Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Quantum Channels: Kraus & Stinespring

CPTP maps, the Choi isomorphism and dilations.

Which linear maps on density matrices are physical? Not all positive ones: the transpose maps states to states, yet no machine implements it. The right answer, complete positivity, has three equivalent faces: Kraus operators, a Stinespring dilation into a larger unitary world, and a single positive matrix, the Choi matrix. This lesson proves they are the same and shows what each is good for.

What you’ll learn

  • Define positive, completely positive and trace-preserving maps, and prove that the transpose is positive but not completely positive.
  • Prove the equivalence of complete positivity, positivity of the Choi matrix, the Kraus form and the Stinespring form, including the trace-preservation conditions.
  • Characterize the freedom in Kraus representations and Stinespring dilations.
  • Use the Choi–Jamiołkowski isomorphism to compute Kraus operators, complementary channels and channel properties.

Why complete positivity?

Let L(H)\mathcal L(\Hil) denote linear operators on H\Hil, with dA=dim⁡HAd_A=\dim\Hil_A and dB=dim⁡HBd_B=\dim\Hil_B. A linear map Φ:L(HA)→L(HB)\Phi:\mathcal L(\Hil_A)\to\mathcal L(\Hil_B) that represents a physical process must send states to states. That makes it positive (X≥0⇒Φ(X)≥0X\ge0\Rightarrow\Phi(X)\ge0) and trace-preserving. But the system may be entangled with a reference RR on which nothing happens. The joint output (idR⊗Φ)(ρRA)(\mathrm{id}_R\otimes\Phi)(\rho_{RA}) must also be a state, for every RR.

DerivationThe transpose is positive but not completely positive
  1. X≥0 ⟹ XT=X‾≥0X\ge0\ \Longrightarrow\ X^{\mathsf T}=\overline{X}\ge0

    For Hermitian XX, transposition equals complex conjugation, which preserves the spectrum. So TT is positive, and trace-preserving.

  2. ∣Φ+⟩ ⁣⟨Φ+∣=12∑i,j∣i⟩ ⁣⟨j∣⊗∣i⟩ ⁣⟨j∣ → id⊗T  12∑i,j∣i⟩ ⁣⟨j∣⊗∣j⟩ ⁣⟨i∣=12F\dyad{\Phi^+}=\tfrac12\sum_{i,j}\ketbra ij\otimes\ketbra ij\ \xrightarrow{\ \mathrm{id}\otimes T\ }\ \tfrac12\sum_{i,j}\ketbra ij\otimes\ketbra ji=\tfrac12F

    Apply it to one half of a Bell pair. The result is half the swap operator F∣a⟩∣b⟩=∣b⟩∣a⟩F\ket{a}\ket{b}=\ket{b}\ket{a}.

  3. F∣Ψ−⟩=−∣Ψ−⟩ ⟹ spec(12F)={12,12,12,−12}F\ket{\Psi^-}=-\ket{\Psi^-}\ \Longrightarrow\ \mathrm{spec}\big(\tfrac12F\big)=\big\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\big\}

    The singlet is antisymmetric, so id⊗T\mathrm{id}\otimes T outputs an operator with a negative eigenvalue: TT is not even 2-positive. ■\blacksquare

The structure theorem

Fix the unnormalized maximally entangled vector ∣Ω⟩=∑i=1dA∣i⟩∣i⟩∈HA⊗HA\ket\Omega=\sum_{i=1}^{d_A}\ket i\ket i\in\Hil_A\otimes\Hil_A and define the Choi matrix

J(Φ)=(id⊗Φ)(∣Ω⟩ ⁣⟨Ω∣)=∑i,j∣i⟩ ⁣⟨j∣⊗Φ(∣i⟩ ⁣⟨j∣) ∈ L(HA⊗HB).J(\Phi)=(\mathrm{id}\otimes\Phi)\big(\dyad\Omega\big)=\sum_{i,j}\ketbra ij\otimes\Phi\big(\ketbra ij\big)\ \in\ \mathcal L(\Hil_A\otimes\Hil_B).
DerivationProof
  1. (1 ⇒ 2). ∣Ω⟩ ⁣⟨Ω∣≥0\dyad\Omega\ge0, and iddA⊗Φ\mathrm{id}_{d_A}\otimes\Phi is positive by assumption, so J(Φ)≥0J(\Phi)\ge0. (Only dAd_A-positivity was used.)

