SUPPLEMENTARY INFORMATION

doi: 10.1038/nPHYS1734

Supplementary Information
The Uncertainty Principle in the Presence of Quantum Memory
Mario Berta,1, 2 Matthias Christandl,1, 2 Roger Colbeck,3, 1, 4 Joseph M. Renes,5 and Renato Renner1
1
Institute for Theoretical Physics, ETH Zurich, 8093 Zurich, Switzerland.
2
Faculty of Physics, Ludwig-Maximilians-Universität München, 80333 Munich, Germany.
3
Perimeter Institute for Theoretical Physics, 31 Caroline Street North, Waterloo, ON N2L 2Y5, Canada.
4
Institute of Theoretical Computer Science, ETH Zurich, 8092 Zurich, Switzerland.
5
Institute for Applied Physics, Technische Universität Darmstadt, 64289 Darmstadt, Germany.

Here we present the full proof of our main result, the uncertainty relation given in Equation (3) of the main
manuscript (Theorem 1 below).
In order to state our result precisely, we introduce a few definitions. Consider two measurements described by
orthonormal bases {|ψj } and {|φk } on a d-dimensional Hilbert space HA (note that they are not necessarily com-
plementary). The measurement processes are then described by the completely positive maps

R : ρ → ψj |ρ|ψj |ψj ψj | and
j

S : ρ → φk |ρ|φk |φk φk |
k

respectively. We denote the square of the overlap of these measurements by c, i.e.

c := max |ψj |φk |2 . (1)
j,k

Furthermore, we assume that HB is an arbitrary finite-dimensional Hilbert space. The von Neumann entropy of
A given B is denoted H(A|B) and is defined via H(A|B) := H(AB) − H(B), where for a state ρ on HA we have
H(A) := −tr(ρ log ρ).
The statement we prove is then
Theorem 1. For any density operator ρAB on HA ⊗ HB ,
1
H(R|B) + H(S|B) ≥ log2 + H(A|B), (2)
c
where H(R|B), H(S|B), and H(A|B) denote the conditional von Neumann entropies of the states (R ⊗ I)(ρAB ),
(S ⊗ I)(ρAB ), and ρAB , respectively.
In the next section, we introduce the smooth min- and max- entropies and give some properties that will be needed
in the proof.
Before that, we show that the statement of our main theorem is equivalent to a relation conjectured by Boileau
and Renes [1].
Corollary 2. For any density operator ρABE on HA ⊗ HB ⊗ HE ,
1
H(R|E) + H(S|B) ≥ log2 . (3)
c
Proof. To show that our result implies (3), we first rewrite (2) as H(RB) + H(SB) ≥ log2 1c + H(AB) + H(B). In the
case that ρABE is pure, we have H(RB) = H(RE) and H(AB) = H(E). This yields the expression H(RE)+H(SB) ≥
log2 1c + H(E) + H(B), which is equivalent to (3). The result for arbitrary states ρABE follows by the concavity of
the conditional entropy (see e.g. [2]).
That (3) implies (2) can be seen by taking ρABE as the state which purifies ρAB in (3) and reversing the argument
above.

I. (SMOOTH) MIN- AND MAX-ENTROPIES—DEFINITIONS

As described above, we prove a generalized version of (2), which is formulated in terms of smooth min- and max-
entropies. This section contains the basic definitions, while Section V B summarizes the properties of smooth entropies
needed for this work. For a more detailed discussion of the smooth entropy calculus, we refer to [3–6].

nature physics | www.nature.com/naturephysics 1

supplementary information doi: 10.1038/nphys1734

We use U= (H) := {ρ : ρ ≥ 0, trρ = 1} to denote the set of normalized states on a finite-dimensional Hilbert space
H and U≤ (H) := {ρ : ρ ≥ 0, trρ ≤ 1} to denote the set of subnormalized states on H. The definitions below apply to
subnormalized states.
The conditional min-entropy of A given B for a state ρ ∈ U≤ (HAB ) is defined as1

Hmin (A|B)ρ := sup Hmin (A|B)ρ|σ ,

σ

where the supremum is over all normalized density operators σ ∈ U= (HB ) and where

Hmin (A|B)ρ|σ := − log2 inf{λ : ρAB ≤ λ 11A ⊗ σB } .

In the special case where the B system is trivial, we write Hmin (A)ρ instead of Hmin (A|B)ρ . It is easy to see that
Hmin (A)ρ = − log2 ρA ∞ and that for ρ ≤ τ , Hmin (A|B)ρ ≥ Hmin (A|B)τ .
Furthermore, for ρ ∈ U≤ (HA ), we define
√
Hmax (A)ρ := 2 log2 tr ρ .

