The Density Operator

You have a quantum state. How do you describe it? If it is pure — perfectly prepared, perfectly known — a ket |\psi\rangle is enough. But as the previous chapter showed, real quantum states are often not pure: a coin flip hid which ket was prepared, a partner qubit was traced out, an environment scrambled the phases. In those cases, no single ket tells the full story. You need the density operator — an object that packages both the quantum state and the classical ignorance about it into one matrix.

This chapter is the formal machinery. Three axioms, one pure-state test, one beautiful geometric picture for qubits (the Bloch ball), and one set of measurement rules that generalises everything you knew about kets. By the end you will be able to look at a 2 \times 2 matrix and say — in seconds — whether it is a valid density operator, whether it is pure or mixed, and what its Bloch vector is.

The three axioms

Strip away every motivation and every analogy, and a density operator \rho is defined by exactly three mathematical conditions.

You will often see "density operator" and "density matrix" used interchangeably. In finite dimensions they are the same thing — any operator on a finite-dimensional space is a matrix once you fix a basis. Physicists lean on "density operator"; mathematicians sometimes prefer "density matrix" to emphasise the matrix form. This article uses both, matching the convention of the source you are most likely to be reading.

Why these three conditions, and not some other three

Each axiom is doing a specific job.

• Hermiticity makes the eigenvalues real. If \rho|u_i\rangle = p_i|u_i\rangle, then p_i \in \mathbb R. Since the p_i will turn out to be probabilities (see below), they had better be real numbers and not complex.
• Positive semi-definiteness makes the eigenvalues non-negative. No "negative probabilities" hiding anywhere. Combined with Hermiticity, every p_i \in [0, \infty).
• Unit trace makes the eigenvalues sum to 1. For Hermitian \rho, \text{tr}(\rho) = \sum_i p_i, so the condition \text{tr}(\rho) = 1 says \sum_i p_i = 1. Combined with the first two, the eigenvalues \{p_i\} form a classical probability distribution.

Why these three conditions and not more or fewer: they are the smallest set that makes \rho represent "a probability distribution over orthogonal pure states" while also being a linear operator. Weaken any of them and you lose a piece of what "quantum state" has to mean. Strengthen (e.g. demand rank 1) and you collapse to pure states only.

Together these axioms give a clean picture: every density operator's spectrum is a classical probability distribution over an orthonormal basis of eigenstates. This is the spectral decomposition:

where \{|u_i\rangle\} is an orthonormal eigenbasis of \rho and p_i \geq 0, \sum_i p_i = 1. So every density operator is an ensemble in its own eigenbasis — the "coin flip" of the previous chapter shows up automatically as the eigenvalue distribution.

Two equivalent ways to write \rho

Density operators admit two natural decompositions, and you will encounter both.

Spectral decomposition — canonical, orthogonal, unique

The spectral decomposition writes \rho in its own eigenbasis:

The eigenstates |u_i\rangle are automatically orthogonal (Hermitian operators have orthogonal eigenvectors). This decomposition is unique up to degeneracy (when two p_i are equal, the corresponding eigenspace is a plane rather than a line, and you can rotate within it).

You can read the spectral decomposition as an ensemble: "with probability p_i, the state is the eigenstate |u_i\rangle." The catch is that this ensemble description is mathematically convenient but not the only way to prepare \rho — see the next decomposition.

Convex-combination decomposition — any set of states, not unique

A second way. \rho can also be written as a convex combination of pure states \{|\psi_i\rangle\} that are not necessarily orthogonal:

Any such expression with these properties produces a valid \rho. And — a deep fact — the same density operator can be written this way in infinitely many ways, with different ensembles \{(q_i, |\psi_i\rangle)\} yielding the same matrix. For instance, the maximally mixed state I/2 can be prepared as a 1/2-1/2 mix of |0\rangle and |1\rangle, or a 1/2-1/2 mix of |+\rangle and |-\rangle, or many other combinations. All are physically indistinguishable because all give the same \rho.

Why this non-uniqueness matters: a density operator encodes only what is physically observable — the statistics of every possible measurement. Two preparations giving the same \rho cannot be told apart by any quantum experiment on the single system. The density operator is an equivalence class of preparations, not a single ensemble. The spectral decomposition picks out one canonical representative (the orthogonal one), but physics does not privilege it.

The pure-state test

Given a matrix that satisfies all three axioms, how do you tell if it represents a pure state (one specific ket) or a mixed state (a genuine distribution over kets)?

