Density Operators for Quantum Information

A density operator describes the information available about a quantum system, whether it is pure, drawn from a preparation ensemble, or part of an entangled register. In a protocol, the essential question is which systems and classical records remain available. This page connects that choice to state evolution, measurement branches, and quantitative comparisons between finite-dimensional states.

Helpful background. Density operators introduces positivity and normalization; tensor products and partial trace explains subsystem descriptions. The Bloch sphere supplies useful one-qubit examples.

Density operators and measurement probabilities

A density operator is the state object used when a quantum-information task involves preparation uncertainty, discarded subsystems, entanglement, noise, measurement branches, or communication ensembles. On a finite-dimensional Hilbert space $\mathcal H$, it satisfies

$$\rho^\dagger=\rho, \qquad \rho\geq0, \qquad \operatorname{Tr}\rho=1.$$

It predicts every measurement probability through

$$\begin{aligned} p(y) &= \operatorname{Tr}(M_y\rho), \\ M_y&\geq0, \qquad \sum_yM_y=I. \end{aligned}$$

and every observable expectation through

$$\langle A\rangle = \operatorname{Tr}(\rho A).$$

The same trace rule applies to several physically different situations. A mixture of prepared signals and the reduced state of an entangled pair can have identical density operators. Their local measurement statistics agree even though their correlations with other systems need not.

Three Layers of Meaning

The operator

At the mathematical layer, $\rho$ is a positive trace-one operator. Positivity means

$$\langle\psi|\rho|\psi\rangle\geq0$$

for every $|\psi\rangle$. Equivalently in finite dimension, all eigenvalues are nonnegative. The trace-one condition normalizes total probability.

The preparation

A laboratory preparation procedure may be described by an ensemble

$$\mathcal E = \{p_x,\rho_x\}_x.$$

If the label $x$ is not retained, the emitted system is described by the average state

$$\bar\rho = \sum_xp_x\rho_x.$$

Different ensembles can have the same average. No measurement on the emitted system alone can determine which ensemble decomposition was used when the density operators agree.

The information available to an observer

If the classical label is retained in a register $X$, the joint classical–quantum state is

$$\rho_{XB} = \sum_x p_x |x\rangle\langle x|_X \otimes \rho_B^x.$$

Discarding $X$ gives

$$\rho_B = \operatorname{Tr}_X\rho_{XB} = \sum_xp_x\rho_B^x.$$

The states $\rho_{XB}$ and $\rho_B$ do not contain the same information. The first preserves which signal was selected; the second describes a receiver who has only the quantum system. This distinction underlies coding, state discrimination, cryptography, and the Holevo information.

Spectra, Rank, and Purity

Every finite-dimensional density operator has a spectral decomposition

$$\begin{aligned} \rho &= \sum_{j=1}^{r} \lambda_j |j\rangle\langle j|, \\ \lambda_j&>0, \qquad \sum_j\lambda_j=1. \end{aligned}$$

where $r=\operatorname{rank}\rho$. This decomposition is orthogonal and is fixed by $\rho$ up to basis choices inside degenerate eigenspaces. It should not be confused with a general preparation ensemble, whose states may be nonorthogonal and whose decomposition is highly nonunique.

A state is pure exactly when

$$\rho^2=\rho,$$

or equivalently

$$\operatorname{Tr}(\rho^2)=1.$$

The purity

$$\gamma(\rho) = \operatorname{Tr}(\rho^2) = \sum_j\lambda_j^2$$

obeys

$$\frac1d \leq \gamma(\rho) \leq 1$$

in dimension $d$. The lower bound is attained by $I/d$.

Purity is a useful diagnostic, but it is not a distance to a target state and it does not identify the noise mechanism. Two states can have equal purity and be far apart. Purities from different Hilbert-space dimensions also have different lower bounds. Pure versus mixed states explains why coherent superpositions and mixtures differ.

For one qubit, $\gamma=(1+\lVert\mathbf r\rVert^2)/2$ connects purity to the Bloch-vector radius. Higher-dimensional state spaces do not form an ordinary Bloch ball.

