Develop mixed states, reduced states and no-signaling using finite-dimensional linear algebra.
§1.9Estimated reading time: 34 min
A state vector ∣ψ⟩ describes a quantum system completely when we know exactly which preparation was performed. Often this knowledge is missing: a device may prepare one of several states at random, or we may hold only one part of a larger system whose other part is out of reach.
In both situations a single unit vector is no longer an adequate description. The correct object is a density operator, also called a density matrix. This page develops density operators from the pure-state Born rule, introduces the partial trace, which produces the state of a subsystem, and uses both tools to prove that local operations cannot transmit information to a distant party.
Conventions
Every Hilbert space is a finite-dimensional complex inner product space H≅Cd, and L(H) denotes the linear operators on it. The inner product ⟨u∣v⟩=i∑uivi is conjugate-linear in the first argument. The adjoint satisfies (A†)ij=Aji, so (∣u⟩⟨v∣)†=∣v⟩⟨u∣. Ensembles and measurements are finite.
For a bipartite system we write HAB=HA⊗HB, with A the first factor and dA=dimHA, dB=dimHB. Product basis vectors are written ∣a,b⟩=∣ab⟩=∣a⟩⊗∣b⟩, the first label belonging to A, and are ordered lexicographically with the A index outermost; for two qubits the order is ∣00⟩,∣01⟩,∣10⟩,∣11⟩. In this order M⊗N is the Kronecker product, whose (a,a′) block of size dB×dB is Maa′N. An operator acting only on B is written IA⊗N.
An operator X is positive semidefinite, written X⪰0, if ⟨v∣X∣v⟩ is real and nonnegative for every v. On a complex space, ⟨v∣X∣v⟩∈R for all v already forces X=X† (expand ⟨v∣(X−X†)∣v⟩=0 at v=u+w and v=u+iw). So every X⪰0 is Hermitian, with an orthonormal eigenbasis and eigenvalues λ=⟨e∣X∣e⟩≥0. Over R this fails: the rotation matrix (01−10) has vTXv=0 for all real v but is not symmetric.
Ensembles and density operators
Suppose a source prepares ∣ψk⟩ with probability pk and the label k is discarded. A tempting description is a superposition such as k∑pk∣ψk⟩. That is wrong: a superposition is a single definite preparation and can show interference that a random choice among preparations does not show. The right object must reproduce the measurement statistics of the random preparation, and nothing more.
Definition 1.9.1
(Ensemble and its density operator)
An ensemble on H is a finite list {(pk,∣ψk⟩)}k=1n with pk≥0, k∑pk=1, and each ∣ψk⟩ a unit vector. The ∣ψk⟩ need not be orthogonal or distinct. Its density operator is
ρ=k=1∑npk∣ψk⟩⟨ψk∣∈L(H).
The intrinsic definition below makes no reference to any ensemble.
Definition 1.9.2
(Density operator, intrinsic form)
An operator ρ∈L(H) is a density operator if ρ⪰0 and trρ=1. The set of density operators is D(H). A density operator ∣ψ⟩⟨ψ∣ with ⟨ψ∣ψ⟩=1 is pure; every other density operator is mixed.
The next theorem shows that the two definitions describe exactly the same set of operators.
Theorem 1.9.3
(Ensembles produce exactly the density operators)
For every ensemble {(pk,∣ψk⟩)}, the operator k∑pk∣ψk⟩⟨ψk∣ is positive semidefinite with trace one.
Conversely, every ρ∈D(H) is the density operator of at least one ensemble, for example its eigenvectors weighted by its eigenvalues.
Proof
(Positivity from squared overlaps, and the spectral ensemble)
where the trace uses tr(∣u⟩⟨w∣)=⟨w∣u⟩, the unit length of each ∣ψk⟩, and k∑pk=1.
(2) Since ρ⪰0 is Hermitian, the spectral theorem gives an orthonormal basis with ρ=i∑λi∣ei⟩⟨ei∣, λi≥0 and i∑λi=trρ=1. So {(λi,∣ei⟩)} is an ensemble with density operator ρ.