Why \text{tr}(\rho^2) = 1 characterises pure states

Start with \rho = |\psi\rangle\langle\psi| for some pure state |\psi\rangle. Compute \rho^2:

Why the middle collapses: \langle\psi|\psi\rangle = 1 because |\psi\rangle is a unit vector. Multiplying by 1 does nothing, so \rho^2 = \rho. Operators satisfying \rho^2 = \rho are called idempotent or, equivalently, projectors. Rank-1 projectors are exactly the pure-state density operators.

Taking the trace: \text{tr}(\rho^2) = \text{tr}(\rho) = 1. So every pure state has purity 1.

Conversely, suppose \text{tr}(\rho^2) = 1. Write \rho in its spectral decomposition with eigenvalues \{p_i\}. Then \rho^2 has eigenvalues \{p_i^2\}, and \text{tr}(\rho^2) = \sum_i p_i^2. You also have \sum_i p_i = 1 (unit trace) and p_i \in [0, 1] (positive semi-definite, unit trace). The only way \sum_i p_i^2 = \sum_i p_i = 1 under these constraints is if exactly one p_i equals 1 and all others are 0 — which is a pure state.

Why \sum p_i^2 = \sum p_i forces rank 1: compare term by term. For p_i \in [0, 1], p_i^2 \leq p_i with equality iff p_i \in \{0, 1\}. So \sum p_i^2 \leq \sum p_i always, and equality requires every p_i to be 0 or 1. Combined with \sum p_i = 1, exactly one p_i is 1 and the rest are 0.

Why the minimum purity is 1/d

The minimum purity is achieved at the maximally mixed state \rho = I/d, for which

For a single qubit (d = 2), this is 1/2. Every single-qubit density matrix has purity somewhere in [1/2, 1]. 1/2 means "maximum uncertainty" (every basis gives 50-50); 1 means "pure state." The purity is the single number that ranks where your state sits on this scale.

The Bloch-vector representation for qubits

For a single qubit there is a uniquely beautiful way to parametrise every density operator: as a real 3-vector inside a unit ball.

The claim

Why this parametrisation works

Any 2\times 2 Hermitian matrix can be expanded in the basis \{I, \sigma_x, \sigma_y, \sigma_z\} of the real vector space of Hermitian 2\times 2 matrices:

Why these four matrices span the Hermitian matrices: the space of 2\times 2 Hermitian matrices is real 4-dimensional (two real numbers on the diagonal plus one complex off-diagonal entry giving two real degrees of freedom = 4 real dimensions). The identity plus three Paulis are four linearly-independent Hermitian matrices, so they span.

Apply the unit-trace condition: \text{tr}(I) = 2 and \text{tr}(\sigma_k) = 0 for each Pauli, so \text{tr}(\rho) = 2a_0. Setting this equal to 1 gives a_0 = 1/2.

Apply the positive semi-definite condition. The eigenvalues of \rho = \tfrac{1}{2}(I + \vec r\cdot\vec\sigma) are \tfrac{1}{2}(1 \pm |\vec r|). Positive semi-definiteness requires both to be non-negative, which gives |\vec r| \leq 1.

Why the eigenvalues are \tfrac{1}{2}(1 \pm |\vec r|): the operator \vec r\cdot\vec\sigma has eigenvalues \pm|\vec r| (it is a sum of r_x \sigma_x + r_y\sigma_y + r_z\sigma_z, and this combination is — up to a unitary rotation — just |\vec r|\sigma_z). Adding I shifts each eigenvalue up by 1; dividing by 2 scales both by half.

So every valid qubit density operator is parameterised by a 3-vector of length \leq 1. The set of all such vectors fills a 3-dimensional ball of radius 1 — the Bloch ball.

The geometry

The Bloch ball picture unifies every qubit state into one geometry.

• Pure states (|\vec r| = 1) sit on the surface — the familiar Bloch sphere. These are the kets you have been using. The north pole \vec r = (0, 0, 1) is |0\rangle; south pole (0, 0, -1) is |1\rangle; equator points (1,0,0), (-1,0,0) are |+\rangle, |-\rangle; (0,1,0), (0,-1,0) are |+i\rangle, |-i\rangle.
• Mixed states (0 < |\vec r| < 1) sit strictly inside the ball. The farther from the surface (i.e., the smaller |\vec r|), the more mixed.
• The maximally mixed state \rho = I/2 sits at the origin \vec r = \vec 0. Complete uncertainty.

