Lecture Notes on the Theory of Open Quantum Systems

Daniel A. Lidar
Departments of Electrical Engineering, Chemistry, and Physics & Astronomy
Center for Quantum Information Science & Technology
University of Southern California, Los Angeles, California 90089, USA

This is a self-contained set of lecture notes covering various aspects of the theory of open quantum system, at
a level appropriate for a one-semester graduate course. The main emphasis is on completely positive maps and
master equations, both Markovian and non-Markovian.

CONTENTS
arXiv:1902.00967v2 [quant-ph] 21 Feb 2020

I. Preface and Acknowledgments 5

II. Review of Quantum Mechanics 6

A. Postulate 1 6
B. Postulate 2 7
C. Postulate 3 7
D. Postulate 4 9
1. Projective (von Neumann) measurements 10
2. Examples of measuring observables 10
3. Expectation value of an observable 11
4. Heisenberg Uncertainty Principle 11
5. Positive Operator Valued Measures (POVMs) 12

III. Density Operators 13

A. Properties of the density operator 14
B. Dynamics of the density operator 15
C. Restatement of the postulates of quantum mechanics 15
D. More on pure and mixed quantum states 16
E. Unitary equivalence in ensembles 17
F. Visualizing the density matrix of a qubit: the Bloch sphere 17

IV. Composite Systems 19

A. Combining a system and a bath 19
B. Partial Trace 20
1. Definition 20
2. State of a quantum subsystem 21
3. Formal justification of using the partial trace to define a subsystem state 22

V. Open System Dynamics 23

A. Kraus Operator Representation 23
B. Normalization and the special case of a single Kraus operator 23
C. The Schrödinger equation as a special case 24

VI. Complete Positivity and Quantum Maps 24

A. Non-selective measurements 24
B. The OSR as a map 25
C. Complete Positivity 25
D. Positive but not Completely Positive: Transpose 26
E. Partial Transpose as a Test for Separability/Entanglement: the PPT criterion 27
F. Kraus OSR as a composition of CP maps 27
1. The assignment map is CP 28
2. The unitary map is CP 28
3. The partial trace is CP 28
G. OSR for a general initial condition? 28
 2

1. General initial states 28

2. Separable states 29
H. The quantum discord perspective 29
1. Quantum Discord 29
2. Zero discord initial states and CP maps 30
I. Equivalence of Quantum Maps 31
1. General conditions for equivalence 31
2. Physical origin of the equivalence 32

VII. Quantum Maps of a Qubit 33

A. Transformation of the Bloch Vector 33
B. Unital Quantum Maps 34
C. The Phase Damping Map 35
D. The Bit Flip Map 35
E. The Depolarizing Map 37
F. Amplitude Damping / Spontaneous Emission 37
G. Generalized (finite temperature) Amplitude Damping/Spontaneous Emission 39

VIII. Quantum Maps from First Principles 40

A. A qubit coupled to a single-qubit bath 40
1. Z ⊗ X coupling 41
2. Z ⊗ Z coupling 41
B. Irreversible open system dynamics: infinite dimensional bath with a continuous density of states 43

IX. Derivation of the Lindblad equation from a short time expansion of quantum maps 45
A. Derivation 45
B. The Markovian evolution operator as a one-parameter semigroup 47
C. Proof that the solution of the Lindblad Equation is a CP map 47
D. Examples 48
1. Just H for a single qubit: the Bloch equations 48
2. Phase Damping for a single qubit 49
3. Amplitude damping / Spontaneous Emission for a single qubit 49

X. The Lindblad equation via coarse graining 50

A. Derivation 50
B. Interaction picture 53
C. Example: the spin-boson model for phase damping 54

XI. Analytical solution of the spin-boson model for phase damping 56

A. Calculation of the Evolution Operator 57
B. Calculation of the Density Matrix 58

XII. Quantum trajectories and unravelling the Lindblad equation 60

A. Method summary 61
B. Equivalent dynamics of the wavefunction 62
1. Naive form 62
2. Telegraph noise form 63
3. Stochastic Schrödinger equation approach 63
4. Comparison between the telegraph noise and Stochastic Schrödinger equation approaches 65
C. Weak measurements 65

XIII. Analytical solution of the general Lindblad equation 65

A. The coherence vector 65
B. Just the non-dissipative part 66
C. Full Lindblad equation for the coherence vector 67
D. Solution for diagonalizable and invertible G 67
E. Solution for general G 68
F. Phase Damping Example 69
 3

XIV. Derivation of the Lindblad equation from the cumulant expansion and coarse graining 70
A. Cumulant expansion 70
B. The second order cumulant 72
C. Why the first order cumulant can be made to vanish 72
D. Derivation of the Lindblad equation 73
E. Complete positivity 74
F. LE from the cumulant expansion and coarse-graining 74
G. Illustration using the spin-boson model for phase damping 75

XV. First-principles derivation of the Lindblad equation from the Born, Markov, and rotating wave approximations 76
A. Setting up 76
B. Born approximation 77
C. Markov approximation and Redfield equation 78
D. Going to the frequency domain 79
E. Rotating Wave Approximation 79
F. The missing steps 81
1. From Born-Markov [Eq. (524)] to the RWA-LE [Eq. (536)] 81
2. Proof of Eq. (539) 81
3. Transformation back to the Schrödinger picture 82
4. Proof that γ(ω) > 0 83

XVI. The Kubo-Martin-Schwinger (KMS) condition and the Gibbs state as a stationary state Lindblad equation 85
A. The KMS condition 85
B. The Gibbs state is a stationary state of the RWA-LE 86
C. Return to equilibrium, quantum detailed balance, and ergodicity under the RWA-LE 88

XVII. Pauli Master Equation 88

XVIII. Lindblad Equation in the Singular Coupling Limit (SCL) 90

A. Derivation 91
B. Examples contrasting the WCL and SCL 92
1. Phase damping when [HS , HSB ] = 0 92
2. Phase damping when [HS , HSB ] ≠ 0 92
C. Example: collective vs independent phase damping 96
1. The collective case 96
2. The independent case 96
3. Contrasting the dephasing rates in the collective and independent cases 97
D. Bounding the Markov approximation error 97
1. Bound on ∥∆1 ∥ 98
2. Bound on ∥∆2 ∥ 99
3. Putting the bounds together 99
E. The RWA-LE is the infinite coarse-graining time limit of the cumulant-LE 99
1. Quick summary 99
2. A useful lemma 100
3. The ω = ω ′ case 100
4. The ω ≠ ω ′ case 101

XIX. The Nakajima-Zwanzig Equation 102

A. Feshbach P-Q partitioning 102
B. Derivation 103
C. From the Nakajima-Zwanzig equation to the Born master equation 104
D. The O(α3 ) term of the Nakajima-Zwanzig master equation 104

XX. The Time Convolutionless (TCL) Master Equation 105

A. Derivation 105
1. Back-propagation 105
2. Solving for the relevant part 105
B. Perturbation theory 106
1. Matching powers of α 106
 4

2. The TCL-ME at second order yields the Redfield equation 107

C. Example: spin-boson model of a qubit in a cavity 108
1. Analytical solution in the 1-excitation subspace 108
D. Jaynes-Cummings model on resonance 111
1. Isolated cavity 112
2. Leaky cavity 112
3. Comparison to TCL-ME, Markov limit, and NZ-ME 113
4. Breakdown of the NZ-ME and TCL-ME expansions for strong coupling 114

XXI. Post Markovian Master Equation 115

A. Measurement interpretation of the Kraus OSR and the Lindblad equation 115
B. Interpolating between the two limits: derivation of the PMME 116
C. Solution of the PMME 118
D. The PMME as a map, and its relation to the TCL-ME 119
E. Complete Positivity of the PMME 119
F. Example of the PMME: phase damping Lindbladian with an exponential kernel 119
G. Experimental determination of the Lindbladian L and kernel k(t) 121

A. Linear algebra background and Dirac notation 122

1. Inner Product 122
2. Orthonormal Bases 122
3. Linear Operators 122
4. Outer Product 123
5. The Cauchy-Schwartz Inequality 124
6. Trace equalities 124
7. Positive operators 124
8. Pauli matrices 124

B. Unitarily invariant norms 125

C. Distance and Fidelity between quantum states 126

1. Total variation distance and quantum distance 126
2. Fidelity Measures 127
3. The distance and fidelity inequality 127
4. Uhlman’s Theorem 127
5. Fidelity for a pure state passing through a noise channel 128
6. Fidelity is invariant under a joint unitary transformation 128
7. Fidelity of Noise channels 129
8. Examples: fidelities of various noise channels 129
a. The pure-dephasing channel 129
b. The depolarizing channel 130

References 131
 5

I. PREFACE AND ACKNOWLEDGMENTS

The theory of open quantum systems is the backbone of nearly all modern research in quantum mechanics and its applications.
The reason is simple: the idealization of an isolated quantum system obeying perfectly unitary quantum dynamics is just that:
an idealization. In reality every system is open, meaning that it is coupled to an external environment. Sometimes these open
system effects are small, but they can almost never be neglected. This is particularly relevant in the field of quantum information
processing, where the existence of a quantum advantage over classical information processing is often derived first from the
idealized, closed system perspective, and must then be re-scrutinized in the realistic, open system setting.

These lecture notes provide a fairly comprehensive and self-contained introduction to the theory of open quantum systems.
They are based on lectures I gave at the University of Southern California as part of a one-semester graduate course on the
topic taught in Fall 2006, Spring 2013, Spring 2017, and Fall 2018. There are several excellent textbooks and monographs
either devoted to or containing the same subject, and these notes are in parts heavily influenced by these works, in particular
the invaluable books by Heinz-Peter Breuer and Francesco Petruccione [1] and by Robert Alicki and Karl Lendi [2]. The
notes do fill in many details not found in the original sources (at times tediously so!), and also draw on various articles and
unpublished materials. I therefore hope that these notes will serve as a useful companion to the textbooks, and will help students
and researchers interested in entering the field in a semester of dedicated study.

The notes were originally typeset by students serving as scribes during the lectures given in 2013 and 2017, and have under-
gone extensive editing and additions since then. I am extremely grateful to all the students who participated in this effort: Chao
Cao, Rajit Chatterjea, Yi-Hsiang Chen, Jan Florjanczyk, Jose Raul Gonzalez Alonso, Anastasia Gunina, Drew Henry, Kung-
Chuan Hsu, Zhihao Jiang, Joshua Job, Hannes Leipold, Milad Marvian, Anurag Mishra, Nicolas Moure Gomez, Siddharth
Muthu Krishnan, Shayne Sorenson, Georgios Styliaris, Christopher Sutherland, Subhasish Sutradhar, Walter Unglaub, Ka Wa
Yip, and Yicong Zheng. I am also very grateful to the students in the 2018 course, who offered numerous additional feedback:
Namit Anand, Mojgan Asadi, Brian Barch, Matthew Kowalsky, Lawrence Liu, Humberto Munoz Bauza, Adam Pearson, Bibek
Pokharel, Evangelos Vlachos, Aaron Wirthwein, Haimeng Zhang, and Zihan Zhao. Finally, I wish to warmly thank Dr. Tameem
Albash and Dr. Jenia Mozgunov, who filled in for me on various occasions, and whose notes I relied on as well.

Of course, all errors, typos, and omissions are mine. The reader is strongly encouraged to send me any corrections at li-
[email protected]. The notes will be updated regularly to reflect these corrections, as well as new material of interest. I apologize
in advance to all the numerous authors whose contributions I did not cite; the field is vast and the intent of these notes is not to
serve as a comprehensive review article. I have certainly not done justice to the literature.

The completion of this work was (partially) supported by the Office of the Director of National Intelligence (ODNI), Intelli-
gence Advanced Research Projects Activity (IARPA), via the U.S. Army Research Office contract W911NF-17-C-0050.
 6

II. REVIEW OF QUANTUM MECHANICS

The introductory material presented here is based on the approach of the excellent textbook by Nielsen & Chuang [3]. There
are four main postulates on which Quantum Mechanics can be built. These four postulates are:
1. Where things happen: Hilbert space
2. Combining Quantum Systems: Tensor product of vectors, matrices and of Hilbert spaces
3. Time Evolution (Dynamics): Schrödinger equation
4. Information extraction: Measurements

A. Postulate 1

“To every Quantum system is associated a state space, i.e, a Hilbert space H.”
A Hilbert space is a vector space equipped with an inner product. The vector spaces that we will be working with, H, can be
defined in the following way, in which C is the field of complex numbers:

H = Cd (1a)
⎛ v0 ⎞
⎜ v ⎟
v=⎜ 1
= {⃗ ⎟ ∣vi  C} (1b)
⎜ ⋮ ⎟
⎝ vd−1 ⎠
Thus for our purposes a vector space can be defined as the set of d-dimensional vectors v⃗, each element of which, vi , is a
complex number. Recall that a vector space has a couple of properties. First, for all vectors v⃗ ∈ H, a⃗ v + bv⃗′ ∈ H, with a, b ∈ C,
i.e., any linear combination of vectors z⃗ is also an element of the vector space V . Second, the vector space must contain the zero
vector, an element that satisfies the condition 0⃗ + v⃗ = v⃗ ∀⃗
v ∈ C.
The postulate means that physical states of a quantum system can be associated to a vector v⃗ ∈ H. We shall use Dirac notation,
in which column vectors are denoted by “kets”: v⃗ ↦ ∣v⟩. In what follows, we will usually assume that the dimension of H is
d−1
finite, and find an orthonormal basis for it. That is, if dim(H) = d, then denote a such a basis {∣k⟩}k=0 . A good (but obviously
non-unique) choice is the standard basis

⎛0⎞
⎜⋮⎟
⎜ ⎟
∣k⟩ = ⎜1⎟ ← k + 1th position (2)
⎜ ⎟
⎜⋮⎟
⎝0⎠

Any vector in the Hilbert space can be expanded in an orthonormal basis as a linear combination
d−1
∣v⟩ = ∑ vk ∣k⟩ , vk ∈ C, (3)
k=0

which quantum physicists often call a superposition. The coefficients vk are called probability amplitudes. The reason is that
2
the probability of a quantum system “being” in a specific state ∣k⟩ is ∣vk ∣ . This latter statement is part of the postulate. The
different orthonormal basis vector ∣k⟩ represent mutually exclusive possibilities, such as the discrete positions of a particle on a
line, or different spin configurations.
2
Of course, in order for the set {∣vk ∣ } to be a proper probability distribution, the probabilities must sum to one. This is the
reason that we need to endow the vector space H with an inner product, i.e., work with Hilbert spaces. To define the inner
product function we first introduce the dual of a ket, called a “bra”. In Dirac notation, row vectors (or bras) are written as ⟨v∣,
†
where by definition ⟨v∣ = ∣v⟩ , where the dagger denotes Hermitian conjugation, i.e., transpose and complex conjugation. Thus
if ∣v⟩ is written as in Eq. (1b) then ⟨v∣ = {v1∗ , v2∗ , . . . , vn∗ }. One reason that Dirac notation is convenient because we can represent
the inner product as a “braket”, i.e.,

⎛ w0 ⎞ d−1
⟨v∣w⟩ ≡ (v0∗ , ⋯, vd−1
∗
) ⎜ ⋮ ⎟ = ∑ vk∗ wk . (4)
⎝wd−1 ⎠ k=0
 7

The normalization condition of the probability distribution can now be written as

d−1 d−1
1 = ∑ ∣vk ∣2 = ∑ vk∗ vk = ⟨v∣v⟩ = ∥∣v⟩∥ ,
2
(5)
k=0 k=0
√
which is to say that every vector ∣v⟩ ∈ H is normalized, i.e., ∥∣v⟩∥ = ⟨v∣v⟩ = 1. Note that an overall phase does not affect
normalization, i.e., ∣v⟩ and eiθ ∣v⟩ have the same norm. In fact we do not distinguish between states that differ only by an overall
phase. Such states form a “ray” in Hilbert space.
Using Dirac notation we can form the outer product of two vectors in the same Hilbert space as follows
∗ ∗
⎛ v0 ⎞ ⎛ v0 w0 ⋯ v0 wd−1 ⎞
∗ ∗
∣v⟩ ⟨w∣ = ⎜ ⋮ ⎟ (w0 , ⋯, wd−1 ) = ⎜ ⋮ ⋱ ⋮ ⎟. (6)
⎝vd−1 ⎠ ⎝vd−1 w0∗ ⋯ vd−1 wd−1∗ ⎠

Additional linear algebra and Dirac notation facts are collected in Appendix A.

B. Postulate 2

“Given two quantum systems with respective Hilbert spaces H1 and H2 the combined quantum system has
associated with it a Hilbert space given by H = H1 ⊗ H2 .”
1 −1 2 −1
Let us define H1 to be the span of {∣vi ⟩}di=0 , and similarly H2 to be the span of {∣wj ⟩}dj=0 . Then we have H defined as the
1 −1,d2 −1
span of {∣vi ⟩ ⊗ ∣wj ⟩}di=0,j=0 . For two states ∣ψ⟩ ∈ H1 , ∣ϕ⟩ ∈ H2 , the tensor product is given by

⎛ ψ0 ϕ0 ⎞
⎛ ψ0 ⎞ ⎛ ϕ0 ⎞ ⎜ ⋮ ⎟
⎜ ⎟
∣ψ⟩ ⊗ ∣ϕ⟩ = ⎜ ⋮ ⎟ ⊗ ⎜ ⋮ ⎟ = ⎜ ψ0 ϕd2 −1 ⎟ (7)
⎜
⎝ψd1 −1 ⎠ ⎝ϕd2 −1 ⎠ ⎜ ⎟
⋮ ⎟
⎝ψd1 −1 ϕd2 −1 ⎠

Note that the underlying Hilbert spaces could represent entirely different physical systems, e.g., the first could be the space of
electron spins, whereas the second could be the space of photon polarizations. We can also define the tensor product between
matrices, i.e., if

⎛ a11 ⋯ a1n ⎞ ⎛b11 ⋯ b1q ⎞

A=⎜ ⋮ ⋱ ⋮ ⎟ B=⎜ ⋮ ⋱ ⋮ ⎟ (8)
⎝am1 ⋯ amn ⎠ ⎝bp1 ⋯ bpq ⎠

then

⎛ a11 b11 ⋯ a1n b1q ⎞

A⊗B =⎜ ⋮ ⋱ ⋮ ⎟, (9)
⎝am1 bp1 ⋯ amn bpq ⎠

i.e., a matrix of dimension mp × nq.

1
For example, let H1 = H2 = C2 and ∣Ψ⟩ = √1 ((∣0⟩ ⊗ ∣0⟩ ) + (∣1⟩ ⊗ ∣1⟩ )) = √1 (1, 0, 0, 1)t ∈ H = H1 ⊗ H2 , where ∣0⟩ = ( )
2 1 2 1 2 2 0
0
and ∣1⟩ = ( ). This example is interesting and important since it represents an entangled state, i.e., a state which cannot be
1
written as a tensor product in the same basis.

C. Postulate 3

“∃ a unitary operator U (t) such that the time evolution of a state is given by

∣ψ(t)⟩ = U (t) ∣ψ(0)⟩ . (10)

 8

Equivalently, the state vector of the system satisfies the Schrödinger equation
i
∣ψ̇(t)⟩ = − ̵ H ∣ψ(t)⟩ (11)
h
with H being a Hermitian operator known as the Hamiltonian.”
̵ = 1, which means that the units of energy and frequency will be the same.
The dot denotes ∂/∂t, and in this course we will set h
We shall show below that the equivalence holds provided U (t) = exp (−iHt) when H is t-independent. In the time dependent
case, the situation is more complicated, and we have

U (t) = T+ e−i ∫0
T
H(t′ )dt′
(12)

where T+ represents Dyson time ordering. This will be discussed later.

To prove the equivalence let us recall a bit of mathematical background. An operator A is normal if A† A = AA† , it is
Hermitian if A† = A, and it is unitary if A† A = I. Clearly, unitary operators and Hermitian operators are also normal.
Theorem 1 (Spectral Theorem). A linear operator A ∶ V → V obeys A† A = AA† (i.e., it is a normal operator) if and only
if A = ∑a λa ∣a⟩ ⟨a∣ for a set of orthonormal basis vectors {∣a⟩} for V , which are also the eigenvectors of A with respective
eigenvalues {λa }.
Using this we can characterize the eigenvalues of Hermitian and unitary operators:
1. Hermitian operators: Applying the spectral theorem we get ∑a λa ∣a⟩ ⟨a∣ = ∑a λ∗a ∣a⟩ ⟨a∣, so that λa = λ∗a . Thus the
eigenvalues are real in this case.
2. Unitary operators: Applying the spectral theorem we get

I = A† A = I = (∑ λ∗a ∣a⟩ ⟨a∣) (∑ λa′ ∣a′ ⟩ ⟨a′ ∣) = ∑ λ∗a λa ∣a⟩ ⟨a∣ = ∑ ∣λa ∣ ∣a⟩ ⟨a∣ = ∑ ∣a⟩ ⟨a∣ ,
2
(13)
a a′ a a a

where the last equality is the spectral representation of the identity operator I (all its eigenvalues are 1). Thus the eigen-
values of a unitary operator are all phases: λa = eiθa where θa ∈ R.
We now define functions of normal operators. If we have a function f ∶ C → C, then we can extend it to the case of normal
operators by defining

f (A) ≡ ∑ f (λa ) ∣a⟩ ⟨a∣ . (14)

a

Note that the function operates only on the eigenvalues.

Let us now prove the equivalence of the two evolution laws. One direction is straightforward, namely assuming Eq. (10) we
easily derive Eq. (11):
d d
∣ψ(t)⟩ = (e−iHt ∣ψ(0)⟩) = −iHe−iHt ∣ψ(0)⟩ = −iHU (t) ∣ψ(0)⟩ = −iH ∣ψ(t)⟩ (15)
dt dt
Note that bringing the term involving the Hamiltonian down from the exponent is justified even for operators, as is easily verified
using the spectral theorem (since H is normal), or directly by differentiating the Taylor expansion of the matrix exponential
(which applies even if A is not normal):

d At d A2 t2 A3 t3 A2 A3 2 A2 t2
(e ) = (I + At + + + ...) = A + (2t) + (3t ) + ... = A(I + At + + ...) = AeAt (16)
dt dt 2! 3! 2! 3! 2!
Now for the other direction, we start with writing the spectral decomposition of the Hamiltonian as H = ∑a λa ∣a⟩ ⟨a∣, and note
also that from the definition (14) we have:

U (t) = e−iHt ⇒ U (t) = ∑ e−iλa t ∣a⟩ ⟨a∣ (17)

a

Now, since the eigenvectors of H are a basis (again, from the spectral theorem), we can decompose ∣ψ(t)⟩ in this basis and write
∣ψ(t)⟩ = ∑a ψa (t) ∣a⟩, so that the left hand side of Eq. (11) becomes
d d
∣ψ(t)⟩ = ∑ ψa (t) ∣a⟩. (18)
dt a dt
 9

As for the right hand side of Eq. (11),

− iH ∣ψ(t)⟩ = −i ∑ λa ∣a⟩ ⟨a∣ ψ(t)⟩ = −i ∑ λa ψa′ (t) ∣a⟩ ⟨a∣a′ ⟩ = −i ∑ λa ψa (t) ∣a⟩. (19)
a aa′ ² a
δaa′

For Eq. (11) to hold, these two need to be equal term by term (from orthonormality of the basis), so that we find
d
ψa (t) = −iλa ψa (t) ⇒ ψa (t) = e−iλa t ψa (0) (20)
dt
Plugging this result into ∣ψ(t)⟩ = ∑a ψa (t) ∣a⟩ and using orthonormality once more we now have:

∣ψ(t)⟩ = ∑ e−iλa t ψa (0) ∣a⟩ = (∑ e−iλa t ∣a⟩ ⟨a∣)(∑ ψa′ (0) ∣a′ ⟩) = e−iHt ∣ψ(0)⟩ = U (t) ∣ψ(0)⟩. (21)
a a a′

This completes the proof.