For real θ, (eiθ∣ψ⟩)(eiθ∣ψ⟩)†=∣ψ⟩⟨ψ∣, so the density operator removes the meaningless global phase, while relative phases survive in the off-diagonal entries. For one qubit, D(C2) is the Bloch ball, treated in the Bloch ball and state discrimination.
Concept lensGeometric
The set of states as a convex set
Geometrically, D(H) is convex: positivity and trace one survive tρ0+(1−t)ρ1 for t∈[0,1]. An ensemble is then a way of writing ρ as a convex combination of pure states, and the pure states are exactly the extreme points, those that admit no nontrivial decomposition into other density operators. The spectral ensemble is one decomposition; for ρ=21I on a qubit there are many, which is why different ensembles can have the same ρ.
The Born rule for density operators
A measurement with outcomes in a finite set is a POVM{Em}: operators with Em⪰0 and m∑Em=I. Projective measurements are the case of mutually orthogonal projectors. We take the pure-state rule Pr[m∣ψ]=⟨ψ∣Em∣ψ⟩ as the postulate and derive everything else with ordinary probability.
Theorem 1.9.4
(Born rule for density operators)
Let the ensemble {(pk,∣ψk⟩)} have density operator ρ, and let {Em} be a POVM. When the label k is unknown, outcome m occurs with probability
Pr[m]=tr(Emρ).
For every ρ∈D(H) these numbers are nonnegative and sum to 1. For a Hermitian observable O, the expectation value is tr(Oρ).
using ⟨ψ∣E∣ψ⟩=tr(E∣ψ⟩⟨ψ∣) (cyclicity for the d×1 and 1×d matrices ∣ψ⟩, ⟨ψ∣) and linearity of the trace. For a general ρ=i∑λi∣ei⟩⟨ei∣ we get tr(Emρ)=i∑λi⟨ei∣Em∣ei⟩≥0, and summing over m with m∑Em=I gives trρ=1. The expectation statement follows from the spectral decomposition O=μ∑μPμ.
The statistics of every measurement depend on the ensemble only through ρ. Two ensembles with the same density operator therefore cannot be told apart by any measurement, even in principle; this is why ρ, and not the ensemble, is called the state.
Worked example 1.9.5
(Three preparations of the maximally mixed qubit)
Let ∣±⟩=(∣0⟩±∣1⟩)/2. The ensemble {(21,∣0⟩),(21,∣1⟩)} gives 21I. So does {(21,∣+⟩),(21,∣−⟩)}, because the off-diagonal entries cancel:
A third preparation uses three non-orthogonal states ∣tk⟩=cosθk∣0⟩+sinθk∣1⟩, θk=2πk/3, each with weight 31. With cos2θ=21(1+cos2θ), sinθcosθ=21sin2θ and k=0∑2cos(4πk/3)=k=0∑2sin(4πk/3)=0, the diagonal sums k∑cos2θk and k∑sin2θk both equal 23 and the off-diagonal sum is 0, so the average is again 21I.
These are physically different procedures, yet by the Born rule for density operators no experiment on the output can tell which one was used.
Common mistake
(Agreement on one measurement is not equality of states)
The mixture {(21,∣0⟩),(21,∣1⟩)} and the pure state ∣+⟩ give the same Z statistics, each outcome with probability 21. They are still different states, 21I=∣+⟩⟨+∣=21(1111), and the X measurement separates them: ∣+⟩ gives outcome + with probability 1, the mixture with probability 21. Only equality of the full operator, off-diagonal entries included, rules out every distinguishing measurement.
Purity
The number tr(ρ2) is the purity of ρ. It is basis-independent and detects exactly whether a state is pure.
Theorem 1.9.6
(Purity criterion)
Let ρ∈D(H) with d=dimH. Then d1≤tr(ρ2)≤1. Moreover tr(ρ2)=1 if and only if ρ is pure, and tr(ρ2)=1/d if and only if ρ=I/d.