The purity has a gorgeous Bloch-vector form:

Pure (|\vec r| = 1): purity = 1. Maximally mixed (|\vec r| = 0): purity = 1/2. Smoothly interpolating for every state in between.

Bloch vector of specific states

A few worked Bloch vectors to anchor the geometry.

• |0\rangle has density matrix \begin{pmatrix}1 & 0 \\ 0 & 0\end{pmatrix}. Matching \tfrac{1}{2}(I + \vec r\cdot\vec\sigma) = \tfrac{1}{2}\begin{pmatrix}1 + r_z & r_x - ir_y \\ r_x + ir_y & 1 - r_z\end{pmatrix} to this matrix: r_z = 1, r_x = r_y = 0. Bloch vector (0, 0, 1) — north pole.
• |+\rangle has density matrix \tfrac{1}{2}\begin{pmatrix}1 & 1 \\ 1 & 1\end{pmatrix}. Match: r_x = 1, r_y = r_z = 0. Bloch vector (1, 0, 0) — +x equator point.
• I/2 has density matrix \tfrac{1}{2}\begin{pmatrix}1 & 0 \\ 0 & 1\end{pmatrix}. Match: \vec r = 0. Bloch vector (0, 0, 0) — centre of the ball.
• A 50-50 mixture of |0\rangle and |+\rangle has density matrix \begin{pmatrix}3/4 & 1/4 \\ 1/4 & 1/4\end{pmatrix}. Match the parameterisation: 1 + r_z = 3/2, so r_z = 1/2; 1 - r_z = 1/2, consistent; r_x - ir_y = 1/2, so r_x = 1/2, r_y = 0. Bloch vector (1/2, 0, 1/2), length |\vec r| = 1/\sqrt{2} \approx 0.71. A mixed state, sitting partway between the centre and the surface.

Measurement and expectation rules in density-operator form

Every measurement rule you know for kets has a density-operator version. Here is the full set, side by side.

Probability of a measurement outcome

For a projective measurement \{P_m\} (with P_m^2 = P_m and \sum_m P_m = I), the probability of outcome m is

Derivation. Substitute \rho = \sum_i p_i|\psi_i\rangle\langle\psi_i|:

Why the trace becomes an inner product: \text{tr}(A|\psi\rangle\langle\psi|) = \langle\psi|A|\psi\rangle for any operator A (cyclic property of the trace — move the bra to the front).

Each \langle\psi_i|P_m|\psi_i\rangle is the probability of outcome m for the pure state |\psi_i\rangle (the usual ket rule). Weighted by the classical probabilities p_i, the sum is the total probability of outcome m for the ensemble — exactly what you want.

For a pure state \rho = |\psi\rangle\langle\psi|, the formula reduces to p(m) = \langle\psi|P_m|\psi\rangle — the original ket rule.

Expectation value of an observable

For an observable A (a Hermitian operator), the expectation value in state \rho is

Derivation. Spectral decomposition A = \sum_m a_m P_m and \rho = \sum_i p_i|\psi_i\rangle\langle\psi_i|:

which is the standard definition of expectation value — outcome times probability, summed.

Post-measurement state

If you obtain outcome m, the post-measurement density operator is

Why the projector sandwiches the state from both sides: for a pure-state ket the update rule is |\psi\rangle \mapsto P_m|\psi\rangle/\sqrt{p(m)}. Writing it as a density operator: (P_m|\psi\rangle)(P_m|\psi\rangle)^\dagger/p(m) = P_m|\psi\rangle\langle\psi|P_m^\dagger/p(m) = P_m\rho P_m/p(m) (using P_m^\dagger = P_m for projectors). The sandwich is the density-operator analogue of applying P_m on both sides of the ket.

The denominator \text{tr}(P_m\rho) = p(m) renormalises the post-measurement state so it still has trace 1.

Unitary evolution

A unitary gate U acts on \rho by conjugation:

Derivation. For a pure state |\psi\rangle \mapsto U|\psi\rangle, the density matrix |\psi\rangle\langle\psi| transforms as

Extend by linearity to any density operator: \rho = \sum_i p_i|\psi_i\rangle\langle\psi_i| \mapsto \sum_i p_i U|\psi_i\rangle\langle\psi_i|U^\dagger = U\rho U^\dagger.

Summary table

Every density-operator formula reduces to the ket formula when you substitute \rho = |\psi\rangle\langle\psi|. The density-operator forms are strict generalisations.

