∈B∞
for any positive map P : B(HA ) → B(HB ). Since B∞ is convex and compact, the supremum
is attained in an extreme point of B∞ , and by Theorem 2.2 we know that these are the unitary
operators. Therefore, we have kP k∞→∞ = kP (U )k∞ for some unitary operator U ∈ U (HA ).
Note that
dA 
1HA U †
  X 
1 λi
MU = = ⊗ |vi ihvi |,
U 1HA λi 1
i=1
Pd
where U = i=1 λi |vi ihvi | is the spectral decomposition of U . Since |λi | 6 1, we find that
MU is separable, and that
P (1HA ) P (U † )
 
(id2 ⊗ P ) (MU ) = > 0.
P (U ) P (1HA )
By Lemma 1.2, we have
1 1
P (U ) = P (1HA ) 2 KP (1HA ) 2 ,
for some K ∈ B(HA ) satisfying kKk∞ 6 1. We conclude that
1
kP (U )k∞ 6 kP (1HA ) 2 k2∞ kKk∞ 6 kP (1HA )k∞ ,
1 1
since kX 2 k∞ = kXk∞
2
for any positive operator X.

5
 The Russo-Dye theorem is sometimes stated in an alternative form, which we point out
for completeness:

Corollary 2.4. For any positive map P : B(HA ) → B(HB ), we have

kP k1→1 = max{kP (|vihv|)k1 : |vi ∈ HA , hv|vi = 1}.

Proof. Since P is positive, we have

kP ∗ (1HB ) k∞ = max{hv|P ∗ (1HB ) |vi : |vi ∈ HA , hv|vi = 1}.

For any |vi ∈ HA , we have

hv|P ∗ (1HB ) |vi = Tr [P (|vihv|)] = kP (|vihv|) k1 ,

since P (|vihv|) ∈ B(HB )+ .

Another important corollary of the Russo-Dye theorem is the following:

Corollary 2.5. The trace norm is contractive under quantum channels, i.e., we have

kT (X)k1 6 kXk1

for every quantum channel T : B(HA ) → B(HB ) and any X ∈ B(HA ).

3 The trace distance

When using the trace norm to measure the distance between quantum states ρ, σ ∈ D(H),
we will often speak about the trace distance kρ − σk1 . Note that kρ − σk1 6 2 with equality if
and only if supp (ρ) ⊥ supp (σ). Some authors define the trace distance as the norm distance
with a factor 1/2 such that the maximum distance of two quantum states is 1 rather than
2. We will occasionally do so as well. The following lemma contains two useful properties of
the trace norm, which we prove in the exercises:

Lemma 3.1 (Some properties of the trace norms). Consider an operator X ∈ B(H) and
normalized vectors |ψi, |φi ∈ H. We have:

1. kXk1 = sup{|hU, XiHS | : U ∈ U (H)}.

R+ .
p
2. ka|ψihψ| − b|φihφ|k1 = (a + b)2 − 4ab|hψ|φi| for any a, b ∈

3.1 Operational interpretation: Quantum state discrimination

The trace distance between quantum states is the quantum analogue of the statistical dis-
tance between probability distributions p, q ∈ P (Σ) given by
X
kp − qk1 = |p(x) − q(x)|.
x∈Σ

As the statistical distance quantifies the maximum probability of discriminating between p

and q in a one-shot setting, it is not surprising that the trace-distance can be interpreted in
the same way. Consider the following scenario:

6
Scenario: State discrimination . Two researchers Alice and Bob are given the following
task visualized in Figure 1. Alice has a device with two buttons labeled “0” and “1”. After
pressing the button “0” the device emits a particle in quantum state ρ0 ∈ D (H), and after
pressing the button “1” the device emits a particle in quantum state ρ1 ∈ D (H). Bob catches
the emitted particle and measures it using some POVM. Then, he tries to guess whether
Alice pressed button 0 or button 1. Assume that Alice presses button “0” with probability
λ ∈ [0, 1] and button “1” with probability 1−λ, then, what is the optimal success probability
of Bob’s guess?

