The Quantum Steering Ellipsoid

Sania Jevtic, Matthew Pusey, David Jennings, and Terry RudolphControlled Quantum Dynamics Theory, Department of Physics, Imperial College London, London SW7 2AZ

(Dated: July 29, 2019)

The quantum ellipsoid of a two qubit state is the set of Bloch vectors that Bob can collapse Alice’squbit to when he implements all possible measurements on his qubit. We provide an elementaryconstruction of the ellipsoid for arbitrary states, and explain how this geometric representation canbe made faithful. The representation provides a range of new results, and uncovers new features,such as the existence of ‘incomplete steering’ in separable states. We show that entanglement can beanalysed in terms of three geometric features of the ellipsoid, and prove that a state is separable ifand only if it obeys a ‘nested tetrahedron’ condition. We provide a volume formula for the ellipsoidin terms of the state density matrix, which identifies exactly when steering is fully three dimensional,and identify this as a new feature of correlations, intermediate between discord and entanglement.

PACS numbers: 03.65.Ta, 03.67.Mn

The Bloch sphere provides a remarkably simple repre-sentation for the state space of the most primitive quan-tum unit - the single qubit system - and results in geomet-ric intuitions that are invaluable in countless fundamentalinformation-processing scenarios. Unfortunately, such adirect three-dimensional depiction of the state space isimpossible for any quantum system of larger dimension.

While the qubit is the primitive unit of quantum in-formation, the two qubit system constitutes the prim-itive unit for the theory of bipartite quantum correla-tions. However, the two qubit state space is already 15-dimensional and possesses a surprising amount of struc-ture and complexity. As such, it is a highly non-trivialchallenge to both faithfully represent its states and toacquire natural intuitions for their properties [1–3].

The phenomenon of steering in quantum states wasfirst uncovered by Schrodinger [4] (and subsequently re-discovered by others [5–7]), who realised that measure-ments on B on the pure entangled state |ψ〉AB could beused to “steer” the state at A into all different convexdecompositions of the reduced state ρA; in particular forrank-one POVM measurements at B we have that A issteered into ensembles of pure states. More significantly,given any ensemble decomposition of ρA into either pureor mixed states there always exists a POVM measure-ment at B that generates that ensemble. In this regardwe say that the steering at A for the pure state |ψ〉AB iscomplete within the Bloch sphere of A. More generally,for a 2-qubit mixed state ρAB it is known [8] that theconvex set of states that A can be steered to is insteadan ellipsoid EA that contains ρA, as in figure 1.

In this Letter we show that all correlation features of atwo qubit state are encoded in its steering ellipsoid andlocal Bloch vectors, and argue that this ellipsoid repre-sentation for two qubits provides the natural generaliza-tion of the Bloch sphere picture, in that it gives a simplegeometric representation of an arbitrary two qubit stateρ in three dimensions, which makes the key properties ofthe state manifest in simple geometric terms.

By adopting the ellipsoid representation we are leadto a range of novel results and fresh insights into quan-

FIG. 1: Ellipsoid representation of a two-qubit state.For any two-qubit state ρAB , the set of states to which Bobcan steer Alice forms an ellipsoid EA in Alice’s Bloch sphere,containing her Bloch vector a. The inclusion of Bob’s Blochvector b determines ρAB up to a choice of basis for Bob, whichcan be fixed by indicating the orientation of EA.

tum correlations for both entangled and separable states.Firstly, we provide an alternative construction to thesteering ellipsoid EA to that in [9], which applies evenwhen EA is degenerate. We then provide a reconstruc-tion of a state ρAB from its geometric data to explore thefaithfulness of the representation. This analysis reveals anew phenomenon of incomplete steering for certain sep-arable quantum states, in which some decompositions ofρA within of the steering ellipsoid EA are inaccessible.

The representation then allows us to decompose entan-glement into simple geometric features, and show how itdepends only on (a) the spatial orientation of the ellip-soid, (b) its distance from the origin and (c) its size. Therepresentation also leads us to establish the surprisingnested tetrahedron condition: a state ρAB is separableif and only if its ellipsoid fits inside a tetrahedron thatitself fits inside the Bloch sphere. We then study the min-imal number of product states in the ensemble of a sep-arable state, which we show is determined solely by thedimension of EA. We then develop a useful volume for-

arX

iv:1

303.

4724

v1 [

quan

t-ph

] 1

9 M

ar 2

013

2

mula that identifies exactly when steering is fully three-dimensional. Finally the work suggests a new feature ofquantum correlations, called obesity, which is neither dis-cord nor separability nor entanglement. We observe thatquantum discord arises from a combination of both theobesity of the state and the orientation of its ellipsoid.

Beyond these new results, we also feel that the methodof compactly depicting any two qubit state ρAB in threedimensions, via a single ellipsoid and two local Blochvectors, should be of interest to a range of researchers inboth the theoretical and experimental quantum sciences.

The Pauli basis. Let σ0 denote the 2×2 identity matrixand σi, i = 1, 2, 3 be the usual Pauli matrices. Anysingle-qubit Hermitian operator E can be written E =12

∑3µ=0Xµσµ, where the Xµ = tr(Eσµ) are real. E is

positive iff X0 ≥ 0 and X20 ≥

∑3i=0X

2i . With X viewed

as a 4-vector, this identifies the set of positive operatorswith the usual forward light cone in Minkowski space.In the case of a qubit state, X0 = 1 and the remainingcomponents form the Bloch vector.

Similarly, we can write a two-qubit state as ρ =14

∑3µ,ν=0 Θµνσµ ⊗ σν . The components of the 4 × 4

real matrix Θ are Θµν = tr(ρσµ⊗σν). As a block matrix

we have Θ =

(1 bT

a T

), where a, b are the Bloch vec-

tors of the reduced states ρA and ρB of ρ and T is a 3× 3 matrix encoding correlations [2]. Now if Bob does ameasurement and obtains a POVM outcome E, then Bobsteers Alice to the state proportional to trB(ρ(σ0 ⊗ E)),with the latter being given by 1

2ΘX in the Pauli basis.The matrix Θ transforms, up to a normalization, un-

der SLOCC operations ρ′ = SA⊗SBρ(SA⊗SB)†, whereSA, SB are invertible complex matrices, as Θ′ = ΛAΘΛTB[9] where ΛA,ΛB are proper orthochronous Lorentztransformations (see the appendices).

In the case of a SLOCC operation affecting only Bob(Θ′ = ΘΛB), the set of states Bob can steer Alice to isunaffected: X is in the forward light cone if and only ifX ′ = ΛBX is, and Θ′X = ΘX ′.

Construction of the quantum steering ellipsoid for agiven state ρ. Steering ellipsoids have been studied forthe representation of SLOCC transformations [8], andmore recently in relation to quantum discord [10].