It follows that for ρ ≤ τ , Hmax (A)ρ ≤ Hmax (A)τ (since the square root is operator monotone).
In our proof, we also make use of an intermediate quantity, denoted H−∞ . It is defined by

H−∞ (A)ρ := − log2 sup{λ : ρA ≥ λ Πsupp(ρA ) } ,

where Πsupp(ρA ) denotes the projector onto the support of ρA . In other words, H−∞ (A)ρ is equal to the negative
logarithm of the smallest non-zero eigenvalue of ρA . This quantity will not appear in our final statements but will
instead be replaced by a smooth version of Hmax (see below and Section V B).
The smooth min- and max-entropies are defined by extremizing the non-smooth entropies over a set of nearby
states, where our notion of nearby is expressed in terms of the purified distance. It is defined as (see [6])

P (ρ, σ) := 1 − F̄ (ρ, σ)2 , (4)

where F̄ (· , ·) denotes the generalized fidelity (which equals the standard fidelity if at least one of the states is
normalized),
  
F̄ (ρ, σ) :=  ρ ⊕ (1 − trρ) σ ⊕ (1 − trσ)1 . (5)
√ √
(Note that we use F (ρ, σ) :=  ρ σ1 to denote the standard fidelity.)
The purified distance is a distance measure; in particular, it satisfies the triangle inequality P (ρ, σ) ≤ P (ρ, τ ) +
P (τ, σ). As its name indicates, P (ρ, σ) corresponds to the minimum trace distance2 between purifications of ρ and σ.
Further properties are stated in Section V A.
We use the purified distance to specify a ball of subnormalized density operators around ρ:

B ε (ρ) := {ρ : ρ ∈ U≤ (H), P (ρ, ρ ) ≤ ε} .

Then, for any ε ≥ 0, the ε-smooth min- and max-entropies are defined by
ε
Hmin (A|B)ρ := sup Hmin (A|B)ρ
ρ ∈Bε (ρ)
ε
Hmax (A)ρ := inf Hmax (A)ρ .
ρ ∈Bε (ρ)

In the following, we will sometimes omit the subscript ρ when it is obvious from context which state is implied.

1 In the case of finite dimensional Hilbert spaces (as in this work), the infima and suprema used in our definitions can be replaced by
minima and maxima. √
2 The trace distance between two states τ and κ is defined by 12 τ − κ1 where Γ1 = tr ΓΓ† .

2 nature physics | www.nature.com/naturephysics

doi: 10.1038/nphys1734
supplementary information

II. OVERVIEW OF THE PROOF

The proof of our main result, Theorem 1, is divided into two main parts, each individually proven in the next
sections.
In the first part, given in Section III, we prove the following uncertainty relation, which is similar to the main result
but formulated in terms of the quantum entropies Hmin and H−∞ .
Theorem 3. For any ρAB ∈ U≤ (HAB ) we have
1
Hmin (R|B)(R⊗I)(ρ) + H−∞ (SB)(S⊗I)(ρ) ≥ log2 + Hmin (AB)ρ .
c
The second part of the proof involves smoothing the above relation and yields the following theorem (see Sec-
tion IV)3 .
Theorem 4. For any ρ ∈ U= (HAB ) and ε > 0,
√ 1 1
5 ε ε ε
Hmin (R|B)(R⊗I)(ρ) + Hmax (SB)(S⊗I)(ρ) ≥ log2 + Hmin (AB)ρ − 2 log2 .
c ε
From Theorem 4, the von Neumann version of the uncertainty relation (Theorem 1) can be obtained as an asymptotic
special case for i.i.d. states. More precisely, for any σ ∈ U= (HAB ) and for any n ∈ N, we evaluate the inequality for
ρ = σ ⊗n where R ⊗ I and S ⊗ I are replaced by (R ⊗ I)⊗n and (S ⊗ I)⊗n , respectively. Note that the corresponding
overlap is then given by

c(n) = max |ψj1 |φk1  . . . ψjn |φkn |2 = max |ψj⊗n |φ⊗n 2 n
k | = c .
j1 ...jn ,k1 ...kn j,k

The assertion of the theorem can thus be rewritten as

1 5√ε n n 1 ε 1 1 ε 2 1
H (R |B )((R⊗I)(σ))⊗n + Hmax (S n B n )((S⊗I)(σ))⊗n ≥ log2 + Hmin (An B n )σ⊗n − log2 .
n min n c n n ε
Taking the limit n → ∞ and then ε → 0 and using the asymptotic equipartition property (Lemma 9), we obtain
H(R|B) + H(SB) ≥ log2 1c + H(AB), from which Theorem 1 follows by subtracting H(B) from both sides.

III. PROOF OF THEOREM 3

In this section we prove a version of Theorem 1, formulated in terms of the quantum entropies Hmin and H−∞ .
 2πij  2πik
We introduce DR = j e d |ψj ψj | and DS = k e d |φk φk | (DR and DS are d-dimensional generalizations of
Pauli operators). The maps R and S describing the two measurements can then be rewritten as
d−1
1 a
R : ρ → D ρD−a
d a=0 R R
d−1
1 b
S : ρ → DS ρDS−b .
d
b=0

We use the two chain rules proved in Section V B (Lemmas 11 and 12), together with the strong subadditivity of the
min-entropy (Lemma 10), to obtain, for an arbitrary density operator ΩA B  AB ,

Hmin (A B  AB)Ω − H−∞ (A AB)Ω ≤ Hmin (B  |A AB)Ω|Ω

≤ Hmin (B  |AB)Ω|Ω
≤ Hmin (B  A|B)Ω − Hmin (A|B)Ω . (6)