  2. J=∑k=1r∣vk⟩ ⁣⟨vk∣,∣vk⟩=(I⊗Kk)∣Ω⟩,  i.e. Kk∣i⟩:=(⟨i∣⊗I)∣vk⟩J=\sum_{k=1}^r\dyad{v_k},\qquad \ket{v_k}=(\id\otimes K_k)\ket\Omega,\ \ \text{i.e. } K_k\ket i:=\big(\bra i\otimes\id\big)\ket{v_k}

    (2 ⇒ 3). Write the positive matrix JJ as a sum of r=rank⁡Jr=\rank J rank-one terms, e.g. eigenvectors scaled by λk\sqrt{\lambda_k}. Every vector of HA⊗HB\Hil_A\otimes\Hil_B is (I⊗K)∣Ω⟩(\id\otimes K)\ket\Omega for a unique operator KK ("vectorization").

  3. Φ(∣i⟩ ⁣⟨j∣)=(⟨i∣⊗I) J (∣j⟩⊗I)=∑kKk∣i⟩ ⁣⟨j∣Kk†\Phi\big(\ketbra ij\big)=\big(\bra i\otimes\id\big)\,J\,\big(\ket j\otimes\id\big)=\sum_kK_k\ketbra ijK_k^\dagger

    Read off the (i,j)(i,j) block of JJ in both expressions; linearity extends it to every XX.

  4. V=∑kKk⊗∣k⟩E⟹Tr⁡E(VXV†)=∑k,l⟨l∣k⟩ KkXKl†=∑kKkXKk†V=\sum_kK_k\otimes\ket k_E\quad\Longrightarrow\quad \Tr_E\big(VXV^\dagger\big)=\sum_{k,l}\braket{l|k}\,K_kXK_l^\dagger=\sum_kK_kXK_k^\dagger

    (3 ⇒ 4). An environment of dimension rr suffices.

  5. (idR⊗Φ)(Y)=Tr⁡E[(IR⊗V) Y (IR⊗V)†] ≥0for Y≥0(\mathrm{id}_R\otimes\Phi)(Y)=\Tr_E\big[(\id_R\otimes V)\,Y\,(\id_R\otimes V)^\dagger\big]\ \ge0\quad\text{for } Y\ge0

    (4 ⇒ 1). Conjugation by an operator preserves positivity, and so does the partial trace, for every reference RR.

  6. Tr⁡Φ(X)=Tr⁡[(∑kKk†Kk)X] ∀X  ⟺  ∑kKk†Kk=I  ⟺  V†V=I,Tr⁡BJ=(∑kKk†Kk)T\Tr\Phi(X)=\Tr\Big[\Big(\textstyle\sum_kK_k^\dagger K_k\Big)X\Big]\ \forall X\iff\sum_kK_k^\dagger K_k=\id\iff V^\dagger V=\id,\qquad \Tr_BJ=\Big(\textstyle\sum_kK_k^\dagger K_k\Big)^{\mathsf T}

    Trace preservation, by cyclicity of the trace. ■\blacksquare

Check your understanding

The phase-flip channel with p=12p=\tfrac12 has Kraus operators {I/2, Z/2}\{\id/\sqrt2,\ Z/\sqrt2\}. Which other family describes the same channel?

  1. {I/2, X/2}\{\id/\sqrt2,\ X/\sqrt2\}
  2. {∣0⟩ ⁣⟨0∣, ∣0⟩ ⁣⟨1∣}\{\dyad0,\ \ketbra01\}
  3. {(I+X)/2, (I−X)/2}\{(\id+X)/2,\ (\id-X)/2\}
  4. {∣0⟩ ⁣⟨0∣, ∣1⟩ ⁣⟨1∣}\{\dyad0,\ \dyad1\}
Show the answer

D. {∣0⟩ ⁣⟨0∣, ∣1⟩ ⁣⟨1∣}\{\dyad0,\ \dyad1\}

Apply the Hadamard matrix u=12(111−1)u=\tfrac{1}{\sqrt2}\begin{pmatrix}1&1\\1&-1\end{pmatrix}: 12(I2±Z2)=I±Z2\tfrac{1}{\sqrt2}\big(\tfrac{\id}{\sqrt2}\pm\tfrac{Z}{\sqrt2}\big)=\tfrac{\id\pm Z}{2}, the projectors onto ∣0⟩\ket0 and ∣1⟩\ket1. "Apply ZZ with probability ½" and "measure in ZZ and forget the result" are the same channel: complete dephasing.

