Quantum information theory (MAT4430) Spring 2021

Lecture 5: Convexity and bipartite entanglement

Lecturer: Alexander Müller-Hermes

We have now introduced the formalism of open quantum systems, which describes general
quantum systems. Before we turn to quantum information processing protocols, we have to
discuss a special feature of quantum systems, which is not present in classical systems:
Entanglement. Roughly speaking, entanglement is a special type of correlation of two (or
more) quantum systems and it serves as a ressource in quantum information processing. In
the exercises, we saw that maximal entanglement enables superdense coding and quantum
teleportation, and we will see more examples throughout the course. Here, we will focus on
entanglement between two quantum systems, which is known as bipartite entanglement. The
case of multipartite entanglement is much more complicated, and will not be covered in this
course. To understand the basics of entanglement theory, we need some basic terminology
for talking about convex cones.

1 Convexity and cones

Let us start with two basic definitions:
Definition 1.1 (Convex sets and cones). Let V denote a real Euclidean space.
1. A set B ⊂ V is called convex if λx + (1 − λ)y ∈ B whenever x, y ∈ B and λ ∈ [0, 1].
2. A set C ⊂ V is called a cone if λx + µy ∈ C for all x, y ∈ C and all λ, µ > 0.
It is clear from these definitions that every cone is a convex set, but not every convex set
is a cone. In particular, non-empty compact convex sets are never cones. It is also clear that
0 ∈ C for every cone C, but obviously 0 is not necessarily contained in a convex set. We call
a convex set B ⊂ V a convex body if it is compact and has non-empty interior. When 0 ∈ /B
for some convex body B ⊂ V, then we can form a cone C ⊂ V by taking
cone (B) = {λx : x ∈ B and λ > 0}.
In this case, B is called a base for the cone C = cone (B). It is easy to show that cone (B)
is a closed cone for every convex body B not containing 0. We will use the notation cone(·)
more generally to denote the conic hull of a set, i.e., the smallest cone containing the set.
This is similar to the convex hull conv(·). We will also need the following concepts:
Definition 1.2 (Extreme points and extremal rays). Let V denote a real Euclidean space
and B ⊂ V a convex set and C ⊂ V a cone.
1. A point z ∈ B is called an extreme point if z = λx+(1−λ)y for x, y ∈ B and λ ∈ (0, 1)
implies that x = y = z.
2. A point x ∈ C generates an extremal ray R+0x ⊆ C if x − y ∈ C for y ∈ C implies
that y ∈ +0 x.R
When B ⊂ V is a convex body not containing 0, then the extremal rays of the cone
C = cone (B) are generated by the extreme points of B. A convex body B can be expressed
as the convex hull1 of its extreme points
B = conv (Ext(B)) ,
1
Note that the closure is not needed here, since we restrict to compact sets in finite-dimensional spaces.

1
which is a special case of the Krein-Milman theorem. We will sometimes need the following
easy facts:
Theorem 1.3. Let B ⊂ V denote a convex body and f : B → R a continuous function.
1. If f is convex, then supx∈B f (x) = maxx∈Ext(B) f (x).

2. If f is concave, then inf x∈B f (x) = minx∈Ext(B) f (x).

The following theorem gives an upper bound on the number of extreme points needed to
express a given point as a convex combination:
Theorem 1.4 (Caratheodory). If B ⊂ V is a convex hull B = conv (S) of a set S ⊆ V, then
any y ∈ B admits a decomposition
XN
y= pi xi ,
n=1

for some probability distribution p ∈ P ({1, . . . , N }) and some x1 , . . . , xN ∈ S with

N 6 dim(V) + 1.

We have the following corollary of Caratheodory’s theorem:

Corollary 1.5. For any compact set S ⊂ V the convex hull conv (S) is compact.
Proof. Let d = dim(V). We will show that conv (S) is sequentially compact, which, in
the Euclidean space V, is equivalent to conv (S) being compact. Consider a sequence
(yn )n∈N ∈ conv (S)N . By Caratheodory’s theorem there are probability distributions p(n) ∈
(n) (n)
P ({1, . . . , d + 1}) and (x1 , . . . , xd+1 ) ∈ S d+1 such that
d+1
(n) (n)
X
yn = pi xi ,
i=1

N
for each n ∈ . Since P ({1, . . . , d + 1}) and S are compact sets, there exist convergent
subsequences (p(nk ) )k∈N ∈ P ({1, . . . , d + 1})N and (xi k ) ∈ S such that
(n )

p(nk ) → p ∈ P ({1, . . . , d + 1})

and
(nk )
xi → xi ∈ S,
as k → ∞, for each i ∈ {1, . . . , d + 1}. We conclude that
d+1
X
ynk → pi xi ∈ conv(S),
i=1

as k → ∞, which finishes the proof.