3 We note that a related relation follows from the work of Maassen and Uffink [7] who derived a relation involving Rényi entropies (the
order α Rényi entropy [8] is denoted Hα ) and the overlap c (defined in (1)). They showed that Hα (R)ρ + Hβ (S)ρ ≥ log2 1c , where
1
α
+ β1 = 2. The case α → ∞, β → 12 yields Hmin (R)ρ + Hmax (S)ρ ≥ log2 1c .

nature physics | www.nature.com/naturephysics 3

supplementary information doi: 10.1038/nphys1734

We now apply this relation to the state ΩA B  AB defined as follows4 :

1 
ΩA B  AB := a
|aa|A ⊗ |bb|B  ⊗ (DR DSb ⊗ 11)ρAB (DS−b DR
−a
⊗ 11) ,
d2
a,b

where {|aA }a and {|bB  }b are orthonormal bases on d-dimensional Hilbert spaces HA and HB  .
This state satisfies the following relations:

Hmin (A B  AB)Ω = 2 log2 d + Hmin (AB)ρ (7)

H−∞ (A AB)Ω = log2 d + H−∞ (SB)(S⊗I)(ρ) (8)
Hmin (B  A|B)Ω ≤ log2 d + Hmin (R|B)(R⊗I)(ρ) (9)
1
Hmin (A|B)Ω ≥ log2 . (10)
c
Using these in (6) establishes Theorem 3. We proceed by showing(7)–(10).
Relation (7) follows because ΩA B  AB is unitarily related to d12 a,b |aa|A ⊗ |bb|B  ⊗ ρAB , and the fact that the
unconditional min-entropy is invariant under unitary operations.
  
To see (8), note that ΩA AB is unitarily related to d12 a |aa|A ⊗ b (S b ⊗ 11)ρAB (S −b ⊗ 11) and that d1 b (S b ⊗
11)ρAB (S −b ⊗ 11) = (S ⊗ I)(ρAB ).
To show inequality (9), note that
1  
ΩB  AB = 2
|bb|B  ⊗ a
(DR DSb ⊗ 11)ρAB (DS−b DR
−a
⊗ 11).
d a
b

To evaluate the min-entropy, define λ such that Hmin (B  A|B)Ω = − log2 λ. It follows that there exists a (normalized)
density operator σB such that
1  
λ 11B  A ⊗ σB ≥ 2
|bb|B  ⊗ a
(DR DSb ⊗ 11)ρAB (DS−b DR
−a
⊗ 11).
d a
b

Thus, for all b,

1  a b
λ 11A ⊗ σB ≥ (DR DS ⊗ 11)ρAB (DS−b DR
−a
⊗ 11),
d2 a

and in particular, for b = 0, we have

1  a −a
λ 11A ⊗ σB ≥ (DR ⊗ 11)ρAB (DR ⊗ 11)
d2 a
1
= (R ⊗ I)(ρAB ).
d
We conclude that 2−Hmin (R|B)(R⊗I)(ρ) ≤ λd, from which (9) follows.
To show (10), we observe that
1  a b
ΩAB = (DR DS ⊗ 11)ρAB (DS−b DR
−a
⊗ 11) = ((R ◦ S) ⊗ I)(ρAB ) .
d2
ab

4 The idea behind the use of this state first appeared in [9].

4 nature physics | www.nature.com/naturephysics

doi: 10.1038/nphys1734
supplementary information

Then,
 

((R ◦ S) ⊗ I)(ρAB ) = (R ⊗ I) |φk φk | ⊗ trA ((|φk φk | ⊗ 11)ρAB )
k

= |φk |ψj |2 |ψj ψj | ⊗ trA ((|φk φk | ⊗ 11)ρAB )
jk
� 
≤ max |φl |ψm |2 |ψj ψj | ⊗ trA ((|φk φk | ⊗ 11)ρAB )
lm
jk
�  
2
= max |φl |ψm | 11A ⊗ trA ((|φk φk | ⊗ 11)ρAB )
lm
k
� 
= max |φl |ψm |2 11A ⊗ ρB .
lm

It follows that 2−Hmin (A|B)((R◦S)⊗I)(ρ) ≤ maxlm |φl |ψm |2 = c, which concludes the proof.

IV. PROOF OF THEOREM 4

The uncertainty relation proved in the previous section (Theorem 3) is formulated in terms of the entropies Hmin
ε ε
and H−∞ . In this section, we transform these quantities into the smooth entropies Hmin and Hmax , respectively, for
some ε > 0. This will complete the proof of Theorem 4.
Let σAB ∈ U≤ (HAB ). Lemma 15 applied to σSB := (S ⊗ I)(σAB ) implies that there exists a nonnegative operator
Π ≤ 11 such that tr((11 − Π2 )σSB ) ≤ 3ε and

ε 1
Hmax (SB)(S⊗I)(σ) ≥ H−∞ (SB)Π(S⊗I)(σ)Π − 2 log2 . (11)
ε
We can assume without loss of generality that Π commutes with the action of S ⊗ I because it can be chosen to be
diagonal in any eigenbasis of σSB . Hence, Π(S ⊗ I)(σAB )Π = (S ⊗ I)(ΠσAB Π), and

tr((11 − Π2 )σAB ) = tr((S ⊗ I)((11 − Π2 )σAB )) = tr((11 − Π2 )σSB ) ≤ 3ε . (12)