Density Operators from Composite Systems

Reduced states

For a bipartite state $\rho_{AB}$, the state available to an observer with access only to $A$ is

$$\rho_A = \operatorname{Tr}_B\rho_{AB}.$$

It is characterized by the requirement

$$\operatorname{Tr} \left[ \rho_{AB} (M_A\otimes I_B) \right] = \operatorname{Tr} \left( \rho_A M_A \right)$$

for every local effect or observable $M_A$. Thus $\rho_A$ contains exactly the statistics of measurements on $A$ alone.

If

$$|\Phi^+\rangle_{AB} = \frac{ |00\rangle+|11\rangle }{\sqrt2},$$

then

$$\rho_A=\rho_B=\frac I2$$

even though $\rho_{AB}=|\Phi^+\rangle\langle\Phi^+|$ is pure. Local mixedness can arise from entanglement, not from classical uncertainty about a hidden local pure state. Partial trace develops the index rules for calculating these reductions.

Discarding is a channel

The map

$$\operatorname{Tr}_B: \mathcal L(\mathcal H_A\otimes\mathcal H_B) \longrightarrow \mathcal L(\mathcal H_A)$$

is a quantum channel. Operationally, “ignore the ancilla,” “do not read the environment,” and “keep only the output register” are partial-trace instructions. They can turn a pure global state into a mixed local state while preserving all local probabilities.

This is why an algorithmic derivation must state which registers are retained. A final state vector for the full circuit is not yet the answer when the task asks for one output register.

Purification

Every density operator can be realized as the reduced state of a pure state on a larger system. From the spectral decomposition,

$$\rho_A = \sum_j\lambda_j|j\rangle\langle j|,$$

one purification is

$$|\Psi\rangle_{AR} = \sum_j \sqrt{\lambda_j} |j\rangle_A|j\rangle_R.$$

Tracing out the reference gives

$$\operatorname{Tr}_R \left( |\Psi\rangle\langle\Psi| \right) = \rho_A.$$

The smallest possible reference dimension is $\operatorname{rank}\rho_A$, and the construction above attains it. Purifications are not unique: their occupied reference subspaces are related by an isometry. This extends to an isometry between the full reference spaces in the smaller-to-larger direction, or to a unitary when their dimensions agree.

Purification is a representation theorem; it does not assert that a particular accessible ancilla was present in the preparation. It is useful for modeling an adversary’s possible side information and for comparing states through Uhlmann’s theorem below. Purification gives the construction and its relation to Schmidt decomposition.

Transformations in Protocols

Closed-system evolution

A unitary transformation acts by conjugation:

$$\rho \longmapsto U\rho U^\dagger.$$

It preserves the spectrum, rank, purity, and von Neumann entropy. A unitary cannot turn a mixed state into a pure state on the same isolated system.

Quantum channels

A quantum channel $\mathcal N$ is a completely positive trace-preserving linear map. In a Kraus representation,

$$\mathcal N(\rho) = \sum_aK_a\rho K_a^\dagger, \qquad \sum_aK_a^\dagger K_a=I.$$

The output is again a density operator. The same framework describes noisy gates, communication, reset, and discarding a subsystem. Quantum channels and noise develops the representation theorems; quantum channels for information processing explains their use in protocols.

Measurement branches are often subnormalized

An instrument consists of completely positive, trace-nonincreasing maps $\mathcal I_m$ whose sum is trace preserving. Outcome $m$ corresponds to

$$\widetilde\rho_m = \mathcal I_m(\rho).$$

The tilde marks a subnormalized positive operator:

$$p(m) = \operatorname{Tr}\widetilde\rho_m.$$

Conditioned on an outcome with nonzero probability, the normalized state is

$$\rho_m = \frac{ \widetilde\rho_m }{ p(m) }.$$

The probability and the conditional state describe different aspects of postselection, heralded entanglement, syndrome extraction, and quantum trajectories. Normalizing each branch too early erases its probability; failing to normalize before using $\rho_m$ as a state gives incorrect conditional expectations.