Figure 1: Alice and Bob at work.

Let us consider the case of some POVM µ : {0, 1} → B(H)+ , which Bob could measure,
i.e., the operators µ(0) and µ(1) are positive semidefinite and add up to 1H . Furthermore,
we assume (without loss of generality) that the outcomes “0” and “1” of this measurement
determine exactly whether he guesses that Alice pressed the corresponding button. Using
the POVM formalism, we compute the success probability as

psucc (µ) = λhµ(0), ρ0 iHS + (1 − λ)hµ(1), ρ1 iHS .

How large can this probability be? The following theorem gives an upper bound and shows
how to achieve it:

Theorem 3.2 (Holevo-Helstrom). Let H denote a complex Euclidean space and ρ0 , ρ1 ∈

D(H) a pair of quantum states. For any λ ∈ [0, 1] and any POVM µ : {0, 1} → B(H)+ , we
have
1 1
λhµ(0), ρ0 iHS + (1 − λ)hµ(1), ρ1 iHS 6 + kλρ0 − (1 − λ)ρ1 k1 .
2 2
The upper bound is achieved by the PVM µopt : {0, 1} → Proj (H) such that µ(0) is the
projector onto supp ((λρ0 − (1 − λ)ρ1 )+ ) (where ·+ denotes the positive part in the Jordan-
Hahn decomposition1 ).

Proof. Any binary POVM µ : {0, 1} → B(H)+ can be written as

1H + X 1H − X
µ(0) = and µ(1) = ,
2 2
for some contraction X ∈ B(H) satisfying kXk∞ 6 1. Inserting this decomposition into the
formula for the success probability shows that
1 1
λhµ(0), ρ0 iHS + (1 − λ)hµ(1), ρ1 iHS = + hX, λρ0 − (1 − λ)ρ1 iHS .
2 2
1
For H ∈ B(H)sa with Jordan-Hahn decomposition H = X1 −X2 with X1 , X2 ∈ B(H)+ we have H + = X1 .

7
Now, we note that

hX, λρ0 − (1 − λ)ρ1 iHS 6 sup |hU, λρ0 − (1 − λ)ρ1 iHS | = kλρ0 − (1 − λ)ρ1 k1 ,
U ∈U (H)

where we used the fact that the unitaries are the extreme points of the k · k∞ -unit ball and
Lemma 3.1. Since the set of unitaries U (H) is compact, the supremum is attained. The
optimal unitary Uopt is the one flipping the signs of the negative eigenvalues of the operator
λρ0 − (1 − λ)ρ1 and we see that

1H + Uopt
µopt (0) = ,
2
which is the projector onto supp ((λρ0 − (1 − λ)ρ1 )+ ).

Scenario: Discriminating states from an ensemble . In the previous scenario, Bob

received a quantum state ρ0 with probability λ and a state ρ1 with probability 1 − λ. Let
us now generalize this scenario to more than two quantum states: Again, Alice is in the
position of a device with n ∈ N
buttons. After pressing a button “i”, the device emits a
particle in quantum state ρi from some set {ρ1 , . . . , ρn } ⊂ D(H) of quantum states. Again,
Bob catches the particle, measures it using a POVM µ : {1, . . . , n} → B(H)+ and guesses
that Alice pressed button j if he received that outcome. Assuming that Alice presses the
buttons according to a probability distribution p ∈ P{1, . . . , n}, what is the optimal success
probability of Bob’s guess?
Again, given a particular POVM µ : {1, . . . , n} → B(H)+ , we can express the success
probability by
Xn
psucc (µ) = pi hµ(i), ρi i.
i=1

