Quantum Channels
Quantum Channels
Quantum Channels are generalizations of Quantum Operations. Since, in general our operations maybe noisy (inherently) and the system we are dealing with maybe open so Quantum Channels us the study of noisy open Quantum Operations.
ρundefinedUρ′\rho \xrightarrow{U}\rho'
The above only holds for closed quantum systems. Thus, in case of noisy open quantum systems we have the following.
ρundefinedEρ′\rho \xrightarrow{\mathcal E} \rho'
Here, E\mathcal E is a combination of:
unitaries (U)(U)
adding systems (ρ→ρ⊗σ)(\rho \to \rho \otimes \sigma)
subtracting systems or partial tracing (ρAB→trBρAB=ρA)(\rho_{AB} \to \text{tr}_B \rho_{AB} = \rho_A)
Review of Density Matrices
Moreover, to deal with noisy open systems we also need to review density matrices which are the general way of representing noisy quantum states.
ρ≡noisy quantum stateE≡noisy open quantum channel\rho \equiv \text{noisy quantum state}\\ \mathcal E \equiv \text{noisy open quantum channel}
Now since, ρ\rho is positive semi-definite thus, 1+∣∣a⃗∣∣2≥0  ⟹  ∣∣a⃗∣∣≤1\frac{1 + ||\vec{a}||}{2} \geq 0 \implies ||\vec{a}|| \leq 1.
This results in the Bloch Ball representation for general mixed state qubits, which serves as a generalization of the Bloch Sphere that serves as representation for pure state qubits. Pure state qubits have eigenvalues of 00 or 11, λ(ρ)=0 or 1\lambda(\rho) = 0\ \text{or}\ 1.
The eigenvalue of the maximally mixed state is I/2{I}/{2} where the state is actually given by the following.
12(∣0⟩⟨0∣+∣1⟩⟨1∣)≡12(∣+⟩⟨+∣+∣−⟩⟨−∣)\frac{1}{2}(|0 \rangle\langle 0| + |1 \rangle\langle 1|) \equiv \frac{1}{2}(|+ \rangle\langle +| + |- \rangle\langle -|)
Quantum Operations
However, in both these above cases, we are still dealing with closed (albeit noisy) for the random case.
Generalized Quantum Operations
We can refer to Generalized Quantum Operations as Quantum Channels (noisy and open) whose definitions we shall rediscover and formalize here.
Now, it should be noted that there are three major equivalent formalisms for Quantum Channels:
Operational (Steinspring)
Mathemtically Simplified (Kraus)
Axiomatic (TPCP maps)
Introduction to Steinspring Representation
For closed systems we have a probabilistic combinations of the above (noisy closed channels).
∣ψ⟩→U∣ψ⟩∣ψ⟩⟨ψ∣→U∣ψ⟩⟨ψ∣U†ρ→Uρ U†|\psi\rangle \to U|\psi\rangle\\ |\psi\rangle\langle\psi| \to U |\psi\rangle\langle\psi|U^\dagger\\ \rho \to U\rho\ U^\dagger
This can be generalized for noisy open channels as combinations of unitaries (U)(U), adding systems together (ρ→ρ⊗σ)(\rho \to \rho \otimes \sigma) and subtracting systems (ρAB→trB(ρAB)=ρA)(\rho_{AB} \to \text{tr}_B(\rho_{AB}) = \rho_A).
But why is partial tracing allowed? Doesn't this "delete" information?
The information has basically flown out of your lab to the heavens (but still exists in nature) and in most cases, you simply can't access it anymore, kinda like your ex-girlfriend.
tr(ρAB)=∑b(IA⊗⟨b∣) ρ (IA⊗∣b⟩)=(IA⊗trB)(ρAB)\text{tr}(\rho_{AB}) = \sum_b (I_A \otimes \langle b|)\ \rho\ (I_A \otimes |b\rangle) = (I_A \otimes \text{tr}_B)(\rho_{AB})
An operator V∈L(X), V:X→YV \in \mathcal{L}(X),\ V:{X \to Y} is called an isometry if and only if ∣∣Vv⃗∣∣=∣∣v⃗∣∣||V\vec{v}|| = ||\vec{v}|| for all v⃗∈X\vec{v} \in {X}. In other words, an isometry is a genralization of norm preserving operators.