For one Kraus operator $M_m$ per branch,

$$\widetilde\rho_m = M_m\rho M_m^\dagger, \qquad \sum_mM_m^\dagger M_m=I.$$

The sum over outcomes,

$$\sum_m\widetilde\rho_m,$$

is the nonselective output state.

Comparing Quantum States

No single scalar answers every comparison question. The task determines the appropriate quantity.

Trace distance

The trace distance is

$$D(\rho,\sigma) = \frac12 \lVert\rho-\sigma\rVert_1.$$

It is a metric, ranges from $0$ to $1$, and controls optimal one-shot state discrimination. For equal priors,

$$P_{\mathrm{guess}} = \frac12 \left[ 1+D(\rho,\sigma) \right].$$

It is contractive under channels:

$$D \left( \mathcal N(\rho),\mathcal N(\sigma) \right) \leq D(\rho,\sigma).$$

Physical processing cannot make two states more distinguishable when the same channel is applied to both and no side information is added. Trace Distance is the formula card with assumptions and convention checks.

Several calculation shortcuts expose its meaning. For commuting states with eigenvalue distributions $p_i$ and $q_i$,

$$D(\rho,\sigma)=\frac12\sum_i|p_i-q_i|.$$

For two pure states,

$$D(|\psi\rangle,|\phi\rangle) =\sqrt{1-|\langle\psi|\phi\rangle|^2},$$

and for qubit Bloch vectors $\mathbf r$ and $\mathbf s$,

$$D(\rho,\sigma)=\frac12\lVert\mathbf r-\mathbf s\rVert_2.$$

The variational characterization

$$D(\rho,\sigma) =\max_{0\le M\le I} \operatorname{Tr}[M(\rho-\sigma)]$$

shows that trace distance is the largest difference in the probability of one measurement event. It also implies, for a bounded observable $A$,

$$|\operatorname{Tr}[A(\rho-\sigma)]| \le2\lVert A\rVert_\infty D(\rho,\sigma).$$

For prior probabilities $p$ and $1-p$, the optimal binary discrimination probability is

$$P_{\mathrm{guess}}^{\mathrm{opt}} =\frac12\left(1+ \lVert p\rho-(1-p)\sigma\rVert_1\right).$$

The equal-prior formula above is its $p=1/2$ specialization.

Fidelity

We use squared Uhlmann fidelity:

$$F(\rho,\sigma) = \left[ \operatorname{Tr} \sqrt{ \sqrt\rho\,\sigma\sqrt\rho } \right]^2.$$

It ranges from $0$ to $1$ and equals $1$ exactly when $\rho=\sigma$. If $\rho=|\psi\rangle\langle\psi|$ is pure,

$$F(\rho,\sigma) = \langle\psi|\sigma|\psi\rangle.$$

Fidelity is not a metric, and some references use the unsquared square-root quantity under the same name. The convention must accompany a reported value. The fidelity formula card collects the conventions.

Useful special cases in the squared convention are

$$F(|\psi\rangle,|\phi\rangle) =|\langle\psi|\phi\rangle|^2,$$ $$F(\rho,\sigma) =\left(\sum_i\sqrt{p_iq_i}\right)^2$$

for commuting states, and

$$F(\rho,\sigma) =\frac12\left[ 1+\mathbf r\cdot\mathbf s +\sqrt{(1-\lVert\mathbf r\rVert^2) (1-\lVert\mathbf s\rVert^2)} \right]$$

for qubits. Fidelity is multiplicative on tensor products and monotone under a common channel:

$$F(\rho_1\otimes\rho_2,\sigma_1\otimes\sigma_2) =F(\rho_1,\sigma_1)F(\rho_2,\sigma_2),$$ $$F(\mathcal N(\rho),\mathcal N(\sigma)) \ge F(\rho,\sigma).$$

Uhlmann’s theorem gives the purification form

$$F(\rho,\sigma) =\max_{|\psi_\rho\rangle,|\psi_\sigma\rangle} |\langle\psi_\rho|\psi_\sigma\rangle|^2,$$

where the maximization is over purifications in a common larger space. Numerically, first check the size of any Hermiticity violation. Symmetrizing roundoff-level errors can be appropriate; a substantial discrepancy requires diagnosing the input or estimator. Reject substantial negative eigenvalues rather than silently clipping them, and compute matrix square roots by positive spectral calculus.