The steering ellipsoid is easiest to understand for stateswith b = 0. Suppose that Bob projects onto some pure

state X =

(1v

)with v = 1. Then

Y =1

2ΘX =

1

2

(1 0T

a T

)(1v

)=

1

2

(1

a+ Tv

), (1)

showing that Bob will obtain that outcome with proba-bility 1/2 and Alice’s Bloch vector will then be a + Tv.The set of all states Alice can end up with is simplythe unit sphere of possible v, shrunk and rotated by Tand translated by a, i.e. an ellipsoid centred at a withorientation and semiaxes given by the eigenvectors andeigenvalues of TTT . The dimension of the ellipsoid is

rank(T ) = rank(Θ) − 1. Points inside the ellipsoid canbe reached via convex combinations of projective mea-surements, and conversely any POVM element for Bobcan be decomposed into a mixture of projections, thusgiving a point within the ellipsoid.

Now consider a general state with b 6= 0. If b = 1then ρ is a product state, and so the steering ellipsoidis a single point. Hence assume that b < 1. Then the

SLOCC operator 1⊗ ρ−12

B that corresponds to a Lorentzboost ΛB by a ‘velocity’ b, transforms ρB to the max-imally mixed state. We refer to this filtered state ρ asthe canonical state on the SLOCC orbit S(ρ) (see ap-pendix). As already noted, SLOCC operations on Bobdo not affect Alice’s steering ellipsoid. Therefore we canfind the parameters of an arbitrary state’s steering el-lipsoid by boosting Θ by ΛB and then reading off theellipsoid parameters. This gives (see appendix) a steer-ing ellipsoid centred at cA = a−Tb

1−b2 , and orientation andsemiaxes lengths si =

√qi given by the eigenvectors and

eigenvalues qi of the ellipsoid matrix

QA =1

1− b2(T − abT

)(1 +

bbT

1− b2

)(TT − baT

).

(2)This together with cA, specify the ellipsoid EA.

To obtain the ellipsoid at B, EB , the roles of A andB are reversed, which amounts to a transposition of thematrix Θ. In terms of its components, this reads as b→a,a→ b, T → TT , and thus EA and EB always have thesame dimensionality, rank(Θ) − 1. This completes theconstruction of the geometric data (EA,a, b) for a givenstate ρ. We next describe the reverse direction, obtainingρAB from an ellipsoid EA and the ectors a and b [8].

Faithfulness of the representation: the reconstructionof ρ from geometric data. Since we are given a and b, allthat remains is to obtain the correlation matrix T as afunction of (QA, cA,a, b). We find that the componentsof the matrix are given by

Tij = (cA)ibj +

3∑k=1

(√QAM)ik tr(

√ρBσk

√ρBσj) (3)

where M is an orthogonal matrix, obtained from Mb =(√QA)−1(a− cA). This specifies M up to an orthogonal

matrix OBb = b, with OB ∈ O(3), which correspondsto the choice of bases at B. This choice of basis can besimply encoded, for example, by providing two contourlines on EA. The derivation is provided in the appendices.

A theorem on “complete” and “incomplete” steering.The steering ellipsoid specifies for which states there is atleast one measurement outcome for Bob that steers Aliceto it. A more subtle question is for which decompositionsof Alice’s reduced state ρA is there a measurement forBob that steers to the decomposition. Clearly a necessarycondition is that all of the states in the decompositionmust be in EA, surprisingly however it turns out thatthis is not sufficient. The situation is captured by thefollowing result.

3

Consider some non-product two-qubit state Θ with el-lipsoids EA and EB . The following are equivalent:

1. (Complete steering) For any convex decompositionof a into states in EA or on its surface, there existsa POVM for Bob that steers Alice to it.

2. Alice’s Bloch vector a lies on the surface of EAscaled down by b.

3. The affine span of EB contains the maximally mixedstate.

The proof is in the appendices. In particular, these con-ditions hold for all non-degenerate ellipsoids (which in-cludes all entangled states) as well as all states whereb = 0. An example of a state where the above conditionsfail is the state ρ = 1

2 (|00〉〈00|+ |1+〉〈1 + |).The three geometric contributions to entanglement.

The Peres-Horodecki criterion [11, 12] asserts that a two-qubit state ρe is entangled if and only if ρTB

e has a neg-ative eigenvalue. It was shown in [9] that at most oneeigenvalue of ρTB

e can be negative, furthermore in [13] itis demonstrated that ρTB

e is full rank for all entangledstates. It follows that det ρTB

e < 0 is a necessary andsufficient condition for entanglement.

Suppose ρ is entangled, then any state in its SLOCCorbit S(ρ) is also entangled [9] and in particular, thecanonical state ρ ∈ S(ρ). Therefore, we can usedet(ρTB ) < 0 ⇔ det(ρTB ) < 0 to show that a quan-tum steering ellipsoid corresponds to an entangled stateif and only if

c4 + c2(1− tr(Q) + nTQn) + h(Q) < 0 (4)

where c = cn and h(Q) = 1 − 8 det√Q + 2 tr(Q2) −

(tr(Q))2 − 2 tr(Q). We have dropped the A,B labels forthe centre c and Q because entanglement is a “symmet-ric” relation. This equation is manifestly invariant underglobal rotations of the ellipsoid, corresponding to localunitaries on the quantum state. It elucidates that entan-glement correlations between the qubits manifest them-selves in three geometric properties: (1) the distance ofEA from the origin, (2) The dimensions of EA, given bysingular values of Q and (3) the “skew” of EA, capturedby the term nTQn, which depends on the alignment ofthe ellipsoid relative to the radial direction.

The nested tetrahedron condition. The condition forentanglement given by equation (4), provides a com-pact algebraic condition for non-separability and uncov-ers contributions from different geometric aspects. How-ever, the steering setting allows us to go further, and cap-ture the distinction between separable and non-separablestates in a remarkably elegant and intuitive way. The re-sult is an single, geometric criterion for separability:

A two qubit state ρ is separable if and only ifits steering ellipsoid EA fits inside a tetrahe-dron that fits inside the Bloch sphere.

Dim. Type Alice’s ellipsoid Bob’s ellipsoid

3 (Ob

ese)

Enta

ngle

d

3 (Ob

ese)

Sep

ara

ble

2 (Panca

ke)

Inco

mple

te

2 (Panca

ke)

Com

ple

te

1 (Nee

dle

)

Inco

mple

te

TABLE I: The classes of steering ellipsoids. The greendot is the reduced state. States with EA being 3-dimensionalhave obesity and can be either entangled or separable. Thestate ρ is separable if and only if EA fits inside a tetrahe-dron inside the Bloch sphere of A. For separable states, theset EA can also be 2-dimensional (a steering pancake), or1-dimensional (a steering needle), or trivially 0-dimensional(not shown). For these cases, steering is either “complete” ifall ensemble decompositions of a in EA are attainable (whenthe span of EB contains 1

21), otherwise the steering is “incom-

plete”. Zero discord occurs only for radial steering needles.