Now, properties of linear isometries:
⟨v∣v⟩=⟨Vv∣Vv⟩=⟨v∣V†V∣v⟩\langle v|v\rangle = \langle Vv|Vv\rangle = \langle v|V^\dagger V|v\rangle
Thus, V†V=IXV^\dagger V = I_{{X}}.
Linear isometries are not always unitary operators, though, as those require additionally that X=YX = Y and VV†=IXV V^\dagger = I_X.
By the Mazur–Ulam theorem, any isometry of normed vector spaces over R\mathbb{R} is affine. Affine transformations are those that preserve lines and parallelism (but not necessarily distances and angles).
Steinspring Representation
Isometries allow us to restate the requirements of a noisy open quantum channel. Now, it can be formed by combinations of isometries and partial trace. Moreover, the purpose of the partial trace is to trash the environment shit out of my lab.
Thus, we have the following formalism.
V:A→B⊗EE:L(A)→L(B)V:A\to B\otimes E\\ \mathcal E : \mathcal L(A) \to \mathcal L(B)
E(ρ)=trE(Vρ V†)\mathcal E(\rho) = \text{tr}_E(V\rho\ V^\dagger)
Partial tracing can be delayed (deferred tracing), an idea similar to deferred measurements. Also, from a philosophical standpoint (Church of the Larger Hilbert Space) you can defer it forever.
Now, the above circuit is equivalent to the below circuit. As a result, inductively, combinations of isometries and partial trace is equivalent to the E\mathcal E operator.
Kraus Operator Picture
We can fix a orthonormal bases {∣e⟩}\{|e\rangle\} for EE, and have {Ve}\{V_e\} such that
V=∑eVe⊗∣e⟩V = \sum_e V_e \otimes |e\rangle
where Ve∈L(A),Ve:A→BV_e \in \mathcal L(A), V_e: A \to B and furthermore
I=V†V=(∑e1Ve1†⊗⟨e1∣)(∑e2Ve2⊗∣e2⟩)I = V^\dagger V = (\sum_{e_1} V^\dagger_{e_1} \otimes \langle e_1|)(\sum_{e_2} V_{e_2} \otimes |e_2\rangle)
such that Ve≠Ve†V_e \neq V_e^\daggerin general and thus, we have the Kraus operator condition (stated below).
I=∑eVe†VeI = \sum_e V_e^\dagger V_e
Thus, we can have a equivalent and simplified formalism of the Steinspring Operator (E)(\mathcal E), where {Ve}\{V_e\} are Kraus operators.
E(ρ)=trE(Vρ V†)=trE(∑e1,e2Ve1ρ Ve2†⊗∣e1⟩⟨e2∣)  ⟹  E(ρ)=∑eVeρ Ve†\mathcal E(\rho) = \text{tr}_E (V \rho\ V^\dagger) = \text{tr}_E (\sum_{e_1, e_2} V_{e_1} \rho\ V_{e_2}^\dagger \otimes |e_1\rangle\langle e_2|)\\ \implies \mathcal E(\rho) = \sum_e V_e\rho\ V_e^\dagger
Conversely, given {Ve}\{V_e\} satisfying Kraus operator conditions, E(ρ)=∑eVeρ Ve†\mathcal E(\rho) = \sum_e {V_e \rho\ V_e^\dagger} is a quantum operator with V=∑eVe⊗∣e⟩V = \sum_e V_e \otimes |e\rangle as the isometry involved in the transformation.