As for convex bodies, we also have that the cone C = cone (B) generated by a convex body
B ⊂ V not containing 0 is the conic hull of the union of its extremal rays. By Caratheodory’s
theorem, every x ∈ C can be written as
N
X
x= xi ,
i=1

for some N 6 dim(V) + 1 and generators xi of extremal rays. When studying a cone it is
often useful to consider its dual:

2
Definition 1.6 (Dual cone). For any cone C ⊂ V in a real Euclidean space V we define the
dual cone C ∗ by
C ∗ = {y ∈ V : hy, xi > 0 for any x ∈ C}.

It is easy to check that C ∗ ⊂ V is a closed cone whenever C is a cone. The most important
result in cone duality is the bipolar theorem. This theorem is a direct consequence of the
hyperplane separation theorem, which we state for convenience in a special case:

/ C. Then, there exists y ∈ C ∗

Theorem 1.7. If C ⊂ V is a non-empty closed cone and z ∈
such that
hy, zi < 0.

For convenience, we will state the proof of this theorem. It needs the following lemma:

Lemma 1.8. Let K ⊂ V be a non-empty closed convex subset of a real Euclidean space V.
Then, there exists a unique xmin ∈ K such that

kxmin k2 = inf kxk2 .

x∈K

Proof. Consider a sequence (xn )n∈N ∈ K N such that limn→∞ kxn k2 = inf x∈K kxk2 =: δ.
Note that
kxn + xm k22 > 4δ 2 ,
for any n, m ∈ N, since (xn + xm)/2 ∈ K. Now, we compute
kxn − xm k22 = 2kxn k22 + 2kxm k22 − kxn + xm k22 6 2kxn k22 + 2kxm k22 − 4δ 2 → 0,

as n, m → ∞. We conclude that (xn )n∈N ∈ K N is a Cauchy sequence and by completeness

of V and closedness of K there is a point xmin ∈ K with kxmin k2 = inf x∈K kxk2 . Assume
now that there is a point x0 ∈ K with kx0 k2 = inf x∈K kxk2 . Then, we have

kxmin − x0 k22 6 2kxmin k22 + 2kx0 k22 − 4δ 2 = 0,

showing that xmin = x0 .

Now, we can prove the hyperplane separation theorem:

Proof of Theorem 1.7. Consider the closed and convex set K = {x − z : x ∈ C} and note
that 0 ∈
/ K. By Lemma 1.8 there exists a vmin ∈ K such that

kvmin k2 = inf kvk2 =: δ,

v∈K

and since 0 ∈
/ K, we have δ > 0. For any v ∈ K and any λ ∈ [0, 1] we have

(1 − λ)vmin + λv = vmin + λ (v − vmin ) ∈ K.

Therefore, we have

δ 2 6 kvmin + λ (v − vmin ) k22 = δ 2 + 2λhvmin , v − vmin i + λ2 kv − vmin k22 .

Since this inequality holds for all λ ∈ [0, 1] (in particular for λ → 0), we conclude that

hvmin , vi > δ 2 ,

for any v ∈ K. By definition of K, this implies that

hvmin , xi > δ 2 + hvmin , zi, (1)

3
for every x ∈ C. Inserting x = 0 into (1) shows that

hvmin , zi 6 −δ 2 < 0.

Moreover, since (1) holds for any λx ∈ C with x ∈ C and any λ > 0, we conclude that

hvmin , xi > 0,

for any x ∈ C. Therefore, we have vmin ∈ C ∗ and the proof is finished.

Now, we can state:

Theorem 1.9 (Bipolar theorem). For any closed cone C ⊂ V in a real Euclidean space V
we have
(C ∗ )∗ = C.

Proof. It is clear that (C ∗ )∗ ⊇ C. For the other direction consider z ∈

/ C. By the hyperplane
∗
separation theorem there exists a y ∈ C such that hy, zi = hz, yi < 0 and we conclude that
z∈/ (C ∗ )∗ .

We will sometimes need a few special properties of cones, which we now prove:

Lemma 1.10 (Interior points). Let C ⊂ V be a closed cone. We have y ∈ int(C ∗ ) if and
only if
hy, xi > 0,
for every x ∈ C \ {0}.