To prove necessity, suppose Alice and Bob share a sep-arable state ρ =

∑ni=1 piαi⊗βi. Since we can always take

n ≤ 4 [14], the Bloch vectors of the αi define a (possiblydegenerate) tetrahedron T within Alice’s Bloch sphere.If Bob measures an outcome E then Alice is collapsed to∑ni=1

pi tr(Eαi)tr(EρB) αi. Hence her new Bloch vector will be a

4

convex combination of the Bloch vectors for the αi — inother words her steering ellipsoid is contained in T .

We prove in the appendix that the non-trivial converseholds: any ellipsoid that fits inside a tetrahedron thatitself fits inside the Bloch sphere must arise from a sepa-rable state, and thus the nested tetrahedron condition isboth necessary and sufficient for separability of the state.

Building further on this geometric understanding, wecan prove the following result. Let ρ be separable andn minimal such that ρ =

∑ni=1 piαi ⊗ βi. Then n =

rank(Θ). Proof: since the Θ matrix for a product statehas rank 1, it is clear that n ≥ rank(Θ). Since any sepa-rable state can be written using 4 (pure) product states[14] we have n ≤ 4. If rank(Θ) = 1 then ρ is a productstate and so n = 1. If rank(Θ) = 2 then EA is a linesegment and we can form a decomposition of ρ using theendpoints of its steering ellipsoid, giving n = 2. The caserank(Θ) = 3 (where EA is an ellipse) is non-trivial, but issolved in the appendix by using the fact that any ellipseinside a tetrahedron inside the unit sphere also fits insidea triangle inside the unit sphere [15].

Volume of the ellipsoid. We can now arrive at a com-pact and useful expression for the volume of the steer-ing ellipsoids, which allows a simple inference as to thetype of steering a given state ρ provides. The volumeof any ellipsoid is proportional to the product of itssemiaxes V = 4π

3 s1s2s3. Therefore the EA has vol-

ume VA = 4π3 |√

detQA|, which may be rewritten as

VA = 4π3|detΘ|

(1−b2)2 . However, it turns out (see appendix)

that |detΘ| = 16|detρ− detρTB |, and so we obtain that

VA =64π

3

∣∣det ρ− det ρTB∣∣

(1− b2)2 . (5)

The EB volume can be calculated from VA via the simple

relation VB = (1−b2)2

(1−a2)2VA, and thus if one side has non-

zero volume then so too must the other. This volume isa witness of entanglement, in the sense any state withVA > V? = 4π/81 must be entangled. This can be seeneither from the nested tetrahedron theorem in the earliersection or from equation (4). With the former one simplyconstructs the largest tetrahedron possible in the Blochsphere, while for the latter one observes that c must van-ish, and then maximizes subject to h(Q) = 0. In bothcases one obtains V?, namely the volume of the Wernerstate on the separable-entangled boundary.

Quantum discord and the ellipsoid. Quantum discordhas received much attention as a measure of the quan-tumness of correlations (see [16] for details) in which zerodiscord for one party roughly corresponds to them pos-sessing a non-disturbing projective measurement. In theappendix, we show that ρ has zero discord for A if andonly if EA is a segment of a diameter, while ρ has zerodiscord for B if and only if EA is one-dimensional and

b = 2|cA−a|lA

, where lA is the length of EA.A new perspective on quantum correlations: obesity.

Consider a game in which Bob must convince Alice that

the qubits they each hold possess more-than-classical cor-relations: he succeeds in his task if he can steer Alice to3 states with linearly independent Bloch vectors. A statethat is a resource for this game is called “obese”: it has anellipsoid with finite volume. Obesity is neither the sameas entanglement nor separability nor discord. All entan-gled states have det ρTB < 0 and so it follows from (F3)that all entangled states have non-zero volume and henceare obese. Equation (F3) also implies that the set of sep-arable states divides into obese and non-obese (skinny)states. All zero discord states are skinny, however one canhave EA being one-dimensional (steering needles) but thestate having positive discord due to EA not being radiallyaligned. In this sense, quantum discord can be viewed asarising from a combination of an obesity component anda skew contribution.

It turns out that VA ∝ |det(T −abT )|, and so an obesestate must have T − abT being full rank. Since abT isrank-one we must have rank(T ) ≥ 2, however we cannotsimply reduce the condition of non-zero obesity to fullrank T , since there exist states with full rank T thathave degenerate steering ellipsoids (ellipses) and so have

zero volume, for example a = b =(

13 ,

13 ,

13

)Tand T =

diag(

13 ,

13 ,

13

). However, if we replace this T with T =

diag(

13 ,

13 , 0)

we obtain an obese state with rank(T ) = 2.

To contrast entanglement, discord and obesity, wecan consider a family of canonical states ρ(θ) =14

(1 + 1

2σz ⊗ 1 +∑ij Tij(θ)σi ⊗ σj

)in which the skew

of the ellipsoid smoothly varies via the elements of T (θ) =Ry(θ)TRTy (θ), with Ry(θ) being a rotation about the

y axis by θ ∈ [0, π) and T = diag(− 920 ,−

310 ,−

310 ).

The ellipsoids have equal volumes and are centred atcA = (0, 0, 1

2 )T . This family of states illustrate an oppos-ing behavior of the discord and concurrence as a functionof skew, see Fig. 2. The entanglement favors an orien-tation in which the longest ellipsoid semiaxis is alignedradially, while discord is maximised when this axis is or-thogonal to the radial direction [22].

Conclusion. The quantum steering ellipsoid providesa faithful representation of any two qubit state and anatural geometric classification of states (as in Table 1).It provides clear and intuitive understanding into theusual key aspects of two qubit states, uncovers surpris-ing new features (such as the nested tetrahedron con-dition, skew & obesity, and incomplete steering) whileprompting novel questions, such as: can we use (4) todefine a class of “least-classical” separable states for fixed(a, b, c)? Can we use the nested tetrahedron conditionto provide a simple construction for the best separableapproximation [17] for a state ρ? What is the characteri-sation of steering ellipsoids for real states? This would beparticularly useful experimentally, acting as a new typeof tomography for two qubits.

5

FIG. 2: Discord (solid) and concurrence (dotted) of the statesρ(θ) as a function of the orientation of the ellipsoid. Entan-glement is maximized when the major axis is radial.

Acknowledgments

We wish to acknowledge Zuzana Gavorova and PeterLewis for their early contributions to this topic. Thiswork was supported by the EPSRC. DJ is funded by theRoyal Commission for the Exhibition of 1851.

This following appendices contain the details for someof the statements and theorems that are presented in themain article.