Choi–Jamiołkowski: channels as states

The map Φ↦J(Φ)\Phi\mapsto J(\Phi) is a linear bijection between maps L(HA)→L(HB)\mathcal L(\Hil_A)\to\mathcal L(\Hil_B) and operators on HA⊗HB\Hil_A\otimes\Hil_B, with inverse

Φ(X)=Tr⁡A[(XT⊗IB) J(Φ)].\Phi(X)=\Tr_A\big[(X^{\mathsf T}\otimes\id_B)\,J(\Phi)\big].

Under it, channels correspond exactly to {J≥0, Tr⁡BJ=IA}\{J\ge0,\ \Tr_BJ=\id_A\}, and τΦ=J(Φ)/dA\tau_\Phi=J(\Phi)/d_A is a genuine state: the output when Φ\Phi acts on half of ∣Φ+⟩\ket{\Phi^+}. Optimizing over channels therefore becomes a semidefinite program, and a channel can be measured by preparing ∣Φ+⟩\ket{\Phi^+}, sending one half through it, and doing state tomography.

DerivationExample: amplitude damping from its Choi matrix
  1. J=∣0⟩ ⁣⟨0∣⊗∣0⟩ ⁣⟨0∣+1−γ (∣0⟩ ⁣⟨1∣⊗∣0⟩ ⁣⟨1∣+∣1⟩ ⁣⟨0∣⊗∣1⟩ ⁣⟨0∣)+∣1⟩ ⁣⟨1∣⊗(γ∣0⟩ ⁣⟨0∣+(1−γ)∣1⟩ ⁣⟨1∣)J=\dyad0\otimes\dyad0+\sqrt{1-\gamma}\,\big(\ketbra01\otimes\ketbra01+\ketbra10\otimes\ketbra10\big)+\dyad1\otimes\big(\gamma\dyad0+(1-\gamma)\dyad1\big)

    Insert Φ(∣i⟩ ⁣⟨j∣)\Phi(\ketbra ij) for the four basis operators.

  2. J=(1001−γ000000γ01−γ001−γ)in the basis ∣00⟩,∣01⟩,∣10⟩,∣11⟩J=\begin{pmatrix}1&0&0&\sqrt{1-\gamma}\\0&0&0&0\\0&0&\gamma&0\\\sqrt{1-\gamma}&0&0&1-\gamma\end{pmatrix}\quad\text{in the basis }\ket{00},\ket{01},\ket{10},\ket{11}

    Tr⁡BJ=I\Tr_BJ=\id, as trace preservation requires.

  3. J=∣v0⟩ ⁣⟨v0∣+∣v1⟩ ⁣⟨v1∣,∣v0⟩=∣00⟩+1−γ ∣11⟩,∣v1⟩=γ ∣10⟩J=\dyad{v_0}+\dyad{v_1},\qquad \ket{v_0}=\ket{00}+\sqrt{1-\gamma}\,\ket{11},\quad \ket{v_1}=\sqrt\gamma\,\ket{10}

    The {∣00⟩,∣11⟩}\{\ket{00},\ket{11}\} block (11−γ1−γ1−γ)\begin{pmatrix}1&\sqrt{1-\gamma}\\\sqrt{1-\gamma}&1-\gamma\end{pmatrix} has determinant 0, so rank⁡J=2\rank J=2: two Kraus operators are necessary and sufficient.

  4. K0=∣0⟩ ⁣⟨0∣+1−γ ∣1⟩ ⁣⟨1∣,K1=γ ∣0⟩ ⁣⟨1∣K_0=\dyad0+\sqrt{1-\gamma}\,\dyad1,\qquad K_1=\sqrt\gamma\,\ketbra01

    Un-vectorize with K∣i⟩=(⟨i∣⊗I)∣v⟩K\ket i=(\bra i\otimes\id)\ket v: the familiar Kraus operators.