Examples of Kraus Operators for Quantum Channels
E(ρ)=Uρ U†→\mathcal E(\rho) = U \rho\ U^\dagger \rightarrow single Kraus operator {U}\{U\}
E(ρ)=∑epeUe ρ Ue†→\mathcal E(\rho) = \sum_e p_e U_e\ \rho\ U_e^\dagger \rightarrow Kraus operators {peUe}\{\sqrt{p_e} U_e\}
tr(ρAB)=∑b(IA⊗⟨b∣) ρ (IA⊗∣b⟩)\text{tr}(\rho_{AB}) = \sum_b (I_A \otimes \langle b|)\ \rho\ (I_A \otimes |b\rangle) where Vb=(IA⊗⟨b∣B)V_b = (I_A \otimes \langle b |_B) thus we also have the following holding true, ∑bVb†Vb=∑bIA⊗∣b⟩⟨b∣=IA⊗IB=I\sum_b V_b^\dagger V_b = \sum_b I_A \otimes |b\rangle\langle b| = I_A \otimes I_B = I.
Ultimate Refrigerator: E(ρ)=∣0⟩⟨0∣\mathcal E (\rho) = |0\rangle\langle 0|, regardless of the initial state ρ\rho.
V0=∣0⟩⟨0∣,V1=∣0⟩⟨1∣V0ρ V0†+V1ρ V1†=∣0⟩⟨0∣V_0 = |0\rangle\langle 0|, V_1 = |0\rangle\langle 1|\\ V_0 \rho\ V_0^\dagger + V_1 \rho\ V_1^\dagger = |0\rangle\langle 0|
Depolarizing Channel: E(ρ)=(1−p)ρ+ρ(I/2)\mathcal E (\rho) = (1 - p)\rho + \rho ({I}/{2}) →\rightarrow simple kind of white noise effect
Amplitude Damping Channel: Eγ(ρ)=K0ρ K0†+K1ρ K1†\mathcal E_\gamma (\rho) = K_0 \rho\ K_0^\dagger + K_1 \rho\ K_1^\dagger where we have the following
K0=[1001−γ] and,  K1=[0γ00]K_0 = \begin{bmatrix} 1 & 0 \\ 0 & \sqrt{1 - \gamma} \\ \end{bmatrix} \text{ and, }\ K_1 = \begin{bmatrix} 0 & \sqrt{\gamma} \\ 0 & 0 \\ \end{bmatrix}
Axiomatic Definitions
By God, properties of E\mathcal E should be as follows:
Linear (assume this or die, basically) and Hermitian preserving
Trace preserving (TP) such that tr(E(ρ))=tr(ρ)=1\text{tr}(\mathcal E(\rho)) = \text{tr}(\rho) = 1
Completely positive (CP) since ρ≥0  ⟹  (E⊗I)(ρ)≥0\rho \geq 0 \implies (\mathcal E \otimes I)(\rho) \geq 0
But why completely positive? Well, because we are dealing with isometries of form A→B⊗EA \to B \otimes E.
Here, completely positive means that if E\mathcal E is applied on a subsystem, then the complete system and the subsystems must remain positive.
Equivalence of TPCP maps and Kraus operators
Equivalence Theorem: If E(ρ)=∑eVeρ Ve†\mathcal E(\rho) = \sum_e V_e\rho\ V_e^\dagger then E\mathcal E is TPCP and converse holds true as well.
This implies that the map N→J(N) N→J(N) is injective. In fact it is known as the Choi-Jamiołkowski isomorphism — a correspondence between quantum channels and quantum states (described by density matrices).
Equivalence of Operators: Given 22 operators E\mathcal E and F\mathcal F with operator elements {Ei}\{\mathcal E_i\} and {Fi}\{\mathcal F_i\} respectively, if E=F\mathcal E = \mathcal F then
Ei=∑juijFj\mathcal E_i = \sum_j u_{ij} \mathcal F_j
Measurements as Quantum Operations
Measurement can be thought of as a quantum operation where the input is any quantum state and the output is classical, E(ρ)=∑xpx∣x⟩⟨x∣\mathcal E(\rho) = \sum_x p_x|x\rangle\langle x| (diagonal density matrix).
