Quantum Channels and Operations

Recall that a density matrix $\rho$ is a positive semidefinite, trace-one operator on a Hilbert space $\mathcal{H}$, representing the most general quantum state, pure or mixed. A closed system evolves unitarily: $\rho \mapsto U\rho U^\dagger$. But most quantum systems are open: they interact with an environment we cannot fully track. We need the most general map $\mathcal{E}: \rho \mapsto \mathcal{E}(\rho)$ consistent with quantum mechanics. This map is called a quantum channel or quantum operation. Three physical requirements pin down exactly which maps are allowed. First, $\mathcal{E}$ must be linear, so that it respects statistical mixtures: if $\rho = p\rho_1 + (1-p)\rho_2$ describes a classical uncertainty about which state we have, then $\mathcal{E}(\rho) = p\,\mathcal{E}(\rho_1) + (1-p)\,\mathcal{E}(\rho_2)$. Second, $\mathcal{E}$ must be trace-preserving: $\mathrm{tr}(\mathcal{E}(\rho)) = \mathrm{tr}(\rho) = 1$, so that $\mathcal{E}(\rho)$ is again a valid state with total probability one (a channel that only sometimes succeeds, e.g. conditioned on a measurement outcome, is instead trace-decreasing, and is renormalized after conditioning). Third — and this is the subtle one — $\mathcal{E}$ must be completely positive (CP). A map is merely positive if it sends positive semidefinite operators to positive semidefinite operators, guaranteeing $\mathcal{E}(\rho) \geq 0$ whenever $\rho \geq 0$. This sounds sufficient, but it is not: it only controls what happens to the system in isolation. Physically, the system of interest may be entangled with a second system $R$ that the channel does not touch. We need $(\mathcal{E} \otimes \mathcal{I}_R)(\rho_{SR}) \geq 0$ for every joint state $\rho_{SR}$ and every auxiliary space $R$, no matter how large. A map with this stronger property for all extensions is called completely positive. The classic counterexample is the transpose map $\rho \mapsto \rho^T$: it is positive on single-qubit states, but $(\mathcal{T} \otimes \mathcal{I})$ applied to an entangled Bell state produces a matrix with a negative eigenvalue. Transposition is a mathematically fine positive map but describes no physical process — this is precisely why CP, not just positivity, is the correct axiom. A linear, trace-preserving, completely positive map is called a CPTP map, and this is the formal definition of a quantum channel. The central structural fact — due to Karl Kraus, building on the mathematical work of Stinespring and Choi — is that CPTP maps have a completely explicit representation in terms of a finite (or countable) set of operators, turning an abstract axiomatic definition into a concrete computational tool.

The Kraus representation. Kraus's theorem states that $\mathcal{E}$ is CPTP on a $d$-dimensional Hilbert space if and only if there exist operators $E_1, \dots, E_m$ (with $m \leq d^2$ needed in general) such that

The $E_k$ are called Kraus operators. The completeness relation $\sum_k E_k^\dagger E_k = I$ is exactly the algebraic statement of trace preservation: $\mathrm{tr}(\mathcal{E}(\rho)) = \mathrm{tr}(\rho \sum_k E_k^\dagger E_k) = \mathrm{tr}(\rho) = 1$.

Where the Kraus operators come from: the Stinespring picture. Every channel can be realized physically as follows. Introduce an environment (ancilla) system $E$ initialized in a pure state $|0\rangle_E$, evolve the joint system unitarily with some $U$ acting on $\mathcal{H}_S \otimes \mathcal{H}_E$, and then discard (trace out) the environment:

This is the Stinespring dilation. Writing $U$'s action in terms of an orthonormal basis $\{|k\rangle_E\}$ of the environment, define $E_k = \langle k|_E \, U \, |0\rangle_E$ (an operator on $\mathcal{H}_S$ alone). A short computation shows this reproduces the Kraus sum exactly, and $\sum_k E_k^\dagger E_k = I$ follows from unitarity of $U$. Conversely, given any Kraus set, one can construct such a $U$ — the two pictures are mathematically equivalent. This is the deep conceptual payoff: every noisy process is unitary evolution with an environment we chose not to look at. Noise is not a departure from unitary quantum mechanics; it is unitary quantum mechanics on a bigger system, viewed through a keyhole.

Non-uniqueness of the Kraus decomposition. The same channel $\mathcal{E}$ can be written with different Kraus sets. If $\{E_k\}$ and $\{F_j\}$ both represent $\mathcal{E}$, they are related by an isometry (extendable to a unitary matrix) $u_{jk}$:

after padding the shorter list with zero operators so both lists have the same length. This freedom reflects the fact that different physical unravelings of the same noise process (different environment measurement bases) give the same net effect on the system.