Proof
(Eigenvalues in the unit interval)
Write ρ=i∑λi∣ei⟩⟨ei∣ with λi≥0 and i∑λi=1, so tr(ρ2)=i∑λi2. Each λi lies in [0,1], and for such λ we have λ2≤λ, with equality exactly when λ∈{0,1}. Hence i∑λi2≤i∑λi=1, with equality exactly when one eigenvalue is 1 and the rest are 0, that is, ρ=∣ei0⟩⟨ei0∣. Conversely, a pure ρ=∣ψ⟩⟨ψ∣ satisfies ρ2=ρ, so tr(ρ2)=1.
For the lower bound, the Cauchy–Schwarz inequality in Rd gives
1=(i∑1⋅λi)2≤di∑λi2,
with equality if and only if (λi) is proportional to (1,…,1), that is, λi=1/d for all i and ρ=I/d.
Positivity puts every eigenvalue in [0,1] and trace one fixes their sum; both conditions are used.
Counterexample mode
Purity one without positivity
Claim: a Hermitian trace-one matrix with tr(τ2)=1 is a pure state. Take τ=diag(32,32,−31) on C3. Then trτ=32+32−31=1 and tr(τ2)=94+94+91=1, yet τ has rank 3, so it is not of the form ∣ψ⟩⟨ψ∣. It is not a state at all: measuring in the computational basis would give outcome 2 with "probability" ⟨2∣τ∣2⟩=−31. The step λ2≤λ fails for λ=−31.
In dimension 2 there is no such example: λ1+λ2=1 and λ12+λ22=1 force 2λ1λ2=0. The repaired statement is the theorem, with ρ⪰0 as a hypothesis.
The partial trace
Let ρAB∈D(HA⊗HB), and suppose Alice acts only on A. Her POVM {Em} acts on the joint system as {Em⊗IB}, with outcome probabilities tr((Em⊗IB)ρAB). We want one operator ρA on HA that reproduces all of these through the ordinary rule tr(EmρA).
For ∣β⟩∈HB, let IA⊗∣β⟩:HA→HA⊗HB send ∣a⟩↦∣a⟩⊗∣β⟩, and let IA⊗⟨β∣ send ∣a⟩⊗∣b⟩↦⟨β∣b⟩∣a⟩, extended linearly. Three properties are used below.
(P1)(IA⊗∣β⟩)†=IA⊗⟨β∣, since ⟨a′⊗b∣a⊗β⟩=⟨a′∣a⟩⟨b∣β⟩=⟨⟨β∣b⟩a′a⟩.
(P2) For an orthonormal basis {∣j⟩} of HB, j∑(IA⊗∣j⟩)(IA⊗⟨j∣)=IA⊗j∑∣j⟩⟨j∣=IA⊗IB.
(P3) For M∈L(HA), (M⊗IB)(IA⊗∣j⟩)=(IA⊗∣j⟩)M, since both send ∣a⟩ to M∣a⟩⊗∣j⟩.
Definition 1.9.7
(Partial trace and reduced state)
Let {∣j⟩}j=1dB be an orthonormal basis of HB. For X∈L(HA⊗HB),
The reduced state of A is ρA=trB(ρAB). Symmetrically, trA(X)=i∑(⟨i∣⊗IB)X(∣i⟩⊗IB) for an orthonormal basis {∣i⟩} of HA, and ρB=trA(ρAB).
For product operators the definition gives trB(M⊗N)=(trN)M, in particular
trB(∣a⟩⟨a′∣⊗∣b⟩⟨b′∣)=⟨b′∣b⟩∣a⟩⟨a′∣,
and by linearity this determines trB completely. In block form, if X is a dA×dA array of dB×dB blocks X(a,a′), then trB(X) has entries trX(a,a′), while trA(X)=a∑X(a,a) is the sum of the diagonal blocks.
Counterexample mode
Orthonormality of the basis is not optional
Claim: any basis of HB may be used in the formula. Use instead the basis {∣0⟩,∣+⟩} of C2, which spans but is not orthonormal, on X=IA⊗∣1⟩⟨1∣:
β∈{0,+}∑⟨β∣1⟩⟨1∣β⟩IA=(0+21)IA=21IA,
whereas the correct value is (tr∣1⟩⟨1∣)IA=IA. Half of the trace is lost because ∣0⟩⟨0∣+∣+⟩⟨+∣=I, so (P2) fails. The formula is valid for orthonormal bases, and the next theorem shows that among those the choice does not matter.