Appendix A: The steering ellipsoid for a general twoqubit state

Previous approaches to representing two qubit stateshave included partitioning the set of all two qubit statesinto SLOCC (Stochastic Local Operations and ClassicalCommunication) equivalence classes [1], which results ina three dimensional representation of a state ρ throughits SLOCC orbit, defined as

S(ρ) := { SA ⊗ SBρ(SA ⊗ SB)†

tr(SA ⊗ SBρ(SA ⊗ SB)†): SA, SB ∈ GL(2,C)}

with GL(2,C) the group of invertible, complex 2 × 2matrices. However replacing ρ with its SLOCC orbit isfar from faithful, and amounts to a highly coarse-grainedrepresentation of the state that erases much of its detail.Another approach is to start with a Pauli basis expan-sion of ρ, which can be converted via a state-dependentchoice of local unitaries on both qubits, to a representa-tion involving three spatial vectors [18]. Again, this isstill not a faithful representation, and more importantlyit is extremely difficult to develop any intuition for whatthe vector representing correlations actually means.

In this section we provide the details for constructingthe steering ellipsoid representation of an arbitrary twoqubit quantum state.

Consider a two-qubit state

Θ =

(1 bT

a T

). (A1)

The matrix Θ transforms, up to a normalization,under SLOCC operations ρ′ = SA ⊗ SBρ(SA ⊗ SB)†

as Θ′ = ΛAΘΛTB [9] where ΛA,ΛB are proper or-thochronous Lorentz transformations given by ΛW =ΥSW ⊗ S∗WΥ†/|detSW |, W ∈ {A,B} and

Υ =1√2

1 0 0 10 1 1 00 i −i 01 0 0 −1

. (A2)

In particular, local unitaries rotate the Pauli basis: a→OAa, b→ OBb, T → OATO

TB , where OA, OB ∈ SO(3).

If b = 1 then Θ must be a product state and no steeringcan occur, so assume otherwise. Define γ = 1/

√1− b2,

and a new state

Θ′ = γΘLb, (A3)

where Lb is a Lorentz boost by b [19]:

Lb =

(γ −γbT−γb 1 + γ−1

b2 bbT

). (A4)

6

The γ in (A3) ensures that Θ′ is normalized: the top-left element is

γ(1 bT

)( γ−γb

)= γ2(1− b2) = 1. (A5)

The boost means Bob’s reduced state is maximallymixed: the top-right block of Θ′ is

γ(1 bT

)( −γbT1 + γ−1

b2 bbT

)=

γ(−γbT + bT + (γ − 1)bT ) = 0T . (A6)

By the above two observations we can write

Θ′ =

(1 0T

a′ T ′

). (A7)

The set of states Bob can steer Alice to will be exactlythe same for Θ and Θ′: Y = 1

2ΘX ⇐⇒ Y = 12Θ′X ′

where X ′ = L−bX/γ and X corresponds to a positiveoperator iff X ′ does because L−b/γ preserves the forwardlight cone.

What is that set of states? Writing X =

(tx

)we see

that without loss of generality we can take t = 1 since theeffect of multiplying X by a positive number is undonewhen normalizing Y . Hence the positivity condition be-comes x ≤ 1. If x < 1 then we could write X as a convexcombination of ones with x = 1, so we restrict attentionto the latter case. So any state y that Alice can steerBob to is a convex combination of states of the formy = a′ + T ′x with x = 1. But this defines a linear imageof the unit sphere displaced by a′, or in other words anellipsoid centred at a′.

Consider the singular value decomposition T ′ =O1DO2. O2 simply rotates the unit sphere and so canbe ignored. D stretches the sphere and thus gives thelengths of the semi-axes of the resulting ellipsoid. O1 ro-tates the semi-axes. We have T ′T ′T = O1D

2OT1 and sothe lengths of the semi-axes can also be found by squarerooting the eigenvalues of T ′T ′T whilst their directionscan be found from its eigenvectors.

Combining (A3) with (A7) we find

a′ = γ(a T

)( γ−γb

)= γ2(a− Tb), (A8)

T ′ = γ(a T

) (−γbT1 + γ−1

b2 bbT)

(A9)

= γ

(−γabT + T +

γ − 1

b2TbbT

). (A10)

And so, after some algebra,

T ′T ′T = γ2(TTT − aaT ) + a′a′T . (A11)

Using −abT(1 + γ−1

b2 bbT)

= −γabT we can alsowrite

T ′ = γ(T − abT

)(1 +

γ − 1

b2bbT

). (A12)

leading to the form in the main text:

T ′T ′T = γ2(T − abT )(1 + γ2bbT )(TT − baT ) =: QA(A13)

Since γLb is invertible we have rank(Θ′) = rank(Θ).By counting linearly independent columns in (A7) wehave rank(Θ′) = rank(T ′) + 1 thus proving that the di-mension of the ellipsoid is rank(Θ)− 1.

Appendix B: Proof of the complete steering theorem

We present an extended formulation of the theorem:Theorem Consider some non-product two-qubit state

Θ =

(1 bT

a T

)with ellipsoids EA and EB . The following

are equivalent:

1. Complete steering: for all convex decompositions ofa into states in EA, there exists a POVM for Bobthat steers Alice to it.

2. Any surface steering: there exists a convex decom-position of a into states on the surface of EA witha POVM for Bob that steers Alice to it.

3. Alice’s Bloch vector lies on the surface of her el-lispoid scaled down by b.

4. The affine span of EB contains the maximally mixedstate.

5.

(10

)∈ range(ΘT ).

6.

(10

)∈ ker(Θ)⊥.

Proof. Preliminaries: Let γ = 1/√

1− b2 and Θ′ =

γΘLb. Then Θ = Θ′ L−b

γ and so 6 is equivalent to(1b

)∈ ker(Θ′)⊥. If we write Θ′ =

(1 0T

a′ T ′

)then we

see that any vector in ker(Θ′) is of the form

(0x

)with

T ′x = 0 and so 6 is equivalent to b ∈ ker(T ′)⊥.EA are the points that can be written a′ + T ′x′ where

x′ ≤ 1. Therefore the surface of EA are the points thatcan be written a′+T ′x′ where x′ = 1 and x′ ∈ ker(T ′)⊥.Hence the scaled down surface is a′ + T ′x′ where x′ = band x′ ∈ ker(T ′)⊥.

From Θ = Θ′ L−b

γ we calculate that a = a′ + T ′b.

1 =⇒ 2: Trivial.2 =⇒ 6: Let yi on the surface of EA form a con-

vex decomposition∑i piyi = a. Since they are on the

surface, we have yi = a′ + T ′x′i where x′i = 1 andx′i ∈ ker(T ′)⊥. Suppose we also have yi = a′ + T ′x′′iwith x′′i ≤ 1. Then x′i − x′′i ∈ ker(T ′), and the only

7

way that the difference between two vectors can be per-pendicular to the longer one is if they are equal. There-

fore 2pi

(1x′i

)is the unique element of the forward light

cone that 12Θ′ maps to pi

(1yi

), and therefore γLb times

these form the only possible POVM elements for Bob.But to be a valid POVM, they must sum to the iden-

tity

(20

), i.e.