Applying Theorem 3 to the operator ΠσAB Π yields

1
Hmin (R|B)(R⊗I)(ΠσΠ) + HR (SB)(S⊗I)(ΠσΠ) ≥ log2 + Hmin (AB)ΠσΠ . (13)
c
Note that ΠσΠ ≤ σ and so

Hmin (AB)ΠσΠ ≥ Hmin (AB)σ . (14)

Using (11) and (14) to bound the terms in (13), we find

ε 1 1
Hmin (R|B)(R⊗I)(ΠσΠ) + Hmax (SB)(S⊗I)(σ) ≥ log2 + Hmin (AB)σ − 2 log2 . (15)
c ε
Now we apply Lemma 18 to ρAB . Hence there exists a nonnegative operator Π̄ ≤ 11 which is diagonal in an eigenbasis
of ρAB such that

tr((11 − Π̄2 )ρAB ) ≤ 2ε (16)

ε
and Hmin (AB)Π̄ρΠ̄ ≥ Hmin (AB)ρ . Evaluating (15) for σAB := Π̄ρAB Π̄ thus gives

ε 1 ε 1
Hmin (R|B)(R⊗I)(ΠΠ̄ρΠ̄Π) + Hmax (SB)(S⊗I)(Π̄ρΠ̄) ≥ log2 + Hmin (AB)ρ − 2 log2 , (17)
c ε
where Π is diagonal in any eigenbasis of (S ⊗ I)(Π̄ρAB Π̄) and satisfies

tr((11 − Π2 )Π̄ρAB Π̄) ≤ 3ε . (18)

nature physics | www.nature.com/naturephysics 5

supplementary information doi: 10.1038/nphys1734

Since ρAB ≥ Π̄ρAB Π̄, we can apply Lemma 17 to (S ⊗ I)(ρAB ) and (S ⊗ I)(Π̄ρAB Π̄), which gives
ε ε
Hmax (SB)(S⊗I)(ρ) ≥ Hmax (SB)(S⊗I)(Π̄ρΠ̄) . (19)

The relation (17) then reduces to

ε 1 ε 1
Hmin (R|B)(R⊗I)(ΠΠ̄ρΠ̄Π) + Hmax (SB)(S⊗I)(ρ) ≥ log2 + Hmin (AB)ρ − 2 log2 . (20)
c ε
Finally, we apply Lemma 7 to (16) and (18), which gives
√
P (ρAB , Π̄ρAB Π̄) ≤ 4ε
√
P (Π̄ρAB Π̄, ΠΠ̄ρAB Π̄Π) ≤ 6ε .

Hence, by the triangle inequality

√ √ √ √
P (ρAB , ΠΠ̄ρAB Π̄Π) ≤ ( 4 + 6) ε < 5 ε .
√
Consequently, (R ⊗ I)(ΠΠ̄ρAB Π̄Π) has at most distance 5 ε from (R ⊗ I)(ρAB ). This implies
√
5 ε
Hmin (R|B)(R⊗I)(ρ) ≥ Hmin (R|B)(R⊗I)(ΠΠ̄ρΠ̄Π) .

Inserting this in (20) gives

√ 1 1
5 ε ε ε
Hmin (R|B)(R⊗I)(ρ) + Hmax (SB)(S⊗I)(ρ) ≥ log2 + Hmin (AB)ρ − 2 log2 ,
c ε
which completes the proof of Theorem 4.

V. TECHNICAL PROPERTIES

A. Properties of the purified distance

The purified distance between ρ and σ corresponds to the minimum trace distance between purifications of ρ and
σ, respectively [6]. Because the trace distance can only decrease under the action of a partial trace (see, e.g., [2]), we
obtain the following bound.
Lemma 5. For any ρ ∈ U≤ (H) and σ ∈ U≤ (H),

ρ − σ1 ≤ 2P (ρ, σ).

The following lemma states that the purified distance is non-increasing under certain mappings.
Lemma 6. For any ρ ∈ U≤ (H) and σ ∈ U≤ (H), and for any nonnegative operator Π ≤ 11,

P (ΠρΠ, ΠσΠ) ≤ P (ρ, σ). (21)

Proof. We use the fact that the purified distance is non-increasing under any trace-preserving completely positive map
(TPCPM) [6] and consider the TPCPM
 
E : ρ → ΠρΠ ⊕ tr( 11 − Π2 ρ 11 − Π2 ).

We have P (ρ, σ) ≥ P (E(ρ), E(σ)), which implies F̄ (ρ, σ) ≤ F̄ (E(ρ), E(σ)). Then,

F̄ (ρ, σ) ≤ F̄ (E(ρ), E(σ))

 
= F (ΠρΠ, ΠσΠ) + (trρ − tr(Π2 ρ))(trσ − tr(Π2 σ)) + (1 − trρ)(1 − trσ)

≤ F (ΠρΠ, ΠσΠ) + (1 − tr(Π2 ρ))(1 − tr(Π2 σ))
= F̄ (ΠρΠ, ΠσΠ),

which is equivalent to the statement of the Lemma.