Composing channels. Channels compose like functions: applying $\mathcal{E}_1$ then $\mathcal{E}_2$ gives a new channel $\mathcal{E}_2 \circ \mathcal{E}_1$, whose Kraus operators are all products $F_j E_k$ over $j,k$. Sequential noisy gates in a real circuit are literally channel composition.

Algorithmic recipe: verifying and applying a channel. ``` INPUT: a candidate set of operators {E_1, ..., E_m} on a d-dimensional space OUTPUT: PASS/FAIL certificate that {E_k} defines a valid quantum channel, and a routine to apply the channel to any density matrix rho STEP 1 (trace preservation check): S <- sum_k E_k^dagger E_k if S != I (within numerical tolerance): FAIL -- not trace preserving; renormalize or reject STEP 2 (complete positivity check via Choi matrix): // Build the Choi matrix: apply (E ⊗ I) to the maximally entangled state // |Phi> = (1/sqrt(d)) sum_i |i>|i> C <- (E ⊗ I) ( |Phi> rho E_k^dagger return sigma // sigma is again a valid density matrix STEP 4 (compose two channels E, F given as Kraus sets): function COMPOSE({E_k}, {F_j}): return { F_j * E_k for all j, k } // apply E first, then F ``` Step 2 uses the Choi–Jamiołkowski isomorphism: a linear map $\mathcal{E}$ is completely positive if and only if its Choi matrix $C_\mathcal{E} = (\mathcal{E} \otimes \mathcal{I})(|\Phi\rangle\langle\Phi|)$, built from the maximally entangled state $|\Phi\rangle = \frac{1}{\sqrt d}\sum_i |i\rangle|i\rangle$, is positive semidefinite. This turns a hard-to-check condition ("positive on all extensions, for all ancilla dimensions") into a single finite eigenvalue computation on one fixed $d^2 \times d^2$ matrix — a dramatic simplification that makes complete positivity checkable in practice.

Correctness of the Kraus representation (proof sketch). The forward direction — that any Kraus sum is CPTP — is a direct check. Linearity and trace preservation are immediate from the formula and the completeness relation. Complete positivity holds because for any extension state $\sigma_{SR} \geq 0$,

and each term $A \sigma_{SR} A^\dagger$ is positive semidefinite whenever $\sigma_{SR}$ is (a standard fact: $\langle \psi | A\sigma A^\dagger|\psi\rangle = \langle \phi|\sigma|\phi\rangle \geq 0$ for $|\phi\rangle = A^\dagger|\psi\rangle$), and a sum of positive semidefinite operators is positive semidefinite. The converse — every CPTP map has a Kraus form — follows from the Choi isomorphism. Since $\mathcal{E}$ is completely positive, its Choi matrix $C_\mathcal{E} \geq 0$, so it has a spectral decomposition $C_\mathcal{E} = \sum_k \lambda_k |v_k\rangle\langle v_k|$ with $\lambda_k \geq 0$. Reshaping each eigenvector $|v_k\rangle \in \mathcal{H} \otimes \mathcal{H}$ into an operator $E_k = \sqrt{\lambda_k}\, \mathrm{reshape}(|v_k\rangle)$ on $\mathcal{H}$ recovers a Kraus set reproducing $\mathcal{E}$ exactly. Since the Choi matrix lives on a $d^2$-dimensional space, it has at most $d^2$ nonzero eigenvalues, which gives the tight bound $m \leq d^2$ on the minimal number of Kraus operators needed for a $d$-dimensional system — for a single qubit, at most 4.

Non-uniqueness, formally. Suppose $\{E_k\}_{k=1}^m$ and $\{F_j\}_{j=1}^m$ (padded with zeros to equal length) both represent $\mathcal{E}$. Their Choi matrices are equal, and two square-root decompositions of the same positive semidefinite matrix are related by an isometry acting on the index that was summed over — this is a standard linear-algebra fact about factorizations $C = AA^\dagger = BB^\dagger$, giving $B=Au$ for unitary $u$. Translating back yields $F_j = \sum_k u_{jk} E_k$ as claimed.

Complexity of the verification algorithm. For a $d$-dimensional system with $m$ Kraus operators:

• Checking $\sum_k E_k^\dagger E_k = I$ costs $O(m d^3)$ (each matrix product and sum is $O(d^3)$, done $m$ times with naive matrix multiplication; $O(m d^{2.37})$ with fast matrix multiplication, though this is rarely the bottleneck in practice).
• Building the Choi matrix costs $O(m d^4)$ time and $O(d^4)$ space, since $C_\mathcal{E}$ is a $d^2 \times d^2$ matrix built from $m$ applications of $E_k$ to a $d^2$-dimensional vector.
• Diagonalizing $C_\mathcal{E}$ to check positivity costs $O(d^6)$ with standard eigenvalue algorithms.