∑i 2pi

(1x′i

)= L−b

γ

(20

)= 2

(1b

). Since

the x′i ∈ ker(T ′)⊥, this implies b ∈ ker(T ′)⊥ which isequivalent to 6.

6 =⇒ 1: Let yi ∈ EA form a convex decomposition∑i piyi = a. Since yi ∈ EA we have yi = a′ + T ′x′i

where x′i ≤ 1. Write x′i = ki + ci where ki ∈ ker(T ′)and ci ∈ ker(T ′)⊥. This implies ci ≤ x′i ≤ 1 and yi =

a′ + T ′ci. So 2pi

(1ci

)are in the forward light cone and

map to pi

(1yi

)under 1

2Θ′. Hence γLb times these are

in the forward line cone and map to pi

(1yi

)under 1

2Θ.

Since∑i piyi = a = a′ + T ′b we have T ′

∑i pici =

T ′b. By construction ci ∈ ker(T ′)⊥ and by assumptionb ∈ ker(T ′)⊥, and so this implies

∑i pici = b. Then∑

i γLb2pi

(1ci

)= 2γLb

(1b

)=

(20

)so we have a valid

POVM.6 =⇒ 3: Immediate from form of scaled down EA and

a in preliminaries.3 =⇒ 6: If a is on the scaled down surface then a′ +

T ′x′ = a′ + T ′b where x′ = b and x′ ∈ ker(T ′)⊥. Hencex′ − b ∈ ker(T ′). The only way the difference betweentwo vectors of the same length can be perpendicular toone of them is if they are the same, and so 6 follows.

4 =⇒ 5: Suppose∑qixi = 0,

∑qi = 1 with xi ∈

EB . Recalling that swapping parties sends Θ → ΘT wesee that there exists Yi in the forward light-cone with(

1xi

)= 1

2ΘTYi. But then ΘT 12

∑i qiYi =

∑i qi

12ΘTYi =∑

i qi

(1xi

)=

(10

).

5 =⇒ 4: Suppose there exists Y with 12ΘTY =

(10

).

If Y =

(ty

)is in the forward light cone (i.e. t ≥ y)

then EB itself contains the maximally mixed state and

we are done. Otherwise, notice that Y1 =

(y − t

0

)and

Y2 =

(yy

)are in the forward light cone. Writing 1

2ΘTYi

as qi

(1xi

)we have xi ∈ EB . Noting that Y =

∑i Yi we

have∑i

12ΘTYi =

(10

), in other words

∑i qi = 1 and∑

i qixi = 0.

5 ⇐⇒ 6: range(AT ) = ker(A)⊥ is a theorem of linearalgebra, which follows straightforwardly from the singu-lar value decomposition.

Appendix C: Reconstructing the state from itssteering ellipsoid and the Bloch vectors

Alice’s steering ellipsoid EA is described by matrix QAand centre cA. One quantum state that corresponds toit is the “canonical” state ρ with Bloch vectors b = 0,a = cA and T-matrix T =

√QAO, with O ∈ O(3) (since

QA = T TT ). Written explicitly it is

ρ =1

4(1 + cA · σ ⊗ 1 +

3∑i,j=1

(√QAO)ijσi ⊗ σj) (C1)

The orthogonal matrix O is not fixed.If in addition to QA, cA we are also given the Bloch

vectors a, b, a corresponding state is

ρ = 1⊗√

2ρB ρ1⊗√

2ρB (C2)

where

ρB =1

2(1 + b · σ) (C3)

This is because√

2ρB is a SLOCC operator for b 6= 0 andAlice’s ellipsoid is invariant under such local operationson Bob’s qubit.

Note that

tr(ρ1⊗ σ) = b, (C4)

where σ is a vector of Paulis, as required. However

tr(ρσ ⊗ 1) = cA +√QAOb (C5)

and the right hand side of this should be a. This con-strains the orthogonal matrix O as a solution to

Ob = (√QA)−1(a− cA) (C6)

We can split up the orthogonal matrix into two orthog-onal rotations: O = MB where M ∈ SO(3) (properly)rotates b to (

√QA)−1(b − cA) about a unit vector per-

pendicular to the plane they lie in, and B ∈ O(3) rotatesabout b, that is Bb = b. The equation (C6) only deter-mines M .

Define the unitary matrix UB as the one correspondingto B, it has the property [UB , ρB ] = 0. Then we maywrite

ρ = 1⊗ UBχ1⊗ U†B (C7)

where

χ = 1⊗ U†B ρ1⊗ UB (C8)

=1

4(1 + cA · σ ⊗ 1 +

3∑i,j=1

(√QAM)ijσi ⊗ σj)(C9)

8

Thus

ρ = 1⊗√

2ρBUB χ1⊗ U†B√

2ρB (C10)

= 1⊗ UB√

2ρB χ1⊗√

2ρBU†B (C11)

since if [UB , ρB ] = 0 then [UB ,√ρB ] = 0.

Therefore the state ρ compatible with QA, cA,a, b is

ρ =1

4

1 + a.σ ⊗ 1 + 1⊗ b.σ +

3∑i,j=1

Tijσi ⊗ σj

(C12)

where T = RB and

Rij = tr(1⊗√

2ρB χ1⊗√

2ρBσi ⊗ σj) (C13)

= (cA)ibj +

3∑k=1

(√QAM)ik tr(

√ρBσk

√ρBσj)(C14)

and so we can reconstruct the state up to a local unitaryB on Bob’s qubit that leaves it invariant.

Appendix D: Steering ellipsoids in a tetrahedron inthe Bloch sphere correspond to separable states

We prove that for any E inside a tetrahedron insidethe Bloch sphere, there is a separable state with ρB = 1

2 Iand E as its steering ellipse. We present the proofs foreach possible dimension of E separately, although eachone is basically a slightly more involved version of theprevious one. Note that in the 0 and 1 dimensional casesthe requirement to fit inside a tetrahedron is triviallysatisfied by any E inside the Bloch sphere.

This result suffices to show that any state with an el-lipse that fits inside a tetrahedron is separable by thefollowing argument. Suppose ρAB has an ellipse that fitsinside the tetrahedron. If b = 1 then ρAB is a prod-uct state and we are done. Otherwise, apply a SLOCCoperator to Bob and obtain ρAB with b = 0, recallingthat SLOCC operators cannot change a state from be-ing entangled to separable. This will leave Alice’s ellipseunchanged whilst moving her reduced state to the centreof her ellipse. Since by the above statement there existsa separable state with the correct ellipse and reducedstates, ρAB must equal the separable state up to a choiceof basis for Bob, and hence must itself be separable.

In fact the separable states constructed below use anumber of product states equal to the dimension of theellipse plus one. Since the SLOCC operator and choice ofbasis for Bob do not affect the number of product statesin a decomposition, we furthermore have that ρAB canbe built using that number of product states.