D. Postulate 4

This is the most controversial postulate, but we will not discuss those issues here and simply assume its validity.
This postulate has two parts: measuring states and measuring operators.
N
1. Measuring States: Quantum measurements are described by a set {Mk }k=1 of measurement operators satisfying the
constraint ∑k Mk† Mk = I.
Given a state ∣ψ⟩ ∈ H, instantaneously after the measurement it becomes,
Mk ∣ψ⟩
∣ψ⟩ ↦ √ ≡ ∣ψk ⟩ , (22)
pk
with probability
pk = ⟨ψ∣Mk† Mk ∣ψ⟩ = ∥Mk ∣ψ⟩ ∥2 ≥ 0. (23)
The measurement outcome is the index k of the state that resulted. The constraint listed in the postulate has the following
origin. Notice that ∑k pk = 1 must be true since pk is a probability, which implies ⟨ψ∣ ∑k Mk† Mk ∣ψ⟩ = 1. Since this is
true for arbitrary ∣ψ⟩ the sum rule follows

∑ Mk Mk = I.
†
(24)
k

2. To every physically measurable quantity is associated an observable, i.e., a Hermitian operator A. A has a spectral
decomposition (since it is Hermitian and hence normal),
A = ∑ λa ∣a⟩ ⟨a∣ , (25)
a

with λa ∈ R since A is Hermitian. The λa ’s, the eigenvalues, are the outcomes of the measurement (hence need to be
real).1
The set of eigenvectors, {∣a⟩} are an orthonormal set. Hence, {Pa ≡ ∣a⟩ ⟨a∣} are projectors (defined below). These are the
measurement operators corresponding to the measurement of this observable. Hence, if the system is in state ∣ψ⟩ before the
observable A is measured, according to Eq. (23) the probability of outcome λa is given by pa = ⟨ψ∣Pa† Pa ∣ψ⟩ = ⟨ψ∣Pa2 ∣ψ⟩ =
⟨ψ∣Pa ∣ψ⟩ = ∣ ⟨ψ∣ a⟩∣2 . Moreover, according to Eq. (22) the state after the measurement is performed and outcome λa is
observed, becomes ∣ψa ⟩ = P√ap∣ψ⟩
a
⟨a∣ψ⟩
= ∣⟨a∣ψ⟩∣ ∣a⟩ = eiθ ∣a⟩, where eiθ is the phase associated with the complex number ⟨a∣ ψ⟩.
We next consider several important special cases of the generalized measurements defined above.

1 It is interesting to ask why physically measurable quantities should be associated with Hermitian operators. Intuitively, since physical measurements pro-
duce real numbers we want to associate an operator with a real spectrum as a physically observable quantity. Moreover, we would like states with dif-
ferent eigenvalues (or different results from the measurement) to be orthogonal. A Hermitian operator satisfies both of these requirements. However,
these justifications admittedly leave something to be desired. For more details see https://physics.stackexchange.com/questions/39602/
why-do-we-use-hermitian-operators-in-qm.
 10

1. Projective (von Neumann) measurements

Projective measurements are a special case of generalized measurements, in which the measurement operators, Mk are Her-
mitian operators called projectors. That is, Mk = Pk , where Pk Pl = δk,l Pk and Pk† = Pk . In particular, Pk2 = Pk . Using this, we
can see that the probability of outcome k, pk = ⟨ψ∣Mk† Mk ∣ψ⟩ = ⟨ψ∣Pk ∣ψ⟩.
1 0
Example: Let ∣ψ⟩ = a ∣0⟩ + b ∣1⟩ where ∣0⟩ = ( ) , ∣1⟩ = ( ) and a, b ∈ C. That is, ∣ψ⟩ ∈ C2 and {∣0⟩ , ∣1⟩} is the standard basis
0 1
for the space. Such a ∣ψ⟩ is called a qubit.
Now, we define measurement operators, M0 = P0 = ∣0⟩⟨0∣ and M1 = P1 = ∣1⟩⟨1∣. We can see that P0,1 2
= P0,1 and P0 P1 = 0.
Hence, this is a set of projective measurements. Thus, the probabilities of outcomes are,
2
p0 = ⟨ψ∣P0 ∣ψ⟩ = ⟨ψ∣0⟩ ⟨0∣ψ⟩ = ∣a∣ , (26a)
2
p1 = ⟨ψ∣P1 ∣ψ⟩ = ⟨ψ∣1⟩ ⟨1∣ψ⟩ = ∣b∣ . (26b)
This shows that the absolute value squared of the amplitudes of a wavefunction when expanding it in an orthonormal basis
provide the probabilities of observing the outcomes corresponding to those basis states. This is sometimes called the Born rule
in quantum mechanics.
Also, using Postulate 4, we can see that the state transformation in the above measurement would be:
⎧
⎪ P0 ∣ψ⟩ 2
with probability p0 = ∣a∣ ,
⎪
∣ψ⟩ ↦ ⎨ P1∣a∣∣ψ⟩ (27)
⎪ 2
⎩ ∣b∣ with probability p1 = ∣b∣ .
⎪

We can easily see that P0 ∣ψ⟩ = ∣0⟩⟨0∣(a ∣0⟩ + b ∣1⟩) = a ⟨0∣0⟩ ∣0⟩ + b ⟨0∣1⟩ ∣0⟩ = a ∣0⟩, where in the last step we have used that
{∣0⟩ , ∣1⟩} is an orthonormal set. Similarly, P1 ∣ψ⟩ = b ∣1⟩. Hence the transformation (27) becomes,
⎧
⎪ 2
⎪ ∣a∣ ∣0⟩ = e a ∣0⟩ with probability p0 = ∣a∣ ,
a iθ
∣ψ⟩ ↦ ⎨ b , (28)
⎪
⎪
2
∣1⟩ = eiθb ∣1⟩ with probability p1 = ∣b∣ .
⎩ ∣b∣

where θa and θb are the arguments of the complex numbers a and b respectively. We can see that the phase factors eiθa,b are
completely arbitrary since they have no influence on the probabilities of the measurement outcomes.
Thus, quantum states are equivalent up to a global phase factor. Because of this, quantum states are rays in a Hilbert space,
since they are not just one vector but an equivalence class of vectors: equivalent up to a global phase.

2. Examples of measuring observables

To illustrate the concept of observables, let’s consider a few examples.

• Pauli matrices: The Pauli matrices and their properties are reviewed in Appendix A 8. Consider, e.g., measuring the Pauli
matrix Z on a qubit ∣ψ⟩ = a ∣0⟩ + b ∣1⟩. Writing the spectral decomposition of Z,

Z = (+1)∣0⟩⟨0∣ + (−1)∣1⟩⟨1∣, (29)

we can see that the set of measurement operators for this observable is {P0 = ∣0⟩⟨0∣, P1 = ∣1⟩⟨1∣}, with outcomes as the
2
corresponding eigenvalues {λ0 = +1, λ1 = −1}. Thus, we obtain λ0 with probability p0 = ⟨ψ∣P0 ∣ψ⟩ = ∣a∣ , and obtain λ1
2
with probability p1 = ⟨ψ∣P0 ∣ψ⟩ = ∣a∣ .
Hence, the action of measuring Z takes ∣ψ⟩ to eiθa ∣0⟩ if the outcome was λ0 ; and to eiθb ∣1⟩ if the outcome was λb .
• Measuring Energy: When measuring energy, the observable we use is simply the Hamiltonian H of the system. Since H
is Hermitian it has a spectral decomposition, We can write H as,

H = ∑ Ea ∣a⟩ ⟨a∣ , (30)

a

where Ea denotes the energy and ∣a⟩ the associated energy eigenstate. So, in our experiment, we measure an energy of
Ea with probability pa = ⟨ψ∣Pa ∣ψ⟩, where Pa = ∣a⟩ ⟨a∣. The post-measurement state is ∣ψa ⟩ = P√ap∣ψ⟩
a
= ⟨a∣ψ⟩
√
pa
∣a⟩, i.e., the
new state is an eigenstate of the Hamiltonian, also sometimes called an energy eigenstate.
 11

Consider the following single-qubit Hamiltonian:

H = ωx σx + ωz σz . (31)

What happens when we measure it in the state ∣ψ⟩? To find √ out we need H’s spectral decomposition, i.e., we need to
diagonalize H. The eigenvalues are easily found to be E± = ± ωx2 + ωz2 , so that H can be written in diagonal form as

H = E− ∣E− ⟩⟨E− ∣ + E+ ∣E+ ⟩⟨E+ ∣, (32)

where ∣E± ⟩ are the corresponding eigenvectors. When H is measured, we find E± with probability p± = ⟨ψ∣E± ⟩⟨E± ∣ψ⟩ =
∣ ⟨ψ∣ E± ⟩∣2 . For example, if the system is prepared in the ground state ∣E− ⟩ (the state with the lower energy), then p− = 1
but p+ = 0. Or,√ if the system is prepared in a uniform superposition of the ground state ∣E− ⟩ and the excited state ∣E+ ⟩,
i.e., ∣ψ⟩ = 1/ 2(∣E− ⟩ + ∣E+ ⟩), then p− = p+ = 1/2.

3. Expectation value of an observable

Given an observable A = ∑a λa ∣a⟩ ⟨a∣, since we obtain λa with probability pa , we can naturally define an expectation value
of this observable in the state ∣ψ⟩ as

⟨A⟩ψ ≡ ∑ λa pa = ∑ λa ⟨ψ∣Pa ∣ψ⟩ (33a)

a a

= ⟨ψ∣ (∑ λa Pa ) ∣ψ⟩ (33b)

a
= ⟨ψ∣A∣ψ⟩ (33c)
= Tr(A ∣ψ⟩ ⟨ψ∣). (33d)

The last equality can be proved as follows: first create an orthonormal basis for the Hilbert space with ∣ψ⟩ as one of the elements
of the basis (say by using the Gram-Schmidt procedure [3][p.66]. That is,

H = Span{∣ψ⟩ = ∣φ0 ⟩ , ∣φ1 ⟩ , ∣φ2 ⟩ , ... ∣φd−1 ⟩}, (34)

where d is the dimension of the Hilbert space and all vectors in the basis are orthonormal. Now,
d−1
Tr(A ∣ψ⟩ ⟨ψ∣) = ∑ ⟨φi ∣A∣ψ⟩ ⟨ψ∣φi ⟩ (35a)
i=0
d−1
= ∑ ⟨φi ∣A∣φ0 ⟩ ⟨φ0 ∣φi ⟩ (35b)
i=0
= ⟨φ0 ∣A∣φ0 ⟩ (35c)
= ⟨ψ∣A∣ψ⟩ , (35d)

where in Eq. (35b) we used the fact that ∣ψ⟩ = ∣φ0 ⟩ is the first element in our basis.
Likewise we can define the standard deviation as follows:
√
2
∆A = ⟨(A − ⟨A⟩ψ ) ⟩ (36)
ψ

4. Heisenberg Uncertainty Principle

In quantum mechanics, an important property of pairs of non-commuting observables is that they cannot be measured with
arbitrary precision simultaneously. What this means is that if we measure the Hermitian operators C and D on ∣ψ⟩ then they
obey the Heisenberg Uncertainty Principle:

1
(∆C) (∆D) ≥ ∣⟨ψ∣[C, D]∣ψ⟩∣ . (37)
2
 12

Let us now show this. Define the Hermitian operators A = C −⟨C⟩, and B = D −⟨D⟩. We can always decompose the expectation
value as a complex number:

⟨ψ∣AB∣ψ⟩ = x + iy, x, y ∈ R. (38)

Note that 2AB = {A, B} + [A, B], where {A, B} is Hermitian (purely real eigenvalues) and [A, B] is anti-Hermitian (purely
imaginary eigenvalues). Therefore
∗
⟨ψ∣{A, B}∣ψ⟩ = ⟨ψ∣{A, B}† ∣ψ⟩ = ⟨ψ∣{A, B}∣ψ⟩ (39a)
∗
⟨ψ∣[A, B]∣ψ⟩ = ⟨ψ∣[A, B] ∣ψ⟩ = − ⟨ψ∣[A, B]∣ψ⟩ ,
†
(39b)

which means that ⟨ψ∣{A, B}∣ψ⟩ is real while ⟨ψ∣[A, B]∣ψ⟩ is purely imaginary. Hence the following must be true:

⟨ψ∣{A, B}∣ψ⟩ = 2x (40a)

⟨ψ∣[A, B]∣ψ⟩ = 2iy. (40b)

Therefore, by using the Cauchy-Schwarz inequality (see Appendix A) in the third line:
2 2
4x2 + 4y 2 = ∣⟨ψ∣{A, B}∣ψ⟩∣ + ∣⟨ψ∣[A, B]∣ψ⟩∣ (41a)
2
= 4 ∣⟨ψ∣AB∣ψ⟩∣ (41b)
≤ 4 ⟨ψ∣A A∣ψ⟩ ⟨ψ∣B B∣ψ⟩
† †
(41c)
= 4 ⟨ψ∣A ∣ψ⟩ ⟨ψ∣B ∣ψ⟩
2 2
(41d)
2 2
= 4 ⟨ψ∣ (C − ⟨C⟩) ∣ψ⟩ ⟨ψ∣ (D − ⟨D⟩) ∣ψ⟩ . (41e)
2
Obviously ∣⟨ψ∣{A, B}∣ψ⟩∣ ≥ 0, and hence:
2 2 2
4 ⟨ψ∣ (C − ⟨C⟩) ∣ψ⟩ ⟨ψ∣ (D − ⟨D⟩) ∣ψ⟩ ≥ ∣⟨ψ∣[A, B]∣ψ⟩∣ (42a)
2
= ∣⟨ψ∣[C, D]∣ψ⟩∣ (42b)

from which the Heisenberg uncertainty principle now follows.

5. Positive Operator Valued Measures (POVMs)

Given a generalized measurement with measurement operators {Mk } we define the elements of a POVM via

Ek = Mk† Mk . (43)
†
The normalization condition then becomes ∑k Ek = I. Clearly, Ek† = Mk† (Mk† ) = Ek , so that the POVM elements are
Hermitian. It is easy to show that the Ek s are moreover positive operators, i.e., that ⟨ψ∣Ek ∣ψ⟩ ≥ 0 is true for every ∣ψ⟩ (for more
2
details on positive operators see Appendix A 7). Indeed, ⟨ψ∣Ek ∣ψ⟩ = ⟨ψ∣Mk† Mk ∣ψ⟩ = ∥Mk ∣ψ⟩∥ ≥ 0. Note that the probability
of outcome k is simply pk = ⟨ψ∣Ek ∣ψ⟩. How about the effect of the measurement Ek on a state ∣ψ⟩? Suppose we are given
an arbitrary set of positive operators {Ek } that satisfy ∑k Ek = I. How do we extend the measurement postulate in this case?
√ It is true that for any operator A, we can always
The answer to this is to use the so called polar decomposition of the operator.
find a unitary U and a positive operator P such that A = U P with P = A† A. If the operator A is invertible, then such a
decomposition is unique and U = AP −1 . In our case, we could use the given POVMs and define for every k
√
Mk = U k E k , (44)

where the Uk ’s are just arbitrary unitaries. In other words, since only the Ek are specified (by assumption), we are free to choose
the Uk ’s, and for every such choice we get a different set of Mk ’s. Hence, we can now write the state after the measurement as
√
Uk Ek ∣ψ⟩
∣ψ⟩ ↦ ∣ψk ⟩ = √ with probability pk = ⟨ψ∣Ek ∣ψ⟩ . (45)
pk

Since Uk ’s are arbitrary (again, since only the Ek ’s were specified), this unitary freedom is a generalization of the freedom to
leave the overall phase of a state unspecified.
 13

To see why POVMs are relevant let’s consider the following example. Suppose we have to play a game. Alice always gives
us one of these two states:
∣ψ1 ⟩ = ∣0⟩ , or (46a)
1
∣ψ2 ⟩ = √ (∣0⟩ + ∣1⟩) ≡ ∣+⟩ . (46b)
2
We do not, a priori, know which state has been handed to us. We do know that it is one of these two states. Our task is
to perform measurements and decide which of the two states we were given. Also, we are not allowed to make an error in
identification, i.e., if we provide an answer, it has to be right. However, we are allowed to proclaim ignorance if we don’t
know the answer. Moreover, we must treat both states equally, i.e., we cannot preferentially identify only one of the states
and proclaim ignorance on the other. What is our strategy? Since these are non-orthogonal states, there is no way distinguish
these two states with complete certainty [3][Box 2.3, p.87]. Suppose we try to do it with projective measurements. Let’s take
the measurement set to be {Mk } = {P0 , P1 }, where Pi = ∣i⟩ ⟨i∣ , i = 0, 1. Suppose that the outcome is the index 0. This can
happen in either of two ways: Alice prepared ∣ψ1 ⟩ or she prepared ∣ψ2 ⟩. The probability that the outcome is 0 given that she
prepared ∣ψ1 ⟩ is p(0∣ψ1 ) = ⟨ψ1 ∣P0 ∣ψ1 ⟩ = ⟨0∣0⟩ ⟨0∣0⟩ = 1. And, the probability that the outcome is 0 given that she prepared ∣ψ2 ⟩
is p(0∣ψ2 ) = ⟨ψ2 ∣P0 ∣ψ2 ⟩ = 1/2. This means that if the outcome is 0 then we cannot know for sure which of the two states Alice
prepared, since both occur with non-vanishing probability. Therefore in this case we must proclaim ignorance. However, note
that it also follows that p(1∣ψ1 ) = 0 and p(1∣ψ2 ) = 1/2 which means that given outcome 1 we know with certainty that Alice
prepared ∣ψ2 ⟩. Thus we cannot satisfy the condition of treating the two states equally. As is easily checked, this will always be
the case with a projective measurement.
Now, let’s try with an intelligent choice of POVMs. Define,
E1 = α∣1⟩⟨1∣, (47a)
E2 = α ∣−⟩ ⟨−∣ , (47b)
E3 = I − E1 − E2 . (47c)
where, ∣−⟩ ≡ √12 (∣0⟩ − ∣1⟩) and α > 0 is an arbitrary parameter which we can optimize later, and which must be chosen so that
E3 > 0. If we do so then this clearly is a set of POVMs, since ∑k Ek = I, and for suitable α, all the Ek ’s are positive. Let us
now compute the probabilities of the 3 possible outcomes,
p(1∣ψ1 ) = ⟨ψ1 ∣E1 ∣ψ1 ⟩ = 0, (48a)
α
p(1∣ψ2 ) = ⟨ψ2 ∣E1 ∣ψ2 ⟩ = , (48b)
2
α
p(2∣ψ1 ) = ⟨ψ1 ∣E2 ∣ψ1 ⟩ = , (48c)
2
p(2∣ψ2 ) = ⟨ψ2 ∣E2 ∣ψ2 ⟩ = 0, (48d)
α
p(3∣ψ1 ) = ⟨ψ1 ∣E3 ∣ψ1 ⟩ = 1 − , (48e)
2
α
p(3∣ψ2 ) = ⟨ψ2 ∣E3 ∣ψ2 ⟩ = 1 − . (48f)
2
So, if we get outcome 1, we can say with certainty that the given state was ∣ψ2 ⟩ and if we get outcome 2, we can say with certainty
that the given state was ∣ψ1 ⟩. With outcome 3, we have no information about the state, i.e., we must proclaim ignorance. But in
two of the three outcomes we have been able to obtain an answer with certainty. So, in order to make the probability of outcome
3 as small as possible (since it yields no information), we have to increase α as much as possible while keeping E3 positive. If
we write out E3 as a matrix and√place the constraint of the eigenvalues of this matrix being positive, it easy to show that the
maximal allowed value of α is 1+√2 2 .

III. DENSITY OPERATORS

We will motivate the study of density operators by considering ensembles of pure quantum states. Suppose, instead of having
a single state vector, we only know that our system is in state ∣ψ1 ⟩ with probability q1 , or in state ∣ψ2 ⟩ with probability q2 , and
so on. In other words, we have an pure state ensemble {qi , ∣ψi ⟩}N
i=1 describing our system.
Now, we would like to understand what happens when we make measurements on this quantum system. Suppose the state
were ∣ψi ⟩ and we measure with a set of measurement operators {Mk }. The measurement transformation would be:
Mk ∣ψi ⟩
∣ψi ⟩ ↦ √ = ∣ψik ⟩ (49)
pk∣i
 14

with probability pk∣i = ⟨ψi ∣Mk† Mk ∣ψi ⟩, which is the probability of outcome k, given a state ∣ψi ⟩.
Now, consider that we did not know what the state was but only that it came from the ensemble {qi , ∣ψi ⟩}N
i=1 . Then the
probability of obtaining the outcome k as a result of the measurement on the ensemble is:
pk = ∑ pk∣i qi (50a)
i

= ∑ qi ⟨ψi ∣Mk† Mk ∣ψi ⟩ (50b)

i

= Tr [Mk† Mk (∑ qi ∣ψi ⟩ ⟨ψi ∣)] . (50c)

i

In Eq. (50c) we define the operator within the parentheses as,

ρ = ∑ qi ∣ψi ⟩ ⟨ψi ∣ . (51)
i

This is called the density matrix or density operator and is a central object in quantum mechanics. The density matrix is
completely equivalent to the pure state ensemble {qi , ∣ψi ⟩}N
i=1 , but it has the advantage of being directly useful for calculations.
Indeed, using the density matrix, Eq. (50c) becomes:
pk = Tr(Ek ρ), (52)
where we have defined Ek ≡ Mk† Mk as the element of a POVM.
What about the state that results after measurement result k has been observed? Suppose that outcome k is observed for a
known initial state ρ = ∑i qi ∣ψi ⟩ ⟨ψi ∣. If we let ∣ψik ⟩ ∶= M√kp∣ψi ⟩ [as in Eq. (49)], then {pi∣k , ∣ψik ⟩}i is the resulting ensemble,
k∣i

where pk∣i is the probability of outcome k given state ∣ψi ⟩. On the other hand, if outcome k was observed, and we don’t know
the initial state, then we should sum over all possible states compatible with outcome k (the states ∣ψik ⟩) with their respective
conditional probabilities pi∣k . Thus, the density operator for result k becomes
ρk = ∑ pi∣k ∣ψik ⟩⟨ψik ∣ (53a)
i

Mk ∣ψi ⟩⟨ψi ∣Mk†

= ∑ pi∣k (53b)
i pk∣i
qi
=∑ Mk ∣ψi ⟩⟨ψi ∣Mk† (53c)
i pk
Mk ρMk†
= (53d)
pk
Mk ρMk†
= , (53e)
Tr [ρMk† Mk ]
where in the third line we used Bayes’ rule Pr(i&k) = Pr(i∣k)Pr(k) = Pr(k∣i)Pr(i), where Pr(i) = qi is the a priori probability
of having state ∣ψi ⟩, and pk is the probability of measurement outcome k, as in Eq. (52). Thus, comparing the pure state case to
the generalized density operator case we observe
Mk ∣ψ⟩ Mk ρMk†
∣ψ⟩ ↦ √ ρ↦ . (54)
pk pk

A. Properties of the density operator

• Unit trace: The trace operation is reviewed in Appendix A 6. The density operator ρ has Tr [ρ] = 1. This property can
easily be seen by the following calculation:
Tr [ρ] = ∑ qi Tr [∣ψi ⟩⟨ψi ∣] = ∑ qi = 1. (55)
i i

• Hermiticity: The density operator ρ is Hermitian. The following line demonstrates this
ρ† = ∑ qi∗ (∣ψi ⟩⟨ψi ∣) = ∑ qi ∣ψi ⟩⟨ψi ∣ = ρ
†
(56)
i i

where we’ve used that probabilities qi are real and projectors formed from outer-products are Hermitian.
 15

• Positive definite: For all vectors ∣ν⟩ ∈ H, the density operator ρ has ⟨ν∣ρ∣ν⟩ ≥ 0:
2
⟨ν∣ρ∣ν⟩ = ∑ qi ∣⟨ψi ∣v⟩∣ ≥ 0, (57)
i

since the qi are all non-negative by virtue of being probabilities. But since Trρ = 1 it clearly must have at least one
eigenvalue that is non-zero. Therefore ρ must be positive, not just positive semi-definite (positive operators are defined in
Appendix A 7).

Note that positivity implies Hermiticity, since an operator is Hermitian iff it has only real eigenvalues. Therefore we don’t
actually need to separately stipulate Hermiticity. Also note that the density operator deserves to be called an operator: it acts as
a transformation between two copies of the Hilbert space, i.e., ρ ∶ H ↦ H.
We define the space of positive, trace-one linear operators acting on H as D(H). Thus

ρ ∈ D(H). (58)

B. Dynamics of the density operator

Recall the two equivalent descriptions of dynamics of the pure quantum state

∣ψ(t)⟩ = U (t) ∣ψ(0)⟩ ⇔ ∣ψ̇⟩ = −iH ∣ψ⟩ (59)

where U (t) and H are related by U (t) = e−iHt . Consider one of the pure states forming the ensemble {qi , ∣ψi ⟩}i . This state will
evolve as

∣ψi (t)⟩ = U (t) ∣ψi (0)⟩ (60)

and the time-evolution of the density operator associated to the ensemble is

ρ(t) = ∑ qi ∣ψi (t)⟩⟨ψi (t)∣ (61a)
i

= ∑ qi U (t)∣ψi (0)⟩⟨ψi (0)∣U † (t) (61b)

i

= U (t)ρ(0)U † (t). (61c)

The Schrödinger equation for the density operator takes a slightly different form however and we can derive it by taking the
time-derivative of the first line above,
∂ ∂
ρ(t) = ∑ qi ∣ψi (t)⟩⟨ψi (t)∣ (62a)
∂t ∂t i
∂ ∂
ρ̇(t) = ∑ qi [( ∣ψi (t)⟩) ⟨ψi (t)∣ + ∣ψi (t)⟩ ( ⟨ψi (t)∣)] . (62b)
i ∂t ∂t
At this point we invoke the Schrödinger equation for pure states while making note that after Hermitian conjugation of the
Schrödinger equation we obtain ⟨ψ̇j ∣ = i ⟨ψj ∣ H. Thus:
ρ̇(t) = ∑ qi (−iH∣ψi (t)⟩⟨ψi (t)∣ + i∣ψi (t)⟩⟨ψi (t)∣H) (63a)
i

= −i [H (∑ qi ∣ψi (t)⟩⟨ψi (t)∣) − (∑ qi ∣ψi (t)⟩⟨ψi (t)∣) H] (63b)

i i
= −i (Hρ − ρH) (63c)
= −i [H, ρ] (63d)
where [⋅, ⋅] represents the commutator of the two operators.

