Quantum information theory (MAT4430) Spring 2021

Lecture 8: The Fidelity

Lecturer: Alexander Müller-Hermes

The goal of this lecture is to introduce the fidelity, which is one of the most important
distance measures in quantum information theory. We will then express this quantity as
convex optimization problems in two different ways. By doing so, it will be easy to prove
the data-processing inequality of the fidelity, i.e., the inequality

F (T (ρ), T (σ)) > F (ρ, σ),

for all positive and trace-preserving maps T : B(HA ) → B(HB ).

1 Definition and basic properties

Let H denote a complex Euclidean space. For quantum states ρ, σ ∈ D(H), we define the
fidelity by
√ √
F (ρ, σ) = k ρ σk1 ,
√
where · denotes the unique positive square root and k · k1 is the trace norm (i.e., the matrix
1-norm). The fidelity measures the distance between the two quantum states, and (using
the definition of the trace norm) it can be written as

1 1
1 
2
F (ρ, σ) = Tr ρ σρ
2 2 .

Note that it is not necessary to have normalized quantum states in the definition of the
fidelity. In general, we define the fidelity
√ √ 1 

1 1 2
F (A, B) = k A Bk1 = Tr A 2 BA 2 ,

for any positive operators A, B ∈ B(H)+ . In the exercises we have shown the following
elementary properties.

Lemma 1.1 (Elementary properties of the fidelity). Let H denote a complex Euclidean space
and consider quantum states ρ, σ ∈ D(H). We have the following properties:

1. We have F (ρ, σ) = F (σ, ρ).

2. We have F (ρ, σ) > 0 with equality if and only if ρσ = 0.

3. We have F (ρ, σ) 6 1 with equality if and only if ρ = σ.

4. We have F (V ρV † , V σV † ) = F (ρ, σ) for any isometry V : H → H0 into another complex

Euclidean space H0 .
p
5. We have F (|ψihψ|, σ) = hψ|σ|ψi for any pure quantum state |ψihψ| ∈ Proj (H).

6. We have F (ρ ⊗ τ, σ ⊗ τ ) = F (ρ, σ) for every quantum state τ ∈ D(H0 ).

Besides these properties, we note that the fidelity is a continuous function in both its
inputs, since the operator square root and the 1-norm are continuous.

1
1.1 Expressing the fidelity as optimization problems
Theorem 1.2. Let H denote a complex Euclidean space. For any A, B ∈ B(H)+ we have

A X†
 
F (A, B) = max{| Tr [X] | : X ∈ B(H), ∈ B(H ⊕ H)+ }.
X B

Proof. By a lemma from the last lecture we have

A X†
 
∈ B(H ⊕ H)+ ,
X B
1 1
if and only if there exists a K ∈ B(H) satisfying kKk∞ 6 1 such that X = B 2 KA 2 .
Therefore, we can write

A X†
 
sup{| Tr [X] | : X ∈ B(H), ∈ B(H ⊕ H)+ }
X B
h 1 1
i
= sup{| Tr B 2 KA 2 | : K ∈ B(H), kKk∞ 6 1}
h 1 1
i
= max{| Tr U A 2 B 2 | : U ∈ U(H)},

where we used, in the last step, that the unitaries are the extreme points of the unit ball of
the operator norm k · k∞ . Using Hölder’s inequality it is easy to see that
h 1 1
i 1 1
| Tr U A 2 B 2 | 6 kA 2 B 2 k1 = F (A, B).

1 1
On the other hand, if A 2 B 2 = V SW is the singular value decomposition with unitaries
V, W ∈ U (H), then the unitary U = W † V † satisfies
h 1 1
i 1 1
| Tr U A 2 B 2 | = kA 2 B 2 k1 = F (A, B).

Together, these two statements finish the proof.

As a simple corollary, we can rewrite this optimization problem slightly:

Corollary 1.3. Let H denote a complex Euclidean space. For A, B ∈ B(H)+ we have

A X†
 
F (A, B) = max{Re (Tr [X]) : X ∈ B(H), ∈ B(H ⊕ H)+ }.
X B

Proof. Clearly, the right-hand-side is less than the fidelity by Theorem 1.2. To show that
they are equal, consider X ∈ B(H) such that

A X†
 
∈ B(H ⊕ H)+ ,
X B

R
and | Tr [X] | = F (A, B). Then, for any α ∈ we have that

e−iα X † A X†
     
A 1H 0 1H 0
= ∈ B(H ⊕ H)+ .
eiα X B 0 eiα 1H X B 0 e−iα 1H

Choosing α ∈ R such that

Re Tr eiα X = | Tr [X] |,
 

finishes the proof.

