Evolution of Rho

You know how a ket evolves. Apply a gate — |\psi\rangle \mapsto U|\psi\rangle. Measure — collapse to the projected component, renormalise. Let a Hamiltonian act for time t — |\psi(t)\rangle = e^{-iHt/\hbar}|\psi(0)\rangle. Three rules, and you can simulate any isolated quantum system on paper.

But the previous chapter introduced \rho, and that changes what you need. If you flip a coin to decide which ket you prepared, no ket describes the full state — only a density matrix does. If the qubit is entangled with an environment you cannot see, again, only \rho. And if you are running a real experiment in an Indian NMR lab at IIT Madras, where 10^{17} molecules carry a thermal equilibrium and the qubit lives as a tiny deviation matrix on top of that background, the object that evolves in time is a density matrix, not a ket.

This chapter writes down how \rho evolves. Four rules, each derived from the ket version, each reducing to the ket version when \rho happens to be pure, each doing new work when \rho is mixed.

Unitary evolution — the sandwich rule

Start with the easy case. You have a state |\psi\rangle. You apply a unitary gate U to it. The new ket is U|\psi\rangle. Its density matrix is

Why the dagger ends up on the right: (U|\psi\rangle)^\dagger = \langle\psi|U^\dagger — when you dagger a product, you reverse the order and dagger each piece. That is how U gets flipped to U^\dagger on the right-hand side of the outer product.

So for a pure state, the rule is: \rho \mapsto U\rho U^\dagger. A mixed state is just a classical mixture of pure states, and linearity carries the rule through to any \rho.

Why the three axioms survive

Each axiom of \rho comes through cleanly.

• Hermiticity. (U\rho U^\dagger)^\dagger = (U^\dagger)^\dagger \rho^\dagger U^\dagger = U\rho U^\dagger, using \rho^\dagger = \rho (Hermitian) and the dagger-reversal rule.
• Positive semi-definiteness. For any state |\phi\rangle: \langle\phi|U\rho U^\dagger|\phi\rangle = \langle U^\dagger\phi|\rho|U^\dagger\phi\rangle \geq 0, because \rho is positive semi-definite on every vector — including the vector U^\dagger|\phi\rangle.
• Unit trace. \text{tr}(U\rho U^\dagger) = \text{tr}(U^\dagger U\,\rho) = \text{tr}(I\rho) = \text{tr}(\rho) = 1, using the cyclic property of the trace.

Why unitarity is exactly the right condition: if you replaced U by a generic invertible matrix M, Hermiticity would still survive, but \text{tr}(M\rho M^\dagger) \neq \text{tr}(\rho) in general — the trace would change — so you would exit the space of valid density matrices. The unitary condition U^\dagger U = I is precisely what keeps you inside.

Derivation from the ensemble

Another way to see the same rule. Take any ensemble decomposition \rho = \sum_i p_i |\psi_i\rangle\langle\psi_i|. Each ket evolves |\psi_i\rangle \mapsto U|\psi_i\rangle, with classical probabilities p_i untouched (because unitaries act on the quantum state, not on the classical probabilities that govern which state you prepared). So

Every ensemble decomposition gives the same answer — as it must, because the same \rho cannot pick different futures based on which of its many equivalent ensemble realisations you happened to write.

Continuous-time unitary evolution — the Liouville-von Neumann equation

A quantum gate is a one-shot unitary. A Hamiltonian is a continuous one — it generates unitary evolution over time. The Schrödinger equation i\hbar\,\tfrac{d}{dt}|\psi\rangle = H|\psi\rangle integrates to |\psi(t)\rangle = U(t)|\psi(0)\rangle with U(t) = e^{-iHt/\hbar}. What is the analogous differential equation for \rho?

Start from \rho(t) = U(t)\rho(0) U(t)^\dagger and differentiate. Using \tfrac{dU}{dt} = -\tfrac{iH}{\hbar} U and \tfrac{dU^\dagger}{dt} = \tfrac{iH}{\hbar}U^\dagger (Hermiticity of H):

Why only one term has H on the left and the other on the right: H and U commute (U = e^{-iHt/\hbar} is a function of H), but H and \rho(0) do not. So H slides through to multiply \rho = U\rho(0)U^\dagger on the outside, landing on the left in the first term and on the right in the second.