For the squared convention, the Fuchs–van de Graaf inequalities are

$$\begin{aligned} 1-\sqrt{F(\rho,\sigma)} &\leq D(\rho,\sigma), \\ D(\rho,\sigma) &\leq \sqrt{1-F(\rho,\sigma)}. \end{aligned}$$

They translate bounds between two different notions of closeness; they do not make fidelity and trace distance interchangeable.

Purity, entropy, and target overlap answer different questions

Purity asks how concentrated the spectrum is. Fidelity asks closeness to a specified target. Trace distance asks distinguishability. Von Neumann entropy,

$$S(\rho) = -\operatorname{Tr} \left( \rho\log_2\rho \right),$$

quantifies spectral uncertainty and has several information-theoretic roles. A high-purity state can have zero fidelity with the desired pure target. A low trace distance to one target says nothing about closeness to another. Quantum entropy develops its information-theoretic uses.

A Source Ensemble Example

Suppose a source emits $|0\rangle$ with probability $p$ and $|+\rangle$ with probability $1-p$, while withholding the label. The receiver’s state is

$$\begin{aligned} \bar\rho &= p|0\rangle\langle0| + (1-p)|+\rangle\langle+| \\ &= \frac12 \begin{pmatrix} 1+p&1-p \\ 1-p&1-p \end{pmatrix}. \end{aligned}$$

Its Bloch vector is

$$\mathbf r=(1-p,0,p),$$

and its purity is

$$\operatorname{Tr}(\bar\rho^2) = 1-p+p^2.$$

At $p=1/2$,

$$\bar\rho = \begin{pmatrix} 3/4&1/4 \\ 1/4&1/4 \end{pmatrix}, \qquad \operatorname{Tr}(\bar\rho^2)=\frac34.$$

For $0<p<1$, the average is mixed because the two possible signals are distinct; the endpoints $p=0,1$ remain pure. It still has coherence in the computational basis because one ensemble member is $|+\rangle$. A mixed state need not be diagonal in the basis chosen for display.

If the source keeps the label $X$, the correct state is instead

$$\rho_{XB} = p|0\rangle\langle0|_X \otimes |0\rangle\langle0|_B + (1-p)|1\rangle\langle1|_X \otimes |+\rangle\langle+|_B.$$

This joint state permits correlations between the label and a later measurement outcome. Averaging to $\bar\rho_B$ intentionally discards those correlations.

A state metric tests a particular input-output pair. It does not by itself test a gate on every input: a reset channel can prepare $|0\rangle$ perfectly while destroying an unknown qubit. Similarly, conditioning on a rare measurement result can make surviving states more distinguishable even though an unconditional channel cannot; the final exercise shows how the success probability restores the comparison.

Numerical and Experimental Validity Checks

For a finite matrix proposed as a density operator, check:

1. Dimensions and basis order: identify every tensor factor and index convention.
2. Hermiticity: verify $\rho^\dagger=\rho$ within a stated tolerance.
3. Trace: verify $\operatorname{Tr}\rho=1$ for a normalized state.
4. Positivity: inspect the smallest eigenvalue or use a positive parameterization.
5. Purity and rank: report them when they matter, without treating numerical rank as tolerance independent.
6. Marginals: check reduced states and known conserved quantities.
7. Probabilities: test that representative effects give values in $[0,1]$.
8. Convergence: vary time steps, truncations, sample counts, or tensor-network tolerances.

