A noisy device can disturb amplitudes, lose energy, or entangle a qubit with an inaccessible environment. Describing only output ensembles for isolated inputs can hide what happens to correlations with another system. A quantum channel acts on density operators and must remain valid when the input is entangled with a reference that the device does not touch.
From an environment to a channel
Let and be finite-dimensional systems. A density operator is positive semidefinite with trace one.
Definition 1.5.1
(Quantum channel)
A linear map from operators on to operators on is trace preserving if for every operator . It is completely positive if, for every finite reference system , the extended map sends positive semidefinite operators on to positive semidefinite operators on . A quantum channel is a completely positive trace-preserving map.
A useful construction is the following.
Definition 1.5.2
(Kraus operators and Kraus maps)
Let be linear maps with
The are Kraus operators, and the map is the Kraus map they define.
Proposition 1.5.3
(Kraus maps are channels)
Every Kraus map is completely positive and trace preserving. Consequently it sends every density operator to a density operator, even on an entangled input with an untouched reference.
Proof
(Conjugation preserves positivity)
For a positive operator on and any vector ,
Thus for every reference , so the map is completely positive. Cyclicity of trace and the normalization identity give .
Proof X-ray
Why the untouched reference is included
The proof uses the two hypotheses in different places. Positivity of the output comes only from the form : each term is a conjugation of the joint operator , so every term is positive and so is their sum, whatever the dimension of . The normalization plays no role in positivity; it is used only in the trace computation. The key move is to test the output against an arbitrary vector of and push the Kraus operators onto as , which turns the output expectation into an expectation of . If is entangled across and , the environment interacts only with but the joint output must still be positive; this is the content of complete positivity, and ordinary positivity on alone does not supply it, as the section on positivity shows.
The Kraus representation is not unique: two lists can induce the same channel.
Proposition 1.5.4
(Isometry with an environment gives Kraus operators)
Let be an isometry, so , and let be an orthonormal basis of . Put . Then , and .
Proof
(Discarding the environment)
Writing the partial trace over in the basis gives . Also .
When , initializing an environment in and applying a unitary on gives the isometry . The standard finite-dimensional representation theorem says every completely positive trace-preserving map admits such a Kraus form; we cite that converse without proving it.
Common mistake
(The completeness relation is K-dagger K, not K K-dagger)
The normalization is . The sum need not equal the identity. For the amplitude-damping operators of the qubit-noise-models section, for , although the channel is trace preserving. Checking the wrong sum would wrongly reject a valid channel.
Two qubit noise models
A bit-flip channel with parameter is
Its Kraus operators are and . If , then
Worked example 1.5.5
(Bit flip on a basis state and on plus)
For we have and , so the output is : a classical mixture of and with weights and . In contrast, for we have , and the output has diagonal entries and off-diagonal entries , so it is unchanged. This agrees with : error probability depends on the state and on what is measured.
Amplitude damping models an excited qubit relaxing to its ground state, with idealized probability . Set
Here . Direct multiplication gives
Worked example 1.5.6
(Amplitude damping on one and on plus)
The state becomes , while is fixed. For input , the output is
Its determinant is , so it is mixed for .
Worked example 1.5.7
(Full strength: damping versus bit flip)
At the operators are and , and the formula gives for every input, since . This is energy relaxation, not a mere random bit flip: every state is sent to , and the map cannot be undone. The bit-flip channel with instead applies the reversible and sends to .
Positivity is not enough
Counterexample mode
Transposition fails on half of a Bell pair
In a fixed computational basis let . Transposition is linear, trace preserving and positive: if , then . But apply only to the first half of . The output is
The normalized antisymmetric vector has expectation in this operator. It is therefore not positive, and is not completely positive. It is not a quantum channel, although it maps every isolated density matrix to another density matrix.
Local noise and no signaling
Proposition 1.5.8
(Local channels leave the reference unchanged)
Let be any joint state, and let a trace-preserving channel act only on . The reference marginal is unchanged:
Proof
(Testing against operators on the reference)
Take Kraus operators and test both sides against any operator on . Cyclicity and the tensor-factor identity allow the sum over to replace by . Equality of all such traces proves equality of the reduced operators.
This is an unconditioned statement: if a particular outcome of an operation with recorded measurement outcomes is selected, the conditional reference state can change, and communicating that outcome can reveal correlations. Without that outcome, local trace-preserving processing cannot signal to .
Common mistake
(Selecting an outcome is not applying a channel)
Let Alice measure on her half of . If she records outcome , Bob's state is , which differs from his earlier state . Averaging over her outcomes, , so without the outcome nothing has changed for Bob. The marginal identity covers only the averaged, trace-preserving operation.
Practice with solutions
Quick check
(What does the bit-flip channel with do to ? To ?)
Apply to each state.
Solution · Answer
It gives ; is unchanged because .
Quick check
(For amplitude damping with , calculate the output on .)
Use .
Solution · Answer
The output is .
Quick check
(Why does positivity of transposition on isolated states not prove that transposition is a quantum channel?)
Which input states does complete positivity also test?
Solution · Answer
Positivity tests only the isolated system, whereas complete positivity also tests entangled inputs. Transposing half of a Bell pair yields expectation on , violating complete positivity.
Quick check
(Alice applies any trace-preserving channel to her half of a Bell pair without sending Bob an outcome. What is Bob's reduced state?)
Use the marginal identity.
Solution · Answer
It remains by the marginal identity. The joint state may nevertheless have changed.