The probabilities, pxp_x, should depend on the state. Further, pxp_x should be a linear function of the density matrix px=tr(Mxρ)p_x = \text{tr}(M_x \rho). From this we can work out the properties that MxM_x should obey:
Normalization: tr(E(ρ))=tr(px)=tr(Mxρ)=1  ⟹  ∑Mx=I\text{tr}(\mathcal E(\rho)) = \text{tr}(p_x) = tr(M_x\rho) = 1 \implies \sum M_x = I
Positive Semi-definite: (E⊗I)(ρ)≥0  ⟹  px≥0(\mathcal E \otimes I)(\rho) \geq 0 \implies p_x \geq 0 thus tr(Mxρ)≥0,∀ Mx\text{tr}(M_x \rho) \geq 0, \forall\ M_x over ρ\rho thus, we have Mx≥0  ⟹  Mx′sM_x \geq 0 \implies M_x's are positive semi-definite   ⟹  Mx=ExEx†\implies M_x = E_x E_x^\dagger
Thus, we have the following:
∑xMx=IMx≥0\sum_x M_x = I\\ M_x \geq 0
These conditions still leave room for noisy measurements, etc.
We can also talk about non-demolition measurements, which do not discard the quantum systems aftermeasurements. Consider the following quantum channel.
E(ρ)=∑eVeρVe†⊗∣e⟩⟨e∣=∑eVeρVe†tr(VeρVe†)⊗tr(VeρVe†)∣e⟩⟨e∣ \mathcal E(\rho) =\sum_e V_e \rho V^\dagger_e \otimes |e\rangle\langle e| = \sum_e \frac{V_e \rho V^\dagger_e} {\text{tr}(V_e\rho V^†_e)} \otimes \text{tr}(V_e \rho V^\dagger_e)|e\rangle\langle e|
This channel has the following interpretation: with probability as
pe=tr(VeρVe†)=tr(VeVe†ρ)=tr(Meρ)p_e = \text{tr}(V_e \rho V_e^\dagger) = \text{tr}(V_e V_e^\dagger\rho) = \text{tr}(M_e\rho)
the state of the system after the application of this channel is ρe=(VeρVe†)/tr(VeρVe†)\rho_e = {(V_e \rho V^\dagger_e)}/{\text{tr}(V_e\rho V^†_e)}. This is similar to having a measurement that outputs the post measured state ρe\rho_e with probability pep_e.
Quantum Norms and Distance Metrics
Motivation: Now that we have defined channels and states of information, how do you differentiate between two items of information? What does it mean to say that information is preserved by some process?
Well for these questions it is necessary to develop distance measures. There is a certain arbitrariness in the way distance measures are defined, both classically and quantum mechanically, and the community of people studying quantum computation and quantum information has found it convenient to use a variety of distance measures over the years. Two of those measures, the trace distance and the fidelity, have particularly wide currency today.
Norm is a distance metric such that
∣cv∣=∣c∣.∣v∣|cv| = |c|.|v|
∣v+w∣≤∣v∣+∣w∣|v+w| \leq |v| + |w|
∣v∣=0|v| = 0 iff v=0v = 0
Examples of norms are as follows
Euclidean (L1L_1)
Manhattan (L2L_2)
LpL_p norm →∣v∣Lp=(∑vip)1/p\to |v|_{L_p} = (\sum v_i^p)^{1/p}
Schatten pp-norms
L1L_1 norm corresponds to probability distributions
L2L_2 norm corresponds to pure states
L∞=max⁡i∣vi∣L_\infty = \max_i {|v_i|}
Similarly, schatten pp-norm :: ∣M∣Sp=∣σ(M)∣Sp|M|_{S_p} = |\sigma(M)|_{S_p} where σ(M)\sigma(M) is a vector of singular values of a matrix, say MM.