Theorem 1.9.8
(Defining identity for local observables)
For all X∈L(HA⊗HB) and M∈L(HA),
tr(trB(X)M)=tr(X(M⊗IB)),
and trB(X) is the unique Y∈L(HA) with tr(YM)=tr(X(M⊗IB)) for all M.
Proof
(Insert the identity on B, then move the local operator)
Rectangular cyclicity says trH(PQ)=trK(QP) for linear maps P:K→H and Q:H→K; it is the identity r,s∑PrsQsr=s,r∑QsrPrs. Insert (P2), apply it with Q=IA⊗⟨j∣, and then use (P3):
For uniqueness, if Y and Y′ both satisfy the identity, then Z=Y−Y′ has tr(ZM)=0 for all M. Taking M=∣a′⟩⟨a∣ gives ⟨a∣Z∣a′⟩=0 for all basis vectors, so Z=0.
Proof X-ray
Where each hypothesis enters the defining identity
The proof has one move: replace the identity on AB by the sum in (P2), then slide M⊗IB past IA⊗∣j⟩ using (P3). Orthonormality is what makes (P2) a true resolution of the identity; the counterexample above shows the formula losing trace without it. Locality of the observable is what makes (P3) available. For a general G on AB nothing can be slid, and no operator on HA alone reproduces tr(XG) for every G; the Bell-pair example below shows this with G=X⊗X. Finite dimension makes every trace a finite sum, so cyclicity is a rearrangement. Uniqueness uses only that the matrix units ∣a′⟩⟨a∣ span L(HA), and that clause is what lets later proofs conclude equality of reduced states from equality of local statistics.
With X=ρAB and M=Em, the identity reads tr((Em⊗IB)ρAB)=tr(EmρA): every local statistic of A is computed from ρA, and by uniqueness ρA is the only operator that does this. The partial trace is therefore the forced answer to the question above, not a convention.
Quick check
(Compute trB and trA of ∣0⟩⟨1∣⊗∣1⟩⟨1∣.)
Use the product formula, and keep the bra index paired with the ket index.
Solution · Answer
Tracing out B gives (tr∣1⟩⟨1∣)∣0⟩⟨1∣=∣0⟩⟨1∣. Tracing out A gives (tr∣0⟩⟨1∣)∣1⟩⟨1∣=⟨1∣0⟩∣1⟩⟨1∣=0. The two partial traces of one operator can be very different.
Theorem 1.9.9
(The partial trace maps states to states)
trB(X) does not depend on the orthonormal basis of HB used to compute it. Moreover, for all X∈L(HA⊗HB):
tr(trBX)=trX;
trB(X†)=(trBX)†;
if X⪰0, then trBX⪰0.
Consequently ρAB∈D(HA⊗HB) implies ρA∈D(HA).
Proof
(Basis-free characterization and diagonal matrix elements)
The right-hand side of the defining identity mentions no basis of HB, and every orthonormal basis yields an operator satisfying it; by uniqueness all choices give the same operator.
(1) Put M=IA in the defining identity. (2) By (P1), the adjoint of (IA⊗⟨j∣)X(IA⊗∣j⟩) is (IA⊗⟨j∣)X†(IA⊗∣j⟩); sum over j. (3) For v∈HA, (IA⊗∣j⟩)∣v⟩=∣v⟩⊗∣j⟩, so
⟨v∣trB(X)∣v⟩=j∑(⟨v∣⊗⟨j∣)X(∣v⟩⊗∣j⟩)≥0,
each term being a diagonal matrix element of X⪰0.
The same argument with ∣v⟩ in HA⊗HR for an auxiliary system R shows that trB is completely positive, hence a quantum channel; complete positivity is the subject of quantum noise channels and complete positivity.
Common mistake
(The partial trace is not multiplicative)
On two qubits, trB((I⊗I)(I⊗I))=2IA but trB(I⊗I)trB(I⊗I)=4IA. What does hold is the module property for local factors, trB((M⊗IB)X(M′⊗IB))=MtrB(X)M′, which follows from (P3) and its adjoint.