1. 0-dimensional

If the steering ellipsoid is a single point r then simplytake ρ with Bloch vector r and let ρAB = ρ⊗ 1

2 I.

2. 1-dimensional

Suppose E is a line segment from r0 to r1. Take ρiwith Bloch vectors ri and let ρAB = 1

2

∑i ρi ⊗ |i〉〈i|.

3. 2-dimensional

If an ellipse fits inside a tetrahedron in the Blochsphere, it also fits inside a triangle in the Bloch sphere[15]. Therefore, suppose an ellipse E fits within a tri-angle in the Bloch sphere whose vertices are {r0, r1, r2}.Without loss of generality we can take the ellipse to betangent to each edge of the triangle, at points {si} wheresi is on the face opposite to ri. Denote the centre of theellipse by c. Clearly there exists unique pi ≥ 0 such that∑i piri = c and

∑i pi = 1.

By the definition of an ellipse, there is an invertibleaffine transformation A that maps E to the unit circlein the (x, z)-plane, centred at the origin. Let ρi haveBloch vectors ri and|ψi〉 be such that the Bloch vector of|ψi〉〈ψi| is −A(si). We claim that the (manifestly sepa-rable) state

ρAB =∑i

piρi ⊗ |ψi〉〈ψi| (D1)

has ρB = 121 and that Alice’s steering ellipsoid for this

state is E . To prove the first part, notice that the Blochvector of ρB is −

∑i piA(si). Since A is affine, the unit

circle will be tangent to the triangle with vertices {A(ri)}at the points {A(si)}, and

∑i piA(ri) = A(c) = 0.

Hence it suffices to prove

Lemma 1. Suppose the triangle with vertices {vi} con-tains the unit circle centered at the origin, and the cir-cle is tangent to each edge of the triangle at the points{ti} (where ti is on the edge opposite vi). Fix pi by therequirements that

∑i pivi = 0 and

∑i pi = 1. Then∑

i piti = 0.

Proof. We use x to represent points on or within thetetrahedron using normalized barycentric co-ordinates(x0, x1, x2) where

∑i xi = 1 and ~x =

∑i xivi. Let A0 be

the volume of the triangle with vertices {x,v1,v2}, A1

be the area of the triangle with vertices {v0,x,v2} andsimilarly for A2. Let A be the area of the original trian-gle (notice A =

∑iAi). Then xi = Ai/A. By definition

the barycentric co-ordinates of the origin are (p0, p1, p2).Let Li be the length of the edge opposite vi, and let

L =∑i Li. By using that the area of a triangle = 1

2(base) × (perpendicular height) and noting that by thetangency assumption the relevant triangles have a per-pendicular height of 1, we obtain that pi = Li/L.

Let M(1)0 = |v0 − t1|, M (2)

0 = |v0 − t2|. In fact M(1)0 =

M(2)0 because they are both the unique length defined by

the requirement of being from a fixed point to a pointon the circle such that the line between them is tangent

9

t0

v0

L0

M1

M2

t1

v1

L1

M2

M0

t2

v2

L2

M0 M1

0

FIG. 3: The various quantities used in proving Lemma 1. Thedashed lines form the three triangles used to show pi = Li/L,the dotted lines indicate their perpendicular heights (whichare equal to the radius of the circle: 1).

to the sphere, so we can write this length simply as M0.Define the other two Mi by a similar argument. All thisis illustrated in Figure 3. Notice that

L0 = M1 +M2, (D2)

L1 = M0 +M2, (D3)

L2 = M0 +M1, (D4)

(D5)

The barycentric co-ordinates of t0, t1 and t2 can now becalculated as

(0,M2,M1)/L0, (D6)

(M2, 0,M0)/L1, (D7)

and

(M1,M0, 0)/L2, (D8)

respectively. Using pi = Li/L and the fact that barycen-tric coordinates respect convex combinations the requiredresult is now immediate.

Suppose Bob projects qubit onto |ψ〉 and the orthog-onal state. Since ρB = 1

21 he will obtain each outcome

with probability 12 . Therefore if he obtains the |ψ〉 out-

come then Alice’s state will be

ρA(|ψ〉) :=trB(ρAB(I ⊗ |ψ〉〈ψ|))

tr(ρB |ψ〉〈ψ|)(D9)

=

∑i piρi |〈ψi|ψ|〉|2

12

(D10)

= 2∑i

piρi |〈ψi|ψ|〉|2 (D11)

Recalling that the Bloch vector of |ψi〉〈ψi| is−A(si), thenif |ψ〉〈ψ| has Bloch vector r then the Bloch vector ofρA(|ψ〉) will be

f(r) := 2∑i

piri1− r · A(si)

2=∑i

piri (1− r · A(si)) .

(D12)Let us extend this expression to all r to define an affinefunction f . The statement that Alice’s steering ellipsoidis E is equivalent to the statement that E is the imageof the unit sphere under f . Since all the A(si) are inthe (x, z)-plane, we have f ((0, 1, 0)) = f(0), i.e. we canthink of f as first projecting onto the (x, z)-plane andthen applying some affine transformation. The image ofthe unit sphere under that projection is the unit disc,and so it suffices to check that E is the image of theunit circle under f . Define g(r) = A(f(r)). Since A isinvertible and maps E to unit circle it suffices to provethat g is the identity on the (x, z)-plane. Since g is thecomposition of two affine functions it is also affine. Bythe definition of the pi, g(0) = A(c) = 0 so g is in factlinear. Hence it suffices to check that g(uj) = uj for somespanning set of vectors {uj}. Since the triangle cannotbe degenerate, its vertex set {rj} span some plane. SinceA is invertible, {A(rj)} must span the (x, z) plane. Fori 6= j, {0,A(si),A(rj)} form a right-angle triangle, and|A(si)| = 1. Therefore A(rj) · A(si) = 1 whenever i 6= j.But

∑i pi(1 − r · A(si)) =

∑i pi − r · (

∑i piA(si)) =

1−r ·0 = 1 (the penultimate equality is from Lemma 1).Hence pi(1−A(rj) · A(si)) = δij and we are done.

4. 3-dimensional

Suppose an ellipsoid E fits within a tetrahedron in theBloch sphere whose vertices are {r0, r1, r2, r3}. Withoutloss of generality we can take the ellipsoid to be tangentto each face of the tetrahedron, at points {si} where siis on the face opposite to ri. Denote the centre of theellipsoid by c. Clearly there exists unique pi ≥ 0 suchthat

∑i piri = c and

∑i pi = 1.