C. Restatement of the postulates of quantum mechanics

We can now summarize the four postulates in terms of the density operator.
 16

1. The state space is the Hilbert-Schmidt space of linear operators ρ such that Tr [ρ] = 1 and ρ > 0. The inner product in the
Hilbert-Schmidt space is defined as Tr[A† B] for any√two operators
√ A and B acting on the same Hilbert space. This inner
product defines a length in the usual way, i.e., ∥ρ∥ = ⟨ρ, ρ⟩ = P . The quantity

P ≡ Tr[ρ2 ] (64)

is called the “purity” of the state ρ. Thus a density matrix can have “length” ≤ 1. A state is called “pure” if P = 1 and
“mixed” if P < 1.
2. State spaces are composed via the tensor product ⊗.
3. Density operators evolve as ρ̇ = −i [H, ρ] under a Hamiltonian H, or equivalently as ρ(t) = U (t)ρ(0)U † (t) where the
unitary U (t) = e−itH .

4. A general measurement operation defined by elements {Mk } results with probability pk = Tr [ρMk† Mk ] in the state
Mk ρMk†
transformation ρ ↦ pk
.
Expectation values are now computable in terms of the ρ as well. Consider an observable A measured for a system in the pure
state ensemble {qi , ∣ψi ⟩}. Previously, in Eq. (33), we showed that the expectation value was ⟨A⟩ψ = ⟨ψ∣A∣ψ⟩ = Tr(A ∣ψ⟩ ⟨ψ∣).
We need to modify this by assigning each pure state ∣ψi ⟩ in the ensemble its weight qi . Thus the new expression for the
expectation value is:

⟨A⟩ρ = ∑ qi ⟨ψi ∣A∣ψi ⟩ = ∑ qi Tr(A ∣ψi ⟩ ⟨ψi ∣) = Tr(A ∑ qi ∣ψi ⟩ ⟨ψi ∣) = Tr(Aρ) = Tr(ρA). (65)
i i i

Likewise, the standard deviation becomes:

√
∆A = ⟨(A − ⟨A⟩ρ )2 ⟩ρ , (66)

and it is not hard to prove the associated uncertainty relation:

1
∆A∆B ≥ ∣⟨[A, B]⟩ρ . (67)
2
To sum up, here is a comparison of the postulates for pure states and density operators:

Pure States General States

State space Hilbert space H Trace-class operator space D
Postulate 1 Tr [ρ] = 1
State ket vector ∣ψ⟩ ∈ H s.t. ⟨ψ∣ψ⟩ = 1 density operator ρ s.t. {
ρ>0
Inner product f (∣µ⟩ , ∣ω⟩) ≡ ⟨µ∣ω⟩, ∀ ∣µ⟩ , ∣ω⟩ ∈ H f (A, B) ≡ Tr [A B], ∀A, B ∈ D
†

Hilbert-Schmidt inner product

Postulate 2 Expansion tensor product ⊗ tensor product ⊗
Dynamics Schrödinger equation: Liouville-von Neumann equation:
Postulate 3
w/ Hamiltonian H d∣ψ(t)⟩
dt
= −iH ∣ψ(t)⟩ dρ(t)
dt
= −i [H, ρ(t)]
Measurement outcome k ∈ K w.p. pk = ⟨ψ∣ Mk Mk ∣ψ⟩ outcome k ∈ K w.p. pk = Tr [Mk ρMk† ]
†
Postulate 4 Mk ρMk†
Mk ∣ψ⟩
w/ meas. ops. {Mk }k∈K ∣ψ⟩ ↦ √
pk
ρ↦ pk

D. More on pure and mixed quantum states

We defined “pure” and “mixed” states above according to the value of the purity P = Tr[ρ2 ] being 1 or < 1. Prior to
introducing the density operator formalism, we had considered quantum states as vectors in the Hilbert space. This formalism
is equivalent to pure state ensembles of the type {1, ∣ψ⟩}, i.e., having only a single element. It is not hard to see that such
special ensembles are “pure” quantum states. The associated density operator is ρ = ∣ψ⟩⟨ψ∣. It is useful to think of pure states as
ensembles with only one member and probability 1.
Any state that is not not pure is by definition mixed. This means that they are described by ensembles of the form {pi , ∣ψ⟩i }
where for all i, 0 < pi < 1. The density operator associated with a mixed ensemble is the mixture of the pure states with their
associated weights [as seen in Eq. (51)].
 17

Note that a pure state is a projector: (∣ψ⟩⟨ψ∣)(∣ψ⟩⟨ψ∣) = ∣ψ⟩⟨ψ∣. Therefore, if a state ρ is pure then ρ2 = ρ. The converse is also
true: ρ2 = ρ implies that ρ is pure. It is easy to check that these conditions are equivalent to the definition in terms of purity P .
We can also define a mixed state ensemble, i.e., a collection of mixed states ρk with associated probabilities pk , as

{pk , ρk } ⇔ ρ = ∑ pk ρk . (68)
k

E. Unitary equivalence in ensembles

When are two pure state ensembles equivalent? Consider for example the two ensembles
{(3/4, 1/4), (∣0⟩ , ∣1⟩)} (69a)
{(1/2, 1/2), (∣a⟩ , ∣b⟩)}, (69b)
where
√ √
∣a⟩ = 3/4 ∣0⟩ + 1/4 ∣1⟩ (70a)
√ √
∣b⟩ = 3/4 ∣0⟩ − 1/4 ∣1⟩ . (70b)
On the face of it, the first of these ensembles represents a biased classical coin (“heads”, or 0, with probability 3/4, tails, or 1
with probability 1/4), whereas the second is quantum in the sense that each state is a superposition state. But are they really
different? It is straightforward to check that in fact the two density matrices corresponding to these two ensembles are equal.
This being the case, there is no measurement that can distinguish them, and that means we must consider them to be the same.
Theorem 1. Two pure state ensembles with the same number of elements2 {qi , ∣ψi ⟩}i and {rj , ∣φj ⟩}j correspond to the same
density operator if and only if there exists a unitary U with entries [U ]ij such that
√ √
qi ∣ψi ⟩ = ∑[U ]ij rj ∣φj ⟩ (71)
j

Proof. We show explicitly the “if” direction of the proof. The complete proof is found in [3][p.104]. Consider the following
mixture,
√ √
∑ qi ∣ψi ⟩⟨ψi ∣ = ∑ ( qi ∣ψi ⟩) (⟨ψi ∣ qi ) (72a)
i i

⎛ √ ⎞ √
= ∑ ∑[U ]ij rj ∣φj ⟩ (∑ ⟨φk ∣ rk [U † ]ik ) (72b)
i ⎝ j ⎠ k

√
=∑ rj rk (∑[U ]ij [U † ]ik ) ∣φj ⟩⟨φk ∣ (72c)
j,k i
√
=∑ rj rk (δjk ) ∣φj ⟩⟨φk ∣ (72d)
j,k

= ∑ rj ∣φj ⟩⟨φj ∣, (72e)

j

where in Eq. (72d) we used the unitarity of U . Thus the two ensembles represent the same density operator.

F. Visualizing the density matrix of a qubit: the Bloch sphere

A qubit is a quantum state ∣ψi ⟩ in a two-dimensional Hilbert space H = C2 = span{∣0⟩ , ∣1⟩} where ∣0⟩ and ∣1⟩ form an
orthonormal basis for H. The density operator for any state in this space is thus of the form ∑i qi ∣ψi ⟩⟨ψi ∣ and can hence be
represented by a 2 × 2 complex matrix of the form

a b
ρ=[ ]. (73)
c d

2 If necessary pad the smaller set with zeroes to make it equal in length to the larger set.
 18

However, applying the properties of density operators can reduce this to an expression of only two variables. First, the unit trace
reduces to the condition d = 1 − a and Hermiticity reduces to the condition that c = b∗ and that a be real. Thus, the density matrix
is completely parametrized by the complex number b and the real number a and takes the form

a b
ρ=[ ]. (74)
b∗ 1 − a
Positivity is the statement that the eigenvalues λ± are non-negative:
∣ρ − λI∣ = 0 ⇒ λ2 − (Trρ)λ + ∣ρ∣ = 0, (75)
i.e., using Trρ = 1:
1 √
λ± = (1 ± 1 − 4∣ρ∣) ≥ 0. (76)
2
This parametrization requires only three parameters and we can thus embed it naturally in three dimensions. Before we proceed
to do this we will decompose the density operator one more time but in a more useful basis.
Recall the Pauli matrices σx , σy , σz , σ0 . Any qubit density matrix can represented by
1 1
ρ= (I + ∑ vi σi ) = (I + v⃗ ⋅ σ
⃗) , (77)
2 i 2
where v⃗ = (vx , vy , vz ) and σ
⃗ = (σx , σy , σz ). In terms of the elements of v⃗, ρ appears as

1 1 + vz vx − ivy
ρ= [ ]. (78)
2 vx + ivy 1 − vz
To relate this to our previous analysis simply let b = 1/2(vx − ivy ) and a = 1/2(1 + vz ). We call v⃗ the Bloch vector. The 2 × 2
matrix we have constructed using the Bloch vector is not, however, necessarily a valid quantum state. Unit trace is guaranteed
by the construction, and positivity can now be made explicit by noting that
1 1
∣ρ∣ = v ∥2 ) ,
(1 − vz2 − (vx2 + vy2 )) = (1 − ∥⃗ (79)
4 4
so that
1
λ± = (1 ± ∥⃗
v ∥) . (80)
2
The two solutions are ∥⃗
v ∥ ≤ 1 and ∥⃗
v ∥ ≥ −1, which is trivially satisfied. Thus if we require positivity, the relevant constraint is

∥⃗
v∥ ≤ 1 . (81)

Let us also relate the magnitude of the Bloch vector to the purity of the quantum state. Recall that a pure quantum state is a
projector and thus ρ2 = ρ for pure states. If we calculate the density operator ρ2 we find
1 1 2
ρ2 = (I + v⃗ ⋅ σ
⃗ ) (I + v⃗ ⋅ σ
⃗ ) = (I + 2⃗ ⃗ + (⃗
v⋅σ ⃗) ) .
v⋅σ (82)
4 4
The term (⃗ ⃗ )2 becomes
v⋅σ
∑ vk vl σk σl . (83)
k,l∈{x,y,z}

⃗ )2 =
v⋅σ
Recall Eq. (A21). Taking the trace and noting that the Pauli matrices are traceless only the δkl term remains. Thus Tr(⃗
2
∥⃗
v ∥ TrI, with TrI = 2, and Eq. (82) yields:
1
(1 + ∥⃗
Trρ2 = v ∥2 ) . (84)
2
From this form it is clear that any unit Bloch vector will make Trρ2 = 1, i.e., a pure state, and Bloch vectors of length less than
1 yield mixed states.
Having gathered the requisite intuition for the geometry at hand, we call the set of all valid Bloch vectors v⃗ the Bloch sphere,
also known as the Poincaré sphere in optics.
Since the Bloch sphere can describe all qubit states and can be embedded in three dimensions it is a useful tool for illustrating
various common qubit states.
 19

FIG. 1. The Bloch sphere is a geometric representation of the collection of all Bloch vectors v⃗ which describe valid qubit density operators.
Thus, the sphere is of radius 1, its surface represents all pure states, and its interior represents all mixed states. In this diagram the blue vector
lies on the surface of the sphere indicating a pure state, whereas the red vector lies in its interior indicating a mixed state.

v = (0, 0, ±1)): The density matrix takes the form

• Z poles (⃗
I ± σz
ρ=
2
1±1 1∓1
= ∣0⟩⟨0∣ + ∣1⟩⟨1∣
2 2
which yields ∣0⟩⟨0∣ for vz = 1 and ∣1⟩⟨1∣ for vz = −1.
v = (±1, 0, 0)): The density matrix takes the form
• X poles (⃗
I ± σx
ρ=
2
1
= (∣0⟩⟨0∣ + ∣1⟩⟨1∣ ± (∣0⟩⟨1∣ + ∣1⟩⟨0∣))
2
1
= (∣0⟩ ± ∣1⟩) (⟨0∣ ± ⟨1∣)
2
which yields ∣+⟩⟨+∣ for vx = 1 and ∣−⟩⟨−∣ for vx = −1.
v = (0, ±1, 0)): The density matrix takes the form
• Y poles (⃗
I ± σy
ρ=
2
1
= (∣0⟩⟨0∣ + ∣1⟩⟨1∣ ± (−i∣0⟩⟨1∣ + i∣1⟩⟨0∣))
2
1
= (∣0⟩ ± i ∣1⟩) (⟨0∣ ± (−i) ⟨1∣)
2
√ √
which yields (∣0⟩ + i ∣1⟩)/ 2 for vy = 1 and (∣0⟩ − i ∣1⟩)/ 2 for vy = −1.
v = (0, 0, 0)): The density matrix takes the form ρ = I/2, the maximally mixed state.
• Center (⃗
Since the dimensionality of this geometric representation goes as d2 − 1 for a d-level system (the density matrix becomes a
d×d matrix, and the trace constraints removes one matrix element), the Bloch sphere is typically only used to represent two-level
systems. As we shall see later on, the Bloch sphere plays an important visualization role in understanding the dynamics of open
quantum systems.

IV. COMPOSITE SYSTEMS

A. Combining a system and a bath

Now that we have discussed in detail the dynamics of a single system, let us consider more complex systems. Consider a two
component system, where we have a subsystem of our interest, A (often we’ll just call it “system”), and the other subsystem is
 20

the bath, B. Together, the system and the bath comprise the lab, or even the entire universe. We can think of A as a quantum
computer, or a molecule, or any other system we’re interested in studying. We shall assume that the total system evolves
according to the Schrödinger equation and that it is described by a density matrix ρ(t). Further, let the subsystem Hilbert spaces
be

HA = span{∣i⟩A } (85)
HB = span{∣µ⟩B } (86)

Here, i goes from 0 to dA − 1, the dimension of the Hilbert space of A, and µ goes from 0 to dB − 1, the dimension of the Hilbert
space of B. Usually, the dimension of the bath, dB → ∞, while dA is finite. By the second postulate, the Hilbert space of the
two system combined is the tensor product of the individual spaces:

H = HA ⊗ HB (87a)
= span{∣i⟩A ⊗ ∣µ⟩B } (87b)

Let us figure out the structure of a density matrix in this combined Hilbert space. We can define a pure state ensemble {∣Ψa ⟩ , qa }
for a set of pure states ∣Ψa ⟩ ∈ H. Each of these states can be expanded in the basis above, i.e.,

∣Ψa ⟩ = ∑ ca;iµ ∣i⟩A ⊗ ∣µ⟩B . (88)

i,µ

Thus, the associated density matrix is:

ρ = ∑ qa ∣Ψa ⟩⟨Ψa ∣ = ∑ qa (∑ ca;iµ ∣i⟩A ⊗ ∣µ⟩B )(∑ c∗a;jν ⟨j∣A ⊗ ⟨ν∣B ). (89)
a a i,µ j,ν

Therefore any density matrix in the combined Hilbert space can be written down as

ρ = ∑ λijµν ∣i⟩A ⟨j∣ ⊗ ∣µ⟩B ⟨ν∣ , (90)

ijµν

where λijµν = ∑a qa ca;iµ c∗a;jν .

ij λµν then ρ = ρA ⊗ ρB , where ρA = ∑iµ λij ∣i⟩A ⟨j∣ and ρB = ∑µν λµν ∣µ⟩B ⟨ν∣. In this case ρ is called a
Note that if λijµν = λA B A B

“factorized” state. Such states exhibit no correlations at all between the A and B subsystems. Clearly, however, this is a special
case and in general, ρ cannot be factored in this manner. When it cannot, the subsystems are correlated. These correlations can
be quantum (due to entanglement), classical, or both.
We are primarily interested in the system A. We thus need to find a way to remove the bath B from our description. To do,
we now define a new operation called partial trace, which effectively averages out of the components of B from the combined
density matrix. The resultant density matrix then describes only A.

B. Partial Trace

1. Definition

The partial trace is a linear operator that maps from the total Hilbert space to the Hilbert space of A, i.e., H ↦ HA , defined as
follows. Consider an operator O = MA ⊗ NB such that O acts on H = HA ⊗ HB . Then

TrB (MA ⊗ NB ) ≡ MA Tr(NB ) (91a)

= MA ∑ ⟨µ∣NB ∣µ⟩ (91b)
µ

= ∑ ⟨µ∣[MA ⊗ NB ]∣µ⟩ (91c)

µ

It is understood in the last line that the basis vectors {∣µ⟩}, which span the space HB , act only on the second Hilbert space.
In other words, the expression ⟨µ∣[MA ⊗ NB ]∣µ⟩ is a partial matrix element, where the matrix element is taken only over the
 21

second factor, and the result is an operator acting on HA . Thus, if O = ∑ij `ij MAi ⊗ NBj , then by linearity:

TrB [O] = ∑ `ij TrB (MAi ⊗ NBj ) (92a)

ij

= ∑ `ij MAi ∑ ⟨µ∣NBj ∣µ⟩ (92b)

ij µ

= ∑ ∑ `ij ⟨µ∣[MAi ⊗ NBj ]∣µ⟩ (92c)

µ ij

= ∑ ⟨µ∣O∣µ⟩ (92d)
µ

For example, when applied to a summand in the expression for ρ:

TrB [∣i⟩A ⟨j∣ ⊗ ∣µ⟩B ⟨ν∣] ≡ ∣i⟩A ⟨j∣ ⟨ν∣µ⟩B . (93)

By linearity,

TrB [ ∑ λijµν ∣i⟩A ⟨j∣ ⊗ ∣µ⟩B ⟨ν∣] = ∑ λijµν ∣i⟩A ⟨j∣ ⟨ν∣µ⟩B (94a)
ijµν ijµν

= ∑ λijµµ ∣i⟩A ⟨j∣ = ∑ λ̄ij ∣i⟩A ⟨j∣ , (94b)

ijµ ij

where in the second line we assumed that {∣µ⟩B } forms an orthonormal basis, and we defined λ̄ij = ∑µ λijµµ . This shows that
taking the partial trace leads to a form that looks like a density matrix for the A subsystem. Of course, we’ll have to verify that
it satisfies the properties of a density matrix (unit trace and positivity). Positivity is more challenging, but unit trace is obvious if
we assume (as we should) that Trρ = 1. For, it is then easy to check that this implies ∑iµ λiiµµ = 1. On the other hand, if we are
to interpret TrB [ρ] as a valid density matrix then Tr(TrB [ρ]) = ∑i λ̄ii should be 1, which it is, since it equals ∑iµ λ̄iiµµ .

2. State of a quantum subsystem

Crucially, we now claim that the density matrix of the subsystem A is given by taking the partial trace of the combined density
matrix with respect to B.

ρA = TrB [ρ] . (95)

This is called the reduced density matrix.

To justify this intuitively, we consider the cases which lie on the two extreme ends of combination of bath and system, viz.
the simplest case of a separable density operator, and the case where system and bath are maximally entangled.

1. Case 1: Consider a case where the states of the bath and system are completely separate, and hence form a tensor product.
In such a case, we expect that the density operator of A obtained by partial trace should be the same as the component
of A contributed in the tensor product. And indeed, clearly, if ρ = ρA ⊗ ρB where both terms in the product are properly
normalized states, then

TrB [ρ] = ρA TrB [ρB ] = ρA (96)

2. Case 2: Consider two qubits that are maximally entangled, that is

1
∣ψ⟩AB = √ (∣0⟩A ∣0⟩B + ∣1⟩A ∣1⟩B ). (97)
2

This means that the state state ∣ψ⟩AB contains no separate information about A or B’s state. The reason is that if we
measure, say, B using the measurement operators {M0 = ∣0⟩⟨0∣, M1 = ∣1⟩⟨1∣}, then we find the outcomes 0 and 1 with
equal probability 1/2, and at the same time the state of A becomes either ∣0⟩ or ∣1⟩, respectively. It is easy to check that
this random outcome remains true for any other choice of measurement operators. This means we gain no knowledge at
 22

all about A or B since the measurement outcome is perfectly random. In terms of the partial trace we find:
ρA = TrB [ρ] = TrB [∣ψ⟩AB ⟨ψ∣] (98a)
1
= TrB [∣0⟩A ⟨0∣ ⊗ ∣0⟩B ⟨0∣ + ∣0⟩A ⟨1∣ ⊗ ∣0⟩B ⟨1∣ + ∣1⟩A ⟨0∣ ⊗ ∣1⟩B ⟨0∣ + ∣1⟩A ⟨1∣ ⊗ ∣1⟩B ⟨1∣]
2
(98b)
1
= [∣0⟩⟨0∣ × 1 + ∣0⟩⟨1∣ × 0 + ∣1⟩⟨0∣ × 0 + ∣1⟩⟨1∣ × 1] (98c)
2
1
= [∣0⟩A ⟨0∣ + ∣1⟩A ⟨1∣] = IA /2 (98d)
2
Therefore, the state of A is an equal probabilistic mixture of the ∣0⟩ and ∣1⟩ states, as expected.

Next, we provide a formal justification.

3. Formal justification of using the partial trace to define a subsystem state

Consider a composite system with the Hilbert space H = HA ⊗ HB . If we had an observable MA on subsystem A, then, the
expectation value of that operator would be given by
⟨MA ⟩ρA = Tr[ρA MA ], (99)
where we used Eq. (65).
̃ = MA ⊗ IB on the entire
However, in the case of this composite system, this measurement is actually of the observable M
system ρ in HA ⊗ HB , where we do nothing (the identity operation) to B. Thus,

̃⟩ = Tr[ρ M
⟨M ̃] (100)
The key idea is that these two operations should correspond to the same physical observation and they should produce the
same number. For the theory to be consistent, we demand that

̃⟩,
⟨MA ⟩ ≡ ⟨M (101)
i.e.,
̃ρ]
Tr[MA ρA ] = Tr[M (102)
It can be shown that this condition is satisfied iff we define ρA ≡ TrB (ρ). We shall prove the theorem in one direction, that is,
if ρA = TrB (ρ), then ⟨M ⟩ = ⟨M̃⟩.

Proof. Let H = HA ⊗ HB = span{∣i⟩A ⊗ ∣µ⟩B }. Then

⟨MA ⟩ = ∑ A ⟨i∣ ρA MA ∣i⟩A (103a)

i
= ∑ A ⟨i∣ TrB [ρ]MA ∣i⟩A (103b)
i
= ∑ A ⟨i∣ ∑ B ⟨µ∣ ρ∣µ⟩B MA ∣i⟩A (103c)
i µ

In going from Eq. (103b) to (103c), we used the expression for the partial trace over operators as given in Eq. (92d). But note
that ρ is an operator acting on the composite system, not just on A, since B ⟨µ∣ ρ∣µ⟩B is a partial matrix element. If we wish to
likewise consider MA as an operator acting on the composite system, then we should extend it to MA ⊗ IB . Also, the correct
order for the product ∣µ⟩B ∣i⟩A , including the tensor product symbol explicitly, is: ∣i⟩A ⊗ ∣µ⟩B . Thus:
⟨MA ⟩ = ∑ ⟨i∣ ⊗ ⟨µ∣ [ρ(MA ⊗ IB )] ∣i⟩ ⊗ ∣µ⟩ (104a)
i,µ

= Tr[ρ(MA ⊗ IB )] (104b)
̃⟩ ,
= ⟨M (104c)
̃⟩.
which shows the desired equality ⟨MA ⟩ = ⟨M
 23

V. OPEN SYSTEM DYNAMICS

In this section we shall find the dynamical evolution of an open quantum system.