Worked examples

Common confusions

• "Density operator and density matrix are different things." In finite dimensions, they are the same: any operator on a finite-dimensional Hilbert space is a matrix once you fix a basis. In infinite-dimensional settings (continuous-variable quantum mechanics, field theory), "density operator" is the more careful term, and matrix representations may only exist in a restricted sense. For qubit and qudit quantum computing, you can use either name.

• "Probabilities in \rho sum to 1 along the diagonal." Partially true. In the computational basis, the diagonal entries of \rho are the probabilities of measuring |0\rangle, |1\rangle, \ldots in that basis, and they do sum to 1 (because \text{tr}(\rho) = 1). In any other basis, the new diagonal entries are different — they are the probabilities for that new basis, and they also sum to 1. The trace is basis-independent; the diagonal's individual values are not.

• "\rho^2 = \rho is the definition of pure." It is equivalent to purity (for any valid density operator). It is shorter than computing the trace of \rho^2 by hand, because if you see \rho^2 = \rho by inspection (for instance, \rho is a recognisable projector), you instantly know it is pure without any further work.

• "Bloch vectors are only for pure states." The Bloch sphere (the surface) indexes pure states; the Bloch ball (the interior plus surface) indexes all qubit states, pure and mixed. Every density operator on a qubit has a unique Bloch vector; pure states are the unit-length ones, mixed states are the shorter ones, and \vec r = 0 is the maximally mixed state.

• "You can read off whether \rho is pure from its matrix entries." Sometimes — if \rho is manifestly a rank-1 projector like |0\rangle\langle 0|, yes. In general, no. You must either compute \text{tr}(\rho^2), or find all eigenvalues and check if only one is non-zero, or (for qubits) compute |\vec r| and check if it equals 1. All three are equivalent tests.

• "Measurement probability is just the diagonal entry." Only for computational-basis projective measurements. For a general projector P_m, the rule is p(m) = \text{tr}(P_m\rho), which is a full trace computation. If P_m = |m\rangle\langle m| for a computational basis state, then \text{tr}(P_m\rho) = \langle m|\rho|m\rangle, which is indeed the (m,m) diagonal entry — but that is a special case.

Going deeper

The convex set of density operators

The set \mathcal D(\mathcal H) of all density operators on a Hilbert space \mathcal H is a convex set: if \rho_1, \rho_2 \in \mathcal D(\mathcal H) and \lambda \in [0, 1], then \lambda\rho_1 + (1-\lambda)\rho_2 is also in \mathcal D(\mathcal H). Direct verification:

• Hermitian: a convex sum of Hermitian operators is Hermitian.
• Positive semi-definite: \langle\phi|(\lambda\rho_1 + (1-\lambda)\rho_2)|\phi\rangle = \lambda\langle\phi|\rho_1|\phi\rangle + (1-\lambda)\langle\phi|\rho_2|\phi\rangle \geq 0.
• Unit trace: \text{tr}(\lambda\rho_1 + (1-\lambda)\rho_2) = \lambda + (1 - \lambda) = 1.

Convex sets have extreme points — points that cannot be written as a non-trivial convex combination of two other points in the set. The extreme points of \mathcal D(\mathcal H) are exactly the pure states. Every mixed state can be written as a convex combination of pure states (by the spectral theorem, using its eigenstates) — the Krein-Milman theorem in action.

For a single qubit, \mathcal D(\mathcal H) is the Bloch ball. Its extreme points are the points on the Bloch sphere's surface — the pure states. Interior points are mixed.

Non-uniqueness of the ensemble and Uhlmann's theorem

The same density operator can be produced by many different ensembles. The classic example: \rho = I/2 for a qubit can be written as

• \tfrac{1}{2}|0\rangle\langle 0| + \tfrac{1}{2}|1\rangle\langle 1| (eigenbasis of \sigma_z),
• \tfrac{1}{2}|+\rangle\langle+| + \tfrac{1}{2}|-\rangle\langle-| (eigenbasis of \sigma_x),
• \tfrac{1}{2}|+i\rangle\langle+i| + \tfrac{1}{2}|-i\rangle\langle-i| (eigenbasis of \sigma_y),
• or infinitely many non-orthogonal three-or-more-state ensembles.