Proof. By definition we have y ∈ int(C ∗ ) if some -ball with  > 0 and center y satisfies
B (y) ⊂ C ∗ . Fix an x ∈ C \ {0} and note that

inf hy + z, xi = hy, xi − kxk2 . (2)

kzk2 61

Therefore, hy, xi > 0 has to be satisified if y ∈ int(C ∗ ). On the other hand, if hy, xi > 0
for every x ∈ C \ {0}, then we can choose  = inf x∈KC hy, xi > 0 where KC = S ∩ C is the
compact set arising as the intersection of the unit sphere

S = {x ∈ V : kxk2 = 1},

and the cone C. By this definition, we have that

kxk2 6 hy, xi,

for any x ∈ C and by (2) we are done.

The following lemma is useful as well:

Lemma 1.11. Let C ⊂ V be a closed cone. The following are equivalent:

1. The cone C is pointed, i.e., C ∩ (−C) = {0}.

2. The cone C ∗ is generating, i.e., we have C ∗ + (−C ∗ ) = V.

3. The cone C ∗ has non-empty interior.

4
Proof. If C ∗ P
is generating, then there exists a basis {y1 , . . . , yd } ⊂ C ∗ and we define the
element e = di=1 yi ∈ C ∗ . Whenever
d
X
he, xi = hyi , xi = 0,
i=1

for some x ∈ C we can conclude that x = 0, and therefore e ∈ int (C ∗ ) by Lemma 1.10. For
any e ∈ int (C ∗ ) we have he, xi > 0 whenever x ∈ C \ {0} and we conclude immediately that
C is pointed. Finally, assume that C ∗ is not generating and let {y1 , . . . , yk } ⊂ C ∗ a maximal
independent set for k < d. Consider

x ∈ span{y1 , . . . , yk }⊥ 6= ∅.

This x ∈ V satisfies
hy, xi = hy, −xi = 0,
for every y ∈ C ∗ and by the bipolar theorem we conclude that both x ∈ C and −x ∈ C.
This shows that C is not pointed.

For any closed and pointed cone C ⊂ V, we may consider y ∈ int(C ∗ ) and define a convex
set B ⊂ C by B = C ∩ H, where

H = {x ∈ V : hy, xi = 1}.

It can be verified that B = C ∩ H is a convex body not containing zero, and hence it is a
base for the cone C. This means that we can always pass between a closed and pointed cone
and its compact base.
We will need a few other results about duality of cones. For cones C1 , C2 ⊂ V we define
their intersection C1 ∩ C2 and their join

C1 ∨ C2 = {x + y : x ∈ C1 , y ∈ C2 }.

These two operations are dual to each other. We summarize this and another fact in the
following lemma:

Lemma 1.12. For closed cones C1 , C2 ⊂ V in a real Euclidean space V we have:

1. If C1 ⊆ C2 , then C1∗ ⊇ C2∗ .

2. We have (C1 ∩ C2 )∗ = C1∗ ∨ C2∗ .

Proof. See exercises.

2 Example: Quantum states and the cone B(H)+

To get more familiar with the terminology from the previous section let us apply it to the set
D (H) of quantum states and the set B(H)+ of positive semidefinite matrices on a complex
Euclidean space H. Note that both of these sets are contained in the space B(H)sa of
selfadjoint operators. In the following, we will consider the real Euclidean space V = B(H)sa
equipped with the Hilbert-Schmidt inner product. The following points can be verified easily:

• The set B(H)+ ⊂ B(H)sa is a closed and pointed cone.

• The cone B(H)+ ⊂ B(H)sa is also generating. Specifically, any operator Z ∈ B(H)sa
can be written as Z = X − Y for X, Y ∈ B(H)+ such that XY = 0. This is called the
Jordan-Hahn decomposition and is an easy consequence of the spectral theorem.

5
 • We have hY, XiHS = Tr [Y X] > 0 for every X ∈ B(H)+ if and only if Y ∈ B(H)+ .
∗
Therefore, we have (B(H)+ ) = B(H)+ and we say that B(H) is selfdual.
∗

• We have 1H ∈ int (B(H)+ ) since Tr [X] = 0 implies X = 0 whenever X ∈ B(H)+ .
• The quantum states D (H) are a convex body and a compact base of the cone B(H)+
and we have
D (H) = B(H)+ ∩ {X ∈ B(H)sa : Tr [X] = 1}.