6 nature physics | www.nature.com/naturephysics

doi: 10.1038/nphys1734
supplementary information

The second inequality is the relation

  
(trρ − tr(Π2 ρ))(trσ − tr(Π2 σ)) + (1 − trρ)(1 − trσ) ≤ (1 − tr(Π2 ρ))(1 − tr(Π2 σ)),
which we proceed to show. For brevity, we write trρ − tr(Π2 ρ) = r, trσ − tr(Π2 σ) = s, 1 − trρ = t and 1 − trσ = u.
We hence seek to show
√ √ 
rs + tu ≤ (r + t)(s + u).
For r, s, t and u nonnegative, we have
√ √  √
rs + tu ≤ (r + t)(s + u) ⇔ rs + 2 rstu + tu ≤ (r + t)(s + u)
⇔ 4rstu ≤ (ru + st)2
⇔ 0 ≤ (ru − st)2 .

Furthermore, the purified distance between a state ρ and its image ΠρΠ is upper bounded as follows.
Lemma 7. For any ρ ∈ U≤ (H), and for any nonnegative operator, Π ≤ 11,
1 
P (ρ, ΠρΠ) ≤ √ (trρ)2 − (tr(Π2 ρ))2 .
trρ
Proof. Note that

√  √ √ √ √
 ρ ΠρΠ1 = tr ( ρΠ ρ)( ρΠ ρ) = tr(Πρ) ,
so we can write the generalized fidelity (see (5)) as

F̄ (ρ, ΠρΠ) = tr(Πρ) + (1 − trρ)(1 − tr(Π2 ρ)) .
For brevity, we now write trρ = r, tr(Πρ) = s and tr(Π2 ρ) = t. Note that 0 ≤ t ≤ s ≤ r ≤ 1. Thus,

1 − F̄ (ρ, ΠρΠ)2 = r + t − rt − s2 − 2s (1 − r)(1 − t).
We proceed to show that r(1 − F̄ (ρ, ΠρΠ)2 ) − r2 + t2 ≤ 0:
  
r(1 − F̄ (ρ, ΠρΠ)2 ) − r2 + t2 = r r + t − rt − s2 − 2s (1 − r)(1 − t) − r2 + t2
� 
≤ r r + t − rt − s2 − 2s(1 − r) − r2 + t2
= rt − r2 t + t2 − 2rs + 2r2 s − rs2
≤ rt − r2 t + t2 − 2rs + 2r2 s − rt2
= (1 − r)(t2 + rt − 2rs)
≤ (1 − r)(s2 + rs − 2rs)
= (1 − r)s(s − r)
≤ 0.
This completes the proof.
Lemma 8. Let ρ ∈ U≤ (H) and σ ∈ U≤ (H)  have eigenvalues ri andsi ordered non-increasingly (ri+1 ≤ ri and
si+1 ≤ si ). Choose a basis |i such that σ = i si |ii| and define ρ̃ = i ri |ii|, then
P (ρ, σ) ≥ P (ρ̃, σ).
Proof. By the definition of the purified distance P (·, ·), it suffices to show that F̄ (ρ, σ) ≤ F̄ (ρ̃, σ).
 √ √
F̄ (ρ, σ) − (1 − trρ)(1 − trσ) =  ρ σ1
√ √
= max Re tr(U ρ σ)
U
√ √
≤ max Re tr(U ρV σ)
U,V
√ √ 
= ri si = F̄ (ρ̃, σ) − (1 − trρ̃)(1 − trσ).
i

The maximizations are taken over the set of unitary matrices. The second and third equality are Theorem 7.4.9 and
Equation (7.4.14) (on page 436) in [10]. Since trρ̃ = trρ, the result follows.

nature physics | www.nature.com/naturephysics 7

supplementary information doi: 10.1038/nphys1734

B. Basic properties of (smooth) min- and max-entropies

Smooth min- and max-entropies can be seen as generalizations of the von Neumann entropy, in the following
sense [5].
Lemma 9. For any σ ∈ U= (HAB ),
1 ε
lim lim H (An |B n )σ⊗n = H(A|B)σ
ε→0 n→∞ n min
1 ε
lim lim Hmax (An )σ⊗n = H(A)σ .
ε→0 n→∞ n

The von Neumann entropy satisfies the strong subadditivity relation, H(A|BC) ≤ H(A|B). That is, discarding
information encoded in a system, C, can only increase the uncertainty about the state of another system, A. This
inequality directly generalizes to (smooth) min- and max-entropies [3]. In this work, we only need the statement for
Hmin .
Lemma 10 (Strong subadditivity for Hmin [3]). For any ρ ∈ S≤ (HABC ),

Hmin (A|BC)ρ|ρ ≤ Hmin (A|B)ρ|ρ . (22)

Proof. By definition, we have

2−Hmin (A|BC)ρ|ρ 11A ⊗ ρBC − ρABC ≥ 0 .

Because the partial trace maps nonnegative operators to nonnegative operators, this implies

2−Hmin (A|BC)ρ|ρ 11A ⊗ ρB − ρAB ≥ 0 .