By the definition of an ellipsoid, there is an invertibleaffine transformation A that maps E to the unit spherecentred at the origin. Let ρi have Bloch vectors ri and|ψi〉 be such that the Bloch vector of |ψi〉〈ψi| is −A(si).We claim that the (manifestly separable) state

ρAB =∑i

piρi ⊗ |ψi〉〈ψi| (D13)

has ρB = 121 and that Alice’s steering ellipsoid for this

state is E . To prove the first part, notice that the Blochvector of ρB is −

∑i piA(si). Since A is affine, the unit

sphere will be tangent to the tetrahedron with vertices{A(ri)} at the points {A(si)}, and

∑i piA(ri) = A(c) =

0. Hence it suffices to prove the following 3-dimensionalanalogue to Lemma 1:

Lemma 2. Suppose the tetrahedron with vertices {vi}contains the unit sphere centered at the origin, and the

10

sphere is tangent each face of the tetrahedron at the points{ti} (where ti is on the face opposite vi). Fix pi by therequirements that

∑i pivi = 0 and

∑i pi = 1. Then∑

i piti = 0.

Proof. We use x to represent points on or within thetetrahedron using normalized barycentric co-ordinates(x0, x1, x2, x3) where

∑i xi = 1 and ~x =

∑i xivi.

Let V0 be the volume of the tetrahedron with vertices{x,v1,v2,v3}, V1 be the volume of the tetrahedron withvertices {v0,x,v2,v3} and so on. Let V be the volumeof the original tetrahedron (notice V =

∑i Vi). Then

xi = Vi/V . By definition the barycentric co-ordinates ofthe origin are (p0, p1, p2, p3).

Let Ai be the area of the face opposite vi, and letA =

∑iAi. By using that the volume of a tetrahedron =

13 (area of base) × (perpendicular height) and noting thatby the tangency assumption the relevant tetrahedra havea perpendicular height of 1, we obtain that pi = Ai/A.

Let A(0)23 be the area of the triangle with vertices

{v2,v3, t0}. Let A(1)23 be the area of the triangle with ver-

tices {v2,v3, t1}. Now we have that |v2 − t0| = |v2 − t1|because they are both the unique length defined by therequirement of being from a fixed point to a point onthe sphere such that the line between them is tangentto the sphere. Similarly |v3 − t0| = |v3 − t1|. Hencethe two triangles are congruent and we can simply writetheir areas as A23. Define the other five Aij by a similarargument. Notice that

A0 = A12 +A13 +A23, (D14)

A1 = A02 +A03 +A23, (D15)

A2 = A01 +A03 +A13, (D16)

A3 = A01 +A02 +A12. (D17)

The barycentric co-ordinates of t0, t1, t2 and t3 can nowbe calculated as

(0, A23, A13, A12)/A0, (D18)

(A23, 0, A03, A02)/A1, (D19)

(A13, A03, 0, A01)/A2, (D20)

and

(A12, A02, A01, 0)/A3 (D21)

respectively. Using pi = Ai/A and the fact that barycen-tric coordinates respect convex combinations the requiredresult is now immediate.

As in the 2-dimensional case we find that if Bobprojects onto the state with Bloch vector r than Alice’sBloch vector is

f(r) :=∑i

piri (1− r · A(si)) . (D22)

Let us extend this expression to all r to define an affinefunction f . The statement that Alice’s steering ellipsoidis E is equivalent to the statement that E is the image ofthe unit sphere under f . Define g(r) = A(f(r)). Since Ais invertible and maps E to unit sphere it suffices to provethat g is the identity. Since g is the composition of twoaffine functions it is also affine. By the definition of thepi, g(0) = A(c) = 0 so g is in fact linear. Hence it sufficesto check that g(uj) = uj for some spanning set of vectors{uj}. Since the tetrahedron cannot be degenerate, itsvertex set {rj} must be spanning. Since A is invertible,{A(rj)} is also spanning. As in the 2-dimensional case,for i 6= j, {0,A(si),A(rj)} form a right-angle triangle,and |A(si)| = 1. Therefore A(rj) · A(si) = 1 wheneveri 6= j. But

∑i pi(1−r·A(si)) =

∑i pi−r·(

∑i piA(si)) =

1−r ·0 = 1 (the penultimate equality is from Lemma 2).Hence pi(1−A(rj) · A(si)) = δij and we are done.

Appendix E: Discord and steering ellipsoids

Below we outline the condition for zero discord for Al-ice from either her or Bob’s ellipsoid.

A state has zero discord for Alice iff her el-lipsoid is a segment of a diameter.

The “only if” part: A general zero discord state for Aliceρ = p|e〉〈e| ⊗ ρ0 + (1− p)|e〉〈e| ⊗ ρ1 has 〈e|e〉 = 0 and

a = te (E1)

b = x (E2)

T = eyT (E3)

where t = 2p − 1, e = 〈e|σ|e〉 and x =tr [(pρ0 + (1− p)ρ1)σ], y = tr [(pρ0 − (1− p)ρ1)σ] [20].

Alice’s steering ellipsoid EA has centre cA =(t−x·y1−x2

)e

and matrix QA = s2Aee

T with

s2A =

1

1− x2

[(y − tx)

T

(1 +

xxT

1− x2

)(y − tx)

](E4)

So EA is a segment of the diameter.The “if” part: suppose we are given EA, a segment of

the diameter. Denote the states endpoints of the ellipsoidρ0 and ρ1. Alice’s state can always be decomposed asρA = qρ0 +(1−q)ρ1. However since all the Bloch vectorsof ρA, ρ0, ρ1 are collinear they will eigendecompose intothe same pair of orthogonal states, call them |ψ〉, |ψ〉.Writing ρi = pi|ψ〉〈ψ| + (1 − pi)|ψ〉〈ψ|, for i = 0, 1 thenρA = p|ψ〉〈ψ| + (1 − p)|ψ〉〈ψ| with p = qp0 + (1 − q)p1.Then the joint state

ρ = p|ψ〉〈ψ| ⊗ β0 + (1− p)|ψ〉〈ψ| ⊗ β1 (E5)

is a zero discord state for Alice with the correct EA andρA for any mixed states on Bob’s side β0, β1.

11

There is zero discord for Alice iff Bob’s ellip-soid is a line segment and the length of Alice’sBloch vector is equal to the distance from thecentre of Bob’s ellipsoid to his Bloch vectordivided by the radius of his ellipsoid.

The “only if” part can easily be checked: it requires

a = |cB−b|sB

. Since, after some algebra,

cB =x− ty1− t2

(E6)

QB =1

(1− t2)2(y − tx)(y − tx)T (E7)

then |cB − b| = t|y−tx|1−t2 = asB .

For the “if” part, let ρ0 and ρ1 be the endpoints ofBob’s ellipsoid, and let the state corresponding to Alice’sBloch vector have eigen-decomposition ρA = p0|ψ0〉〈ψ0|+p1|ψ1〉〈ψ1|. Then the joint state

ρ = p0|ψ0〉〈ψ0| ⊗ ρ0 + p1|ψ1〉〈ψ1| ⊗ ρ1 (E8)

has zero discord for Alice, the correct Bloch vector forAlice and the correct ellipsoid for Bob. If necessary swap-ping ρ0 and ρ1, it also has the right Bloch vector for Bob.Bob’s ellipsoid EB is invariant under local unitaries on Al-ice’s qubit, so Alice and Bob’s actual state is thereforeequivalent to ρ up to this transformation, which preservesdiscord.