Using the spectral decomposition we can show the following theorem characterizing the
extreme points of D (H):
Theorem 2.1. The extreme points of D (H) are the pure states |ψihψ| ∈ Proj (H).
Note that the spectral decomposition gives a much better bound than Caratheodory’s
theorem on the number of pure states needed to express a given quantum state as a convex
combination.

3 Separability and entanglement

We start with the central definition of this lecture:
Definition 3.1 (Separability). For Euclidean spaces HA and HB we define the cone
Sep (HA , HB ) = cone{XA ⊗ YB : XA ∈ B(HA )+ , YB ∈ B(HB )+ } ⊂ B(HA ⊗ HB )sa .
We call a positive operator XAB ∈ B(HA ⊗ HB )+ separable if XAB ∈ Sep (HA , HB ), and
otherwise we will call XAB entangled.
We will call a quantum state ρAB ∈ D(HA ⊗ HB ) separable if ρAB ∈ Sep (HA , HB )
and otherwise it is called entangled. Entanglement plays a fundamental role in quantum
information theory, and it can be seen as a resource enabling many useful quantum protocols
such as teleportation or superdense coding. Such protocols are often based on entangled pure
states (e.g., maximally entangled states), and there are many open problems related to the
generation of pure entangled quantum states from mixed entangled quantum states.
To motivate the notion of entanglement further let us mention the following simple obser-
vation showing that local measurements of separable quantum states is inherently classical:
Theorem 3.2 (Local hidden variables). For Euclidean spaces HA and HB consider a pair
of POVMs {Pn }N + M +
n=1 ⊂ B(HA ) and {Qm }m=1 ⊂ B(HB ) . For any separable quantum state
ρAB ∈ Sep (HA , HB )∩D(HA ⊗HB ), there exists a probability distribution r ∈ P ({1, . . . , K})
and conditional probability distributions pA (·|k) ∈ P ({1, . . . , N }) and q B (·|k) ∈ P ({1, . . . , M })
such that
XK
Prob (n, m) = r(k)pA (n|k)q B (m|k).
k=1

Proof. The separable quantum state ρAB ∈ Sep (HA , HB ) ∩ D(HA ⊗ HB ) can be written as
K
X
ρAB = r(k)σkA ⊗ τkB ,
k=1

for a probability distribution r ∈ P ({1, . . . , K}) and quantum states σkA ∈ D(HA ) and
τkB ∈ D(HB ) for any k ∈ {1, . . . , K}. Now, we define the conditional probability distributions
by
pA (n|k) = Tr Pn σkA and q B (m|k) = Tr Qm τkB ,
   

for any k ∈ {1, . . . , K}, and the proof is finished.

6
 The previous theorem shows that the correlations observed in the outcome statistics of
local measurements of a separable quantum state are classical. In the 1930s it was proposed
that the weird effects of quantum theory can be explained by so-called hidden variables,
i.e., that there are classical quantities underlying the physical reality, which are unknown
(or hidden) and influence the outcome of experiments. In particular, such hidden variables
might become classically correlated and then influence the outcomes of local measurements
of an entangled pure state so that such outcomes come out with the same value. The
previous theorem shows that local measurements of separable states indeed follow such a
local hidden variable model with the hidden variable being the value of k. However, we will
see later that many2 entangled quantum states give rise to measurement statistics under
local measurements that are incompatible with any local hidden variable model.

4 Positive maps and entanglement witnesses

We will now compute the dual cone of Sep (HA , HB ). For this we will need the Choi-
Jamiolkowski isomorphism and the set of positive maps, i.e., linear maps P : B(HA ) →
B(HB ) such that P (B(HA )+ ) ⊆ B(HB )+ . Let us start with two elementary observations:

Lemma 4.1 (Decompositions of operators). For a Euclidean space H consider an operator

Z ∈ B(H). Then, there exist positive operators X1 , X2 , X3 , X4 ∈ B(H)+ such that

Z = X1 − X2 + i(X3 − X4 ).

Proof. For Z ∈ B(H) define

H1 = (Z + Z † )/2 ∈ B(H)sa and H2 = i(Z † − Z)/2 ∈ B(H)sa ,

which are also called the real and imaginary part of Z. It is easy to check that

Z = H1 + iH2 .

Now, we can use the Jordan-Hahn decomposition to show find positive operators X1 , X2 ∈
B(H)+ satsifying H1 = X1 − X2 and positive operators X3 , X4 ∈ B(H)+ satsifying H2 =
X3 − X4 . These operators give rise to the decomposition from the theorem.