So the classical cost of verifying or simulating a channel scales polynomially in $d$ but the Choi matrix's $d^2\times d^2$ size means the eigenvalue step dominates and scales like $d^6$ — cheap for one or two qubits ($d=2,4$: $d^6 = 64, 4096$), but this exponential blow-up in the number of qubits ($d = 2^n \Rightarrow d^6 = 2^{6n}$) is exactly why simulating generic open quantum systems classically is intractable, foreshadowing why genuinely quantum evolution (including noisy evolution) is believed to escape efficient classical simulation — a theme that leads directly into the definition of The Complexity Class BQP as decision problems solvable by poly-size quantum circuits, where the channel formalism is what makes "efficient" and "physically realizable" precisely defined even for noisy, non-unitary quantum computers.

Honesty about tradeoffs. The Kraus representation is computationally convenient but physically non-unique — you cannot ask "which unraveling is really happening" without additional physical input (e.g., a specific continuous monitoring scheme), since all Kraus sets related by a unitary mixing give indistinguishable predictions for the system alone. The Choi matrix test is exact and complete, but its $d^4$ memory footprint makes it impractical for large multi-qubit systems; approximate methods (Pauli transfer matrices, randomized benchmarking) are used instead in real device characterization, trading exactness for scalability.

We trace the amplitude damping channel, which models spontaneous emission: an excited qubit state $|1\rangle$ decays to $|0\rangle$ with probability $\gamma$ per application, emitting a photon into the environment.

Setup. Kraus operators:

Check completeness.

sum $= I$. Good — this is a valid trace-preserving map for any $\gamma \in [0,1]$.

Input state. Take $\rho = |+\rangle\langle+|$ where $|+\rangle = \frac{1}{\sqrt2}(|0\rangle+|1\rangle)$, i.e.

Let $\gamma = 0.3$, so $\sqrt{1-\gamma} = \sqrt{0.7} \approx 0.8367$ and $\sqrt\gamma = \sqrt{0.3} \approx 0.5477$.

Step-by-step trace.

Interpretation. The output density matrix

has trace $0.65+0.3499 \approx 1.0$ (matches, up to rounding) confirming trace preservation. Compare to the input: the population of $|1\rangle$ dropped from $0.5$ to $0.3499$ — some of it decayed to $|0\rangle$ — and the off-diagonal coherence shrank from $0.5$ to $0.4183$, since amplitude damping also degrades coherence as a side effect of the population decay (this is why $T_1$ decay processes also contribute to $T_2$ dephasing in real hardware specs). Eigenvalues of the output are approximately $0.973$ and $0.027$, both non-negative, confirming $\mathcal{E}(\rho) \geq 0$ as required by positivity — a necessary (not sufficient) sanity check that the computation is consistent with $\mathcal{E}$ being CPTP. As a second, contrasting case, apply the completely dephasing channel $\mathcal{E}'(\rho) = (1-p)\rho + p\, Z\rho Z$ with $p=0.5$ (Kraus operators $E_0=\sqrt{0.5}\,I$, $E_1=\sqrt{0.5}\,Z$) to the same $\rho$. Direct computation gives

the maximally mixed state — total decoherence, with populations unchanged but all coherence destroyed. Contrasting the two channels side by side shows how amplitude damping mixes population change with coherence loss, while pure dephasing destroys coherence while leaving populations exactly fixed — a distinction that matters enormously when choosing an error-correcting code tailored to a device's dominant noise mechanism.

Quantum error correction codes are designed and analyzed entirely in the language of channels: a code "corrects" a noise channel $\mathcal{N}$ if there exists a recovery channel $\mathcal{R}$ with $\mathcal{R} \circ \mathcal{N}$ acting as the identity on the code space, and the celebrated Knill–Laflamme conditions are stated directly on the Kraus operators of $\mathcal{N}$. Hardware vendors (IBM, Google, Rigetti, IonQ) report gate and readout errors via process tomography, which reconstructs a device's actual Kraus operators (or equivalently its Choi matrix) from experimental data, feeding directly into calibrated noise models used in compilers and error-mitigation software. Quantum key distribution security proofs bound an eavesdropper's information gain by optimizing over all CPTP attacks consistent with observed statistics — the completely-positive axiom is what makes "worst-case adversary" a mathematically well-posed, exhaustive search rather than a guess. Continuous-variable and photonic quantum computing platforms describe photon loss, thermal noise, and detector inefficiency as CPTP channels acting on Gaussian states. Finally, the entire framework of fault-tolerant thresholds — how much physical noise a quantum computer can tolerate while still simulating an ideal circuit efficiently — is a statement about which noisy channels admit efficient correction, a question that underlies the operational, noise-tolerant definition of The Complexity Class BQP and connects back to the role of Quantum Entropy and von Neumann Entropy in quantifying how much information a channel irreversibly leaks to its environment.

Ask the tutor: How does the choice of Kraus decomposition affect quantum trajectory simulations versus the average (master equation) evolution of an open quantum system?