The Bell pair and the classical mixture
Worked example 1.9.10
(Reduced states of the Bell pair)
Let ∣Φ+⟩=21(∣00⟩+∣11⟩), a unit vector because ∣00⟩,∣11⟩ are orthonormal. Its coefficients are real, so
Write each term as ∣a⟩⟨a′∣⊗∣b⟩⟨b′∣ and apply the product formula. The two diagonal terms give ∣0⟩⟨0∣ and ∣1⟩⟨1∣; the cross terms give ⟨1∣0⟩∣0⟩⟨1∣=0 and ⟨0∣1⟩∣1⟩⟨0∣=0. Hence ρA=21I.
Block check: the 2×2 blocks are 21(1000), 21(0010), 21(0100), 21(0001), with traces 21,0,0,21, so ρA=21I; the sum of the diagonal blocks gives ρB=21I.
The global purity is 1, while tr(ρA2)=21, the minimum 1/d. So the reduced state of a pure state need not be pure. Complete knowledge of the whole coexists with maximal ignorance of each part, which never happens classically: a marginal of a deterministic joint distribution is deterministic.
Now compare the separable state σAB=21(∣00⟩⟨00∣+∣11⟩⟨11∣), a fair coin that prepares ∣0⟩∣0⟩ or ∣1⟩∣1⟩. The same computation gives trBσAB=trAσAB=21I. The difference
Δ=ρAB−σAB=21(∣00⟩⟨11∣+∣11⟩⟨00∣)
consists only of coherences, and the defining identity gives tr((M⊗IB)Δ)=tr(MtrBΔ)=0 for every M. No measurement on one side alone separates the two states; a correlated measurement is needed.
Worked example 1.9.11
(Z⊗Z versus X⊗X statistics)
In Z⊗Z, outcome (a,b) has probability ⟨ab∣ρ∣ab⟩, a diagonal entry. Both ρAB and σAB have diagonal (21,0,0,21), so both give Pr[00]=Pr[11]=21 and ⟨Z⊗Z⟩=1. This measurement is blind to Δ, since ⟨ab∣Δ∣ab⟩=0.
In X⊗X, expand ∣++⟩=21(∣00⟩+∣01⟩+∣10⟩+∣11⟩) and ∣−−⟩=21(∣00⟩−∣01⟩−∣10⟩+∣11⟩), so 21(∣++⟩+∣−−⟩)=∣Φ+⟩. For ρAB, Pr[++]=Pr[−−]=21 and ⟨X⊗X⟩=1. For σAB, each branch is a product state, so for s,t∈{+,−}
and ⟨X⊗X⟩σ=0. As an operator check, X⊗X swaps ∣00⟩ and ∣11⟩, so tr((X⊗X)σAB)=0 and tr((X⊗X)Δ)=21(1+1)=1.
The mixture is perfectly correlated only in the basis in which it was prepared; the Bell pair is perfectly correlated in both. The extra correlation lives in Δ and is revealed by X⊗X, an observable not of the form M⊗IB.
Reduced states therefore do not determine the joint state: ρAB, σAB and 41I⊗I all have reduced states 21I on both sides, yet ⟨X⊗X⟩ separates ρAB from the others and ⟨Z⊗Z⟩, equal to 1 for σAB and 0 for 41I⊗I, separates the remaining pair. For general bipartite pure states, the reduced states are analyzed through the Schmidt decomposition.
No-signaling
Alice holds A and Bob holds B of a shared ρAB. If Bob's choice of local operation could change what Alice observes, he could send her a message with no physical carrier between them. The theorem below rules this out for trace-preserving operations; general channels and their Kraus form are developed in quantum noise channels.
Definition 1.9.12
(Kraus operation on B)
A finite family {Kk}⊂L(HB) with the completeness relationk∑Kk†Kk=IB defines a trace-preserving operation on B, acting on the joint system as
N(X)=k∑(IA⊗Kk)X(IA⊗Kk†).