This implies that 2−Hmin (A|B)ρ|ρ ≤ 2−Hmin (A|BC)ρ|ρ , which is equivalent to the assertion of the lemma.
The chain rule for von Neumann entropy states that H(A|BC) = H(AB|C) − H(B|C). This equality generalizes
to a family of inequalities for (smooth) min- and max-entropies. In particular, we will use the following two lemmas.
Lemma 11 (Chain rule I). For any ρ ∈ S≤ (HABC ) and σC ∈ S≤ (HC ),

Hmin (A|BC)ρ|ρ ≤ Hmin (AB|C)ρ − Hmin (B|C)ρ .

Proof. Let σC ∈ S≤ (HC ) be arbitrary. Then, from the definition of the min-entropy we have

ρABC ≤ 2−Hmin (A|BC)ρ|ρ 11A ⊗ ρBC

≤ 2−Hmin (A|BC)ρ|ρ 2−Hmin (B|C)ρ|σ 11AB ⊗ σC .

This implies that 2−Hmin (AB|C)ρ|σ ≤ 2−Hmin (A|BC)ρ|ρ 2−Hmin (B|C)ρ|σ and, hence Hmin (A|BC)ρ|ρ ≤ Hmin (AB|C)ρ|σ −
Hmin (B|C)ρ|σ . Choosing σ such that Hmin (B|C)ρ|σ is maximized, we obtain Hmin (A|BC)ρ|ρ ≤ Hmin (AB|C)ρ|σ −
Hmin (B|C)ρ . The desired statement then follows because Hmin (AB|C)ρ|σ ≤ Hmin (AB|C)ρ .
Lemma 12 (Chain rule II). For any ρ ∈ S≤ (HAB ),

Hmin (AB)ρ − H−∞ (B)ρ ≤ Hmin (A|B)ρ|ρ .

Note that the inequality can be extended by conditioning all entropies on an additional system C, similarly to
Lemma 11. However, in this work, we only need the version stated here.
Proof. From the definitions,

ρAB ≤ 2−Hmin (AB) 11A ⊗ Πsupp(ρB )

≤ 2−Hmin (AB) 2H−∞ (B) 11A ⊗ ρB .

It follows that 2−Hmin (A|B)ρ|ρ ≤ 2−Hmin (AB) 2H−∞ (B) , which is equivalent to the desired statement.
The remaining lemmas stated in this appendix are used to transform statements that hold for entropies Hmin and
ε ε
HR into statements for smooth entropies Hmin and Hmax . We start with an upper bound on H−∞ in terms of Hmax .

8 nature physics | www.nature.com/naturephysics

doi: 10.1038/nphys1734
supplementary information

Lemma 13. For any ε > 0 and for any σ ∈ S≤ (HA ) there exists a projector Π which is diagonal in any eigenbasis
of σ such that tr((11 − Π)σ) ≤ ε and
1
Hmax (A)σ > H−∞ (A)ΠσΠ − 2 log2 .
ε

Proof. Let σ = i ri |ii| be a spectral decomposition of σ where the eigenvalues ri are ordered non-increasingly
(ri+1 ≤ ri ). Define the projector Πk := i≥k |ii|. Let j be the smallest index such that tr(Πj σ) ≤ ε and define
Π := 11 − Πj . Hence, tr(Πσ) ≥ tr(σ) − ε. Furthermore,
√ √ −1
tr σ ≥ tr(Πj−1 σ) ≥ tr(Πj−1 σ)Πj−1 σΠj−1 ∞2 .

We now use tr(Πj−1 σΠj−1 ) > ε and the fact that Πj−1 σΠj−1 ∞ cannot be larger than the smallest non-zero
eigenvalue of ΠσΠ,5 which equals 2−H−∞ (A)ΠσΠ . This implies
√ 
tr σ > ε 2H−∞ (A)ΠσΠ .

Taking the logarithm of the square of both sides concludes the proof.
Lemma 14. For any ε > 0 and for any σ ∈ S≤ (HA ) there exists a nonnegative operator Π ≤ 11 which is diagonal in
any eigenbasis of σ such that tr((11 − Π2 )σ) ≤ 2ε and
ε
Hmax (A)σ ≥ Hmax (A)ΠσΠ .
ε
Proof. By definition of Hmax (A)σ , there is a ρ ∈ B ε (σ) such that Hmax
ε
(A)σ = Hmax (A)ρ . It follows from Lemma 8
that we can take ρ to be diagonal in any eigenbasis of σ. Define

ρ := ρ − {ρ − σ}+ = σ − {σ − ρ}+

where {·}+ denotes the positive part of an operator. We then have ρ ≤ ρ, which immediately implies that Hmax (A)ρ ≤
Hmax (A)ρ . Furthermore, because ρ ≤ σ and because ρ and σ have the same eigenbasis, there exists a nonnegative
operator Π ≤ 11 diagonal in the eigenbasis of σ such that ρ = ΠσΠ. The assertion then follows because

tr((11 − Π2 )σ) = tr(σ) − tr(ρ ) = tr({σ − ρ}+ ) ≤ ρ − σ1 ≤ 2ε ,

where the last inequality follows from Lemma 5 and P (ρ, σ) ≤ ε.