A. Kraus Operator Representation

Consider a system S and bath B, such that they have a joint unitary evolution given by U (t) = e−iHt . The initial joint state is
ρ(0). Then, by Schrödinger’s equation,
ρ(t) = U (t)ρ(0)U † (t) (105)
As the density operator of the bath is positive and normalized, it has a spectral decomposition in an orthonormal basis with
non-negative eigenvalues. Hence
ρB (0) = ∑ λν ∣ν⟩⟨ν∣ (106)
ν

where λν are the eigenvalues (probabilities) and {∣ν⟩} are the corresponding orthonormal eigenvectors.
The state of the system is then found by performing a partial trace over the bath, i.e.,
ρS (t) = TrB [ρ(t)]. (107)
We can perform the partial trace in the orthonormal basis of bath eigenstates, i.e.,
ρS (t) = TrB [U (t)ρ(0)U † (t)] (108a)
= ∑ ⟨µ∣U (t)ρ(0)U (t)∣µ⟩
†
(108b)
µ

Let us now assume that the initial state is completely decoupled, that is
ρ(0) = ρS (0) ⊗ ρB (0). (109)
Then
ρS (t) = ∑ ⟨µ∣[U (t)ρS (0) ⊗ ∑ λν ∣ν⟩⟨ν∣U † (t)]∣µ⟩ (110a)
µ ν
√ √
=∑ λν ⟨µ∣U (t)∣ν⟩B ρS (0) λν ⟨ν∣U † (t)∣µ⟩B (110b)
µν

= ∑ Kµν (t)ρS (0)Kµν

†
(t). (110c)
µν

The system-only operators {Kµν } are called the Kraus operators and are given by
√
Kµν (t) = λν ⟨µ∣U (t)∣ν⟩ (111)
(note the partial matrix element, leaving us with an operator acting on the system), and the equation defining the evolution of
the system in terms of Kraus operator is called the Kraus Operator Sum Representation (OSR)

ρS (t) = ∑ Kµν (t)ρS (0)Kµν

†
(t) (112)
µν

This is a pivotal result; as we shall see it includes the Schrödinger equation as a special case.

B. Normalization and the special case of a single Kraus operator

The system state should be normalized at all times, so we demand

Tr[ρS (t)] = 1 (113a)
= Tr[∑ Kµν (t)ρS (0)Kµν
†
(t)] (113b)
= ∑ Tr[Kµν (t)ρS (0)Kµν
†
(t)] (113c)
= ∑ Tr[Kµν
†
(t)Kµν (t)ρS (0)] (113d)
= Tr[∑ Kµν
†
(t)Kµν (t)ρS (0)] (113e)
 24

It is easy to check that the equation is satisfied if ∑ Kµν

†
(t)Kµν (t) = I. However, this condition is not necessary. Thus the
system state is guaranteed to be normalized provided the Kraus operators satisfy the following identity,
∑ Kµν (t)Kµν (t) = I .
†
(114)
µν

This criterion can be verified for our definition of Kraus operators, given by Eq. (111).
∑ Kµν Kµν = ∑ λν ⟨µ∣U (t)∣ν⟩ ⟨ν∣U (t)∣µ⟩
† †
(115a)
µν µν

⎛ ⎞
= ∑ λν ⟨ν∣U † (t) ∑ ∣µ⟩⟨µ∣ U (t)∣ν⟩ (115b)
ν ⎝ µ ⎠
= ∑ λν ⟨ν∣ν⟩ (115c)
ν
= ∑ λν = 1 (115d)
Thus, such a set of Kraus operators preserves normalization.
Note that when there is just a single Kraus operator, the normalization condition (114) forces it to be unitary, which is just the
case of closed system evolution! We can see more explicitly how this comes about, as follows.

C. The Schrödinger equation as a special case

√
Assume that U = US ⊗ UB . In this special case the Kraus operators become Kµν = US λν ⟨µ∣UB ∣ν⟩ ≡ cµν US . It’s easy to
see that the sum rule normalization condition implies ∑µν c∗µν cµν = 1, since now ∑µν Kµν
†
Kµν = ∑µν c∗µν cµν US† US = I. Thus:

ρS (t) = ∑ cµν US (t)ρS (0)c∗µν US† (t) = US (t)ρS (0)US† (t) , (116)
µν

which is unitary, Schrödinger-like dynamics. Hence, the Kraus operator sum representation is more general than the Schrödinger
equation, because it contains the latter as a special case.

VI. COMPLETE POSITIVITY AND QUANTUM MAPS

We have seen [Eq. (112)] that the evolution of the state ρS of an open quantum system can be expressed as unitary evolution
of the composite system+bath, followed by a partial trace, which leads to the Kraus operator sum representation (Kraus OSR):
ρS (t) = TrB [U (t)(ρS ⊗ ρB )U † (t)] = ∑ Kα (t)ρS (0)Kα† (t) , (117)
α

where we have collected the earlier µν indices into a single index: α = (µν). From now on let us drop the S subscript since
we’ll be focusing on the system alone. We’ll reintroduce it as necessary.

A. Non-selective measurements

Let us observe that the OSR represents more than dynamics. It can also capture measurements. Specifically, consider mea-
surement operators {Mk } with ∑k Mk† Mk = I. Recall that a state subjected to this measurement maps to

Mk ρMk†
ρ ↦ ρk = (118)
Tr [Mk ρMk† ]

with probability pk = Tr [Mk ρMk† ]. Consider the case where we perform this measurement but do not learn the outcome k.
What happens to ρ after this measurement? In this case
ρ z→ ⟨ρ⟩ = ∑ pk ρk = ∑ Mk ρMk† (119)
k k

which we recognize as a non-selective measurement. This last form is in the Kraus operator-sum representation with the Kraus
operators Mk . Thus, we can encapsulate the non-selective measurement postulate in the operator-sum formalism.
Since both dynamics and measurements are captured by the OSR, and there are no other quantum processes according to our
postulates, this suggests that the OSR is truly fundamental. It thus deserves further scrutiny.
 25

B. The OSR as a map

It is useful to think of the OSR as a map (or synonymously a process or channel) Φ from the initial to the final system state,
i.e.,

ρ(t) = Φ[ρ(0)] ↔ Φ ∶ ρ(0) ↦ ρ(t) , (120)

where Φ[X] ≡ ∑α Kα XKα† . Note that Φ is an operator acting on operators, sometimes called a superoperator. While we
started with vectors ∣v⟩ in a Hilbert space HS , and moved the density operators ρ ∶ HS ↦ HS belonging to the space of positive
trace-class operators D(HS ), the map Φ ∶ D ↦ D belongs to D[D(HS )], as we shall see shortly. In terms of dimensions, if
dim(HS ) = d, then dim(D(HS )) = d2 , and dim(D[D(HS )]) = d4 , reflecting the fact that vectors are of dimension d × 1,
density matrices of dimension d × d, and quantum maps of dimension d2 × d2 .
It will prove to be profitable to adopt an even more abstract point of view, and seek to determine the key properties that any
such map possesses. We can easily identify three properties by inspection:

1. Trace Preserving:

Tr[Φ(ρ)] = ∑ Tr(Kα ρKα† ) = ∑ Tr(Kα† Kα ρ) = Tr(∑ Kα† Kα ρ) = Tr(ρ) , (121)

α α α

where we used the fact that ∑α Kα† Kα = I. Thus the map Φ is trace-preserving.
2. Linear:
By direct substitution we find:

Φ(aρ1 + bρ2 ) = ∑ Tr(Kα aρ1 Kα† ) + ∑ Tr(Kα bρ2 Kα† ) = a ∑ Tr(Kα ρ1 Kα† ) + b ∑ Tr(Kα ρ2 Kα† ) = aΦ(ρ1 ) + bΦ(ρ2 )
α α α α
(122)
for any scalars a and b. Thus the map Φ is linear.
3a. Positivity:
This property means that Φ maps positive operators to positive operators. Assume the operator A > 0, i.e., it has only
non-negative eigenvalues, not all zero. Note that any density matrix ρ must be positive, and we can write A = ∑i λi ∣i⟩⟨i∣
where all λi ≥ 0 (the spectral decomposition of A).
In order to demonstrate that Φ(A) > 0 it is sufficient show that ⟨ν∣Φ(A)∣ν⟩ ≥ 0 for all ∣ν⟩ ∈ HS , since this means in
particular that the eigenvalues of Φ(A) are all non-negative. Let ∣wa ⟩ = Kα† ∣ν⟩. Then:

⟨ν∣Φ(A)∣ν⟩ = ∑⟨ν∣Kα AKα† ∣ν⟩ = ∑⟨wa ∣A∣wa ⟩ = ∑ λi ∣⟨wa ∣i⟩∣2 . (123)

α α ai

On the right hand side it is clear that each term in the sum is positive. Therefore Φ(A) > 0, and Φ itself is a positive map.

The Kraus OSR satisfies these three properties, but does every map that satisfy the same properties have a Kraus OSR? The
answer is negative. It turns out that we must modify and strengthen the positivity property into “complete positivity”.

C. Complete Positivity

The map Φ is a completely positive (CP) map. It maps positive operators to positive operators (is “positivity preserving”), and
(k)
moreover, it can be shown that even Φ ⊗ IR is positive for all k, where k is the dimension of an ancillary Hilbert space HR ,
and IR denotes the identity (super-)operator on HR . Conversely, every CP map can be represented as a Kraus OSR.
More formally, let B(H) denote the space of linear operators acting on the Hilbert space H, i.e., X ∶ H ↦ H is equivalent
to X ∈ B(H). Let A ∈ B(HS ⊗ HR ), where HS denotes the system space and HR is some auxiliary space with dimension k.
Assume that A > 0. Denote by IR the identity map on B(HR ) [i.e., IR (V ) = IV I for all V ∈ B(HR )]. Also, let Φ ∈ B[B(HS )],
i.e., Φ ∶ B(HS ) ↦ B(HS ).

3b. Complete Positivity

If (Φ ⊗ IR )(A) > 0 ∀k, then Φ is called a completely positive (CP) map. If in addition Tr[Φ(X)] = Tr(X) ∀X ∈ B(HS )
then Φ is called a completely positive trace preserving (CPTP) map.
 26

Note that when k = 1, complete positivity reduces to ordinary positivity.

It turns out that conditions 1, 2, 3b are necessary and sufficient for the Kraus OSR. That is:
Theorem 2. A map Φ has a Kraus operator sum representation [i.e., Φ(X) = ∑α Kα XKα† with ∑α Kα† Kα = I] iff it is trace
preserving, linear, and completely positive.
Let us prove one direction of this theorem: that the Kraus OSR is completely positive (we already showed trace preservation
and linearity). To this end, note that if Φ has a Kraus OSR then
(Φ ⊗ IR )(A) = ∑(Kα ⊗ IR )(A)(Kα† ⊗ IR ) . (124)
α

To prove that Φ is CP we need to show that Φ ⊗ IR is positive for all d = dim(HR ). Indeed:
⟨ν∣ (Φ ⊗ IR )(A) ∣ν⟩ = ∑ ⟨ν∣ (Kα ⊗ IR )A(Kα† ⊗ IR ) ∣ν⟩ = ∑⟨wα ∣A∣wα ⟩ > 0 , (125)
α α

where we defined ∣wα ⟩ = (Kα† ⊗ IR ) ∣ν⟩, where now ∣ν⟩ ∈ HS ⊗ HR , and we drew upon the fact that A > 0, as in Eq. (123).
The key feature of the Kraus OSR that makes it a completely positive map is having the same operator (Kα ) on both sides.
For example, something like ∑αβ Kα XKβ† is not a CP map, and the proof of positivity as in Eq. (125) would clearly not have
worked.
To prove the reverse direction, that all maps that satisfy conditions 1, 2, 3b have a Kraus OSR, is more challenging and requires
a tool known as the Choi decomposition [4].
From now on we define a quantum map (or quantum channel) as a map that is (1) trace preserving, (2) linear, (3) completely
positive. This definition is motivated by the fact that we know that such maps have a Kraus OSR, and that the Kraus OSR arises
both from the physical prescription of unitary evolution followed by partial trace, and from (non-selective) measurements.

D. Positive but not Completely Positive: Transpose

Do maps that are positive but not completely positive exist? The answer is affirmative. The canonical example is the elemen-
tary transpose map T .
Given a real basis {∣i⟩} for HS , the action of the transpose on the basis elements is: T (∣i⟩⟨j∣) = ∣j⟩⟨i∣ (for a real basis this is
the same as Hermitian conjugation). For example, for a 2 × 2 matrix:
a b a c
T ∶( ) = a∣0⟩⟨0∣ + b∣0⟩⟨1∣ + c∣1⟩⟨0∣ + d∣1⟩⟨1∣ z→ a∣0⟩⟨0∣ + b∣1⟩⟨0∣ + c∣0⟩⟨1∣ + d∣1⟩⟨1∣ = ( ). (126)
c d b d
Claim 1. T is a positive map.
Proof. To prove the claim it suffices to show that the eigenvalues of X and T (X) are the same for any X ∈ HS [since then
in particular their sign is preserved, so if X > 0 then also T (X) > 0]. The eigenvalues of X are found by solving for the
roots of its characteristic polynomial: p(X) = det(X − λI). Now, since the determinant is invariant under elementary row and
column operations, it is invariant under transposition. Therefore det(X − λI) = det[T (X − λI)] = det[T (X) − λT (I)] =
det[T (X) − λI], i.e., p(X) = p[T (X)].
Is T also completely positive? To test this we need to check if any extension T p ≡ T ⊗ IR of T is also positive. This extension
is called the partial transpose, and its action on any basis element of B(HS ⊗ HR ) is as follows:
T p (∣i⟩⟨j∣ ⊗ ∣µ⟩⟨ν∣) = ∣j⟩⟨i∣ ⊗ ∣µ⟩⟨ν∣ . (127)
To prove that T is not a CP map, it suffices to find a counterexample. Indeed, consider the pure state ρ = ∣ψ⟩⟨ψ∣, where
∣ψ⟩ = √12 (∣0⟩S ∣0⟩R + ∣1⟩S ∣1⟩R ). Then:
1
T p (ρ) = (T ⊗ I)[∣0S 0R ⟩⟨0S 0R ∣ + ∣00⟩⟨11∣ + ∣11⟩⟨00∣ + ∣11⟩⟨11∣] (128a)
2
⎛1 0 0 0 ⎞
1 1⎜0 0 1 0 ⎟
= (∣00⟩⟨00∣ + ∣10⟩⟨01∣ + ∣01⟩⟨10∣ + ∣11⟩⟨11∣) = ⎜ ⎟.
2⎜ ⎟ (128b)
2 ⎜0 1 0 0 ⎟
⎝0 0 0 1 ⎠

The eigenvalues of this matrix are ( 12 , 12 , 12 , − 12 ), and the existence of a negative eigenvalue shows that T is not a CP map, since
T p >/ 0. Therefore T does not have a Kraus OSR, and is not a quantum map. Note furthermore that this means that a maximally
entangled two-qubit state has a negative partial transpose. This observation motivates the study of the partial transpose as a tool
for entanglement testing.
 27

E. Partial Transpose as a Test for Separability/Entanglement: the PPT criterion

Consider a separable (thus by definition unentangled) state ρ = ∑i pi ρAi ⊗ ρi , where the pi are probabilities and the ρi and
B A

ρB
i are quantum states (positive, normalized). The state ρ obviously arises from the mixed state ensemble {ρi ⊗ ρi , pi }, in
A B

which every element is a tensor product state. Mixing such states classically does not generate any entanglement between A and
B, hence the definition.
Applying the partial transpose yields:

T p (ρ) = (T ⊗ I)(ρ) = ∑ pi T (ρA

i ) ⊗ ρi = ∑ pi σi ⊗ ρi .
B A B
(129)
i i

Since the transpose does not change the eigenvalues, σiA ≡ T (ρA i ) is also a valid quantum state, and hence T (ρ) is another
p

separable quantum state. In particular, this shows that every separable state has a positive partial transpose (PPT). In other
words, separability implies PPT. Conversely, a negative partial transpose (NPT) implies entanglement. This means that PPT is a
necessary condition for separability.
Is PPT also sufficient for separability? It turns out that this is the case only for the 2 × 2 (two qubits) or 2 × 3 (qubit and qutrit)
cases. I.e., only in these cases a state is separable iff it has a positive partial transpose (PPT) (conversely, is entangled iff it has a
NPT) [5, 6]. Indeed, we saw in the previous subsection that a (maximally) entangled state has NPT.
In higher dimensions the PPT criterion it is still necessary but no longer sufficient. In such higher dimensions there are
examples of so-called “bound-entangled” states that have PPT but are not separable [7].
As an example of the use of the PPT criterion consider the Werner states:

I
ρ = p∣Ψ− ⟩⟨Ψ− ∣ + (1 − p) (130)
4
√
where ∣Ψ− ⟩ is a maximally entangled singlet state: ∣Ψ− ⟩ = (∣01⟩ − ∣10⟩)/ 2. This represents a family of quantum states
parametrized by the probability p of being in the singlet state as opposed to the maximally mixed state.
Its density matrix in the standard basis is

⎛1 − p 0 0 0 ⎞
1⎜ 0 p + 1 −2p 0 ⎟
ρ= ⎜ ⎟,
4⎜ ⎟ (131)
⎜ 0 −2p p + 1 0 ⎟
⎝ 0 0 0 1 − p⎠

and the partial transpose

⎛1 − p 0 0 −2p ⎞
1 ⎜ 0 p+1 0 0 ⎟
T p (ρ) = ⎜⎜
⎟.
⎟ (132)
4⎜ 0 0 p+1 0 ⎟
⎝ −2p 0 0 1 − p⎠

The eigenvalues of this matrix are (1−3p)/4 and (threefold) (1+p)/4. Therefore, the state is entangled for p > 1/3 and separable
for p ≤ 1/3 (for p = 1/3 all eigenvalues are non-negative so PPT).

F. Kraus OSR as a composition of CP maps

The Kraus OSR is a actually a composition of three other maps:

Φ = TrB ○ U ○ A, (133)

where (i) A is the “assignment map” which associates to every initial system state ρS (0) a fixed bath state ρB (0), i.e.,
A[ρS (0)] = ρS (0) ⊗ ρB (0); (ii) U is the unitary evolution superoperator, i.e., U[X] = U XU † ; (iii) TrB is the usual par-
tial trace operator. This is depicted in Fig. 2.
Let us show that each of these three maps is, in turn, CP.
 28

ρS (0) ⊗ ρB (0) U [ρS (0) ⊗ ρB (0)] U †

U

A TrB

Φ
ρS (0) ρS (t)

FIG. 2. A commutative diagram showing that the quantum map Φ can be viewed as a composition of three maps.

1. The assignment map is CP

The map A is from D(HS ) to D(HS ⊗ HB ). To prove that it is CP we need to consider positive operators A ∈ D(HS ⊗ HR ).
Thus, writing A = ∑ir λir ∣i⟩S ⟨i∣ ⊗ ∣r⟩R ⟨r∣ and ρB = ∑µ λµ ∣µ⟩B ⟨µ∣, with λir , λµ ≥ 0:

⟨v∣ (A ⊗ IR )(A) ∣v⟩ = ∑ λir ⟨v∣ [∣i⟩S ⟨i∣ ⊗ ρB ⊗ ∣r⟩R ⟨r∣] ∣v⟩ (134a)
ir
= ∑ λir λµ ⟨v∣ [∣i⟩S ⟨i∣ ⊗ ∣µ⟩B ⟨µ∣ ⊗ ∣r⟩R ⟨r∣] ∣v⟩ (134b)
irµ

= ∑ λir λµ ∣⟨v∣iµr⟩∣2 ≥ 0. (134c)

irµ

2. The unitary map is CP

This is obvious since the unitary map is a special case of a Kraus OSR having such a single Kraus operator U .

3. The partial trace is CP

To demonstrate that the partial trace, TrB ∶ ρSB ↦ ρS , is CP, we can perform a direct calculation like we did for the assignment
map. However, instead we can also directly demonstrate that it has a Kraus OSR (since this is a sufficient condition for CPness).
Consider the following explicit Kraus operators for the partial trace map:

Kα = IS ⊗ ⟨α∣ , (135)

where {∣α⟩} denotes the elements of some chosen basis for the bath Hilbert space. This choice is motivated by the fact that the
partial trace leaves the system alone but sandwiches the bath between basis states.
Applying the map Φ = {Kα } to an arbitrary system-bath state ρSB = ∑ijµν λijµν ∣i⟩⟨j∣ ⊗ ∣µ⟩⟨ν∣ written in the same basis for
the bath, and noting that TrB (ρSB ) = ∑ijµνα λijµν ∣i⟩⟨j∣⟨α∣µ⟩⟨ν∣α⟩ = ∑ijα λijαα ∣i⟩⟨j∣, we find the following:

⎛ ⎞
Φ(ρSB ) = ∑ Kα ρSB Kα† = ∑ IS ⊗ ⟨α∣ ∑ λijµν ∣i⟩⟨j∣ ⊗ ∣µ⟩⟨ν∣ IS ⊗ ∣α⟩ = ∑ λijαα ∣i⟩⟨j∣ = TrB (ρSB ) , (136)
α α ⎝ijµν ⎠ ijα

as desired. Thus, the partial trace has Kraus elements as given in Eq. (135), and is CP.

G. OSR for a general initial condition?

What would happen if we were to relax the initial condition? Will we still get a CP map?

1. General initial states

Using a general orthonormal basis for the joint Hilbert space we can always write

ρ(0) = ∑ λijαβ ∣i⟩⟨j∣ ⊗ ∣α⟩⟨β∣. (137)

ijαβ
 29

The corresponding initial state of the system is

ρS (0) = TrB [ρ(0)] = ∑ λijαα ∣i⟩⟨j∣ . (138)
ijα

If we go through the same steps as in the derivation of the Kraus OSR, we have, with {∣µ⟩} now representing the same bath
basis as {∣α⟩}:
ρS (t) = ∑ ⟨µ∣U (t)ρ(0)U † (t)∣µ⟩ (139a)
µ

= ∑ λijαβ ⟨µ∣U (t)∣i⟩⟨j∣ ⊗ ∣α⟩⟨β∣U † (t)∣µ⟩ (139b)

µijαβ

= ∑ ⟨µ∣ U (t) ∣α⟩ λijαβ ∣i⟩⟨j∣ ⟨β∣U † (t)∣µ⟩ (139c)

µijαβ

⎛ ⎞ ⎛ ⎞
= ∑ ⟨µ∣ U (t) ∣α⟩ ∑ λijαα ∣i⟩⟨j∣ ⟨α∣U † (t)∣µ⟩ + ∑ ⟨µ∣ U (t) ∣α⟩ ∑ λijαβ ∣i⟩⟨j∣ ⟨β∣U † (t)∣µ⟩ . (139d)
µ,α ⎝ ij ⎠ µij,α≠β ⎝ ij ⎠
The first summand in Eq. (139d) has Kraus operators ⟨µ∣ U (t) ∣α⟩ and may look fine. However, because of the sum over α we
cannot factor out ρS (0). Moreove, the second summand in Eq. (139d) in addition involves off-diagonal terms λijαβ that do not
appear in ρS (0) [Eq. (138)]. Clearly, we cannot factor out ρS (0), so we do not even get a map from ρS (0) to ρS (t).

2. Separable states

What if we consider separable states,

ρ(0) = ∑ pi ρiS ⊗ ρiB (140)
i

where ρiS and ρiB are themselves states of the system and bath? For such a state the initial system state is ρS (0) = TrB [ρ(0)] =
∑i pi ρiS . Let’s decompose each bath state as
ρiB = ∑ λνi ∣νi ⟩⟨νi ∣ (141)
νi

and try again:

ρS (t) = ∑ ∑ ∑ pi λνi ⟨µ∣U (t)ρiS ⊗ ∣νi ⟩⟨νi ∣U † (t)∣µ⟩ (142a)
µ i νi

= ∑ ∑ ∑ λνi ⟨µ∣U (t) ∣νi ⟩ pi ρiS ⟨νi ∣ U † (t)∣µ⟩ . (142b)

µ i νi

We can move the sum over i inside if we first assume that all ρiB commute, i.e., are diagonal in the same basis so that νi = ν ∀i,
for then ρiB (0) = ∑ν λiν ∣ν⟩⟨ν∣ and hence
ρS (t) = ∑ ⟨µ∣U (t) ∣ν⟩ ∑ λiν pi ρiS ⟨ν∣ U † (t)∣µ⟩ , (143)
µ,ν i

but this still doesn’t allow us to extract the initial system state ∑i pi ρiS . To accomplish this we may moreover assume that
eigenvalues are the same, i.e., λiν = λν ∀i. If we do so we find ρS (t) = ∑µ,ν λν ⟨µ∣U (t) ∣ν⟩ ∑i pi ρiS ⟨ν∣ U † (t)∣µ⟩, and this
involves a map acting on ρS (0) = ∑i pi ρiS as desired, but we haven’t gained anything: this is the case if ρiB = ρB ∀i, i.e., we’re
back to Eq. (109) again.