Combine: \tfrac{d\rho}{dt} = -\tfrac{i}{\hbar}(H\rho - \rho H) = -\tfrac{i}{\hbar}[H, \rho], where [H, \rho] = H\rho - \rho H is the commutator. Rearranging:

The equation is named after two people. Joseph Liouville wrote down a very similar equation for classical phase-space distributions in 1838 — it governs how a classical probability cloud flows under Hamilton's equations. John von Neumann — working at Göttingen in the 1920s — saw the quantum analogue: the commutator [H, \rho] plays the role of the classical Poisson bracket, and the rest is the same structure. The classical distribution f(\vec q, \vec p, t) and the quantum density matrix \rho(t) satisfy almost the same equation, with one replacement: \{H, f\} \to \tfrac{1}{i\hbar}[H, \rho].

Why this is more than cosmetic: classical statistical mechanics and quantum statistical mechanics share a geometric structure — both describe the flow of probability (classical or quantum) under a generator. This is why so much of classical intuition (conservation laws, invariant measures, ergodicity) carries over to the quantum setting. The Liouville-von Neumann equation is the bridge.

Stationary states and thermal equilibrium

One immediate consequence of i\hbar\,\tfrac{d\rho}{dt} = [H, \rho]: if [H, \rho] = 0, then \tfrac{d\rho}{dt} = 0 and \rho is stationary in time. Any \rho that is a function of H (for example, \rho \propto e^{-\beta H}, the Gibbs thermal state at inverse temperature \beta) automatically commutes with H and therefore does not evolve. This is the quantum version of "equilibrium distributions do not evolve" — the quantum thermal state is a fixed point of the Liouville-von Neumann flow.

Measurement evolution — collapse, recorded

Measurement is where things get interesting. The ket rule you know from the projective measurement chapter is

where \{P_m\} is a complete set of orthogonal projectors. Translate to a density matrix. If outcome m is observed:

Why P_m^\dagger becomes P_m: projectors are Hermitian, P_m^\dagger = P_m, so the dagger in the outer product drops. And why the denominator is \text{tr}(P_m\rho): that is the density-matrix form of p(m) = \langle\psi|P_m|\psi\rangle, which you derived in the density operator chapter.

Linearity extends this from pure states to any \rho.

Measurement evolution — collapse, not recorded

The genuinely new feature of density-matrix evolution is this: what if you perform the measurement but throw away the outcome? Classically this is unremarkable — if a coin flips somewhere and nobody tells you the result, your best description is the 50-50 probability distribution you had before anyone flipped. In quantum mechanics, the same logic gives a deep new phenomenon: pure states can become mixed.

Average over the outcomes, weighted by their probabilities:

The p(m) factors cancel beautifully. The result is just the sum \sum_m P_m\rho P_m, each projector sandwiching \rho, with no extra weights.

Why pure states become mixed

Try it on |+\rangle\langle+| in the computational basis. With P_0 = |0\rangle\langle 0| and P_1 = |1\rangle\langle 1|:

Compute P_0 \rho P_0 — pick out the top-left 1\times 1 block, pad with zeros: \tfrac{1}{2}\begin{pmatrix}1 & 0 \\ 0 & 0\end{pmatrix}. Compute P_1 \rho P_1: pick the bottom-right block: \tfrac{1}{2}\begin{pmatrix}0 & 0 \\ 0 & 1\end{pmatrix}. Sum: \tfrac{1}{2}\begin{pmatrix}1 & 0 \\ 0 & 1\end{pmatrix} = \tfrac{I}{2}.

A pure state — |+\rangle\langle+|, Bloch vector (1, 0, 0), purity 1, on the surface of the Bloch sphere — has become the maximally mixed state I/2 — Bloch vector (0, 0, 0), purity 1/2, the centre of the ball. The off-diagonal entries have been wiped out. This is decoherence in one line: measure in the computational basis, throw away the outcome, and the coherent superposition becomes a classical probability distribution.