Lemma 15. For any ε > 0 and for any σ ∈ S≤ (HA ) there exists a nonnegative operator Π ≤ 11 which is diagonal in
any eigenbasis of σ such that tr((11 − Π2 )σ) ≤ 3ε and

ε 1
Hmax (A)σ ≥ H−∞ (A)ΠσΠ − 2 log2 .
ε
Proof. By Lemma 14, there exists a nonnegative operator Π̄ ≤ 11 such that
ε
Hmax (A)σ ≥ Hmax (A)Π̄σΠ̄
¯ such that
and tr((11 − Π̄2 )σ) ≤ 2ε. By Lemma 13 applied to Π̄σ Π̄, there exists a projector Π̄
1
Hmax (A)Π̄σΠ̄ ≥ H−∞ (A)ΠσΠ − 2 log2
ε
¯ )Π̄σ Π̄) ≤ ε, where we defined Π := Π̄
and tr((11 − Π̄ ¯ Π̄. Furthermore, Π̄, Π̄
¯ and, hence, Π, can be chosen to be diagonal
in any eigenbasis of σ. The claim then follows because
¯ Π̄2 )σ) = tr((11 − Π̄2 )σ) + tr((11 − Π̄
tr((11 − Π2 )σ) = tr((11 − Π̄ ¯ )Π̄σ Π̄) ≤ 3ε.

5 If ΠσΠ has no non-zero eigenvalue then H−∞ (A)ΠσΠ = −∞ and the statement is trivial.

nature physics | www.nature.com/naturephysics 9

supplementary information doi: 10.1038/nphys1734

10

Lemma 16. Let ε ≥ 0, let σ ∈ S≤ (HA ) and let M : σ → i |φi φi |φi |σ|φi  be a measurement with respect to an
orthonormal basis {|φi }i . Then
ε ε
Hmax (A)σ ≤ Hmax (A)M(σ) .

Proof. The max-entropy can be written in terms of the (standard) fidelity (see also [4]) as

Hmax (A)σ = 2 log2 F (σA , 11A ).

Using the fact that the fidelity can only increase when applying a trace-preserving completely positive map (see, e.g.,
[2]), we have

F (σA , 11A ) ≤ F (M(σA ), M(11A )) = F (M(σA ), 11A ) .

Combining this with the above yields

Hmax (A)σ ≤ Hmax (A)M(σ) , (23)

which proves the claim in the special case where ε = 0.

To prove the general claim, let HS and HS  be isomorphic to HA and let U be the isometry from HA to span{|φi S ⊗
|φi S  }i ⊆ HS ⊗ HS  defined by |φi A → |φi S ⊗ |φi S  . The action of M can then equivalently be seen as that of U
 † 
followed by the partial trace over HS  . In particular, defining σSS  := U σA U , we have M(σA ) = σS .

Let ρ ∈ S(HSS  ) be a density operator such that
ε
Hmax (S)ρ = Hmax (S)σ (24)

and

P (ρSS  , σSS

) ≤ ε . (25)

(Note that, by definition, there exists a state ρS that satisfies (24) with P (ρS , σS ) ≤ ε. It follows from Uhlmann’s
theorem (see e.g. [2]) and the fact that the purified distance is non-increasing under partial trace that there exists an
extension of ρS such that (25) also holds.)
 
Since σSS  has support in the subspace span{|φi S ⊗ |φi S  }i , we can assume that the same is true for ρSS  . To see

this, define Π as the projector onto this subspace and observe that trS  (ΠρSS  Π) cannot be a worse candidate for the
ε
optimization in Hmax (S)σ : From Lemma 8, we can take ρS to be diagonal in the {|φi } basis, i.e. we can write

ρS = λi |φi φi |,
i

where λi ≥ 0. We also write


ρSS  = cijkl |φi φj | ⊗ |φk φl |,
ijkl

for some coefficients cijkl . To ensure ρS = trS  ρSS  , we require k cijkk = λi δij . Consider then
 

trS  (ΠρSS  Π) = trS  

cijij |φi φj | ⊗ |φi φj |
ij

= ciiii |φi φi |.
i

It follows that trS  (ΠρSS  Π) ≤ ρS (since k ciikk = λi and ciikk ≥ 0) and hence we have
ε
Hmax (S)trS (Πρ Π) ≤ Hmax (S)ρ .
SS 

Furthermore, from Lemma 6, we have

P (ΠρSS  Π, σSS
    
 ) = P (ΠρSS  Π, ΠσSS  Π) ≤ P (ρSS  , σSS  ) ≤ ε,

10 nature physics | www.nature.com/naturephysics

doi: 10.1038/nphys1734
supplementary information

11

from which it follows that

trS  (ΠρSS  Π) ∈ B ε (σS ).

We have hence shown that there exists a state ρSS  satisfying (24) and (25) whose support is in span{|φi S ⊗ |φi S  }i .
We can thus define ρA := U † ρSS  U so that ρS = M(ρA ) and hence (24) can be rewritten as
ε
Hmax (A)M(ρ) = Hmax (A)M(σ) ,

and (25) as

P (ρA , σA ) ≤ ε .