H. The quantum discord perspective

1. Quantum Discord

In classical information theory there are two equivalent ways to define the mutual information between two random variables
X and Y :
I(Y ∶ X) = H(Y ) + H(X) − H(X, Y ) (144a)
J(Y ∶ X) = H(Y ) − H(Y ∣X) , (144b)
 30

where H(X) = − ∑i pi log(pi ) is the Shannon entropy associated with X, with pi = Pr(xi ) being the probability of X assuming
the value xi . The quantity H(X, Y ) is the entropy of the joint distribution, and H(Y ∣X) is the entropy of Y conditioned on
X. The equivalence follows directly from Bayes’ rule [the joint probability satisfies p(y, x) = p(y∣x)p(x), where p(y∣x) is the
conditional probability], which implies that H(X, Y ) = H(Y ∣X) + H(X), and hence that I(Y ∶ X) = J(Y ∶ X).
In the quantum case, measuring system X generally affects system Y if the joint state ρXY is correlated, so the asymmetry
inherent in the second expression J(Y ∣X) means that there is the potential for a different outcome from the symmetric first
expression I(Y ∣X). This observation forms the basis for the definition of the quantum discord, IQ (Y ∶ X) − JQ (Y ∶ X). Let
us thus define the quantum mutual information expressions IQ (Y ∶ X) and JQ (Y ∶ X).
First, we need the quantum von Neumann entropy associated with a state ρ:

S(ρ) = −Tr[ρ log(ρ)] . (145)

Then

IQ (Y ∶ X) = S(ρY ) + S(ρX ) − S(ρXY ) , (146)

where ρXY is the total state of systems X and Y , ρY = TrX ρXY , and ρX = TrY ρXY . The second mutual information JQ
arises from first measuring X. Assume that this is done using a projective measurement with projectors {Πi }, acting only on
X. Then the post-measurement state obtained in case i is ρY ∣Πi ≡ Πi ρXY Πi /pi , where pi = Tr[Πi ρXY ] is the probability of
case i. Let us associate an entropy to this state: S(ρY ∣Πi ). The entropy conditioned non-selectively on the entire measurement is
S(Y ∣{Πi }) = ∑i pi S(ρY ∣Πi ), and the conditional entropy is the minimum over all possible measurements, since we’re interested
in maximizing the mutual information: S(Y ∣X) = min{Πi } S(Y ∣{Πi }). Explicitly:

S(Y ∣X) = min ∑ pi S(Πi ρXY Πi /pi ) , pi = Tr[Πi ρXY ] . (147)

{Πi } i

With this, we are ready to define the second quantum mutual information:

JQ (Y ∶ X) = S(Y ) − S(Y ∣X) . (148)

Generally, JQ (Y ∶ X) ≠ IQ (Y ∶ X). We thus define the quantum discord [8] as

D(ρXY ) = IQ (Y ∶ X) − JQ (Y ∶ X) . (149)

D(ρXY ) = 0 only for zero-discord states (by definition), which are states that have no quantum correlations at all. Note
that separable states can have non-zero discord [i.e., JQ (Y ∶ X) ≠ IQ (Y ∶ X)], which means that they have some quantum
correlations despite being a convex combination of product states. However, it is not hard to show that a special class of
separable states does have zero discord. Such states are known as zero-discord states, and they are of the form

ρSB (0) = ∑ pi Πi ⊗ ρiB , (150)

i

where the Πi are projectors, i.e., Πi Πj = δij Πi . This initial state would be the result of a non-selective projective measurement
of the system with measurement operators {Πi } (you can easily check that the state is invariant under a non-selective projective
measurement with the same set of measurement operators, which is the property we expect from the state after a first projective
measurement; see subsection VI A for non-selective measurements).

2. Zero discord initial states and CP maps

It turns out that zero-discord states do allow us to generalize the assumption of a factorizable initial state [Eq. (109)] in the
derivation of the Kraus OSR [9]. Let’s assume that the initial state is of the form given in Eq. (150). Thus the system state
becomes

ρS (t) = ∑ ⟨µ∣ U (t) ∑ pi Πi ⊗ ρiB U † (t) ∣µ⟩ (151a)

µ i
√ √
= ∑ ⟨µ∣ U (t) ρiB pi Πi ⊗ ρiB U † (t) ∣µ⟩ (151b)
µ,i
√ √
∑ ⟨µ∣ U (t) ρiB ∣ν⟩ pi Πi ⊗ ⟨ν∣ ρiB U † (t) ∣µ⟩ , (151c)
µν,i
 31

where we used the fact that ρiB is a positive operator to take its square root, and inserted a bath identity operator ∑ν ∣ν⟩⟨ν∣ in the
last line. Let’s define
√
Diµν ≡ ⟨µ∣ U (t) ρiB ∣ν⟩ , (152)

and note that this is a system-only operator. Now, we can always write Diµν = ∑m Dmµν δim . Inserting this into the last equation
we have

ρS (t) = ∑ pi (∑ Dmµν δim ) Πi (∑ Dnµν

†
δin ) (153a)
µν,i m n

= ∑ pi (∑ Dmµν δim Πi ) (∑ Πi δin Dnµν

†
), (153b)
µν,i m n

where we used Π2i = Πi . Next, note that δim Πi = Πm Πi and δin Πi = Πi Πn , which allows us to replace the δ’s by Π’s:

ρS (t) = ∑
†
pi Dmµν Πm Πi Πn Dnµν . (154)
µν,i,m,n

We can now move the sum over i inside so it is performed first. Thus, we have

ρS (t) = ∑ Dmµν Πm (∑ pi Πi ) Πn Dnµν

†
, (155)
µν,m,n i

and using Eq. (150) we recognize the middle term as the initial system state: ρS (0) = ∑i pi Πi . We can also define new Kraus
operators as Kµν = ∑m Dmµν Πm . This then gives us a proper Kraus OSR:

ρS (t) = ∑ Kµν ρS (0)Kµν

†
. (156)
µν

It turns out that there are also discordant states that give rise to CP maps, and even entangled states. Read about generalizations
in Refs. [10–12].

I. Equivalence of Quantum Maps

Given two quantum maps, a natural question is under which conditions they are equivalent. As an example, consider the
two single-qubit quantum maps defined by the following two sets of Kraus operators: Φ = {K0 = √12 I, K1 = √12 σ z } and
Ψ = {L0 = ∣0⟩⟨0∣, L1 = ∣1⟩⟨1∣}. Note that Φ can be interpreted as the map the flips the phase or leaves the state alone with equal
probability, while Ψ can be interpreted as a non-selective measurement in the σ z basis. Thus, a priori it seems that the two maps
describe very different physical processes. Nevertheless, it’s easy to show that the two maps are identical,3 i.e., ∀ρ
I I σz σz
√ ρ √ + √ ρ √ = ∣0⟩⟨0∣ρ∣0⟩⟨0∣ + ∣1⟩⟨1∣ρ∣1⟩⟨1∣. (157)
2 2 2 2

1. General conditions for equivalence

What is the general condition such that two maps are equivalent? The following theorem provides the answer:

Theorem 2. Consider the maps produced by the following two sets of Kraus operators Φ = {Kα }, Ψ = {Lβ }: ρ′ = ∑ Kα ρKα†
and ρ′′ = ∑ Lβ ρL†β . Then

∀ρ ∶ ρ′ = ρ′′ ⇐⇒ ∃ a unitary operator, u ∶ s.t. Kα = ∑ uαβ Lβ . (158)

β

3 Simply write ρ as a general 2 × 2 matrix and note that ZρZ flips the sign of the off-diagonal elements, so that both Φ and Ψ erase ρ’s off-diagonal elements.
 32

Proof. Here we prove the “if” direction, i.e., assume that such a unitary exists; then

ρ′ = ∑ (∑ uαβ Lβ )ρ(∑ u∗αβ ′ L†β ′ ) (159a)

α β β′

= ∑ Lβ ρL†β ′ ∑[u† ]β ′ α [u]αβ = ∑ Lβ ρL†β ′ [u† u]β ′ β (159b)

ββ ′ α ββ ′

= ∑ Lβ ρL†β ′ δββ ′ (159c)

ββ ′

= ∑ Lβ ρL†β = ρ′′ . (159d)

β

In the example above the relation between the operators is:

1 1
K0 = √ (L0 + L1 ) , K1 = √ (L0 − L1 ), (160)
2 2

1 1
so the unitary is u = √1 ( ).
2 1 −1

2. Physical origin of the equivalence

Where does this unitary equivalence between Kraus operators come from? To see this intuitively, note that in deriving the
Kraus operators, after evolving with a unitary operator acting on both the system and the bath, we trace out the bath, so the Kraus
operators should remain equivalent under the change of basis of the bath. Let us show that this “gauge freedom” gives rise to the
unitary equivalence between different sets of Kraus operators. As we shall see, we need to be a bit careful in accounting for the
presence of the square-root of the eigenvalue of the bath density matrix in the definition of the Kraus operators.
Let us write Eq. (158) as

Kµν (t) = ∑ uµνηξ Lηξ (t), (161)

ηξ

where we have let α = (µν) and β = (ηξ). In terms of the explicit form of the Kraus operators this becomes
√ √
λν ⟨µ∣ U ∣ν⟩ = ∑ uµνηξ λξ ⟨η∣ U ∣ξ⟩ . (162)
ηξ

Let us now assume that

uµνηξ ≡ vµη wξν = ⟨µ∣ v ∣η⟩ ⟨ξ∣ w ∣ν⟩ , (163)

where v and w are both unitary. We can then show that the matrix u is unitary:

[u† u]αβ = ∑[u† ]αγ [u]γβ = ∑ u∗γα uγβ = ∑ u∗µ′ ν ′ µν uµ′ ν ′ ηξ = ∑ (vµ′ µ wνν ′ )∗ vµ′ η wξν ′ (164a)
γ γ µ′ ν ′ µ′ ν ′

= ∑ vµ∗′ µ vµ′ η ∑ wνν

∗
′ wξν ′ = δµη δξν = δαβ , (164b)
µ′ ν′

where we used the unitarity of v and w in the penultimate equality.

Plugging this expression for uµνηξ into Eq. (162) gives:
√ √
⟨µ∣ U λν ∣ν⟩ = ∑ λξ ⟨µ∣ v∣η⟩⟨η∣U ∣ξ⟩⟨ξ∣w ∣ν⟩ (165a)
ηξ
√
= ⟨µ∣ v[∑ ∣η⟩⟨η∣]U [∑ λξ ∣ξ⟩⟨ξ∣]w ∣ν⟩ (165b)
η ξ
√
= (⟨µ∣ v)U ( ρB w ∣ν⟩), (165c)
 33

i.e., the gauge freedom giving rise to the unitary equivalence between to sets of Kraus operators is:

⟨µ∣ ↦ ⟨µ∣ v (166a)

√ √
λν ∣ν⟩ ↦ ρB w ∣ν⟩ . (166b)

Eq. (166a) simply expresses the freedom to apply a unitary transformation on the bath basis vectors before taking the partial
trace (which we did by sandwiching inside ⟨µ∣ ⋯ ∣µ⟩). Eq. (166b) tells us that we can also apply a second unitary
√ transformation
√
on the eigenstates of ρB (i.e., ∣ν⟩ ↦ w ∣ν⟩), but that in general we should also replace the eigenvalue term λν by ρB . To
understand the latter, note that in deriving the Kraus OSR we can also proceed as follows:

√ √
ρ′S = TrB [U ρS ⊗ ρB U † ] = ∑ ⟨µ∣ [U ρS ⊗ ( ρB ∑ ∣ν⟩⟨ν∣ ρB ) U † ] ∣µ⟩ (167a)
µ ν
√ √
= ∑ (⟨µ∣ U ρB ∣ν⟩) ρS (⟨ν∣ ρB U † ∣µ⟩) , (167b)
µν

√
which means that the Kraus operators we derived originally by using ρB ’s spectral decomposition, ⟨µ∣ U λν ∣ν⟩, are equivalent
√
to Kraus operators of the form ⟨µ∣ U ρB ∣ν⟩. In other words, the spectral decomposition was just one of infinitely many
equivalent ways to decompose ρB . We recover the spectral decomposition if we choose the basis {∣ν⟩} in Eq. (167) as the
eigenbasis of ρB .

VII. QUANTUM MAPS OF A QUBIT

In this section, by focusing on the case of one qubit, we will develop a geometric picture of the action of quantum maps. The
main tool that will allow us to do this is the Bloch sphere representation.
Recall that the density matrix of a qubit may be written as ρ = 12 (I + v⃗ ⋅ σ
⃗ ) where σ
⃗ = (σx , σy , σz ) and v⃗ = (vx , vy , vz ) ∈ R3 is
the Bloch vector. In this way, a single-qubit state may be thought of as a point in or on the unit sphere in R3 —the Bloch sphere.
States with ∥⃗v ∥ = 1 lie on the surface of the sphere and correspond to pure states of the form ρ = ∣ψ⟩⟨ψ∣. Points on the interior of
the sphere correspond to mixed states with purity P = Tr[ρ2 ] < 1.

A. Transformation of the Bloch Vector

What happens when a quantum map acts on a single qubit? As a map of the density matrix, Φ ∶ ρ ↦ ρ′ . At the same time
ρ must be expressible in terms of a new Bloch vector v⃗′ , where ρ′ = 12 (I + v⃗′ ⋅ σ
′
⃗ ). We shall show that ρ ↦ ρ′ is equivalent to
mapping the Bloch vector

v⃗ ↦ v⃗′ = M v⃗ + c⃗ (168)

for some real 3 × 3 matrix M and a vector c⃗ ∈ R3 . This is an affine transformation. Before proving Eq. (168), let us decompose
M in a way that will reveal more of the geometric aspects of the transformation.
√ the polar decomposition, which allows us to write any square matrix A as A = U ∣A∣, where U is a unitary matrix and
Recall
∣A∣ ≡ A† A is Hermitian (since clearly its eigenvalues are real), a generalization of the polar representation of a complex number
z = eiθ ∣z∣. If A is a real matrix, U becomes real-unitary, i.e., orthogonal, and ∣A∣ becomes real-Hermitian,
√ i.e., symmetric. So,
for our 3 × 3 real matrix M we can write M = OS, for orthogonal O and symmetric S = M † M . S causes deformation by
scaling along the directions of the eigenvectors by a factor of the corresponding eigenvalues. O is a rotation matrix. Now we
may interpret the action of a quantum map on a qubit state as mapping the Bloch vector according to

v⃗ ↦ v⃗′ = OS⃗
v + c⃗, (169)

as a shift by c⃗, a deformation by S and a rotation by O. Because the Bloch sphere represents the set of possible Bloch vectors,
we may view the Kraus map acting on a qubit as a transformation of the Bloch sphere that displaces its center by c⃗ and turns the
sphere into an angled ellipsoid.
To prove Eq. (168), we plug the Bloch vector representation of ρ into the quantum map:

1 1
ρ′ = ∑ Kα ρKα† = ∑ Kα (I + v⃗ ⋅ σ
⃗ )Kα† = (∑ Kα Kα† + ∑ vj Kα σj Kα† ) . (170)
α 2 α 2 α αj
 34

To isolate the components of v⃗′ we multiply both sides by σi and take the trace, while remembering that the Pauli matrices are
all traceless and satisfy Eq. (A22). Thus, Eq. (170) becomes

1
Tr(ρ′ σi ) = [∑ Tr(Kα Kα† σi ) + ∑ vj Tr(Kα σj Kα† σi )] , (171)
2 α αj

On the other hand, using ρ′ = 21 (I + v⃗′ ⋅ σ

⃗ ) and Eq. (A22) again:

1
Tr(ρ′ σi ) = [Tr(σi ) + ∑ vj′ Tr(σj σi )] = 0 + vi′ . (172)
2 j

Equating Eqs. (171) and (172) we thus have

vi′ = ci + ∑ Mij vj , (173)

j

where
1
Mij = ∑ Tr(σi Kα σj Kα )
†
(174a)
2 α
1
ci = ∑ Tr(σi Kα Kα† ) . (174b)
2 α

Moreover, using the Hermiticity of the Pauli matrices and properties of the trace [Eq. (A19)]:

∗ 1 1 1
Mij = ∑ Tr(σi Kα σj Kα ) = ∑ Tr(Kα σj Kα σi ) = ∑ Tr(σi Kα σj Kα ) = Mij ,
† † † †
(175)
2 α 2 α 2 α

i.e., M is real. Likewise,

1 1 1
c∗i = ∑ Tr(σi Kα Kα ) = ∑ Tr(Kα Kα σi ) = ∑ Tr(σi Kα Kα ) = ci ,
† † † †
(176)
2 α 2 α 2 α

so c⃗ ∈ R3 . This proves Eq. (168).

B. Unital Quantum Maps

Returning temporarily to the general (beyond a single qubit) case, a quantum map is said to be unital if it maps the identity
operator to itself, i.e.:

Definition 1. Φ is unital if Φ(I) = I. Otherwise it is non-unital.

Since a quantum map always has a Kraus OSR, we find that unital quantum maps satisfy

∑ Kα Kα = I ,
†
(177)
α

in addition to the trace-preservation constraint ∑α Kα† Kα = I.

Note that if Φ is unital, so that Eq. (177) holds, then

1 1
ci = Tr(σi ∑ Kα Kα† ) = Tr(σi ) = 0 (unital case) . (178)
2 α 2

⃗
Conversely, if Φ is non-unital, then c⃗ ≠ 0.
Note that, as is clear from Eq. (172), M is associated purely with the transformation of v⃗⋅⃗
σ under the map, while c⃗ is associated
purely with the transformation of I under the map. This observation will help us read off M and c⃗ in the examples we study
below.
 35

C. The Phase Damping Map

The phase damping map is:

Φ(ρ′ ) = pρ + (1 − p)ZρZ , (179)

√ √
where Z ≡ σz , so the Kraus operators are K0 = pI and K1 = 1 − pZ. This map can be understood as
⎧
⎪
⎪ρ w/ prob. p
ρ ↦ ρ′ = ⎨ (180)
⎪
⎪ w/ prob. 1 − p
⎩ZρZ
Using our general result, Eq. (174) we have in this case:
1 1
ci = ∑ Tr(σi Kα Kα ) = [pTr(σi ) + (1 − p)Tr(σi )] = 0
†
(181)
2 α 2

[in agreement with the fact that the phase damping map is unital; recall Eq. (178)], and:
1 1 1
Mij = ∑ Tr(σi Kα σj Kα ) = [pTr(σi σj ) + (1 − p)Tr(σi Zσj Z)] = pδij + (1 − p)Jij ,
†
(182)
2 α 2 2

where Jij ≡ Tr(σi Zσj Z). Written explicitly the matrix J is:

⎛ Tr(XZXZ) Tr(XZY Z) Tr(XZZZ) ⎞ ⎛ Tr(−I) Tr(σ) Tr(σ) ⎞

J =⎜ ⎟ ⎜ ⎟
⎜ Tr(Y ZXZ) Tr(Y ZY Z) Tr(Y ZZZ) ⎟ = ⎜ Tr(σ) Tr(−I) Tr(σ) ⎟ = diag(−2, −2, 2) , (183)
⎝ Tr(ZZXZ) Tr(ZZY Z) Tr(ZZZZ) ⎠ ⎝ Tr(σ) Tr(σ) Tr(I) ⎠

where σ denotes a Pauli matrix. Thus,

⎛2p − 1 0 0⎞
⎜
M = diag[p − (1 − p), p − (1 − p), p + (1 − p)] = ⎜ 0 2p − 1 0⎟
⎟, (184)
⎝ 0 0 1⎠

and

v⃗′ = M v⃗ = [(2p − 1)vx , (2p − 1)vy , vz ] .

t
(185)

The corresponding transformation of the Bloch sphere is shown in Fig. 3. There is no shift of the Bloch sphere, while there is a
rescaling along the vx and vy directions by a factor of (2p − 1), and all points on the vz axis are fixed. The map has two fixed
pure states, the north and south poles of the Bloch sphere, ∣0⟩⟨0∣ and ∣1⟩⟨1∣. For p = 1, the Bloch sphere remains unchanged.
Because p is a probability, −1 ≤ 2p − 1 ≤ 1. Hence the scaling factor can take negative values,
√ corresponding to a rotation by π
about the vz axis. To see why, let us use the polar decomposition to write M = OS, where S = M † M = diag(∣2p−1∣, ∣2p−1∣, 1).
Therefore the rotation matrix must be O = diag(sign(2p − 1), sign(2p − 1), 1) = (±1, ±1, 1). When 2p − 1 < 0, O is a rotation by
π about the vz axis.
The purity [Eq. (64)] of the transformed state is
1 1
P ′ = Tr[(ρ′ )2 ] = (1 + ∥⃗
v ′ ∥2 ) = [1 + (2p − 1)2 (vx2 + vy2 ) + vz2 ] ≤ P . (186)
2 2
Thus the purity always decreases under the phase damping channel, except for the states on the vz axis (with vx = vy = 0), whose
purity is invariant.

D. The Bit Flip Map

The bit flip map is:

⎧
⎪
⎪ρ w/ prob. p
ρ ↦ ρ′ = ⎨ (187)
⎪
⎪ w/ prob. 1 − p
⎩XρX
 36

0.5

pauli Z 0

−0.5

−1
1
0.5 1
0 0.5
0
−0.5
−0.5
pauli Y −1 −1
pauli X

FIG. 3. The Bloch sphere become an ellipsoid after transformation by the phase damping channel. The invariant states are those on the σz
axis. The major axis has length 2, the minor axis has length 2(2p − 1).

In the computational basis, the bit flip map acts like a classical error channel, flipping bits at random. The phase damping map is
purely quantum in the same basis, since of course the notion of a phase is not classical. However, mathematically the two maps
are essentially identical. We can guess that since the phase flip map leaves the vz axis alone and shrinks the Bloch sphere in the
(vx , vy ) plane, the bit flip map will leave vx axis alone and shrinks the Bloch sphere in the (vy , vz ) plane. To confirm this, let
us use a more direct approach than the one we used for the phase flip map.
Using ρ = 21 (I + v⃗ ⋅ σ
⃗ ), we have:
1
ρ ↦ ρ′ = pρ + (1 − p)XρX = (I + p⃗ ⃗ + (1 − p)X v⃗ ⋅ σ
v⋅σ ⃗ X) . (188)
2
The key point is now that
X(⃗ ⃗ )X = X(vx X + vy Y + vz Z)X = vx X − vy Y − vz Z .
v⋅σ (189)
This shows that vx is unchanged, but the sign of both vy and vz is flipped. Had we studied the phase damping map instead, we
would have seen that vz is unchanged, but the sign of both vx and vy is flipped. We now have:
1 1
ρ′ = (I + vx X + (2p − 1)vy Y + (2p − 1)vz Z) = (I + v⃗′ ⋅ σ
⃗) . (190)
2 2
Thus, we find that the bit flip channel transforms v⃗ as:
v⃗ ↦ v⃗′ = [vx , (2p − 1)vy , (2p − 1)vz ] = M v⃗ + c⃗ , (191)
where
⎛1 0 0 ⎞
⎜
M = ⎜0 2p − 1 0 ⎟ ⎟ (192a)
⎝0 0 2p − 1⎠
c⃗ = 0̄ . (192b)
 37

Geometrically, this corresponds to the exact same deformation of the Bloch sphere as depicted in Fig. 3, but with the vx and vz
axes interchanged. If we replace X with Y in Eq. (187) we have the “bit-phase flip channel,” where the roles of the vx and vy
axes is interchanged.

E. The Depolarizing Map

The depolarizing map acting on a qubit either takes the state to the maximally mixed state with probability p, or leaves the
state unchanged with probability 1 − p:
⎧
⎪
⎪1I w/ prob. p
ρ ↦ ρ′ = ⎨ 2 . (193)
⎪
⎪ w/ prob. 1 − p
⎩ρ
Thus, with probability p, all the information held in the state is erased. Equivalently,
I
ρ′ = p + (1 − p)ρ . (194)
2
Clearly, this is also a unital map. However, note that it is not in Kraus OSR form. To put it in Kraus OSR form, note that

ρ + XρX + Y ρY + ZρZ = 2I , (195)

which we can prove easily using the same idea as in Eq. (189):

Y (⃗ ⃗ )Y = −vx X + vy Y − vz Z
v⋅σ (196a)
Z(⃗ ⃗ )Z = −vx X − vy Y + vz Z ,
v⋅σ (196b)

so that

∑ σi (⃗ ⃗ )σi = 0 .
v⋅σ (197)
i

Thus we may write the map as:

1 3 p
ρ ↦ ρ′ = p (ρ + XρX + Y ρY + ZρZ) + (1 − p)ρ = (1 − p)ρ + (XρX + Y ρY + XρZ) , (198)
4 4 4
from which we see that the Kraus operators are
√ √
3 p
K0 = 1 − pI , Ki = σi for i = 1, 2, 3 . (199)
4 4
The analysis is particularly straightforward in terms of the Bloch vector:
I 1−p I 1−p 1
ρ′ = p + (I + v⃗ ⋅ σ
⃗) = + ⃗ = (I + v⃗′ ⋅ σ
v⃗ ⋅ σ ⃗) , (200)
2 2 2 2 2
which implies that v⃗′ = (1 − p)⃗
v , so that

M = (1 − p)I (201a)
c⃗ = 0̄ . (201b)

This corresponds to the Bloch sphere shrinking uniformly to a radius of 1 − p, as illustrated in Fig. 4. The only invariant state is
the fully mixed state (the origin, v⃗ = 0̄). Every other state loses purity as it becomes more mixed.