Appendix F: Volume formula for the steeringellipsoid

The volume of any ellipsoid is proportional to the prod-uct of its eigenvalues V = 4π

3 s1s2s3. Therefore the EAhas volume VA = 4π

3 |√

detQA|, which may be rewrittenas

VA =4π

3

∣∣det(T − abT )∣∣

(1− b2)2 =

4π

3

|detΘ|(1− b2)2

. (F1)

To express this in terms of the density matrix ρ, weuse the equation Θ = 2ΥρRΥT [9], where the unitarymatrix Υ is given in equation (A2) and R denotes a

reshuffling operation: if ρ =∑1i,j=0 ρij;kl|ij〉〈kl| then

ρR =∑1i,j=0 ρik;jl|ij〉〈kl|. We also require a curious re-

lation that holds for any 4×4 or 9×9 matrix M . For anysuch M we have that

detM = detMTB − det(MTB )R. (F2)

Applying this relation to (F1), together with the reshuf-fled form of Θ, we obtain

VA =64π

3

∣∣det ρ− det ρTB∣∣

(1− b2)2 . (F3)

Appendix G: The steering ellipsoid zoo

In this section we illustrate the main types of ellipsoid.

1. Entangled states

For every pure entangled state the ellipsoid coincideswith the Bloch sphere. When the state is mixed andentangled, the ellipsoid does not satisfy the tetrahedralcondition because, loosely speaking, the ellipsoid is eithertoo big (with volume V ≥ V? > 0 ) or too near (large c)to the surface of the Bloch sphere, see figure 4. Everyentangled state is completely steerable.

FIG. 4: A generic entangled state ρAB : both ellipsoids EA andEB are always full rank and neither can be inscribed within atetrahedron within the Bloch sphere.

2. Separable states with full-dimensional ellipsoids

Separable states admit a convex decomposition interms of product states, and have “more classical” cor-relations. Steering is still possible, however the steeringellipsoids necessarily obey the tetrahedral condition, asin figure 5.

If the state has a three dimensional EA then it hasnon-zero obesity and non-zero discord, and furthermore,it can be written as a mixture of just four product states.Such states are also completely steerable.

FIG. 5: A generic separable state ρAB , where both ellipsoidsEA and EB are full rank fit inside a tetrahedron.

12

FIG. 6: A Bell-diagonal state: the ellipsoid is centred at theorigin and its semiaxes are given by the three singular valuesof T . The vector of these singular values t = (t1, t2, t3) lives ina tetrahedron with vertices at (1,1,-1),(1,-1,1),(-1,1,1),(-1,-1,-1), and when it is inside an octahedron inside of this tetrahe-dron, then the state is necessarily separable. This defines theset of ellipsoids that fit inside the nested tetrahedron (theseare not the same tetrahedra).

FIG. 7: A Werner state. The steering ellipsoid EA is a spherecentered at the origin. The ellipsoid fits inside a tetrahedronwhen its radius is less than 1

3and thus the state is separable.

3. Steering pancakes

The set of states that Bob can steer Alice to may be-come degenerate, and form a two-dimensional set. This“steering pancake” will not only fit inside a tetrahedron,but will fit within a triangle that is inscribed within theBloch sphere as shown in 8. Recall that we have a novel

FIG. 8: A generic separable state ρAB , where both ellipsoidsEA and EB are steering pancakes.

feature for some steering pancakes (and steering needles)of incomplete steering. For steering pancakes we havecomplete steering of qubit A if and only if the affine spanof EB contains the origin of the Bloch sphere.

4. Steering needles

The steering can become even more degenerate, andthe steering set collapse to a one-dimensional line seg-ment, or “steering needle”. These states include perfectlyclassical (doubly zero-discord states) with needles beingradial (figure 9), but also includes non-zero discord statesfor which either one or both of the steering needles EAand EB is not radial (figure 10). Being radial is indicatedby dashed lines on the figures, which depict diameters.

FIG. 9: A doubly zero discord state, where both EA and EBare radial line segments.

FIG. 10: A state with zero discord at B, but non-zero discordat A. We have EB being a radial line segment, but EA is nota radial line segment.

[1] J. E. Avron, G. Bisker, and O. Kenneth, J. Math. Phys.48, 102107 (2007).

[2] R. Horodecki and M. Horodecki, Phys. Rev. A 54, 1838

(1996).

[3] I. Bengtsson and K. Zyczkowski, Geometry of Quan-tum States: An Introduction to Quantum Entanglement

13

(Cambridge University Press, 2006).[4] E. Schrodinger, Proc. Cam. Phil. Soc. 31, 555 (1935).[5] N. Gisin, Helvetica Physica Acta 62, 363 (1984).[6] L. P. Hughston, R. Jozsa, and W. K. Wootters, Phys.

Lett. A 183, 14 (1993).[7] R. W. Spekkens and T. Rudolph, Quantum Inform.

Compu. 2, 66 (2002).[8] F. Verstraete, PhD Thesis (2002).[9] F. Verstraete, J. Dehaene, and B. DeMoor, Phys. Rev. A

64, 010101 (2001).[10] M. Shi, W. Yang, F. Jiang, and J. Du, J. Phys. A: Math.

Theor. 44, 415304 (2011).[11] A. Peres, Phys. Rev. Lett. 77, 1413 (1996).[12] M. Horodecki, P. Horodecki, and R. Horodecki, Phys.

Lett. A 223, 1 (1996).[13] A. Sanpera, R. Tarrach, and G. Vidal, Phys. Rev. A 58,

826 (1998).[14] W. K. Wootters, Phys. Rev. Lett. 80, 2245 (1998).

[15] M. Vidrighin (2012), URL http://mathoverflow.net/

questions/109153.[16] H. Ollivier and W. H. Zurek, Phys. Rev. Lett. 88, 017901

(2001).[17] M. Lewenstein and A. Sanpera, Phys. Rev. Lett. 80, 2261

(1998).[18] R. Horodecki and M. Horodecki, Phys. Rev. A 54, 1838

(1996).[19] R. Frosch, Four-tensors, the Mother Tongue of Classical

Physics (vdf Hochschulverlag AG, 2006).[20] B. Dakic, V. Vedral, and C. Brukner, Phys. Rev. Lett.

105, 190502 (2010).[21] C. S. Mingjun Shi, Fengjian Jiang and J. Du, N. J. Phys.

13, 073016 (2011).[22] Discord can be calculated analytically when θ = 0, π

2,

and at these points ρ is an X-state [21] and this agreeswith our findings.