S1=∑singular valuesS2=∑singular values2S∞=max⁡i∣singular values∣S_1 = \sum \text{singular values}\\ S_2 = \sqrt {\sum \text{singular values}^2}\\ S_\infty = \max_i |\text{singular values}|
In case of density matrices (in general to positive semi-definite Hermitian operators), we have singular values = eigen values\text{singular values = eigen values} thus, σ(M)=λ(M)\sigma(M) = \lambda(M).
∣X∣Sp=∣σ(X)∣Lp=(∑iσip)1/p|X|_{S_p} = |\sigma(X)|_{L_p} = (\sum_i \sigma_i^p)^{1/p}
If X≥0X \geq 0, then σ(X)=λ(X)\sigma(X) = \lambda(X) then, ∣X∣S1=trX  ⟹  ∣ρ∣S1=tr(ρ)|X|_{S_1} = tr X \implies |\rho|_{S_1} {= \text{tr} (\rho)}.
Thus, ρ\rho is a density matrix   ⟺  ρ≥0\iff \rho \geq 0 and ∣ρ∣S1=0|\rho|_{S_1} = 0.
Let us consider a simple system with {M,I−M}\{M, I-M\} as measurement operators.
Then, we have
M≥0,I−M≥0  ⟺  M≤I  ⟺  0≤M≤I  ⟺  ∣M∣S∞≤1M \geq 0, I - M \geq 0 \iff M \leq I\\ \iff 0 \leq M \leq I\\ \iff |M|_{S_\infty} \leq 1
But why is it that two objects with different norms can co-exist in the same operator framework? The answer lies in the notion of duality.
Given a norm ∣.∣|.| there exists a dual norm such that, ∣x∣∗=max⁡∣y∣≤1∣⟨x,y⟩∣|x|_* = \max_{|y| \leq 1} |\langle x,y\rangle|.
L2L_2 is dual to L2L_2
L1L_1 and L∞L_\infty  are dual to each other
S2S_2 is dual to S2S_2
S1S_1 and S∞S_\infty  are dual to each other
Introduction to Trace Norm
T(ρ,σ)=12∣ρ−σ∣S1=12 tr ∣ρ−σ∣T(\rho, \sigma) = \frac{1}{2}|\rho - \sigma|_{S_1} = \frac{1}{2}\ \text{tr}\ |\rho - \sigma|
Then, ∀M\forall M where MM denotes measurement,
∣tr(Mρ)−tr(Mσ)∣=∣tr(M(ρ−σ))∣≤∣M∣S∞∣ρ−σ∣S1≤∣ρ−σ∣S1=2T(ρ,σ)|\text{tr}(M\rho) - \text{tr}(M\sigma)| = |\text{tr}(M (\rho - \sigma))| \leq |M|_{S_\infty} |\rho -\sigma|_{S_1} \leq |\rho -\sigma|_{S_1} = 2T(\rho, \sigma)
In fact, we can tighten this inequality even further to obtain
∣tr(Mρ)−tr(Mσ)∣≤T(ρ,σ)↓T(ρ,σ)=max⁡M∣tr M(ρ−σ)∣↓T(E(ρ)−E(σ))≤max⁡M∣tr M(ρ−σ∣=T(ρ,σ)|\text{tr}(M\rho) - \text{tr}(M\sigma)| \leq T(\rho, \sigma)\\ \Big\downarrow\\ T(\rho, \sigma) = \max_M |\text{tr}\ M(\rho-\sigma)|\\ \Big\downarrow\\ T(\mathcal E(\rho) - \mathcal E(\sigma)) \leq \max_M |\text{tr}\ M(\rho-\sigma| = T(\rho, \sigma)
and finally we get the required inequality
T(E(ρ),E(σ))≤T(ρ,σ)T(\mathcal E(\rho), \mathcal E(\sigma)) \leq T(\rho, \sigma)
Thus, I can never increase the distance of two states, no matter what. So, basically when you wanna protect against noise all we are doing is slowing down the rate of noise (indistinguishable nonsense).