F. Amplitude Damping / Spontaneous Emission

Spontaneous emission (SE) is the process by which an atom, nucleus, etc., undergoes a transition from a higher state of energy
to a lower state of energy, thus releasing energy to the bath (relaxation). This could through the release of a photon, a phonon,
or some other elementary excitation. If the bath is at temperature T = 0, as we assume in this subsection, then the system cannot
absorb energy, so the reverse process of excitation does not occur. We shall deal with it in the next subsection.
 38

0.5

pauli Z 0

−0.5

−1
1
0.5 1
0 0.5
0
−0.5
−0.5
pauli Y −1 −1
pauli X

FIG. 4. The Bloch sphere transformed by the depolarizing channel. As p → 1, all states converge to the fully mixed state at the origin.

We consider a single qubits, with a ground state ∣0⟩ and an excited state ∣1⟩. Thus the map Φ is:
∣0⟩ ↦ ∣0⟩ with probability 1 (202a)
∣1⟩ ↦ ∣0⟩ with probability p (202b)
Let us find the Kraus operators for this process. One Kraus operator is obvious: the transition from the excited state to the ground
state is given by
√
K1 = p∣0⟩⟨1∣ . (203)
The second Kraus operator should keep the ground state in place, i.e., contains ∣0⟩⟨0∣. But this isn’t enough, since the normal-
ization condition must be satisfied, and it’s easy to check that it isn’t if these are our Kraus operators. Instead, let us add an
unspecified matrix A and find out its form from the normalization condition. Thus:

1 a
K0 = ∣0⟩⟨0∣ + A = ( ), (204)
b c

and the normalization condition K0† K0 + K1† K1 = I becomes:

1 + ∣b∣2 a + b∗ c 1 0
( ∗ ) + p∣1⟩⟨0∣0⟩⟨1∣ = ( ). (205)
a + bc∗ ∣a∣2 + ∣c∣2 0 1
On equating the upper √left entries we get b = 0, which implies from the off-diagonal entries that a = 0. Equating the bottom right
entries then yields c = 1 − p. Thus:

1 0
K0 = ( √ ). (206)
0 1−p
√
The (perhaps curious) 1 − p component expresses the fact that not observing an emission event (imagine a detector for the
emitted photons) increases the likelihood that the system is in its ground state, but we cannot know this with certainty since
 39

the emission event might yet arrive in the future. We will see this more clearly later when we discuss quantum trajectories in
Sec. XII.
We can now directly derive M and c⃗. Since ρ′ = ∑1α=0 Kα [ 21 (I + v⃗ ⋅ σ
⃗ )] Kα† , the most direct way to do this is to map I and
v⃗ ⋅ σ
⃗ via the Kraus OSR and read off M and c⃗. Starting with I, we have:

1 0 1+p 0
I ↦ K0 K0† + K1 K1† = ( ) + p∣0⟩⟨1∣1⟩⟨0∣ = ( ) = I + pZ. (207)
0 1−p 0 1−p
Thus SE is not a unital map. Since c⃗ captures the mapping of I, we see that

c⃗ = (0, 0, p) . (208)

Next, v⃗ ⋅ σ
⃗ ↦ K0 (⃗ ⃗ )K0† + K1 (⃗
v⋅σ ⃗ )K1† . It is simple to check by explicit matrix multiplication that
v⋅σ
√
K0 XK0† + K1 XK1† = 1 − pX (209a)
√
K0 Y K0† + K1 Y K1† = 1 − pY (209b)
K0 ZK0† + K1 ZK1† = (1 − p)Z (209c)

We thus arrive at the following M matrix:

√
⎛ 1−p 0 0 ⎞
⎜ √
M =⎜ 0 1−p 0 ⎟ ⎟. (210)
⎝ 0 0 1 − p⎠

The geometric meaning of the spontaneous emission map is now clear. The center (0, 0, 0) ↦ (0, 0, p), and the Bloch sphere
is compressed more along the vz -axis than along the vx and vy -axes. In other words, all points on the Bloch sphere move closer
to its north pole, which is the ground state. If p = 1 then the entire Bloch sphere is compressed to a single point, the north pole.
The latter is a fixed point of the map. To see this, note that

Φ(∣0⟩⟨0∣) = K0 ∣0⟩⟨0∣K0† + K1 ∣0⟩⟨0∣K1† = ∣0⟩⟨0∣ + 0 = ∣0⟩⟨0∣ (211)

G. Generalized (finite temperature) Amplitude Damping/Spontaneous Emission

If the qubit is able to absorb energy from the bath (since the latter is at a temperature T > 0), then the reverse process, of
excitation from the ground state to the excited state, is also possible. To account for this let us assume that the spontaneous
emission process of the previous subsection occurs with probability q, while the reverse process occurs with probability 1 − q.
Then the Kraus operators for the SE event become

√ 1 0
K0 = q( √ ) (212a)
0 1−p
√
K1 = qp∣0⟩⟨1∣ . (212b)

The Kraus operators for the reverse process are simply:

√ √
1−p 0
K2 = 1−q( ) (213a)
0 1
√
K3 = (1 − q)p∣1⟩⟨0∣ . (213b)

Thus:

1+p 0 1−p 0
I ↦ K0 K0† + K1 K1† + K2 K2† + K3 K3† = q ( ) + (1 − q) [( ) + p∣1⟩⟨0∣0⟩⟨1∣] (214a)
0 1−p 0 1
= q(I + pZ) + (1 − q)(I − pZ) = I + (2q − 1)pZ , (214b)

which shows that

c⃗ = (0, 0, (2q − 1)p) . (215)

 40

As for the M matrix it is again simple to check by explicit matrix multiplication that
3 √
∑ Ki XKi = 1 − pX
†
(216a)
i=0
3 √
∑ Ki Y Ki = 1 − pY
†
(216b)
i=0
3
∑ Ki ZKi = (1 − p)Z ,
†
(216c)
i=0

i.e., M is unchanged and is still given by Eq. (210).

Thus the only effect of allowing relaxation is to modify the center of the deformed Bloch sphere, which is now positioned at
(0, 0, (2q − 1)p). This corresponds to a new fixed point, ρeq = diag(q, 1 − q):

3
Φ(ρeq ) = ∑ Ki ρeq Ki† = ρeq . (217)
i=0

Note that the case q = 1/2 is unital (it corresponds to c⃗ = ⃗0) and has a fixed point the fully mixed state. Also note that when
q < 1/2 the new center is at (0, 0, −∣2q − 1∣p), which corresponds to a preference for the excited state rather than the ground state.

VIII. QUANTUM MAPS FROM FIRST PRINCIPLES

So far we postulated the form of certain quantum maps. Let us now consider examples where we can analytically derive the
Kraus operators from first principles.

A. A qubit coupled to a single-qubit bath

Consider a system of two qubits, such that the first qubit is the system (HS ) and the second is the bath (HB ). Consider also the
interaction Hamiltonian HSB = λσSα ⊗ σB β
where α, β ∈ {x, y, z}. The system qubit is initially in the pure state ρS (0) = ∣ψ⟩⟨ψ∣,
∣ψ⟩ = a ∣0⟩ + b ∣1⟩, written in the computational basis (eigenbasis of σ z ). The initial state of the bath is mixed:

λ0 0
ρB (0) = λ0 ∣0⟩⟨0∣ + λ1 ∣1⟩⟨1∣ = [ ], (218)
0 λ1

where λ1 = 1 − λ0 . There are 4 Kraus operators:

√ β
λ0 ⟨0∣ e−iλtσS ⊗σB ∣0⟩
α
K00 = (219a)
√ β
−iλtσS
α
⊗σB
K01 = λ1 ⟨0∣ e ∣1⟩ (219b)
√ β
−iλtσS
α
⊗σB
K10 = λ0 ⟨1∣ e ∣0⟩ (219c)
√ β
−iλtσS
α
⊗σB
K11 = λ1 ⟨1∣ e ∣1⟩ (219d)

Let θ ≡ λt. Recall now that if A2 = I then eiθA = cos θI + i sin θA (which can be easily checked by Taylor expansion). Therefore
β
eiθσS ⊗σB = cos θ ⋅ IS ⊗ IB + i sin θ ⋅ σSα ⊗ σB
α
β
, (220)

and hence (for a general µ, ν ∈ {0, 1})

√
Kµν = λν {cos θδµν ⋅ IS − i sin θ ⟨µ∣ σB
β
∣ν⟩ ⋅ σSα } . (221)

The system then evolves according to the Kraus map

ρS (t) = ∑ Kµν (t)∣ψ⟩⟨ψ∣Kµν

†
(t) . (222)
µν
 41

1. Z ⊗ X coupling

Consider first HSB = λZS ⊗ XB . In this case, we can use Eq. (221) to find
√
K00 = λ0 cos θ ⋅ I (223a)
√
K11 = λ1 cos θ ⋅ I (223b)
√
K01 = −i λ1 sin θ ⋅ σ z (223c)
√
K10 = −i λ0 sin θ ⋅ σ z (223d)

The density matrix for this map evolves under the action of these 4 Kraus operators:

ρS (t) = ∑ Kµν (t)ρS (0)Kµν

†
(t) (224a)
µν
√ 2 √ 2 √ 2 √ 2
= ( λ0 cos θ) ∣ψ⟩⟨ψ∣ + ( λ1 cos θ) ∣ψ⟩⟨ψ∣ + ( λ0 sin θ) σ z ∣ψ⟩⟨ψ∣σ z + ( λ1 sin θ) σ z ∣ψ⟩⟨ψ∣σ z (224b)
= cos θ∣ψ⟩⟨ψ∣ + sin θ ⋅ σ ∣ψ⟩⟨ψ∣σ
2 2 z z
(224c)
∣a∣2 ab∗ cos(2θ)
=[ ∗
], (224d)
a b cos(2θ) ∣b∣2

where we used the fact that λ0 + λ1 = 1.

Can we relate this result to the phase damping map discussed in Sec. VII C? This seems plausible since in both cases the
system is affected by a Z operator. In the phase damping case we have

∣a∣2 (2p − 1)ab∗

ρS (t) = Φ[ρS (0)] = p∣ψ⟩⟨ψ∣ + (1 − p)Z∣ψ⟩⟨ψ∣Z = [ ∗
], (225)
(2p − 1)a b ∣b∣2

which we would like to equate with Eq. (224d). Clearly, this requires 2p − 1 = f (θ), so that

1 + f (θ)
p= , (226)
2
and the phase damping map has as a physical origin the model given by HSB = λZS ⊗ XB .
Why did λ0 and λ1 drop out? The intuitive reason is that by having the bath qubit subject to σ x , its ∣0⟩ and ∣1⟩ state are
constantly flipped, which also interchanges λ0 and λ1 , so it is as if they are averaged to 1/2.
There is much more to say about this result, but first let us consider another case, which will turn out to subsume this one.

2. Z ⊗ Z coupling

Consider the interaction Hamiltonian H = λσSz ⊗ σB

z
. For this choice, since σ z is diagonal, only the K00 and K11 Kraus
operators are non-zero and have the form
√ √ e−iθ 0
K00 = λ0 (cos θ ⋅ IS − i sin θ ⋅ σ z ) = λ0 [ ] (227a)
0 eiθ
√ √ eiθ 0
K11 = λ1 (cos θ ⋅ IS + i sin θ ⋅ σ z ) = λ1 [ ] (227b)
0 e−iθ

Altogether, the pure state ∣ψ⟩ under each of these operators becomes

√ e−iθ 0 a √ ae−iθ
K00 ∣ψ⟩ = λ0 [ iθ
] [ ] = λ0 [ iθ ] (228a)
0 e b be
√ aeiθ
K11 ∣ψ⟩ = λ1 [ ] (228b)
be−iθ
 42

Therefore:
ae−iθ aeiθ
] [ a∗ eiθ b∗ e−iθ ] + λ1 [ −iθ ] [ a∗ e−iθ b∗ eiθ ]
ρS (t) = λ0 [
iθ
be be
⎡ ∣a∣2 ab∗ (λ0 e−2iθ + λ1 e2iθ ) ⎤⎥
⎢
=⎢ ∗ ⎥ (229)
⎢ a b (λ0 e + λ1 e
2iθ −2iθ
) ∣b∣2 ⎥
⎣ ⎦
where the diagonal elements have again been simplified with the use of the fact that λ0 + λ1 = 1. Defining
f (θ) = λ0 e−2iθ + λ1 e2iθ , (230)
yields
∣a∣2 ab∗ f (θ)
ρS (t) = [ ∗ ∗
]. (231)
a bf (θ) ∣b∣2
The previous example, HSB = Z ⊗X, is now seen to be a special case of this one, where λ0 = λ1 = 1/2 [for then f (θ) = cos(2θ)],
so everything we discuss next applies to it as well.
Note that the diagonal elements (“population”) haven’t changed under time evolution and yet the off-diagonal elements (“co-
herence”) are modulated by the periodic function f . This is like elastic scattering where no energy is exchanged and only relative
phases are impacted. More precisely, this is a dephasing process, although in our case, the phase coherence recurs periodically.
The period of f is τ = π/λ.
Consider the purity P = Tr(ρ2 ):4
⎡ 2⎤
⎢ ∣a∣2 ab∗ f (θ) ⎥⎥
P = Tr ⎢⎢( ∗ ∗ ) ⎥ (232a)
⎢ a bf (θ) ∣b∣2 ⎥
⎣ ⎦
∣a∣4 + ∣a∣2 ∣b∣2 ∣f ∣2 ...
= Tr ( ) (232b)
... ∣b∣ + ∣a∣2 ∣b∣2 ∣f ∣2
4

= ∣a∣4 + ∣b∣4 + 2∣a∣2 ∣b∣2 ∣f ∣2 . (232c)

Thus, this function is periodic with period τP = π/(2λ) since f appears squared in the expression. Since ∣f ∣ = 2
λ20 + λ21 +
2λ0 λ1 cos(4θ) we have
2
min ∣f ∣2 = λ20 + λ21 − 2λ0 λ1 = (λ0 − λ1 ) (233a)
θ
2
max ∣f ∣2 = λ20 + λ21 − 2λ0 λ1 = (λ0 + λ1 ) = 1, (233b)
θ
so that the minimum and maximum values of the purity are
min P = ∣a∣4 + ∣b∣4 + 2∣ab∣2 ∣λ0 − λ1 ∣2 (234a)
θ
max P = ∣a∣4 + ∣b∣4 + 2∣ab∣2 = (∣a∣2 + ∣b∣2 )2 = 1 . (234b)
θ

The purity achieves a minimum of 1/2 when the bath qubit is in √ a maximally mixed state, ρB = IB /2 (so that λ0 = λ1 = 1/2),
and when the system qubit is an equal superposition, ∣a∣ = ∣b∣ = 1/ 2.5
For short times t ≪ τP the purity decays quadratically. This is typical of non-Markovian decay, as we will see later (in
contrast, Markovian decay is always exponential, i.e., it starts out linearly). One might also write this inequality as a weak
coupling limit λ ≪ π/t, which suggests that in this limit the purity appears to be only decaying (i.e., there is no time for a
recurrence). However, if the coupling between the system and the bath is strong, that is λ ≫ 1, then we may not necessarily
resolve the oscillations in purity and instead measure an average purity significantly lower than 1. In both these limits an observer
would conclude that the state of the system is mixed, even though it started out pure.
Can we relate this model to the phase damping map discussed in Sec. VII C? Clearly, this requires 2p − 1 = f (θ) = f ∗ (θ).
Thus, equality only holds subject to the additional constraint that f (θ) is real. The constraint that f (θ) = f ∗ (θ) requires that
λ1 = λ2 = 1/2, i.e., the initial bath state is I/2. Therefore this Hamiltonian model is more general than the phase damping map.
This is because the bath operator ZB in the former does not affect the bath state ρB (0) = diag(λ0 , λ1 ), with which it commutes.
This keeps λ0 and λ1 in play, unlike the previous case where they were averaged out.

4 We can obtain the same result using the formula P = 21 (1+∥⃗v ∥2 ), as follows: 12 (vx −ivy ) = ab∗ f and 21 (vx +ivy ) = a∗ bf ∗ , so that vx2 +vy2 = 4∣a∣2 ∣b∣2 ∣f ∣2 .
Also, vz = ∣a∣ − ∣b∣ , and 1 = (∣a∣ + ∣b∣ ) . Adding all this up gives Eq. (232c).
2 2 2 2 2

√ note that ∣a∣ + ∣b∣ = ∣a∣ + (1 − ∣a∣ ) = 2∣a∣ − 2∣a∣ + 1 = 2x − 2x + 1 with x = ∣a∣ ; this is minimized at 4x − 2 = 0, i.e., x = 1/2, or
5 To see this 4 4 4 2 2 4 2 2 2

∣a∣ = 1/ 2.
 43

B. Irreversible open system dynamics: infinite dimensional bath with a continuous density of states

Our previous example involved a finite-dimensional bath, and we saw that the purity in this case is periodic. This reflect
reversibility, which is a general characteristic of the finite dimensional case. To exhibit irreversibility we shall investigate an
infinite-dimensional bath, but as we shall see, one additional ingredient (a continuum) will be needed as well.
Assume the system is either a qubit or a quantum harmonic oscillator (QHO). We will work out both cases. The system-bath
interaction Hamiltonian has one of the following forms:

HSB = λσSz ⊗ n̂B (235a)

HSB = λn̂S ⊗ n̂B , (235b)

where n̂S is the number operator satisfying n̂S ∣n⟩ = n ∣n⟩ for n = 0, 1, ..., ∞. The total Hamiltonian is H = HSB + HB , where
we have set HS = 0 for simplicity. We assume that the bath is itself a QHO with Hamiltonian
∞
HB = ∑ Eν ∣ν⟩⟨ν∣ , (236)
ν=0

̵ ≡ 1).
where n̂B ∣ν⟩ = ν ∣ν⟩ and n̂B is the number operator, and Eν are QHO energies: Eν = ω(ν + 12 ) (where as before we set h
We assume that the initial state of the bath is a Gibbs state:
1 −βHB 1 ∞ −βEν ∞
ρB (0) = e = ∑e ∣ν⟩⟨ν∣ ≡ ∑ λν ∣ν⟩⟨ν∣ , (237)
Z Z ν=0 ν=0

where β = kB1T and λν = Z1 e−βEν are the eigenvalues of the bath density matrix. The denominator is the partition function:
Z = Tr[e−βHB ] = ∑∞ν=0 e
−βEν
.
Using the Hamiltonians in Eq. (235) and again defining θ ≡ λt, the joint unitary evolution operator becomes

U (t) = e−itH = e−itHSB e−itHB , (238)

where we have used the fact that [HSB , HB ] = 0. Thus:

∞
σz
U (t) = exp [−iθ { } ⊗ n̂B ] exp [−itIS ⊗ ∑ Eν ∣ν⟩⟨ν∣] (239a)
n̂S ν=0
∞
σz
= ∑ exp [−iθ { } ⊗ n̂B ] exp (−itEν IS ) ⊗ ∣ν⟩⟨ν∣ (239b)
ν=0 n̂S
∞
σz
= ∑ exp (−itEν ′ IS ) exp [−iθ { } ν ′ ] ⊗ ∣ν ′ ⟩⟨ν ′ ∣ , (239c)
ν ′ =0 n̂S

where in the last equality we used n̂B ∣ν⟩ = ν ∣ν⟩.

Taking the partial matrix element with respect to the bath, we find:
∞
σz
⟨µ∣ U (t) ∣ν⟩ = ∑ exp (−itEν ′ IS ) exp [−iθ { } ν ′ ] ⟨µ∣ν ′ ⟩⟨ν ′ ∣ν⟩ (240a)
ν ′ =0 n̂S
∞
σz
= ∑ exp (−itEν ′ IS ) exp [−iθ { } ν ′ ] δµν ′ δν ′ ν (240b)
ν ′ =0 n̂S
σSz
= exp [−itEν IS ] exp [−iθ { } µ] δµν . (240c)
n̂S
√
Thus, the Kraus operators Kµν (t) = λν ⟨µ∣ U (t) ∣ν⟩ can be written as
√ σz
Kµν (t) = λν exp [−iθν { }] δµν , (241)
n̂
where we dropped the S subscripts since it is now clear that the remaining operators act only on the system, and also dropped the
term exp [−itEν I] (whose origin was HB ), since it acts as an overall phase and will drop out once we apply Kµν (t)[⋅]Kµν †
(t).
 44

Let us write the initial system density matrix as:

1 or ∞
ρS (0) = ∑ rmn ∣m⟩⟨n∣, (242)
m,n=0

where we expanded the density matrix in the eigenvectors of the σ z or n̂ operator, with the upper limits being 1 or ∞, respectively.
In the case where the system is a qubit, we have, using Eq. (112):
1
ρS (t) = ∑ rmn ∑ λν e−iθνσ ∣m⟩⟨n∣eiθνσ
z z
(243a)
m,n=0 ν
1
= ∑ rmn ∑ λν e−iθν(−1) ∣m⟩⟨n∣eiθν(−1) .
m n
(243b)
m,n=0 ν

Let us rewrite this as

1
ρS (t) = ∑ rmn ∣m⟩⟨n∣gn,m (θ), (244)
m,n=0

where
∞ n
−(−1)m ]νθ
gn,m (θ) ≡ ∑ λν ei[(−1) . (245)
ν=0

The diagonal terms gm,m = ∑∞ ν=0 λν = 1 are constant, and therefore they do not evolve in this case. Let us focus next on the case
in which both system and bath are QHO’s. We then have, using Eq. (112):

ρS (t) = ∑ rmn ∑ λν e−iθν n̂ ∣m⟩⟨n∣eiθν n̂ (246a)

m,n ν

= ∑ rmn ∑ λν e−iθνm ∣m⟩⟨n∣eiθνn . (246b)

m,n ν

Let us rewrite this as

ρS (t) = ∑ rmn ∣m⟩⟨n∣fn−m (θ), (247)
m,n

where

fx (θ) ≡ ∑ λν eixνθ . (248)

ν

Note that f0 (θ) = ∑ν λν = 1, so the state of the system at time t can be split into diagonal (population) and off-diagonal
(coherence) terms:
ρS (t) = ∑ rnn ∣n⟩⟨n∣ + ∑ rmn ∣m⟩⟨n∣fn−m (λt) (249)
n m≠n

The population term is time-independent, i.e., is the same as in ρS (0). The coherence term is time-dependent and is affected by
the coupling to the bath. Its behavior is completely determined by the modulation function f , which can be computed explicitly
by performing the geometric sum:

1 ∞ e− 2 βω ∞ −βων ixνθ
1

fx (θ) = ∑ e−βω(ν+ 2 ) eixνθ =

1
∑e e (250a)
Z ν=0 Z ν=0
e− 2 βω ∞ ν
1
−(βω−ixθ)
= ∑q , q≡e (250b)
Z ν=0
e− 2 βω 1
1

= , (250c)
Z 1−q

where convergence of the infinite series is guaranteed since ∣q∣ = e−βω < 1 due to βω > 0.
Note that fx (θ) = fx (θ + 2π/x), i.e., f is periodic, with period T (x) = 2π/(λx). Each off-diagonal element ∣m⟩⟨n∣ in
Eq. (249) thus has a different period τmn = 2π/(λ∣m − n∣). This suggests that we might have an example of irreversible
 45

decoherence [decay of the off-diagonal elements of ρS (t)], if ρS (t) isn’t periodic. But is it? Periodicity requires there to be a
time τ̃ that is simultaneously divisible by all periods τmn (i.e., all such periods fit an integer number of times into τ̃ ). Clearly,
τ̃ = 2π/λ is just such a time: τ̃ /τmn = ∣m − n∣. Thus ρS (t) is periodic after all, with a period of 2π/λ, and we do not have
irreversibility.
Note that the qubit-system case is just a special case of the QHO-system. To see this observe that g00 = g11 = ∑ν λν = 1, and
∗
g01 = g10 = ∑ν λν e2iθν = f2 (θ).
To better understand the emergence of irreversibility, we thus consider a modified model, where we introduce a mode density
Ω(ν) (a standard trick in condensed matter physics; consider, e.g., the Debye model). We thus replace the sum by an integral
over ν, and write
1 ∞
dνe−βω(ν+ 2 ) eixθν Ω(ν).
1
fx (θ) = ∫ (251)
Z 0
If Ω(ν) = ∑∞ ′
ν ′ =0 δ(ν − ν ) then we recover Eq. (250a). The modified model has the following mode density:

Ω0 if νc ≥ ν ≥ 0;
Ω(ν) = { , (252)
0 otherwise.

i.e., it has a continuous set of modes with a high-mode cutoff of νc . The cutoff is physically well-motivated: it reflects the fact
that any physical model must have a highest but finite accessible energy. Then:

Ω0 νc
e−βω(ν+ 2 ) eixθν dν
1
fx (θ) = ∫ (253a)
Z 0
Ω0 e− 2 βω
1
νc
= ∫ e−(βω−ixθ)ν dν (253b)
Z 0
Ω0 e− 2 βω e−(βω−ixθ)νc − 1
1

= (253c)
Z −βω + ixθ

The numerator is periodic just like in the previous case, so the same comments apply. However, the denominator contains a
(n − m)λt dependence (the xθ term), which shows that the coherences decay irreversibly as 1/t, with the decay being faster for
off-diagonal elements that are farther apart.
We have thus seen how an infinite-dimensional bath with a continuous mode density can result in a decay which is truly
irreversible. The decay of the off-diagonal elements is often called decoherence, since it refers to the gradual disappearance of
coherence, the name given to the off-diagonal elements. This is not an entirely satisfactory definition of decoherence, since it is
obviously basis dependent. We shall give a more careful definition later.