Lemma 4.2. For Euclidean spaces HA and HB and any positive map P : B(HA ) → B(HB )
we have:

1. The Choi operator CP ∈ B(HA ⊗ HB ) is selfadjoint.

2. The adjoint P ∗ : B(HB ) → B(HA ) is a positive map.

Proof. By Lemma 4.1 any Z ∈ B(HA ) admits the decomposition Z = X1 − X2 + i(X3 − X4 )

with X1 , X2 , X3 , X4 ∈ B(HA )+ . Now, we note that

P (Z)† = (P (X1 ) − P (X2 ) + i(P (X3 ) − P (X4 )))†

= P (X1 ) − P (X2 ) − i(P (X3 ) − P (X4 )) = P (Z † ),
2
but not all!

7
for any Z ∈ B(HA ). We conclude that
 †
dA
(CP )† = 
X
|iA ihjA | ⊗ P (|iA ihjA |)
i,j=1
dA
|jA ihiA | ⊗ P (|iA ihjA |)†
X
=
i,j=1
dA
X
= |jA ihiA | ⊗ P (|jA ihiA |) = CP .
i,j=1

This shows the first statement.

For the second statement note that Tr [XY ] > 0 for any X ∈ B(H)+ holds if and only if
Y ∈ B(H)+ , which is equivalent to the aforementioned fact that the cone B(H)+ is selfdual
for any Euclidean space H. Now, note that

Tr [XP (Y )] = Tr [P ∗ (X)Y ] = Tr [Y P ∗ (X)] > 0

for any X ∈ B(HB )+ and any Y ∈ B(HA )+ whenever the linear map P : B(HA ) → B(HB )
is positive.

Theorem 4.3 (Block-positive cone). For Euclidean spaces HA and HB we define the cone
of block-positive operators by

BP (HA , HB ) = {CP : P : B(HA ) → B(HB ) positive map} ⊂ B(HA ⊗ HB )sa .

Then, we have
Sep (HA , HB )∗ = BP (HA , HB ) .
Proof. By Lemma 4.2, adjoints of positive maps are again positive and we find that

hCP , XA ⊗ YB iHS = h(idA ⊗ P ) (ωAB ) , XA ⊗ YB iHS = hωAB , XA ⊗ P ∗ (YB )iHS > 0,

for any positive map P : B(HA ) → B(HB ), any XA ∈ B(HA )+ and any YB ∈ B(HB )+ .
Since the tensor products XA ⊗ YB for XA ∈ B(HA )+ and YB ∈ B(HB )+ generate the cone
Sep (HA , HB ), we conclude that Sep (HA , HB )∗ ⊇ BP (HA , HB ).
For the other direction consider WAB ∈ Sep (HA , HB )∗ ⊂ B(HA ⊗ HB )sa . By the Choi-
Jamiolkowski isomorphism there exists a linear map Q : B(HA ) → B(HB ) such that WAB =
CQ . As before, we can compute

hWAB , XA ⊗ YB iHS = hωAB , XA ⊗ Q∗ (YB )iHS = Tr XA

 T ∗ 
Q (YB ) ,

where we used the necklace identities in the final equality. By assumption, we have
 T ∗ 
hWAB , XA ⊗ YB iHS = Tr XA Q (YB ) > 0

for any XA ∈ B(HA )+ and any YB ∈ B(HB )+ . Therefore, we conclude that Q∗ : B(HB ) →
B(HA ) and also Q : B(HA ) → B(HB ) are positive maps, which implies that WAB ∈
BP (HA , HB ). This finishes the proof.

It is not so easy to decide whether operators belong to the cone BP (HA , HB ). The
following lemma sometimes helps a little with this:
Lemma 4.4. For XAB ∈ B(HA ⊗ HB )sa the following are equivalent:
1. We have XAB ∈ BP (HA , HB ).

8
 2. For every |xi ∈ HA and every |yi ∈ HB we have

(hx| ⊗ hy|)XAB (|xi ⊗ |yi) > 0.

Proof. This lemma is simply the fact that the operators |xihx| ⊗ |yihy| for |xi ∈ HA and
|yi ∈ HB generate the extremal rays of Sep (HA , HB ).