Small negative eigenvalues can come from floating-point roundoff, statistical linear inversion, or a genuinely nonphysical model. Silently clipping them to zero and renormalizing changes the state and can bias metrics. A repair procedure should be justified, documented, and propagated into uncertainty.

For infinite-dimensional systems, density operators are positive trace-class operators with trace one. Expectations of unbounded observables require additional finite-moment conditions. A finite basis cutoff is an approximation: small omitted probability alone does not bound the error in an unbounded observable. The carrier and mode examples illustrate this distinction.

Exercises

1. Test a candidate density matrix

Consider

$$\rho = \begin{pmatrix} 0.6&0.5 \\ 0.5&0.4 \end{pmatrix}.$$

It is Hermitian and has trace one. Is it a valid density operator?

Solution

Positivity remains to be checked. The determinant is

$$\det\rho = (0.6)(0.4)-(0.5)^2 = -0.01.$$

A positive semidefinite $2\times2$ matrix cannot have negative determinant. Equivalently, this matrix has Bloch vector

$$\mathbf r=(1,0,0.2)$$

with

$$\lVert\mathbf r\rVert=\sqrt{1.04}>1.$$

Its eigenvalues are

$$\lambda_\pm = \frac{ 1\pm\sqrt{1.04} }{2},$$

so $\lambda_-<0$. The matrix is not a physical density operator despite being Hermitian and trace one.

2. Analyze a nonorthogonal source ensemble

Set $p=1/2$ in the source example. Verify the average density matrix and purity. Find the receiver’s best probability of guessing the source label from one signal, assuming equal priors and a compulsory guess on every run.

Solution

Substitution gives

$$\begin{aligned} \bar\rho &= \frac12|0\rangle\langle0| + \frac12|+\rangle\langle+| \\ &= \begin{pmatrix} 3/4&1/4 \\ 1/4&1/4 \end{pmatrix}. \end{aligned}$$

Squaring and tracing gives

$$\operatorname{Tr}(\bar\rho^2) = \frac34.$$

The signal states are nonorthogonal:

$$\lvert\langle0|+\rangle\rvert^2 = \frac12.$$

Their trace distance is $D=\sqrt{1-1/2}=1/\sqrt2$, so the Holevo–Helstrom formula gives

$$P_{\mathrm{guess}}^{\mathrm{opt}} =\frac{1+1/\sqrt2}{2}\simeq0.853553.$$

An optimal measurement projects onto the positive and negative eigenspaces of the difference of the two signal density operators. Perfect identification on every run is impossible without additional information. An unambiguous-discrimination protocol can instead make error-free conclusive identifications, but must allow and report inconclusive outcomes.

3. Reduce a Bell state

For

$$|\Phi^+\rangle = \frac{|00\rangle+|11\rangle}{\sqrt2},$$

compute $\rho_A=\operatorname{Tr}_B|\Phi^+\rangle\langle\Phi^+|$ and its purity.

Solution

Expanding the projector gives two diagonal terms and two cross terms. The cross terms vanish under the trace over $B$ because $\langle0|1\rangle=0$. Thus

$$\rho_A = \frac12 \left( |0\rangle\langle0| + |1\rangle\langle1| \right) = \frac I2.$$

Its purity is

$$\operatorname{Tr}(\rho_A^2) = \frac12.$$

The global state is pure while each one-qubit marginal is maximally mixed. The mixed marginal reflects entanglement with the other qubit. Purity itself is not an additive conserved quantity to be divided between subsystems and correlations.

4. Construct a purification

Let

$$\rho_A = q|0\rangle\langle0| + (1-q)|1\rangle\langle1|.$$

For $0\le q\le1$, construct a purification on two qubits and verify it by tracing out the reference.