IX. DERIVATION OF THE LINDBLAD EQUATION FROM A SHORT TIME EXPANSION OF QUANTUM MAPS

Just as the Hamiltonian is the generator of unitary evolution, we may ask if there is a generator for open system dynamics. By
this we mean that the solution of the differential equation ρ̇ = Lρ is a quantum map, and L plays the role of a generator. In this
section we will see how to find such a generator for very short evolution times using just a short time expansion of the Kraus
OSR. We will then postulate that the same generator applies for all times (a type of Markovian approximation), and thus arrive
at a “master equation” of the form ρ̇ = Lρ that generates a quantum map. The generator L is called the Lindbladian, and the
master equation is the Lindblad equation, whose special form guarantees complete positivity (i.e., that the evolution it generates
is a quantum map).

A. Derivation

By Taylor expansion around t = 0 we have:

ρ(dt) = ρ(0) + ρ̇∣0 dt + O(dt2 ) . (254)

On the other hand, the Kraus OSR tells us that:

ρ(dt) = ∑ Kα (dt)ρ(0)Kα† (dt) . (255)

α
 46

Let’s try to find the Kraus operators that make these two equations agree up to O(dt). Clearly, to get the ρ(0) term in Eq. (254)
one of the Kraus operators must contain the identity operator. Thus, let us write

K0 = I + L0 dt , (256)

so that

K0 ρ(0)K0† = ρ(0) + [L0 ρ(0) + ρ(0)L†0 ]dt + O(dt2 ) . (257)

This contributes one term of order dt, but there must be more (since as we know a Kraus OSR with a single Kraus operator is
equivalent to unitary evolution). Thus, we can pick all other Kraus operators as
√
Kα = dtLα , α ≥ 1 , (258)

so that

Kα ρ(0)Kα† = Lα ρ(0)L†α dt . (259)

Let us now enforce the normalization condition ∑α=0 Kα† Kα = I, up to O(dt):

I = K0† K0 + ∑ Kα† Kα = I + dt (L0 + L†0 + ∑ L†α Lα ) + O(dt2 ) . (260)

α≥1 α≥1

Without loss of generality we can decompose the general operator L0 into a Hermitian and anti-Hermitian part: L0 = A − iH,
with A = A† and H = H † . Thus, Eq. (260) tells us that to O(dt):
1
A=− ∑ L Lα .
†
(261)
2 α≥1 α

Plugging all this back into the Kraus OSR, Eq. (255), we find:

ρ(dt) = K0 ρ(0)K0† + ∑ Kα ρ(0)Kα† (262a)

α≥1

= ρ(0) + (A − iH)dtρ(0) + ρ(0)(A + iH)dt + ∑ Lα ρ(0)L†α dt + O(dt2 ) (262b)

α≥1

= ρ(0) − i[H, ρ(0)]dt + {A, ρ(0)}dt + ∑ Lα ρ(0)L†α dt + O(dt2 ) (262c)

α≥1
1 †
= ρ(0) − i[H, ρ(0)]dt + ∑ (Lα ρ(0)L†α − {L Lα , ρ(0)}) dt + O(dt2 ) . (262d)
α≥1 2 α

Therefore:
ρ(dt) − ρ(0) 1
ρ̇(t)∣0 = lim = −i[H, ρ(0)] + ∑ (Lα ρ(0)L†α − {L†α Lα , ρ(0)}) . (263)
dt→0 dt α≥1 2

This is almost
√ the form of the master equation we are after. Note that the √ operators Lα are not dimensionless, but must have
units of 1/ time. To make them dimensionless, let us replace them by γα′ Lα , where γα′ has units of 1/time, so that the new
√ √
Lα are dimensionless. Substituting this into Eq. (263) only generates the combinations γα′ γα′∗ = ∣γα′ ∣ ≡ γα ≥ 0. Thus:
1 †
ρ̇(t)∣0 = −i[H, ρ(0)] + ∑ γα (Lα ρ(0)L†α − {L Lα , ρ(0)}) . (264)
α≥1 2 α

This result is valid as a short time expansion near t = 0. We now make an extra, very significant assumption:
Assumption 1. Eq. (264) is valid for all times t > 0.
This is essentially the Markovian limit, which states (informally) that there is no memory in the evolution, as manifested by
the fact that the evolution “resets” every dt. It is motivated in part by the observation that if we limit our attention just to ρ̇(t)∣0 =
−i[H, ρ(0)], then we already know that this replacement is valid, i.e., that we can indeed replace this with ρ̇(t) = −i[H, ρ(t)]
for all t, since this is just the Schrödinger equation. With this we finally arrive at the Lindblad equation:
dρ 1
= −i[H, ρ(t)] + ∑ γα (Lα ρ(t)L†α − {L†α Lα , ρ(t)}) ≡ Lρ . (265)
dt α 2
 47

The generator of the evolution, L, is called the Lindbladian. The Lα are called the Lindblad operators. The operator H is
Hermitian and will be interpreted later as the Hamiltonian of the system (plus a correction called the Lamb shift). The form of
the dissipative part of the Lindbladian, also known as the dissipator, is:
1
LD [⋅] = ∑ γα (Lα ⋅ L†α − {L†α Lα , ⋅}) , γα ≥ 0 . (266)
α 2

We can now define decoherence in a basis-independent manner. Decoherence is what happens when LD ≠ 0. In this case the
evolution of the density matrix is governed not only by the Schrödinger component −i[H, ⋅] (responsible for unitary evolution),
but also by the dissipator, which gives rise to non-unitary evolution.
The positivity of the Lindblad rates (they have units of 1/time) is a direct consequence of complete positivity. Conversely, it
guarantees that the map generated by the Lindblad equation (265) is CP, as we will show in Sec. IX C. As our derivation shows,
this map has Kraus operators given by

K0 = I + (−iH + A)dt (267a)

√ √
Kα = γα Lα dt , α ≥ 1 . (267b)

B. The Markovian evolution operator as a one-parameter semigroup

The formal solution of the (Lindblad) equation ρ̇(t) = Lρ is

ρ(t) = eLt ρ(0) ≡ Λt ρ(0) , (268)

where Λt is called the Markovian evolution operator (it is also a quantum map). The set {Λt }t≥0 forms a one-parameter
semigroup. The one-parameter part is clear: the set depends only on the time t, once the Lindblad generator L is fixed. The
reason this is a semi-group is that the superoperators Λt only satisfy three of the four properties of a group:
1. Identity operator: Λ0 = I.
2. Closed under multiplication: Λt Λs = eLt eLs = eL(t+s) = Λt+s .
3. Associative: (Λt Λs )Λr = Λt (Λs Λr ).
However, not every element has an inverse: as we shall see, complete positivity forces all the eigenvalues of L to be non-positive,
so that the map Λt is contractive, corresponding to exponential decay. This means that Λ∞ has at least one zero eigenvalue, so
it does not possess an inverse. We shall shortly see this in examples.

C. Proof that the solution of the Lindblad Equation is a CP map

The argument we use to prove that the solution of the Lindblad Equation is a CP map is essentially the reverse of that presented
in Sec. IX A, plus a proof that the concatenation of CP maps (and in particular of a CP map with itself) is still a CP map.
Let us start from the Lindblad equation and let A ≡ − 21 ∑α≥1 L†α Lα :

1 †
ρ̇(t) = Lρ(t) = −i[H, ρ(t)] + ∑ (Lα ρ(t)L†α − {L Lα , ρ(t)}) (269a)
α≥1 2 α
1 †
= −i[H, ρ(t)] + ∑ (Lα ρ(t)L†α − {L Lα , ρ(t)}) (269b)
α≥1 2 α
= −i[H, ρ(t)] + {A, ρ(t)} + ∑ Lα ρ(t)L†α (269c)
α≥1

= (A − iH)ρ(t) + ρ(t)(A + iH) + ∑ Lα ρ(t)L†α . (269d)

α≥1

Now define K0 ≡ [I + (A − iH)dt] + O[(dt)2 ], where in the end we will take the dt → 0 limit to remove any residual O[(dt)2 ]
terms. Then:

K0 ρ(t)K0† = [I + (A − iH)dt]ρ(t)[I + (A + iH)dt] + O[(dt)2 ] (270a)

= ρ(t) + [(A − iH)dt]ρ(t) + ρ(t)[(A + iH)dt] + O[(dt) ] . 2
(270b)
 48
√
Thus, using Eq. (269d), Eq. (270b), and defining Kα ≡ Lα dt:

ρ(t + dt) = ρ(t) + ρ̇(t)dt + O[(dt)2 ] (271a)

√ √
= ρ(t) + [(A − iH)dt]ρ(t) + ρ(t)[(A + iH)dt] + O[(dt) ] + ∑ (Lα 2
dt)ρ(t)(L†α dt) (271b)
α≥1

= ∑ Kα ρ(t)Kα† + O[(dt)2 ] (271c)

α≥0
≡ Φ[ρ(t)] , (271d)

which is in Kraus OSR form. However, to prove that this is a valid quantum map we still need to show that the set {Kα }α≥0
satisfies the normalization condition. Indeed, we have:

∑ Kα Kα = I + dt (2A + ∑ Lα Lα ) + O(dt ) = I + O(dt ) ,

† † 2 2
(272)
α≥0 α≥1

where in the first equality we used the Hermiticity of A and H, and in the second equality we used the definition of A.
Thus, we have shown that in the dt → 0 limit the map Φ [Eq. (271d)] is a quantum map from ρ(t) to ρ(t + dt). Let
dt = limn→∞ t/n, and consider the concatenated sequence of maps limn→∞ Φ○n [ρ(0)] = Φ[Φ[⋯Φ[ρ(0)]]] = Λt ρ(0), which
is clearly equivalent to the solution of the Lindblad equation [since it maps ρ(0) → ρ(dt) → ρ(2dt) → ⋯ → ρ(t)], i.e., if
ρ̇ = Lρ(t), with L the Lindbladian of Eq. (269a), then Λt = eLt . Since we have shown that Φ is a CP map, it remains to be
shown that a concatenation of quantum maps is still a quantum map. This is true, since if Φ1 and Φ2 are quantum maps then

Φ2 ○ Φ1 (ρ) = Φ2 [Φ1 (ρ)] = ∑ Kβ′ Φ1 (ρ)Kβ′ = ∑ Kβ′ Kα (ρ)Kα† Kβ′ = ∑ Kγ′′ ρKγ′′ ,
† † †
(273)
β αβ γ

where Kγ′′ = Kβ′ Kα , and ∑γ Kγ′′ Kγ = ∑α Kα† (∑β Kβ′ Kβ )Kα = ∑α Kα† Kα = I as required.
† †

D. Examples

1. Just H for a single qubit: the Bloch equations

Consider the Lindblad equation with all γα = 0, i.e., ρ̇ = −i[H, ρ]. This is just the Schrödinger equation written for density
matrices (also known as the Liouville-von Neumann equation). Let us solve it for the case of a single qubit. We can always write
⃗ = (hx , hy , hz ) ∈ R3 (since the Pauli matrices with identity form a basis over R4 for all 2 × 2
H = h0 I + ∑i∈{x,y,z} hi σ i , with h
matrices). Thus, using ρ = 2 (I + v⃗ ⋅ σ
1
⃗ ):

i i ⃗ × v⃗) ⋅ σ
− i[H, ρ] = − ∑ hi [σ , v⃗ ⋅ σ
⃗] = − hi vj [σ i , σ j ] = εijk hi vj σ k = (h ⃗.
i
∑ ∑ (274)
2 i∈{x,y,z} 2 i,j∈{x,y,z} i,j,k∈{x,y,z}

Since ρ̇ = 12 (v⃗˙ ⋅ σ
⃗ ), we find

v̇ ⋅ σ ⃗ × v⃗) ⋅ σ
⃗ = 2(h ⃗, (275)

which are three coupled first order differential equations for the components of v⃗. These are known as the Bloch equations,
and their solution has the Bloch vector v⃗ rotating around the vector h ⃗ with a frequency equal to 2∥h∥,
⃗ as is easily checked.
For example, consider a rotation about the vx axis, i.e., let h ⃗ = (h, 0, 0). Then Eq. (275) becomes: v̇x = 0, v̇y = −2hvz , and
v̇z = 2hvy . Differentiating again gives v̈y = −4h2 vy . The solution of these equations is

vx (t) = vx (0) (276a)

vy (t) = vy (0) cos(2ht) − vz (0) sin(2ht) (276b)
vz (t) = vz (0) cos(2ht) + vy (0) sin(2ht) . (276c)

The general case follows from this one by a reorientation of the axes to align with what we called the vx axis in the solution
above.
 49

2. Phase Damping for a single qubit

We already encountered the phase damping model in the Kraus OSR setting in Sec. VII C. Let us now study a Lindblad
equation model that generates the same map.
Let L1 = σ z = Z, γ1 = γ, γα≥2 = 0, and H = 0. Thus,
1
ρ̇(t) = γ(ZρZ † − {Z † Z, ρ}) = γ(ZρZ − ρ) . (277)
2

Using ρ = 12 (I + v⃗ ⋅ σ
⃗ ), the left-hand side evaluates to ρ̇ = 21 v⃗˙ ⋅ σ
⃗ . For the right hand side ZρZ = 12 (I − vx X − vy Y + vz Z), and
we thus arrive at:
1
(v˙x X + v˙y Y + v˙z Z) = −γ (vx X + vy Y ) . (278)
2
Equating the two sides componentwise (multiply both sides by X, Y , or Z, and take the trace) gives:

v̇x = −2γvx Ô⇒ vx (t) = vx (0)e−2γt (279a)

−2γt
v̇y = −2γvy Ô⇒ vy (t) = vy (0)e (279b)
v̇z = 0 Ô⇒ vz (t) = vz (0) . (279c)

We can see that, since γ ≥ 0, the map is contractive, and the Bloch sphere collapses to the vz -axis exponentially fast with time.
In the limit t → ∞, this simply projects every state directly to the vz -axis, which is manifestly uninvertible.
We can now match the Lindblad√ equation solution to the Kraus OSR result from Sec. VII C, where we found the Kraus
√
operators K0 = pI and K1 = 1 − pZ, and found that the Bloch vector is mapped to

v⃗′ = ((2p − 1)vx (0), (2p − 1)vy (0), vz (0)) ≡ (vx (t), vy (t), vz (t)) . (280)

The Lindblad phase damping result and the Kraus OSR thus have exactly the same effect provided we identify
1
2p − 1 = e−2γt Ô⇒ p(t) = (1 + e−2γt ) . (281)
2
The probability in this model approaches 1/2 in the limit t → ∞.
If we now allow H ≠ 0, i.e., solve the full Lindblad equation ρ̇ = −i[H, ρ] + γ(ZρZ † − ρ), then the result in Sec. IX D 1 shows
that this gives rise to a rotating Bloch ellipsoid that is simultaneously shrinking exponentially along its principal axis.

3. Amplitude damping / Spontaneous Emission for a single qubit

Likewise, we can construct a Lindblad equation for amplitude damping, which we encountered as a quantum map in
Sec. VII D.
Let L1 = σ − = ∣0⟩⟨1∣ = (σ + )† , γ1 = γ, γα≥2 = 0, and H = 0. Plugging these into the Lindblad equation we get:

1 1
ρ̇(t) = γ (σ − ρσ + − {σ + σ − , ρ}) = γ (∣0⟩⟨1∣ρ∣1⟩⟨0∣ − {∣1⟩ ⟨0∣0⟩ ⟨1∣ , ρ}) (282)
2 2

Using ρ = 21 (I + v⃗ ⋅ σ
⃗ ) we find, for the right-hand side:

1 1
∣0⟩⟨1∣ [ (I + vx X + vy Y + vz Z)] ∣1⟩⟨0∣ = (∣0⟩⟨0∣ − vz ∣0⟩⟨0∣) (283a)
2 2
1 1
− ∣1⟩⟨1∣(I + vx X + vy Y + vz Z) = − (∣1⟩⟨1∣ + (vx + ivy )∣1⟩⟨0∣ − vz ∣1⟩⟨1∣) (283b)
4 4
1 1
− (I + vx X + vy Y + vz Z)∣1⟩⟨1∣ = − (∣1⟩⟨1∣ + (vx − ivy )∣1⟩⟨0∣ − vz ∣1⟩⟨1∣) . (283c)
4 4
Adding up all these terms gives:
1 1 1
(Z − vx X − vy Y − vz Z) , (284)
2 2 2
 50

which we need to equate with 12 v⃗˙ ⋅ σ

⃗ . Therefore:
1 1
v̇x = − γvx , v̇y = − γvy , v̇z = −γ(vz − 1) . (285)
2 2
The last of these is solved by writing dvz /(vz − 1) = −γdt and integrating, to give ln(vz − 1) = −γt + c, i.e., vz (t) = c′ e−γt + 1,
so that c′ = vz (0) − 1. Thus:

vx (t) = vx (0)e−γt/2 (286a)

−γt/2
vy (t) = vy (0)e (286b)
−γt
vz (t) = 1 + [vz (0) − 1]e . (286c)
As t → ∞, vx , vy → 0 and vz → 1. This represents a contraction of the Bloch sphere to the north pole state ∣0⟩⟨0∣. Eq. (286) also
show that the contraction rate is twice as high along the vz axis than the vx and vy axes.
√ Now recall that in our √
Kraus OSR treatment of amplitude damping (Sec. VII D) we √ operators K0 = ∣0⟩⟨0∣ +
√ had the Kraus
1 − p∣1⟩⟨1∣ and K1 = p∣0⟩⟨1∣, and found that the Bloch vector was mapped to v⃗′ = ( 1 − pvx (0), 1 − pvy (0), (1−p)vz (0)+
p). The Lindblad amplitude damping result and the Kraus OSR thus have exactly the same effect provided we identify p =
1 − e−γt . Thus, the probability of a transition from the excited state to the ground state increases exponentially with time, and in
the limit t → ∞ we have p → 1.
Note that this dynamical description is not unique, as the Kraus map only fixes the discrete mapping from the initial to the
final state, and there are many dynamical descriptions which will recreate the mapping. Markovian dynamics is only one of the
possible evolutions.

X. THE LINDBLAD EQUATION VIA COARSE GRAINING

In this section we provide an alternative analysis leading to the Lindblad equation. The derivation is longer than the one
we saw in Sec. IX, but provides additional insight and generalizability. Our analysis follows Ref. [13], with some changes of
notation as well as clarifications and minor corrections.

A. Derivation

Let us start again with the Kraus OSR, and recall that the Kraus operators act on HS , i.e., Kα ∈ B(HS ). Let us introduce a
d2 −1
fixed (time-independent) operator basis {Fi }i=0
S
for B(HS ), where dS = dim(HS ), such that F0 = I. We can then expand the
Kraus operators in this basis:
d2S −1
Kα (t) = ∑ biα (t)Fi , (287)
i=0

where biα are the time-dependent elements of a (rectangular) d2S × d2B -dimensional matrix b, and dB = dim(HB ). Then the
Kraus OSR becomes:
ρ(t) = ∑ Kα (t)ρ(0)Kα† (t) = ∑ χij (t)Fi ρ(0)Fj† (288a)
α ij

= χ00 (t)ρ(0) + ∑[χ0i (t)ρ(0)Fi† + χi0 (t)Fi ρ(0)] + ∑ χij (t)Fi ρ(0)Fj† , (288b)
i>0 i,j>0

where
χij (t) = ∑ biα (t)b∗jα (t) , (289)
α

i.e., χ = bb† . It follows immediately that χ is positive semidefinite: ⟨v∣ χ ∣v⟩ = ∥b† ∣v⟩ ∥2 ≥ 0. Note that χ is a d2S × d2S matrix.
Now consider the normalization condition:
I = ∑ Kα† (t)Kα (t) = ∑ χij (t)Fj† Fi (290a)
α ij

= χ00 (t)I + ∑(χ0i (t)Fi† + χi0 (t)Fi ) + ∑ χij (t)Fj† Fi . (290b)

i>0 i,j>0
 51

We can use this to eliminate the χ00 ρ(0) term from Eq. (288b). Multiply Eq. (290b) first from the right by 12 ρ(0), then from the
left, and add the resulting two equations:
1 1
ρ(0) = χ00 (t)ρ(0) + ∑ [χ0i (t) (Fi ρ(0) + ρ(0)Fi ) + χi0 (t) (Fi ρ(0) + ρ(0)Fi )] + ∑ χij (t) {Fj Fi , ρ(0)} .
† † †
(291)
2 i>0 2 i,j>0

Subtracting this from Eq. (288b) yields:

1 1
ρ(t) − ρ(0) = ∑ [χi0 (t) (Fi ρ(0) − ρ(0)Fi ) − χ0i (t) (Fi ρ(0) − ρ(0)Fi )] + ∑ χij (t) (Fi ρ(0)Fj − {Fj Fi , ρ(0)}) .
† † † †
2 i>0 i,j>0 2
(292)
Let us now define
i
Q(t) ≡ ∑ χj0 (t)Fj − χ0j (t)Fj ,
†
(293)
2 j>0

and note that Q = Q† , i.e., Q is Hermitian. Then we can rewrite Eq. (292) as:
1
ρ(t) − ρ(0) = −i[Q(t), ρ(0)] + ∑ χij (t) (Fi ρ(0)Fj† − {F † Fi , ρ(0)}) . (294)
i,j>0 2 j

This obviously resembles the Lindblad equation, but it relates the state at t = 0 to the state at some arbitrary later time t, i.e., it
still represents a quantum map. Indeed, everything we have done so far is exact and we have simply rewritten the Kraus OSR
in a fixed operator basis. As a first step towards getting this closer to standard Lindblad form, let us diagonalize the χ matrix,
which will allow us to rewrite the double sum in Eq. (294) as a single sum. We have already noted that χ ≥ 0, so that it can be
diagonalized via some unitary matrix u: γ̃ = uχu† , where γ̃ is diagonal and positive semidefinite. Define Lk = ∑j>0 u∗kj Fj , so
that, using the unitarity of u:

Fi = ∑ uki Lk , (295)
k>0

where the sum over k > 0 excludes L0 = I. Thus, again using the unitarity of u:

∑ χij Fi ρ(0)Fj = ∑ Lk ρ(0)Ll ∑ uki χij (u )jl = ∑ γk Lk ρ(0)Lk

† † † †
(296a)
i,j>0 k,l>0 i,j>0 k>0

∑ χij Fj Fi = ∑ Ll Lk ∑ uki χij (u )jl = ∑ γk Lk Lk ,

† † † †
(296b)
i,j>0 k,l>0 i,j>0 k>0

where γk ≥ 0 are the eigenvalues of χ. We can now rewrite Eq. (294) as:
1 †
ρ(t) − ρ(0) = −i[Q(t), ρ(0)] + ∑ γk (t) (Lk ρ(0)L†k − {L Lk , ρ(0)}) . (297)
k>0 2 k

This is as far as we can go towards the Lindblad equation without introducing an approximation.
Let us now take a step back and introduce a generator for the exact quantum map. I.e., let us write ρ(t) = Λ(t, 0)[ρ(0)],
where
t
L(s)ds
Λ(t, 0) = T+ e∫0 . (298)

Let τ denote a short time interval, where the meaning of short will become clear momentarily. We define a “coarse-grained”
generator Lj as follows:

1 (j+1)τ
Lj = ∫ L(s)ds . (299)
τ jτ
t
∫0 L(s)ds = ∑j=0 Lj provided t = nτ , so that
1 n−1
Then τ

n−1
Lj
Λ(t, 0) = T+ eτ ∑j=0 . (300)

We now make a (strong) assumption:

 52

Assumption 2. The coarse-grained generators belonging to different time intervals commute:

[Lj , Lk ] = 0 ∀j, k . (301)

This assumption amounts to there being no memory of the evolution from one interval to the next.6 Under this assumption,
which we can also understand as a Markovian approximation, the time-ordered exponential becomes a product of exponentials:
n−1 n−1
Λ(t, 0) = ∏ eτ ∑ Lj ≡ ∏ Λj . (302)
j=0 j=0

Thus, ρj+1 = Λj [ρj ], where ρj ≡ ρ(jτ ), or, after Taylor expansion:

ρj+1 − ρj
ρj+1 = (I + τ Lj + O(τ 2 )) ρj Ô⇒ = Lj ρj (303)
τ
where we dropped the higher order corrections subject to the following, additional assumption:
Assumption 3.