Using this and (23), we conclude that

ε ε
Hmax (A)M(σ) = Hmax (A)M(ρ) ≥ Hmax (A)ρ ≥ Hmax (A)σ .

Lemma 17. Let ε ≥ 0, and let σ ∈ S≤ (HA ) and σ  ∈ S≤ (HA ). If σ  ≤ σ then

ε ε
Hmax (A)σ ≤ Hmax (A)σ .

Proof. By Lemma 16, applied to an orthonormal measurement M with respect to the eigenbasis of σ, we have
ε ε
Hmax (A)σ ≤ Hmax (A)M(σ ) .

Using this and the fact that M(σ  ) ≤ M(σ) = σ, we conclude that it suffices to prove the claim for the case where
σ  and σ are diagonal in the same basis.
ε
By definition, there exists ρ such that P (ρ, σ) ≤ ε and Hmax (A)ρ = Hmax (A)σ . Because of Lemma 8, ρ can
be assumed to be diagonal in an eigenbasis of σ. Hence, there exists an operator Γ which is diagonal in the same
eigenbasis such that ρ = ΓσΓ. We define ρ := Γσ  Γ for which ρ ≥ 0 and tr(ρ ) ≤ tr(ρ) ≤ 1. Furthermore, since
ρ ≤ ρ, we have
ε
Hmax (A)ρ ≤ Hmax (A)ρ = Hmax (A)σ .

Because σ  and σ can be assumed to be diagonal in the same basis, there exists a nonnegative operator Π ≤ 11 which
is diagonal in the eigenbasis of σ (and, hence, of Γ and ρ) such that σ  = ΠσΠ. We then have

ρ = Γσ  Γ = ΓΠσΠΓ = ΠΓσΓΠ = ΠρΠ .

Using the fact that the purified distance can only decrease under the action of Π (see Lemma 6), we have

P (ρ , σ  ) = P (ΠρΠ, ΠσΠ) ≤ P (ρ, σ) ≤ ε .

ε
This implies Hmax (A)σ ≤ Hmax (A)ρ and thus concludes the proof.
Lemma 18. For any ε ≥ 0 and for any (normalized) σ ∈ S= (HA ), there exists a nonnegative operator Π ≤ 11 which
is diagonal in any eigenbasis of σ such that tr((11 − Π2 )σ) ≤ 2ε and
ε
Hmin (A)σ ≤ Hmin (A)ΠσΠ .

Proof. Let ρ ∈ B ε (σ) be such that Hmin (A)ρ = Hmin ε

(A)σ . It follows from Lemma 8 that
we can take ρ to be diagonal

in an eigenbasis |i of σ. Let ri (si ) be the list of eigenvalues of ρ (σ) and define σA = i min(ri , si )|ii|. It is easy
to see that there exists a nonnegative operator Π ≤ 11 such that σ  = ΠσΠ. Since σ  ≤ ρ, we have
ε
Hmin (A)ΠσΠ = Hmin (A)σ ≥ Hmin (A)ρ = Hmin (A)σ .

Furthermore, tr((11 − Π2 )σ) = tr(σ − σ  ) = i: si ≥ri (si − ri ) ≤ σ − ρ1 . The assertion then follows because, by
Lemma 5, the term on the right hand side is bounded by 2P (σ, ρ) ≤ 2ε.

nature physics | www.nature.com/naturephysics 11

supplementary information doi: 10.1038/nphys1734

12

[1] Renes, J. M. & Boileau, J.-C. Conjectured strong complementary information tradeoff. Physical Review Letters 103,
020402 (2009).
[2] Nielsen, M. A. & Chuang, I. L. Quantum Computation and Quantum Information (Cambridge University Press, 2000).
[3] Renner, R. Security of Quantum Key Distribution. Ph.D. thesis, ETH Zürich (2005). URL http://arxiv.org/abs/
quant-ph/0512258.
[4] König, R., Renner, R. & Schaffner, C. The operational meaning of min- and max-entropy. IEEE Transactions on
Information Theory 55, 4337–4347 (2009).
[5] Tomamichel, M., Colbeck, R. & Renner, R. A fully quantum asymptotic equipartition property. IEEE Transactions on
information theory 55, 5840–5847 (2009).
[6] Tomamichel, M., Colbeck, R. & Renner, R. Duality between smooth min- and max-entropies (2009). URL http://arxiv.
org/abs/0907.5238.
[7] Maassen, H. & Uffink, J. B. Generalized entropic uncertainty relations. Physical Review Letters 60, 1103–1106 (1988).
[8] Rényi, A. On measures of information and entropy. In Proceedings 4th Berkeley Symposium on Mathematical Statistics
and Probability, 547–561 (1961).
[9] Christandl, M. & Winter, A. Uncertainty, monogamy and locking of quantum correlations. IEEE Transactions on
Information Theory 51, 3159–3165 (2005).
[10] Horn, R. A. & Johnson, C. R. Matrix Analysis (Cambridge University Press, 1985).

12 nature physics | www.nature.com/naturephysics