Theorem 1.9.13
(No-signaling)
Let ρAB∈D(HA⊗HB) and let N be a Kraus operation on B with k∑Kk†Kk=IB. Then trB(N(ρAB))=ρA. Consequently, Alice's outcome probabilities for every POVM are the same whether Bob applies N, any other such operation, or nothing.
Proof
(Test against every local observable of Alice)
By the uniqueness clause of the defining identity, it suffices to show tr((M⊗IB)N(ρAB))=tr((M⊗IB)ρAB) for every M∈L(HA). By cyclicity on HA⊗HB and factorwise multiplication, (IA⊗Kk†)(M⊗IB)(IA⊗Kk)=M⊗Kk†Kk, so
Taking M=Em gives the statement about Alice's probabilities.
Proof X-ray
The anatomy of no-signaling
Locality enters when (IA⊗Kk†)(M⊗IB)(IA⊗Kk) collapses to M⊗Kk†Kk; if Bob's operators also acted on A, M could not be pulled out. Completeness is used exactly once, in the last equality, and it is the whole reason the result holds. The uniqueness clause turns equal statistics for all M into equal reduced operators, so no basis computation is needed. Nothing else is used: not spatial separation, not timing, and no property of ρAB, so the result holds for entangled and unentangled states alike.
The Kraus family also describes a measurement by Bob with outcome k. Knowing k, outcome k occurs with probability pk=tr(Kk†KkρB), and for pk>0 the conditional state is ρAB(k)=(IA⊗Kk)ρAB(IA⊗Kk†)/pk, with Alice's conditional reduced state ρA(k)=trBρAB(k), which can differ from ρA. Not knowing k, a zero-probability term is the zero operator, so k:pk>0∑pkρAB(k)=N(ρAB) and, by the theorem, k∑pkρA(k)=ρA. If Alice later measures {Em}, the joint distribution is p(m,k)=tr((Em⊗Kk†Kk)ρAB); its marginal k∑p(m,k)=tr(EmρA) is independent of Bob's choice, while the conditional p(m∣k) may depend on k but is usable only by someone who learns k, through a separate classical channel as in teleportation.
Worked example 1.9.14
(Bob measures Z on a Bell pair)
Take ρAB=∣Φ+⟩⟨Φ+∣, K0=∣0⟩⟨0∣, K1=∣1⟩⟨1∣; completeness holds since K0†K0+K1†K1=I. Then (I⊗K0)∣Φ+⟩=21∣00⟩ and (I⊗K1)∣Φ+⟩=21∣11⟩, so p0=p1=21 and Alice's conditional states are ρA(0)=∣0⟩⟨0∣ and ρA(1)=∣1⟩⟨1∣, both pure and both different from 21I. Without k, her state is 21∣0⟩⟨0∣+21∣1⟩⟨1∣=21I=ρA, unchanged. Moreover N(ρAB)=σAB: Bob's unrecorded Z measurement erases exactly the coherence Δ without altering either reduced state.
Common mistake
(Post-selection is not a trace-preserving operation)
If Bob keeps only the runs with outcome 0, the single operator K=∣0⟩⟨0∣ violates completeness, and the renormalized state gives Alice ∣0⟩⟨0∣=21I. This is not signaling. Selecting the kept runs requires knowing the outcomes; Alice's raw stream still has reduced state 21I until someone who knows the outcomes tells her which runs to discard. The proof breaks exactly at the step that uses k∑Kk†Kk=IB.
Exercises
Exercise 1.9.15
(Let c∈C and ρ=(21cc21). Show that ρ is a density operator if and only if ∣c∣≤21, compute tr(ρ2), and find a unit vector ∣ψ⟩ with ρ=∣ψ⟩⟨ψ∣ whenever ρ is pure.)
Foundation
Solution(4)
1
Reduce to an eigenvalue condition
The off-diagonal entries are conjugates, so ρ is Hermitian, and trρ=1. It remains to decide when ρ⪰0.
Because: A positive semidefinite operator on a complex space is Hermitian, and a Hermitian matrix is positive semidefinite exactly when its eigenvalues are nonnegative.
2
Find the eigenvalues
Key step
The characteristic polynomial is (21−λ)2−∣c∣2, so λ±=21±∣c∣. Both are nonnegative if and only if ∣c∣≤21.