  5. V∣0⟩=∣0⟩B∣0⟩E,V∣1⟩=1−γ ∣1⟩B∣0⟩E+γ ∣0⟩B∣1⟩EV\ket0=\ket0_B\ket0_E,\qquad V\ket1=\sqrt{1-\gamma}\,\ket1_B\ket0_E+\sqrt\gamma\,\ket0_B\ket1_E

    The Stinespring isometry is a beam splitter that leaks the excitation into the environment. Tracing out BB instead gives the complementary channel, which is amplitude damping with parameter 1−γ1-\gamma; this symmetry decides its quantum capacity.

PropertyKraus formChoi matrix
Trace-preserving∑kKk†Kk=I\sum_kK_k^\dagger K_k=\idTr⁡BJ=IA\Tr_BJ=\id_A
Unital, Φ(I)=I\Phi(\id)=\id∑kKkKk†=I\sum_kK_kK_k^\dagger=\idTr⁡AJ=IB\Tr_AJ=\id_B
Unitary channela single unitary K1=UK_1=Urank one, J=∣v⟩ ⁣⟨v∣J=\dyad{v} with ∣v⟩=(I⊗U)∣Ω⟩\ket v=(\id\otimes U)\ket\Omega
Entanglement-breakingcan be chosen rank one, Kk=∣bk⟩ ⁣⟨ak∣K_k=\ketbra{b_k}{a_k}JJ separable (Horodecki, Shor & Ruskai, 2003)
Qubit Pauli channel, probabilities pPp_PpP P\sqrt{p_P}\,PBell-diagonal, eigenvalues 2pP2p_P
Predict before you look

What is the largest number of Kraus operators that could ever be necessary for a channel from a qutrit (dA=3d_A=3) to a qubit (dB=2d_B=2)?

Show the answer

6 operators

The minimal number is rank⁡J(Φ)\rank J(\Phi), and JJ acts on a space of dimension dAdB=6d_Ad_B=6. The bound is attained, e.g., by the channel that discards the input and outputs I/2\id/2: its Choi matrix I3⊗I2/2\id_3\otimes\id_2/2 has full rank 6. So a qubit-to-qubit channel never needs more than a two-qubit environment.

Distances, composition and the frontier

Channels compose (Ψ∘Φ\Psi\circ\Phi), tensor (Φ⊗Ψ\Phi\otimes\Psi) and convexly combine, and the Choi picture tracks each operation. Two further constructions recur throughout quantum Shannon theory. The complementary channel Φc(X)=Tr⁡B VXV†\Phi^c(X)=\Tr_B\,VXV^\dagger describes what leaks to the environment; it is unique up to an isometry on EE. A channel is degradable if Φc=D∘Φ\Phi^c=\mathcal D\circ\Phi for some channel D\mathcal D (Bob can simulate Eve), and antidegradable if Φ=A∘Φc\Phi=\mathcal A\circ\Phi^c (Eve can simulate Bob). Amplitude damping is degradable for γ≤12\gamma\le\tfrac12 and antidegradable for γ≥12\gamma\ge\tfrac12.

Key takeaways

  • Physical maps must be completely positive: they must stay positive when acting on part of an entangled system. The transpose is positive but not CP, since (id⊗T)∣Φ+⟩ ⁣⟨Φ+∣=F/2(\mathrm{id}\otimes T)\dyad{\Phi^+}=F/2 has eigenvalue −12-\tfrac12.
  • CP   ⟺  J(Φ)≥0  ⟺  \iff J(\Phi)\ge0\iff Kraus form   ⟺  \iff Stinespring form; TP   ⟺  ∑kKk†Kk=I  ⟺  V†V=I  ⟺  Tr⁡BJ=IA\iff\sum_kK_k^\dagger K_k=\id\iff V^\dagger V=\id\iff\Tr_BJ=\id_A. The minimal number of Kraus operators is rank⁡J≤dAdB\rank J\le d_Ad_B.
  • Kraus families are unique up to isometric mixing, and Stinespring dilations up to an isometry on the environment.
  • Choi–Jamiołkowski turns channels into (normalized) states, making channel optimization a semidefinite program; complementary channels and degradability follow directly from the Stinespring picture.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

What is the minimal number of Kraus operators of the completely depolarizing qubit channel ρ↦(Tr⁡ρ) I/2\rho\mapsto(\Tr\rho)\,\id/2?