In the following, we will denote by opt ({pi , ρi }ni=1 ) the optimal success probability achievable
by Bob’s measurement, when Alice chooses the states ρi with probability pi . The collection
of probabilities and corresponding operators {pi , ρi }ni=1 is also called an ensemble of quantum
states.
Contrary to the previous scenario, it is more difficult to analyze this expression. In
particular, there is no easy closed formula for the optimal guessing probability. Instead, the
optimal guessing probability can be expressed as a particular convex optimization problem
known as a semidefinite program (or SDP for short). Such optimization problems can be
solved efficiently (i.e., their runtime scales at most polynomial in the size of the problem
and the inverse of the desired accuracy). We will not go into the details of this, but instead
prove the following remarkable fact, that there actually is a very simple measurement Bob
can do, which achieves a pretty good success probability:

Theorem 3.3 (Barnum and Knill’s pretty good measurement). Consider a set {ρ1 , . . . , ρn } ⊂
D(H) of quantum states on a complex Euclidean space H. For any probability distribution
p ∈ Pn , we can define a POVM µ : {1, . . . , n} → B(H)+ by
1 1 1
µ(i) = ρ− 2 pi ρi ρ− 2 + Π ,
n ker(ρ)
Pn −1 denotes the Moore-Penrose pseudo-inverse. This POVM
where ρ = i=1 pi ρi , and ρ
satisfies the inequality
n
X
pi hµ(i), ρi iHS > (opt ({pi , ρi }ni=1 ))2 .
i=1

8
Proof. By positivity, we have ker (ρ) ⊆ ker (ρi ) and hence im (ρi ) ⊆ im (ρ) for any i ∈
{1, . . . , n}. For any X ∈ B(H)+ , we have
1 1 1 1
hX, ρi iHS = hX, Pim(ρ) ρi Pim(ρ) iHS = hρ 4 Xρ 4 , ρ− 4 ρi ρ− 4 iHS ,

where ρ−1 denotes the Moore-Penrose pseudo-inverse. Applying the Cauchy-Schwarz in-
equality, we find that
1 1 1 1
hX, ρi iHS 6 kρ 4 Xρ 4 k2 kρ− 4 ρi ρ− 4 k2 . (2)
Consider now a POVM ν : {1, . . . , n} → B(H)+ . We have
n n
X X 1 1 1 1
pi hν(i), ρi iHS 6 kρ 4 ν(i)ρ 4 k2 kρ− 4 pi ρi ρ− 4 k2
i=1 i=1
n
!1/2 n
!1/2
1 1
− 14 1
X X
6 kρ ν(i)ρ
4 4 k22 kρ pi ρi ρ− 4 k22 ,
i=1 i=1

where we used (2) for the first inequality and the Cauchy-Schwarz inequality for vectors in
the second inequality. Using that 0 6 ν(i) 6 1H , we find that
1 1
h 1 1
i
kρ 4 ν(i)ρ 4 k22 = Tr ν(i)ρ 2 ν(i)ρ 2 6 Tr [ρν(i)] ,

and therefore
n n
X 1 1 X
kρ ν(i)ρ
4 4 k22 6 Tr [ρν(i)] = Tr [ρ] = 1.
i=1 i=1

We conclude that !1/2

n n
− 14 1
X X
pi hν(i), ρi iHS 6 kρ pi ρi ρ− 4 k22 ,
i=1 i=1

for any POVM ν : {1, . . . , n} → B(H)+ . Consider now the pretty good measurement
µ{1, . . . , n} → B(H)+ defined in the statement of the theorem. We can compute
1 1 1 1 1 1
pi hµ(i), ρi iHS = hρ− 4 pi ρi ρ− 4 , ρ− 4 pi ρi ρ− 4 iHS = kρ− 4 pi ρi ρ− 4 k22 .

Finally, if ν : {1, . . . , n} → B(H)+ is an optimal measurement, then we conclude that

n n
!1/2 n
!1/2
− 14 1
X X X
opt ({pi , ρi }ni=1 ) = pi hν(i), ρi iHS 6 kρ pi ρi ρ− 4 k22 = pi hν(i), ρi iHS ,
i=1 i=1 i=1

which finishes the proof.