2
Theorem 1.4. Let H denote a complex Euclidean space. For A, B ∈ B(H)+ we have
1 1
F (A, B) = inf{ hA, Y iHS + hB, Y −1 iHS : Y ∈ B(H)++ }.
2 2
Proof. Note that  
Y −1H
∈ B(H ⊕ H)+
−1H Y −1
for any Y ∈ B(H)++ . If
A X†
 
∈ B(H ⊕ H)+ ,
X B
for some X ∈ B(H), then we have

A X†
  
Y −1H −1
h i
0 6 Tr = hA, Y i HS + hB, Y i HS − Tr [X] − Tr X† ,
X B −1H Y −1

and, by Corollary 1.3, we conclude that

1 1
inf{ hA, Y iHS + hB, Y −1 iHS : Y ∈ B(H)++ } > F (A, B),
2 2
for all A, B ∈ B(H)+ . To show that the infimum coincides with the fidelity, we start with
the case where A, B ∈ B(H)++ are invertible. Then, we can define

1
 1 1
1 1
Y = A− 2 A 2 BA 2 A− 2 ∈ B(H)++
2

and check that  1 

1 1 2
hA, Y iHS = Tr [AY ] = Tr A BA
2 2 = F (A, B),

and

1
 1  1 1
1 −2

1 1
1 
hB, Y −1 iHS = Tr BY −1 = Tr BA 2 A 2 BA 2
  2
A 2 = Tr A 2 BA 2 = F (A, B).

We conclude that the infimum is attained and coincides with the fidelity in this case.
For general A, B ∈ B(H)+ we will use a continuity argument. For every  > 0 and any
Y ∈ B(H)++ we have
1 1 1 1
hA, Y iHS + hB, Y −1 iHS 6 hA + 1H , Y iHS + hB + 1H , Y −1 iHS .
2 2 2 2
Taking the infimum over Y ∈ B(H)++ on both sides and using the previous argument, we
find that
1 1
inf{ hA, Y iHS + hB, Y −1 iHS : Y ∈ B(H)++ } 6 F (A + 1H , B + 1H ).
2 2
Taking the limit  → 0 finishes the proof.

In the exercises you will be asked to prove the following corollary:

Corollary 1.5 (Alberti’s theorem). Let H denote a complex Euclidean spaces and ρ, σ ∈
D(H) two quantum states. We have

F (ρ, σ)2 = inf{hρ, Y iHS hσ, Y −1 iHS : Y ∈ B(H)++ }.

3
1.2 Data-processing inequality and joined concavity of the fidelity
Theorem 1.6 (Data-processing inequality for the fidelity). For any positive and trace-
preserving map P : B(HA ) → B(HB ) we have

F (P (ρ), P (σ)) > F (ρ, σ),

for all quantum states ρ, σ ∈ D(HA ).

Proof. Without loss of generality we may assume that P ∗ (X) is invertible for any X ∈
B(HA )+ . If this is not the case, we can consider the positive maps P : B(HA ) → B(HB )
given by P (X) = (1 − )P (X) +  Tr [X] 1HB for 0 <  6 1 instead, and in the end of the
proof take the limit  → 0 using continuity of the fidelity.
By Theorem 1.4 we have
1 1
F (P (ρ), P (σ)) = inf{ hP (ρ), Y iHS + hP (σ), Y −1 iHS : Y ∈ B(H)++ }
2 2
1 1
= inf{ hρ, P (Y )iHS + hσ, P ∗ (Y −1 )iHS : Y ∈ B(H)++ }
∗
2 2
1 1
> inf{ hρ, P ∗ (Y )iHS + hσ, P ∗ (Y )−1 iHS : Y ∈ B(H)++ }
2 2
> F (ρ, σ).

Here, we have used Choi’s inequality from the exercises in the second-to-last step (note that
P ∗ is unital whenever P is trace-preserving) and Theorem 1.4 for the final inequality.

Next, we can prove the joined concavity of the fidelity:

Theorem 1.7 (Joined concavity). For quantum states ρ1 , ρ2 , σ1 , σ2 ∈ D(H) and λ ∈ [0, 1]
we have

F ((1 − λ)ρ1 + λρ2 , (1 − λ)σ1 + λσ2 ) > (1 − λ)F (ρ1 , σ1 ) + λF (ρ2 , σ2 ).

Proof. Consider the quantum states

ρ = (1 − λ)ρ1 ⊗ |1ih1| + λρ2 ⊗ |2ih2| ∈ D(H ⊗ C2),

and
σ = (1 − λ)σ1 ⊗ |1ih1| + λσ2 ⊗ |2ih2| ∈ D(H ⊗ C2),
and the completely positive and trace-preserving partial trace map TrB = idH ⊗ Tr. By
Theorem 1.6 we have
F (TrB (ρ) , TrB (σ)) > F (ρ, σ).
Now, note that

F (TrB (ρ) , TrB (σ)) = F ((1 − λ)ρ1 + λρ2 , (1 − λ)σ1 + λσ2 ),

and
√ √ √ √ √ √
F (ρ, σ) = k ρ σk1 = k(1 − λ) ρ1 σ1 ⊗ |1ih1| + λ ρ2 σ2 ⊗ |2ih2|k1
√ √ √ √
= (1 − λ)k ρ1 σ1 k1 + λk ρ2 σ2 k1
= (1 − λ)F (ρ1 , σ1 ) + λF (ρ2 , σ2 ).