Operators WAB ∈ BP (HA , HB ) \ B(HA ⊗ HB )+ are sometimes called entanglement

witnesses since they “witness” the entanglement in some entangled operators. The following
corollary (which is implicit already in the proof of the previous theorem) makes this point
of view a bit more explicit:

Corollary 4.5 (Entanglement witnesses). Consider a selfadjoint operator XAB ∈ B(HA ⊗

HB )sa for Euclidean spaces HA and HB . The following are equivalent:

1. The operator XAB is separable.

2. For any WAB ∈ BP (HA , HB ) we have hWAB , XAB iHS > 0.

3. For any positive map P : B(HB ) → B(HA ) we have (idA ⊗ P ) (XAB ) ∈ B(HA ⊗HA )+ .

Proof. The equivalence of 1. and 2. follows immediately from Theorem 4.3. It is also clear
that
(idA ⊗ P ) (XA ⊗ YB ) = XA ⊗ P (YB ) ∈ B(HA ⊗ HA )+ ,
for any positive map P : B(HB ) → B(HA ) and any XA ∈ B(HA )+ and any YB ∈ B(HB )+ .
Therefore, 1. implies 3.. To see that 3. implies 2. note that

hWAB , XAB iHS = hωAB , (idA ⊗ P ) (XAB )iHS = hΩAB | (idA ⊗ P ) (XAB ) |ΩAB i > 0,

for the positive map P : B(HB ) → B(HA ) defined such that WAB = CP ∗ .

We finish this section with a few elementary observations about the convex structures of
the cones Sep (HA , HB ) and BP (HA , HB ):

• Clearly, the cones Sep (HA , HB ) and BP (HA , HB ) are closed and pointed.

• By duality we conclude that both cones are generating as well and therefore they have
non-empty interiors.

5 The positive partial transpose criterion

How can we tell whether a quantum state is entangled or not? In general, this question
turns out to be difficult. It has been shown that deciding whether a quantum state ρAB ∈
C C
D( dA ⊗ dB ) is -close to the set of separable quantum states is an NP-hard problem in the
dimensions dA and dB and if  is scaling like the inverse of a polynomial in dA and dB . It is
unlikely3 that this problem can be solved efficiently on a classical (or quantum) computer.
Even though the separability problem is not efficiently solvable, there are neccessary
conditions for separability, which can be quite helpful. The easiest conditions are derived
from Corollary 4.5 by choosing a particular positive map P : B(HB ) → B(HA ). One of the
most well-studied choices is given by the transpose map:
3
and would yield a 1.000.000$ prize

9
Definition 5.1 (Partial transpose). For a selfadjoint operator XAB ∈ B(HA ⊗ HB )sa we
define its (right) partial transpose by
Γ
XAB = (idA ⊗ ϑB ) (XAB ) .

We say that XAB has positive partial transpose or that XAB is PPT if
Γ
XAB ∈ B(HA ⊗ HB )+ .

The set of PPT operators4 will be denoted by

Γ
PPT (HA , HB ) = {XAB ∈ B(HA ⊗ HB )sa : XAB ∈ B(HA ⊗ HB )+ }.

By Corollary 4.5 we can show that a quantum state is entangled by showing that it is not
PPT. An example of this is the (unnormalized) maximally entangled state ωd ∈ B( d ⊗ d )+ C C
where it can be checked that

ωdΓ = Cϑ = Fd ∈/ B(Cd ⊗ Cd)+.

F C C
The operator d ∈ B( d ⊗ d ) arising as the Choi operator of the transpose map is also
F
called the flip operator due to its property d (|vi ⊗ |wi) = |wi ⊗ |vi for any |vi, |wi ∈ d . C
More generally, we note the following theorem:

Theorem 5.2. We have Sep (HA , HB ) ⊆ PPT (HA , HB ) ∩ B(HA ⊗ HB )+ .

Remarkable, there is a converse of the previous theorem in small dimensions:

C C C C
Theorem 5.3. We have Sep dA , dB = PPT dA , dB ∩ B( dA ⊗ dB )+ if and only



C C
if
(dA , dB ) ∈ {(2, 2), (2, 3), (3, 2)},
or if dA = 1 or dB = 1.

The case (dA , dB ) = (2, 2) is commonly attributed to Størmer (although he disagrees

with this), and the cases (dA , dB ) ∈ {(2, 3), (3, 2)} are due to Woronowicz. The proofs of
these two cases are very different and we will not present the proof of Woronowicz’ result.
However, recently an “easy” method has been discovered by Aubrun and Szarek for how to
prove the case of (dA , dB ) = (2, 2). We will do most of this proof in the exercises and the
lecture.

4
Note that some authors use the term PPT to mean positive and having positive partial transpose, which
we do not!

10