Why this is not a violation of anything: no quantum information has been created or destroyed — it has been transferred. The superposition in |+\rangle (a coherent 1/\sqrt 2-1/\sqrt 2 split between |0\rangle and |1\rangle) has been replaced by classical ignorance (a 1/2-1/2 mixture), because the act of measurement created a correlated record somewhere (in your apparatus, your brain, the universe). Ignoring that record is the act that turns "coherent superposition" into "classical mixture." The full joint state of system-plus-record is still pure.

Generalised measurements — POVMs

Projective measurements are the ones where \{P_m\} are genuine orthogonal projectors. The most general kind of quantum measurement drops the orthogonality and allows any set of operators \{M_m\} with \sum_m M_m^\dagger M_m = I. The probability of outcome m and the post-measurement state are

The set \{E_m = M_m^\dagger M_m\} is called a POVM — Positive Operator-Valued Measure — and it is the most general description of a quantum measurement allowed by the formalism. Every projective measurement is a POVM (with M_m = P_m and E_m = P_m), but not every POVM is projective. POVMs arise naturally when a projective measurement on a larger system is restricted to a subsystem — the POVM is the "shadow" of the projective measurement that lives only on the system you can see.

POVMs matter for two reasons. First, they describe weak measurements and unsharp measurements used in real experiments. Second, they are the right tool for distinguishing non-orthogonal states — you cannot do that with a projective measurement alone, but a POVM with more than two outcomes can. This chapter introduces them only as a name; the full treatment waits for a later chapter.

Summary — four rules on one card

Every rule is a direct density-matrix translation of the corresponding ket rule — with one exception. The "outcome discarded" rule has no pure-state analogue, because kets simply cannot describe the mixed state that results. This is not a bug in the ket formalism; it is the reason density matrices exist.

Worked examples

Open systems and the Lindblad master equation

The Liouville-von Neumann equation describes a closed system — one isolated from its environment. Real experiments are not closed. A superconducting qubit in a dilution refrigerator couples weakly to vibrational modes of the chip and to stray electromagnetic noise. A trapped ion couples to the background blackbody radiation. An NMR nuclear spin couples to the random tumbling of surrounding water molecules. In every case, the system alone is not governed by a Hamiltonian — because tracing out the environment breaks the unitary evolution.

The most general continuous-time evolution of an open quantum system (under reasonable assumptions — Markovianity, trace preservation, complete positivity) is captured by the Lindblad master equation:

where the anticommutator \{A, B\} = AB + BA. The operators \{L_k\} are called Lindblad operators and they encode the different dissipation and decoherence channels available to the environment. The first term is the familiar unitary flow; the second term is the non-unitary part that describes energy loss (T_1 processes), pure dephasing (T_2 processes), and every other physical form of decoherence.

You will not derive this equation in this chapter — the derivation is a careful exercise in tracing out an environment under the Born-Markov approximation and belongs in the dedicated Lindblad master equation chapter. The point here is to flag what the equation does: it is the open-system replacement for the Liouville-von Neumann equation. When the dissipation is turned off (L_k = 0 for all k), the Lindblad equation reduces to Liouville-von Neumann and closed-system dynamics is recovered. When the dissipation is turned on, a pure state smoothly evolves into a mixed state over a timescale set by \|L_k\|^2.

Indian NMR quantum computing groups at TIFR Mumbai have been working with the Lindblad equation since the 1990s — NMR is the quintessential open-system platform, with T_1 and T_2 times on the order of seconds for nuclear spins, much longer than the millisecond gate times. The Lindblad formalism is how theorists analyse NMR pulse sequences in the presence of relaxation, and it is the language in which every decoherence-free subspace and dynamical decoupling protocol is expressed.

Common confusions

• "U\rho U^\dagger requires U unitary." Yes — this is the point. For a generic matrix M, the expression M\rho M^\dagger is still Hermitian and positive semi-definite, but its trace is not 1 in general, so it is not a valid density matrix. Only U^\dagger U = I guarantees trace preservation. Dropping the unitary condition lands you in the territory of quantum channels (next chapter), where you sandwich with multiple operators and require \sum_k K_k^\dagger K_k = I to restore trace.