4
1.3 Uhlmann’s theorem
The following lemma characterizes the fidelity between two quantum states by using their
purifications:

Theorem 1.8 (Uhlmann’s theorem). Let HA , HE denote complex Euclidean spaces and
ρ, σ ∈ D(HA ) two quantum states. For any normalized vector |ψEA i ∈ HE ⊗ HA satisfying
TrE [|ψEA ihψEA |] = ρ, we have

F (ρ, σ) = max{|hψEA |φEA i| : TrE [|φEA ihφEA |] = σ}.

Proof. Given normalized vectors |ψEA i, |φEA i ∈ HE ⊗ HA such that

TrE [|ψEA ihψEA |] = ρ, and TrE [|φEA ihφEA |] = σ,

we can use the data-processing inequality from Theorem 1.6 (for P = TrE ) to conclude

|hψEA |φEA i| = F (|ψEA ihψEA |, |φEA ihφEA |) 6 F (ρ, σ).

To show that it is attained, consider a unitary U ∈ U (HA ) such that

 √ √ 
F (ρ, σ) = Tr U ρ σ .

Writing |ψEA i = vec(X) for some X ∈ B(HE , HA ) we find that ρ = XX † . Using the singular
value decomposition X = V SW we find that
 √ √ 
F (ρ, σ) = Tr U ρ σ
h √ √ i
= Tr U V S 2 V † σ
h √ i
= Tr U V SV † σ
h √ i
= Tr U V W (W † SV † ) σ
h √ i
= Tr U V W X † σ
√
= hX, σU V W iHS = hψEA |φEA i,

where we introduced √

|φEA i := vec σU V W .
We can check that
√ √
TrE [|φEA ihφEA |] = ( σU V W )( σU V W )† = σ,

and the proof is finished.

1.4 Relationship between fidelity and trace distance

The following theorem is quite useful:

Theorem 1.9 (Fuchs-van de Graaf inequalities). Let H denote a complex Euclidean space
and ρ, σ ∈ D(H) two quantum states. Then, we have
r
1 1
1 − kρ − σk1 6 F (ρ, σ) 6 1 − kρ − σk21 .
2 4

5
Proof. Clearly, the statement from the theorem is equivalent to
p
2 − 2F (ρ, σ) 6 kρ − σk1 6 2 1 − F (ρ, σ)2 . (1)

The proof will show this statement. For the second inequality of (1), we use Uhlmann’s the-
orem to find normalized vectors |ψEA i, |φEA i ∈ HE ⊗ HA satisfying F (ρ, σ) = |hψEA |φEA i|.
Then, we have

kρ − σk1 = k TrE [|ψEA ihψEA |] − TrE [|φEA ihφEA |] k1 6 k|ψEA ihψEA | − |φEA ihφEA |k1 ,

by the Russo-Dye theorem. Using a result from the exercises, we conclude that the right-
hand-side equals
p p
k|ψEA ihψEA | − |φEA ihφEA |k1 = 2 1 − |hψEA |φEA i|2 = 2 1 − F (ρ, σ)2 .

For the other inequality in (1) let us restrict first to the case where ρ and σ are invertible.
Recall, the operator
 1 1 1
− 12 1
ρ− 2 ∈ B(H)++
2
Y =ρ ρ 2 σρ 2
from the proof of Theorem 1.4 satisfying σ = Y ρY , and

hY, ρiHS = hY −1 , σiHS = F (ρ, σ).

By the spectral decomposition there are normalized vectors |vi i ∈ H and numbers λi > 0 for
any i ∈ {1, . . . dim(H)} such that
dim(H)
X
Y = λi |vi ihvi |.
i=1

After defining probability distributions p, q ∈ P ({1, . . . , dim(H)}) by

pi = hvi |ρ|vi i and qi = hvi |σ|vi i,

we compute
dim(H) dim(H)
X √ √ X p p
pi qi = hvi |ρ|vi i hvi |Y ρY |vi i
i=1 i=1
dim(H)
X
= λi hvi |ρ|vi i = hY, ρiHS = F (ρ, σ).
i=1

Finally, note that the linear map T : B(H) → B(H) given by

dim(H)
X
T (X) = hvi |X|vi i|vi ihvi |,
i=1

is a quantum channel, and we have

dim(H)
X √ √
kρ − σk1 > kT (ρ) − T (σ)k1 = kp − qk1 > 2 − 2 pi qi = 2 − 2F (ρ, σ),
i=1

where we used first the Russo-Dye theorem and in the last step an exercise from sheet 1.
Finally, for general quantum states ρ, σ ∈ D(H), we may consider the invertible quantum
states ρ , σ ∈ D(H) given by
1H
ρ = (1 − )ρ +  ,
dim(H)

6
and
1H
σ = (1 − )σ +  ,
dim(H)
for  ∈ (0, 1). From the previous argument, we have

2 − 2F (ρ , σ ) 6 kρ − σ k1

for any  ∈ (0, 1) and taking the limit  → 0 finishes the proof.