Solution

One purification is

$$|\Psi\rangle_{AR} = \sqrt q\,|0\rangle_A|0\rangle_R + \sqrt{1-q}\,|1\rangle_A|1\rangle_R.$$

The projector contains two diagonal and two cross terms. Tracing over $R$ removes the cross terms:

$$\begin{aligned} \operatorname{Tr}_R |\Psi\rangle\langle\Psi| &= q|0\rangle\langle0| + (1-q)|1\rangle\langle1| \\ &= \rho_A. \end{aligned}$$

Applying any unitary to $R$ produces another purification of the same $\rho_A$.

5. Compare commuting states

Let

$$\rho = \operatorname{diag}(0.8,0.2), \qquad \sigma = \operatorname{diag}(0.5,0.5).$$

Compute their trace distance and squared fidelity, then check the Fuchs–van de Graaf bounds.

Solution

Because the states commute, the trace distance reduces to the classical total-variation distance:

$$\begin{aligned} D(\rho,\sigma) &= \frac12 \left( |0.8-0.5|+|0.2-0.5| \right) \\ &=0.3. \end{aligned}$$

The squared fidelity is

$$\begin{aligned} F(\rho,\sigma) &= \left( \sqrt{0.8\cdot0.5} + \sqrt{0.2\cdot0.5} \right)^2 \\ &= \left( \sqrt{0.4}+\sqrt{0.1} \right)^2 \\ &=0.9. \end{aligned}$$

The bounds give

$$1-\sqrt{0.9} \leq 0.3 \leq \sqrt{0.1},$$

which is true numerically:

$$0.051\ldots \leq 0.3 \leq 0.316\ldots.$$

6. Postselection and distinguishability

Consider qutrit states

$$\rho=(1-\varepsilon)|0\rangle\langle0|+\varepsilon|1\rangle\langle1|, \qquad \sigma=(1-\varepsilon)|0\rangle\langle0|+\varepsilon|2\rangle\langle2|,$$

where $0<\varepsilon<1$. Use the two projective instrument branches with Kraus operators $P=|1\rangle\langle1|+|2\rangle\langle2|$ and $I-P$. Calculate the initial trace distance, the probability of the $P$ branch, and its conditional output states. Does their increased distance violate contractivity?

Solution

The difference is $\rho-\sigma=\varepsilon\operatorname{diag}(0,1,-1)$, so $D(\rho,\sigma)=\varepsilon$. The successful subnormalized outputs are $\varepsilon|1\rangle\langle1|$ and $\varepsilon|2\rangle\langle2|$. Each has trace $\varepsilon$; after normalization the states are orthogonal and have distance one.

There is no contradiction. Conditioning includes division by a success probability, so the normalized successful branch is not a trace-preserving linear channel on arbitrary inputs. The other branch has probability $1-\varepsilon$ and gives the same state $|0\rangle$ for both preparations. If both outcomes and their flags are retained, the total trace distance is still $\varepsilon$. A rare perfect result is different from an unconditional perfect result.

Continue to quantum entropy for information measures, state tomography for inference from data, and quantum instruments for general measurement updates.

References

• Fano, U. “Description of states in quantum mechanics by density matrix and operator techniques.” Reviews of Modern Physics 29, 74–93 (1957). doi:10.1103/RevModPhys.29.74.
• Fuchs, C. A., and J. van de Graaf. “Cryptographic distinguishability measures for quantum-mechanical states.” IEEE Transactions on Information Theory 45, 1216–1227 (1999). doi:10.1109/18.761271; author manuscript.
• Helstrom, C. W. Quantum Detection and Estimation Theory. Academic Press, 1976.
• Kraus, K. States, Effects, and Operations. Springer, 1983. doi:10.1007/3-540-12732-1.
• Nielsen, M. A., and I. L. Chuang. Quantum Computation and Quantum Information, 10th anniversary edition. Cambridge University Press, 2010. doi:10.1017/CBO9780511976667.
• Uhlmann, A. “The transition probability in the state space of a *-algebra.” Reports on Mathematical Physics 9, 273–279 (1976). doi:10.1016/0034-4877(76)90060-4.
• Watrous, J. The Theory of Quantum Information. Cambridge University Press, 2018. Chapters 2–3. Author’s manuscript and errata.