• "Measurement takes pure states to mixed states — always." Only if you discard the outcome. If you record the outcome, the post-measurement state is still pure (it is an eigenstate of the observable). The transition pure \to mixed happens when you average over outcomes you chose not to look at. The mixing is a statement about your information, not about the quantum state in isolation.

• "Schrödinger's equation is |\psi\rangle \mapsto H|\psi\rangle. So the density-matrix version is \rho \mapsto H\rho." No. The Schrödinger equation is i\hbar\,d|\psi\rangle/dt = H|\psi\rangle — a first-order time derivative, not a matrix product. The density-matrix version is i\hbar\,d\rho/dt = [H, \rho] — a commutator of H with \rho, which is essential. Dropping the commutator breaks Hermiticity of \rho and violates trace preservation.

• "The Liouville-von Neumann equation is classical, not quantum." Joseph Liouville's 1838 equation is classical (for phase-space distributions). Von Neumann's 1920s quantum analogue uses the commutator in place of the Poisson bracket, and it governs density matrices, not classical distributions. The name "Liouville-von Neumann" honours both — the structural parallel is real, but the quantum version is a genuinely quantum equation.

• "Lindblad evolution is just decoherence slowly." Lindblad dynamics includes decoherence (the off-diagonal part) and dissipation (the diagonal part — energy exchange with the environment). A qubit coupled to a cold bath will lose energy until it reaches the thermal ground state; a qubit with only pure-dephasing noise will keep its populations but lose off-diagonal coherences. Both are Lindblad; they correspond to different choices of the operators \{L_k\}.

• "Unitary evolution preserves purity; anything else reduces it." Unitary evolution preserves \text{tr}(\rho^2), yes. Projective measurement with outcome recorded also preserves purity (the post-measurement state is an eigenstate, which is pure). Only unread measurements and open-system dynamics can genuinely reduce purity. Purity is conserved by every "clean" rule and destroyed by information loss.

Going deeper

Derivation sketch of the Lindblad master equation

The Lindblad master equation can be derived by starting from a closed system-plus-environment pair, tracing out the environment, and applying the Born-Markov approximations (weak system-environment coupling and short environment correlation time). Here is the skeleton.

Let H_{\text{tot}} = H_S \otimes I_E + I_S \otimes H_E + \lambda V_{SE} be the total Hamiltonian, with V_{SE} the system-environment interaction. Start from the Liouville-von Neumann equation for the full density matrix \rho_{SE}. Move to the interaction picture. Integrate formally. Take the partial trace over the environment. Apply the Born approximation (the environment stays close to its initial state) and the Markov approximation (the environment forgets its past faster than the system evolves). Do a secular approximation to drop rapidly-oscillating terms. What survives is the Lindblad form:

The operators L_k are built from matrix elements of V_{SE} and spectral densities of the environment. The Hamiltonian H_S^{\text{eff}} picks up a small Lamb-shift correction from the environment's back-action. Göran Lindblad (1976) and, independently, Gorini, Kossakowski, and Sudarshan (1976) proved the general theorem: every Markovian completely-positive trace-preserving semigroup has this exact form. So writing down a Lindblad equation is not an ad-hoc modelling choice — it is the unique form compatible with the axioms of Markovian open-system dynamics.

The full derivation is in Breuer and Petruccione's The Theory of Open Quantum Systems (Oxford, 2002), which is the standard reference for open-system evolution.

The Heisenberg picture for density matrices

You are used to the Schrödinger picture — states evolve, observables stay fixed. The Heisenberg picture flips this: states stay fixed, observables evolve. For a density matrix \rho and an observable A, the expectation value \langle A\rangle = \text{tr}(A\rho) can be written in either picture:

• Schrödinger: \rho(t) = U(t)\rho(0) U(t)^\dagger, A fixed. \langle A\rangle_t = \text{tr}(A\,\rho(t)).
• Heisenberg: \rho fixed, A(t) = U(t)^\dagger A\,U(t). \langle A\rangle_t = \text{tr}(A(t)\,\rho) = \text{tr}(U^\dagger A U \rho).