3
Compute the purity
tr(ρ2)=(21+∣c∣)2+(21−∣c∣)2=21+2∣c∣2,
which agrees with i,j∑∣ρij∣2=41+41+∣c∣2+∣c∣2.
4
Identify the pure case
ρ is pure exactly when ∣c∣=21. Write c=21e−iφ and ∣ψ⟩=21(∣0⟩+eiφ∣1⟩). Then ∣ψ⟩⟨ψ∣=21(1eiφe−iφ1)=ρ; the bra carries e−iφ, which places c in the (0,1) entry.
Because: The purity criterion: tr(ρ2)=1 exactly for pure states.
Exercise 1.9.16
(Let ∣ψ⟩=a,b∈{0,1}∑Cab∣a⟩A∣b⟩B with a,b∑∣Cab∣2=1. Show that ρA=CC† and ρB=CTC, that tr(ρA2)=1−2∣detC∣2, and apply this to ∣ψ⟩=31(∣00⟩+∣01⟩+∣10⟩).)
So ρB is the transpose of C†C, which in general differs from C†C.
Because: trB contributes ⟨b′∣b⟩=δbb′ and trA contributes δaa′.
3
Express the purity through the determinant
Key step
For a 2×2 density operator with eigenvalues λ1,λ2, tr(ρA2)=(λ1+λ2)2−2λ1λ2=1−2detρA, and detρA=detCdetC=∣detC∣2. So ρA is pure exactly when detC=0, that is, when C=xyT has rank one and ∣ψ⟩=∣x⟩⊗∣y⟩ is a product state.
4
Apply to the example
Here C=31(1110) with detC=−31, and ρA=ρB=31(2111) because C is real and symmetric. The purity is 91(4+1+1+1)=97=1−2⋅91. The state is entangled, since detC=0, but not maximally, since the purity exceeds 21.
Exercise 1.9.17
(Fix γ∈[0,1]. Bob applies to his half of ∣Φ+⟩ the operation with K0=(1001−γ) and K1=(00γ0). Verify completeness, find p0,p1 and Alice's conditional states, check that trBN(ρAB)=21I, and compute Bob's state after the operation.)
Core
Solution(5)
1
Verify completeness
K0†K0=diag(1,1−γ) and, since K1=γ∣0⟩⟨1∣, K1†K1=γ∣1⟩⟨1∣. The sum is I.
The unnormalized reduced operators are 21diag(1,1−γ) and 2γ∣1⟩⟨1∣. Normalizing,
ρA(0)=2−γ1(1001−γ),ρA(1)=∣1⟩⟨1∣(γ>0),
both different from 21I when γ>0.
Because: The previous exercise gives ρA=CC† with C=21diag(1,1−γ) for the first branch.
4
Average without the outcome
Key step
trBN(ρAB)=21diag(1,1−γ)+2γdiag(0,1)=21I=ρA, as no-signaling requires. Bob's choice of γ cannot encode a message.
5
Compute Bob's state and reconcile
Tracing out A gives 21diag(1,1−γ)+2γ∣0⟩⟨0∣, so ρB′=21(1+γ001−γ), which has decayed toward ∣0⟩. There is no contradiction: a local operation may change the system it acts on and the correlations, but not the other party's reduced state.
Summary
A finite-dimensional state is a positive semidefinite, trace-one operator; ensembles produce exactly these, and the spectral ensemble realizes each one.
Born probabilities are tr(Emρ), so ensembles with the same ρ are physically indistinguishable.
The purity satisfies 1/d≤tr(ρ2)≤1, with 1 exactly for pure states; positivity is essential.
The partial trace, computed in any orthonormal basis, is the unique operator with tr(trB(X)M)=tr(X(M⊗IB)) for all M; it is basis-independent, positive and trace-preserving.
The Bell pair and the classical mixture share reduced states and Z⊗Z statistics but differ in X⊗X, through coherences the partial trace discards.
Completeness k∑Kk†Kk=IB leaves Alice's reduced state unchanged; conditional states can change, but using them requires Bob's outcomes, sent classically.