  1. 1
  2. 2
  3. 4
  4. 3
Show the answer

C. 4

J=I2⊗I2/2J=\id_2\otimes\id_2/2 has rank 4, so 4 operators are necessary, e.g. 12{I,X,Y,Z}\tfrac12\{\id,X,Y,Z\} or the four 12∣i⟩ ⁣⟨j∣\tfrac{1}{\sqrt2}\ketbra ij.

Question 2

Given the Choi matrix J=∑ij∣i⟩ ⁣⟨j∣⊗Φ(∣i⟩ ⁣⟨j∣)J=\sum_{ij}\ketbra ij\otimes\Phi(\ketbra ij), how do you recover Φ(X)\Phi(X)?

  1. Φ(X)=Tr⁡A[(XT⊗I) J]\Phi(X)=\Tr_A\big[(X^{\mathsf T}\otimes\id)\,J\big]
  2. Φ(X)=Tr⁡B[(I⊗X) J]\Phi(X)=\Tr_B\big[(\id\otimes X)\,J\big]
  3. Φ(X)=J (X⊗I) J†\Phi(X)=J\,(X\otimes\id)\,J^\dagger
  4. Φ(X)=Tr⁡A[(X⊗I) J]\Phi(X)=\Tr_A\big[(X\otimes\id)\,J\big]
Show the answer

A. Φ(X)=Tr⁡A[(XT⊗I) J]\Phi(X)=\Tr_A\big[(X^{\mathsf T}\otimes\id)\,J\big]

Tr⁡A[(XT⊗I)J]=∑ij⟨j∣XT∣i⟩ Φ(∣i⟩ ⁣⟨j∣)=∑ijXijΦ(∣i⟩ ⁣⟨j∣)=Φ(X)\Tr_A[(X^{\mathsf T}\otimes\id)J]=\sum_{ij}\bra jX^{\mathsf T}\ket i\,\Phi(\ketbra ij)=\sum_{ij}X_{ij}\Phi(\ketbra ij)=\Phi(X).

Question 3

A proposed qubit "noise model" has a trace-preserving Choi matrix with eigenvalues {1.2, 0.9, 0, −0.1}\{1.2,\ 0.9,\ 0,\ -0.1\}. What follows?

  1. It is a valid channel with three Kraus operators
  2. It is not completely positive, so no physical process implements it, although it may still map every single-qubit state to a state
  3. It is not positive on any input
  4. It is valid once renormalized so that the eigenvalues sum to 2
Show the answer

B. It is not completely positive, so no physical process implements it, although it may still map every single-qubit state to a state

Φ\Phi is CP iff J(Φ)≥0J(\Phi)\ge0. A negative eigenvalue rules out a Kraus representation, so acting on half of an entangled pair would produce a non-state. Positivity on single-qubit inputs is a weaker property that such a map can still have (as the transpose does).

Question 4

A qubit channel has Kraus rank 3. What is the smallest environment for a Stinespring dilation V:HA→HB⊗HEV:\Hil_A\to\Hil_B\otimes\Hil_E?

  1. A single qubit, dim⁡HE=2\dim\Hil_E=2, always suffices
  2. dim⁡HE=4\dim\Hil_E=4 is the minimum for every qubit channel
  3. No finite environment works
  4. dim⁡HE=3\dim\Hil_E=3, the Kraus rank
Show the answer

D. dim⁡HE=3\dim\Hil_E=3, the Kraus rank

V=∑kKk⊗∣k⟩V=\sum_kK_k\otimes\ket k needs one environment dimension per Kraus operator, and the minimal number of Kraus operators is the Choi rank. Fewer dimensions would give a Kraus representation with fewer operators, contradicting minimality.

Further reading

  • W. F. Stinespring, "Positive functions on C-algebras", Proceedings of the American Mathematical Society* 6, 211 (1955).
  • A. Jamiołkowski, "Linear transformations which preserve trace and positive semidefiniteness of operators", Reports on Mathematical Physics 3, 275 (1972).
  • M.-D. Choi, "Completely positive linear maps on complex matrices", Linear Algebra and its Applications 10, 285 (1975).
  • K. Kraus, States, Effects, and Operations: Fundamental Notions of Quantum Theory, Lecture Notes in Physics 190 (Springer, 1983).
  • J. Watrous, The Theory of Quantum Information, Chapter 2 (Cambridge, 2018).