The two are equivalent by the cyclic property of the trace: \text{tr}(A U\rho U^\dagger) = \text{tr}(U^\dagger A U\,\rho). The Heisenberg-picture evolution equation for A is

which has a sign opposite to the Liouville-von Neumann equation for \rho — a well-known feature that sometimes bites students. The sign flip is consistent with the two pictures being related by time-reversal of the evolution (one evolves forward, the other effectively backward).

The Heisenberg picture is the natural setting for Hamiltonian simulation, the derivation of Ehrenfest's theorem, and a huge chunk of quantum field theory. For quantum computing, the Schrödinger picture tends to dominate — you evolve the state, you measure the state — but the Heisenberg-picture identities are useful when writing observables as explicit functions of time.

Entropy production under Lindblad dynamics

The von Neumann entropy S(\rho) = -\text{tr}(\rho\log\rho) is conserved by unitary evolution. Under Lindblad dynamics, it is not conserved — entropy can increase (system becomes more mixed) or decrease (system becomes less mixed, with the environment absorbing the entropy difference). The Spohn inequality gives a concrete bound: for a Lindblad evolution with a thermal fixed point \rho_\beta at inverse temperature \beta, the relative entropy S(\rho\|\rho_\beta) monotonically decreases, and the entropy production rate is

This is the quantum version of the second law of thermodynamics. It holds for every Markovian Lindblad evolution and underlies the thermodynamic consistency of quantum open-system models.

Numerical integration of Lindblad equations

For a d-dimensional Hilbert space, \rho has d^2 real parameters (constrained to d^2 - 1 by unit trace), and the Lindblad equation is a linear ODE on this parameter space. Standard integrators (RK4, Crank-Nicolson) work — but for large d (say, n = 10 qubits gives d = 1024, d^2 = 10^6), direct integration becomes expensive.

Two popular alternatives. Quantum trajectories (or the Monte Carlo wave-function method): unravel the Lindblad equation into stochastic pure-state trajectories, each undergoing conditional non-unitary evolution punctuated by random "jumps" from the operators L_k. Average many trajectories to recover \rho(t). Each trajectory lives on a d-dimensional ket (not a d^2-dimensional \rho), which is a big memory saving. Vectorisation: treat \rho as a d^2-dimensional vector and the Lindblad superoperator as a d^2\times d^2 matrix, then apply standard linear-ODE methods.

The QuTiP library in Python implements both approaches and is the default tool for Lindblad simulations in the quantum-computing research community. It is used extensively by the Indian quantum-computing research groups at IIT Madras, IISc Bangalore, and TIFR Mumbai.

Where this leads next

• Kraus Representation — the next chapter. The most general one-shot operation on \rho, including unitaries, measurements, and noise, all in one framework.
• Quantum Channels — the category of completely-positive trace-preserving maps, the formal home of every physically-realisable quantum operation.
• Lindblad Master Equation — the continuous-time, open-system generalisation of the Liouville-von Neumann equation in full detail.
• Decoherence Introduction — the physics of how coherent superpositions become classical mixtures, explained through the off-diagonal-entry argument of this chapter.
• Density Operator — the previous chapter, where \rho was formally defined.
• Projective Measurement — the ket-level treatment of measurement, which this chapter generalises to \rho.

References

1. Wikipedia, Lindbladian — definitions, properties, and derivation sketch of the Lindblad master equation.
2. Nielsen and Chuang, Quantum Computation and Quantum Information, §8.2 (the operator-sum representation and density-matrix evolution) — Cambridge University Press.
3. John Preskill, Lecture Notes on Quantum Computation, Ch. 3 — theory.caltech.edu/~preskill/ph229.
4. Wikipedia, Von Neumann equation — the Liouville-von Neumann equation and its classical counterpart.
5. Heinz-Peter Breuer and Francesco Petruccione, The Theory of Open Quantum Systems — summary and chapter preview at Wikipedia: Open quantum system.
6. John Watrous, The Theory of Quantum Information (2018), Ch. 2 — cs.uwaterloo.ca/~watrous/TQI.