All describe the same \rho = I/2 and are physically indistinguishable. Uhlmann's theorem — a foundational result — characterises exactly when two ensembles produce the same density matrix: they do iff they are related by a partial isometry on a purifying space. Equivalently, two ensembles \{(p_i, |\psi_i\rangle)\} and \{(q_j, |\phi_j\rangle)\} give the same \rho iff there exists a matrix U with U U^\dagger = I (identity on the subspace of the first ensemble) such that \sqrt{p_i}|\psi_i\rangle = \sum_j U_{ij}\sqrt{q_j}|\phi_j\rangle. The proof uses the purification construction from the why density matrices chapter.

Trace distance and fidelity — measuring closeness of states

How different are two density operators? The two workhorse metrics are:

Trace distance:

where |A| = \sqrt{A^\dagger A}. It ranges from 0 (identical states) to 1 (orthogonal pure states). It has the operational interpretation: if a referee randomly hands you \rho or \sigma (each with probability 1/2) and you must guess which, your optimal success probability is (1 + T(\rho, \sigma))/2.

Fidelity:

Ranges from 0 (orthogonal) to 1 (identical). Fidelity is symmetric and unitarily invariant, and it reduces to |\langle\psi|\phi\rangle| for pure states — the overlap magnitude you already know.

These two metrics are related by Fuchs-van de Graaf inequalities:

so they measure the same intuitive notion (how different two states are) but in slightly different operational ways. Quantum error-correction protocols use trace distance to bound logical-error probabilities; teleportation and state verification experiments use fidelity to quantify success.

Von Neumann entropy

A density operator encodes a probability distribution (its spectrum), and every probability distribution has a Shannon entropy. The quantum analogue is the von Neumann entropy:

where the second equality is in the spectral decomposition \rho = \sum_i p_i|u_i\rangle\langle u_i|. Properties:

• S(\rho) = 0 iff \rho is pure. (No uncertainty about the state.)
• S(\rho) is maximised at \rho = I/d, giving S(I/d) = \log d. (Maximum uncertainty.)
• S(\rho) is invariant under unitary evolution: S(U\rho U^\dagger) = S(\rho). (Unitaries do not create or destroy information.)
• For bipartite pure states |\psi\rangle_{AB}, S(\rho_A) = S(\rho_B) — the Schmidt number — measures the entanglement between A and B.

Von Neumann entropy is the quantum analogue of Shannon entropy, and it underpins quantum information theory — source coding, channel capacities, entanglement measures, the second law of thermodynamics in the quantum setting. Part 13 will develop it properly; for now, the fact that \rho has an entropy at all is itself non-trivial and a direct consequence of the probability-distribution-on-eigenvalues picture.

An experimental note — density matrices in Indian NMR labs

Indian NMR quantum computing groups at IIT Madras and TIFR Mumbai have long experience with density-matrix tomography — the experimental procedure of measuring enough observables on an ensemble to reconstruct \rho. The NMR sample is \sim 10^{17} molecules in thermal equilibrium; the density matrix encoding the qubit algorithm sits on top of the large equilibrium background as a small deviation density matrix proportional to \sigma_z (initially) and manipulated by radio-frequency pulses. Tomography reads out the full 2^n \times 2^n matrix for n-qubit experiments by measuring expectation values of an informationally complete set of Paulis. This is where density-matrix theory stops being a formal abstraction and becomes the daily working object of experimental quantum computing — the language in which results are reported and fidelities quoted.

Where this leads next

• Why density matrices — the motivating chapter on when kets fail and why \rho is unavoidable.
• Density-matrix properties — deeper properties, spectral decomposition, rank, and operations.
• Von Neumann entropy — the quantum analogue of Shannon entropy on density operators.
• The Bloch ball — the geometry of qubit states in three dimensions.
• Pauli X, Y, Z — the three matrices that, together with the identity, form the basis of the Bloch-vector representation.
• Density matrices preview — an earlier, more informal introduction in Part 3.

References

1. Wikipedia, Density matrix — definitions, properties, spectral decomposition, purity.
2. Nielsen and Chuang, Quantum Computation and Quantum Information (2010), §2.4 (the density operator) — Cambridge University Press.
3. John Preskill, Lecture Notes on Quantum Computation, Ch. 3 — theory.caltech.edu/~preskill/ph229.
4. Wikipedia, Bloch sphere — the Bloch-vector representation, extending to the Bloch ball for mixed states.
5. John Watrous, The Theory of Quantum Information (2018), Ch. 2 — cs.uwaterloo.ca/~watrous/TQI.
6. Wikipedia, Quantum state — general framing of pure and mixed states, including references to Petz's Quantum Information Theory and Quantum Statistics. </content> </invoke>