τ ∥Lj ∥ ≪ 1 ∀j . (304)

Note that Eq. (485) sets an upper bound on τ in terms of the largest eigenvalue of the coarse-grained Lindblad generator.
This eigenvalue determines the fastest timescale for the system evolution (we’ll see later that these eigenvalues are all possible
differences of energies, i.e., they correspond to transition frequencies). Thus, Eq. (485) can also be interpreted as stating that the
coarse-graining timescale should be small compared to the timescale over which ρj changes.
Eq. (303) implies that, in particular, for j = 0:

ρ(τ ) − ρ(0)
= L0 [ρ(0)] . (305)
τ
Lemma 1.

χij (0) = δi0 δj0 . (306)

Proof. Using U (t) = e−iHt , we have for the Kraus operators:

Kα (0) = b0α (0)I + ∑ biα (0)Fi (307a)

i>0
√ √
= λν ⟨µ∣ U (0) ∣ν⟩ = λν δµν I , (307b)

so that [recall that α = (µν)]

√
biα (0) = λν δµν δi0 . (308)

Therefore

χij (0) = ∑ biα (0)b∗jα (0) = ∑ λν δi0 δj0 , (309)

α ν

which proves the lemma, since ∑ν λν = 1.

It follows immediately that χ(0) is already diagonal, and its eigenvalues are γ0 (0) = 1 and γk>0 (0) = 0. It also follows
immediately from Eq. (293) that Q(0) = 0.
Now define
1 (j+1)τ
⟨X⟩j ≡ ∫ X(s)ds . (310)
τ jτ

6 It is an interesting open problem to derive rigorous conditions for this to hold from first principles.
 53

Then
ρ(τ ) − ρ(0)
⟨ρ̇⟩0 = (311a)
τ
Q(τ ) − Q(0) Q(τ )
⟨Q̇⟩0 = = (311b)
τ τ
γk (τ ) − γk (0) γk (τ ) − δk0
⟨γ̇k ⟩0 = = . (311c)
τ τ
We can therefore rewrite Eq. (297) as:
ρ(τ ) − ρ(0) Q(τ ) γk (τ ) − δk0 1
= −i[ , ρ(0)] + ∑ (Lk ρ(0)L†k − {L†k Lk , ρ(0)}) , (312)
τ τ k>0 τ 2
which must equal L0 [ρ(0)] by Eq. (305). Hence, we can read off L0 :
1 †
L0 [X] = −i[⟨Q̇⟩0 , X] + ∑ ⟨γ̇k ⟩0 (Lk XL†k − {L Lk , X}) , (313)
k>0 2 k
This generator is precisely in Lindblad form. However, it only connects ρ(0) to ρ(τ ). In order to connect ρ(jτ ) to ρ((j + 1)τ )
we may now postulate that the same generator form remains valid, i.e., that
1 †
Lj [X] = −i[⟨Q̇⟩j , X] + ∑ ⟨γ̇k ⟩j (Lk XL†k − {L Lk , X}) ∀j , (314)
k>0 2 k
which we can do as long as Eq. (301) is satisfied. The simplest way to ensure this is to demand that in fact
Lj = L0 ∀j . (315)
This is again the Markovian limit, where there is no memory of the previous evolution segment. If, instead, we keep the more
general form of Eq. (314) [again, subject to Eq. (301)], then we have a time-dependent Markovian process, where the generator
is allowed to change over time, as long as these changes are uncorrelated between different time-segments.
Retaining the time-independent Markovian form of Eq. (315), and further replacing ⟨ρ̇⟩j by ρ̇ (another approximation, that
becomes exact in the limit τ → 0), we finally have the following result for the coarse-grained Lindblad equation, representing a
time-independent Markovian limit:
1 †
ρ̇(t) = −i[⟨Q̇⟩0 , ρ(t)] + ∑ ⟨γ̇k ⟩0 (Lk ρ(t)L†k − {L Lk , ρ(t)}) (316)
k>0 2 k

One point remains, which is to show that the coefficients ⟨γ̇k ⟩0 are non-negative, which is a requirement for complete positivity
of the map generated by the Lindblad equation. To show this, note that
1 τ 1
⟨γ̇k ⟩0 = ∫ γ̇k (t)dt = (γk (τ ) − γk (0)) . (317)
τ 0 τ
We already know that γk (t) ≥ 0 ∀t (recall that these are the eigenvalues of χ), so we need to show that nothing is spoiled by
subtracting γk (0). But, this is true since we already showed above that γk>0 (0) = 0. Thus, Eq. (317) shows that the coefficients
are all non-negative, as required for the Lindblad equation.

B. Interaction picture

As a brief digression, let us review the interaction picture, in preparation for the example we shall study in the next subsection.
Consider a (time-dependent) Hamiltonian H of the form:
H(t) = H0 (t) + V (t) . (318)
The unitary evolution operators satisfy:
dU (t)
= −iH(t)U (t) (319a)
dt
dU0 (t)
= −iH0 (t)U0 (t) . (319b)
dt
Define the interaction picture propagator with respect to H0 via:
Ũ (t) = U0† (t)U (t, 0) . (320)
 54

Claim 2. Ũ satisfies the Schrödinger equation

dŨ (t)
= −iH̃(t)Ũ (t), (321)
dt
with the interaction picture Hamiltonian

H̃(t) = U0† (t)V (t)U0 (t). (322)

Proof. Differentiate both sides of Eq. (320), while making use of Eqs. (318), (321) and (322):

dŨ (t) d [U0 (t)U (t)]

†
= = U̇0† U + U0† U̇ = iU0 H0 U + U0† (−iHU )
dt dt
= iU0 H0 U − iU0† (H + V )U0 Ũ = −iU0† V U0 Ũ = −iH̃ Ũ . (323)
The initial conditions of the equations are also the same [U (0) = I], thus Eqs. (320)-(322) describe the propagator generated by
H(t).

To make contact with open quantum systems, let V = HSB and H0 = HS + HB . Then U0 = e−itHS ⊗ e−itHB . We can now
transform the Schrödinger picture density matrix to the interaction picture via ρ̃SB (t) = U0† (t)ρSB (t)U0 (t), and if we write
HSB = ∑a λa Sa ⊗ Ba (Sa and Ba are system-only and bath-only operators, respectively), then H̃SB (t) = ∑a λa Sa (t) ⊗ Ba (t),
where Sa (t) = eitHS Sa e−itHS and Ba (t) = eitHB Ba e−itHB . This interaction picture density matrix satisfies

ρ̃SB (t) = Ũ (t)ρSB (0)Ũ † (t) (324)

(note that the Schrödinger picture and the interaction picture coincide at t = 0).
At this point everything we’ve shown for quantum maps and the Lindblad equation carries through with appropriate modifi-
cations. The Kraus OSR in the interaction picture becomes

ρ̃(t) = TrB [ρ̃SB (t)] = ∑ K̃α (t)ρ(0)K̃α† (t) (325)

α

where the interaction picture Kraus operators are

√
K̃α (t) = λν ⟨µ∣ Ũ (t) ∣ν⟩ . (326)

The interaction picture Lindblad equation, replacing Eq. (316), becomes:

˙ ˙ , ρ̃(t)] + ⟨γ̃˙ ⟩ (L ρ̃(t)L† − 1 {L† L , ρ̃(t)}) ,

= −i[⟨Q̃⟩
ρ̃(t) 0 ∑ k 0 k k k (327)
k>0 2 k

where Q̃ = Q − HS and γ̃k are the eigenvalues of the interaction picture χ-matrix χ̃ = b̃b̃† , with b̃ the expansion matrix of the
interaction picture Kraus operators: K̃α (t) = ∑i b̃iα (t)Fi .

C. Example: the spin-boson model for phase damping

To illustrate the predictions of the coarse-grained Lindblad equation, we consider the spin-boson model for phase damping of
a single qubit, described by the Hamiltonian

H = HS + HB + HSB (328a)
1
HS = − gZ , HB = ∑ ωk (nk + 1/2) , HSB = Z ⊗ (∑ λk bk + λ∗k b†k ) , (328b)
2 k k

where nk = b†k bk and bk are the bosonic number and annihilation operator for mode k, respectively ([bk , b†l ] = δkl I). Here
HSB describes
√ coupling of the qubit phase to the position x of each oscillator; recall that quantization means replacing x by
(b + b† ) / 2mω (where m is the oscillator mass), so that
√
λk ∝ 1/ ωk , (329)
 55

GHtL

Exact
2
Τ=2Ωc

Τ=8Ωc
1.5
Τ=32Ωc

0.5

Ωc t
2 4 6 8 10 12 14

FIG. 5. Comparison of the exact solution of the spin-boson model for single-qubit phase damping to the result obtained from the coarse-grained
Markovian master equation. Plotted are the arguments Γ(t) of the exponentials in Eq. (342). Straight lines correspond to the Markovian
solution, which intersects the exact solution (thick line) at t = τ , as seen from Eqs. (342a) and (342b). The bosonic bath density of states is
represented by the Debye model [Eq. (341)]. The results shown correspond to C = 0.05 and ωc = 1. Reproduced from Ref. [13].

a relation we will need later. In the interaction picture, it is easy to show that:7

H̃SB (t) = Z ⊗ (∑ λk e−iωk t bk + λ∗k eiωk t b†k ) . (330)

k

Assume that the bath is initially in a thermal Gibbs state at inverse temperature β = 1/T : ρB (0) = e−βHB /Z [Eq. (237)], and
let ⟨X⟩B ≡ Tr(XρB ). It is then a standard exercise to show that
1
⟨b†k bl ⟩B = δkl , ⟨b†k ⟩B = ⟨bk ⟩B = ⟨bk bl ⟩B = ⟨b†k b†l ⟩B = 0 . (331)
eβωk −1
Using this, it can be shown that the coarse-grained, interaction picture Lindblad equation Eq. (327) becomes [13]:
˙
ρ̃(t) = γ(τ ) (Z ρ̃(t)Z − ρ̃(t)) , (332)
˙ = 0, L = Z, and there are no other Lindblad operators (as should be obvious from the form of H above), and where
i.e., ⟨Q̃⟩ 0 1 SB

γ(τ ) = π ∑ ∣λk ∣2 coth(βωk /2)δ̄(ωk , τ ) (333)

k

is the dephasing rate, where we have defined

1
δ̄(ω, τ ) ≡ τ sinc2 (ωτ /2) . (334)
π
We already encountered Eq. (332) in Sec. IX D 2, and as we saw there its solution for the coherence (off-diagonal elements) is

ρ̃01 (t) = e−2γ(τ )t ρ01 (0) . (335)

As we shall see in Sec. XI, the spin-boson model we are considering here has an exact analytical solution. The exact solution
for the coherence is:
(e)
ρ̃01 (t) = e−2γ(t)t ρ01 (0) . (336)

This allows us to compare the Markovian result to the exact one, and better understand the condition the coarse-graining
timescale τ must satisfy. The only difference between the two is the argument of γ: τ versus t. However, this is a very

7 Some basic quantum mechanics would make this process very simple. Note that [b, n] = b gives bn = (n + 1)b. And therefore we would have ben = en+1 b.
 56

significant difference, since while the Markovian solution represents irreversible exponential decay, the exact solution is oscil-
latory: γ(t)t ∼ ∑k sin2 (ωk t). In order to observe closer agreement, we must once again invoke a continuous density of states
γ(ω), as we did in Sec. VIII B [recall Eq. (252)], which results in irreversible decay also in the case of the exact solution. Doing
so replaces Eq. (333) by
ωc
γ(τ ) = π ∫ Ω(ω)∣λ(ω)∣2 coth(βω/2)δ̄(ω, τ )dω , (337)
0

where we assumed that Ω(ω) has a high-frequency cutoff at ωc . Now note that δ̄ behaves similarly to the Dirac-δ function:
∞
∫ δ̄(ω, τ )dω = 1 , lim δ̄(ω, τ ) = δ(ω) , (338)
0 τ →∞

i.e., it is sharply peaked at ω = 0, and the peak becomes sharper as τ grows. The peak width is ∼ 1/τ . This suggests under
ω
what condition γ(t) ≈ γ(τ ), such that the exact and Markovian solutions agree: τ ≫ 1/ωc . The reason is that then ∫0 c captures
nearly all the area under δ̄(ω, τ ), whereas in the opposite case (τ ≲ 1/ωc ), most of the area under δ̄(ω, τ ) is not captured by the
same integral. Thus, assuming τ ≫ 1/ωc , δ̄(ω, τ ) effectively behaves as a Dirac-delta function, and if we assume in addition
that t > τ , then certainly also δ̄(ω, t) behaves as a Dirac-δ function. Thus, assuming
t > τ ≫ 1/ωc , (339)
we have
ωc
γ(τ ) ≈ γ(t) ≈ π ∫ Ω(ω)∣λ(ω)∣2 coth(βω/2)δ(ω)dω , (340)
0

so that the exact and Markovian cases agree. This is borne out numerically as well. Assume a Debye model, so that

ω 2 for ω < ωc
Ω(ω) ∝ { , (341)
0 for ω ≥ ωc
and that ∣λ(ω)∣2 ∝ ω −1 , in accordance with Eq. (329). In the high-temperature limit coth(βω/2) ∝ ω −1 , so that in all we have
ωc
ρ̃01 (t) ∝ exp (−Ctτ ∫ dωsinc2 (ωτ /2)) (342a)
0
ωc
(e)
ρ̃01 (t) ∝ exp (−Ct2 ∫ dωsinc2 (ωt/2)) , (342b)
0

where C is the temperature-dependent coupling-strength, with dimensions of frequency. Figure 5 shows the argument of the
exponentials in Eq. (342), Γ(t), for the exact solution and for the coarse-grained Lindblad equation, corresponding to different
values of the course-graining time-scale, τ . The curves corresponding to the Markovian solutions are just straight lines, as they
all describe simple exponential decays. It is clear that the Markovian solutions cannot account for the initial transition period,
but for sufficiently large τ (in units of the bath cutoff time 1/ωc ) the Lindblad result approximates the exact solution very well
at large times.
To summarize, the Markovian approximation gives reliable results for times greater than the coarse-graining time-scale, which
in turn must be greater than the inverse of the bath high-frequency cut-off. It does not account for the initial (Zeno-like) time
evolution.

XI. ANALYTICAL SOLUTION OF THE SPIN-BOSON MODEL FOR PHASE DAMPING

We present the analytical solution of the spin-boson model for pure dephasing. The derivation is based on [13, 14].
The model is the same as the one we considered in Sec. X C, except that we will consider a system of multiple qubits (indexed
by i). Starting from the interaction picture system-bath Hamiltonian [generalizing Eq. (330)]:
∗
H̃SB (t) = ∑ Zi ⊗ [λik e−iωk t ak + (λik ) eiωk t a†k ] , (343)
i,k

we want to find the system density matrix

ρ̃(t) = TrB [ρ̃tot (t)] = TrB [Ũ (t)ρ(0) ⊗ ρB (0)Ũ † (t)] , (344)
where
t
Ũ (t) = T+ exp [−i ∫ H̃(τ )dτ ] . (345)
0
 57

A. Calculation of the Evolution Operator

Note that H̃(t) does not commute with itself at different times, which is why we need the time-ordered product:

[H̃(t), H̃(t′ )] = i ∗ −i(ωk t−ωk′ t )

∑ Zi Zi′ ⊗ λk (λk′ ) e [ak , a†k′ ] + Zi Zi′ ⊗ (λik )∗ λik′ ei(ωk t−ωk′ t ) [a†k , ak′ ]
′ ′ ′ ′
i
(346a)
i,i′ ;k,k′
∗
= 2i ∑ Zi Zi′ ∑ I [λik (λik ) e−iωk (t−t ) ] ⊗ IB
′ ′
(346b)
i,i′ k

where we used the canonical bosonic commutation relations [ak , a†k′ ] = − [a†k , ak′ ] = Iδkk′ , [ak , al ] = [a†k , a†l ] = 0. Note that
further,
[[H̃(t), H̃(t′ )] , H̃(t′′ )] = 0. (347)
This means that we can use the Baker-Hausdorf formula exp(A + B) = exp(−[A, B]/2) exp(A) exp(B) (valid if [[A, B], A] =
[[A, B], B] = 0) to calculate U (t). To do so note the generalization
1
exp (∑ An ) = ( ∏ exp (− [An , An′ ])) (∏ exp(An )) , (348)
n n<n′ 2 n

which is valid if every second-order commutator vanishes. To apply this result for our case let us formally discretize the integrals
and denote Hn ≡ −iH̃(n∆t). We let ∆t = t/N and take the limit N → ∞. Then:
t N
U (t) = T+ exp [−i ∫ H̃(τ )dτ ] = T+ lim exp [ ∑ Hn ∆t] (349a)
0 ∆t→0 n=0
1 2
= lim ∏ exp (− [Hn , Hn′ ] (∆t) ) ∏ exp(Hn ∆t) (349b)
∆t→0 n<n′ 2 n
1 2
= lim ∏ (1 − [Hn , Hn′ ] (∆t) ) ∏ (1 − Hn ∆t) (349c)
∆t→0 n<n′ 2 n
1 2
= lim [1 − ∑ [Hn , Hn′ ] (∆t) ] [1 − ∑ Hn ∆t] (349d)
∆t→0 2 n<n′ n
1 2
= lim exp (− ∑ [Hn , Hn′ ] (∆t) ) exp (∑ Hn ∆t) (349e)
∆t→0 2 n<n′ n
1 t t1 t
= exp [ ∫ dt1 ∫ dt2 [H̃(t2 ), H̃(t1 )]] exp [−i ∫ H̃(τ )dτ ] . (349f)
2 0 0 0

Note that in the second line we enforced time-ordering by keeping n < n′ . To go from the third to the fourth line we kept the
lowest relevant order in each term, inherited from the second line. Note how in the last line the time-ordering is implemented
via t2 ≤ t1 . We find:
t
− i∫ H̃(τ )dτ = ∑ Zi ⊗ ∑ ((αki )∗ ak − αki a†k ) , (350)
0 i k

where
∗
(λik ) (eiωk t − 1)
αki (t) = . (351)
ωk
Now, since
t t1 t e−iωk t1 − 1 1 − eiωk t + iωk t
∫ dt1 ∫ dt2 e−iωk (t2 −t1 ) = ∫ dt1 eiωk t1 = , (352)
0 0 0 −iωk ωk2
we have, using Eq. (346b):
i t t1 ∗ t t1
j′ −iωk (t2 −t1 )
− ∫ dt1 ∫ dt2 [H̃(t2 ), H̃(t1 )] = ∑ Zj Zj ′ ∑ I [λk (λk ) ∫ dt1 ∫ dt2 e ] ⊗ IB
j
(353a)
2 0 0 jj ′ k 0 0

′ ∗ 1 − eiωk t + iωk t
= ∑ Zj Zj ′ ∑ I [λjk (λjk ) ] ⊗ IB . (353b)
jj ′ k ωk2
 58

Therefore, defining
′ ∗ eiωk t − iωk t − 1
fjj ′ (t) ≡ ∑ I [λjk (λjk ) ], (354)
k ωk2
we can write the first term in Eq. (349f) as follows:
1 t t1
exp [ ∫ dt1 ∫ dt2 [H̃(t2 ), H̃(t1 )]] = ei ∑jj′ fjj′ (t)Zj Zj′ ⊗ IB . (355)
2 0 0

Note that this is an operator acting non-trivially just on the system, and is a global phase for the case of a single qubit. Its action
is, however, non-trivial for multiple qubits (it represents a Lamb shift).
Since the ak operators commute for different modes we have as our final simplified result for the evolution operator:
Ũ (t) = eif (t) ∏ exp [Zi ⊗ (αki (t)ak − αki (t)∗ a†k )] . (356)
i,k

B. Calculation of the Density Matrix

Now recall the definition of the coherent states. These are eigenstates of the annihilation operator:
a∣α⟩ = α∣α⟩ . (357)
They are minimum-uncertainty states in a harmonic potential, and can be expanded as
∞
αn
∣α⟩ = e−∣α∣ /2
2
∑ √ ∣n⟩ (358)
n=0 n!
where ∣n⟩ are number (Fock) states. The completeness relation for coherent states is:
1
∫ d α ∣α⟩⟨α∣ = 1
2
(359)
π
where the integration is over the entire complex plane. They are useful in our context since they are created by the displacement
operator
D (α) ≡ exp (αa† − α∗ a) = D(−α)† (360)
acting on the vacuum state:
D (α) ∣0⟩ = ∣α⟩, (361)
which is clearly related to U (t). We will need the result:
αβ ∗ − α∗ β
D (α) D (β) = exp D(α + β), (362)
2
which is easily derived from D (α) = exp (αa† − α∗ a), [a, a† ] = 1, and the Baker-Hausdorf formula exp(A + B) =
exp(−[A, B]/2) exp(A) exp(B) (again, valid if [[A, B], A] = [[A, B], B] = 0).
Now let Rik (t) ≡ αki (t)a†k − αki (t)∗ ak and consider exp [Zi ⊗ Rik (t)]:
∞ ∞
R2n R2n+1
exp [Z ⊗ R] = IS ⊗ ∑ +Z ⊗ ∑ = IS ⊗ cosh R + Z ⊗ sinh R (363a)
n=0 (2n)! n=0 (2n+!)!
1 1
= IS ⊗ [D (α) + D (−α)] + Z ⊗ [D (α) − D (−α)] = ∣0⟩⟨0∣ ⊗ D (α) + ∣1⟩⟨1∣ ⊗ D (−α) . (363b)
2 2
This shows that depending on whether the field is coupled to the qubit ∣0⟩ or ∣1⟩ state, the field acquires a different displacement.8
The evolution operator can thus be written as:
U (t) = ei ∑jj′ fjj′ (t)Zj Zj′ ∏ [∣0⟩i ⟨0∣ ⊗ D (αki ) + ∣1⟩i ⟨1∣ ⊗ D (−αki )] . (364)
i,k

8 Note that this is the source of the dephasing the qubits undergo, since when acting on a superposition state of a qubit, the qubit and field become entangled:
exp [σz ⊗ R] (a∣0⟩ + b∣1⟩)∣β⟩ = a∣0⟩ ⊗ D (α) ∣β⟩ + b∣1⟩ ⊗ D (−α) ∣β⟩ = e(αβ a∣0⟩ ⊗ ∣α + β⟩ + e−(αβ b∣1⟩ ⊗ ∣β − α⟩ .
∗
−α∗ β)/2 ∗
−α∗ β)/2
 59

Now assume that the bosonic bath is in thermal equilibrium:

−1
1 −βHB e−βωk /2 1
ρB = e = [∏ ] exp (−β ∑ ωk (Nk + )) = ∏ ρB,k , (365a)
k 1−e
Z −βωk 2
k k

where

1
ρB,k = exp (−βωk Nk ) , (366)
⟨Nk ⟩

and the mean boson occupation number is given by the Bose-Einstein distribution:

1
⟨Nk ⟩ = . (367)
eβωk −1

As shown in [15], p.122-3, this can be transformed into the coherent-state representation, with the result:

1 ∣αk ∣2
ρB,k = ∫ d αk exp (−
2
) ∣αk ⟩⟨αk ∣ . (368)
π⟨Nk ⟩ ⟨Nk ⟩

For simplicity let us from now on consider the case of a single qubit. It suffices to calculate the evolution of each of the four
pure states ∣x⟩⟨y∣, where x, y ∈ {0, 1}, separately. Thus

ρx,y (t) = TrB [U (t)∣x⟩⟨y∣ ⊗ ρB (0)U † (t)]

= TrB [∏ [∣0⟩⟨0∣ ⊗ D (αk ) + ∣1⟩⟨1∣ ⊗ D (−αk )] ∣x⟩⟨y∣ ⊗ ∏ ρB,m ∏ [∣0⟩⟨0∣ ⊗ D† (αl ) + ∣1⟩⟨1∣ ⊗ D† (−αl )]] .
k m l

The terms in the three products match one-to-one for equal indices, so we can write everything as a product over a single index
k. Using Tr(A ⊗ B) = TrA × TrB to rearrange the order of the trace and the products, and D† (−α) = D (α), we have:

ρx,y (t) = δx,0 δy,0 ∣0⟩⟨0∣ ⊗ ∏ Tr [D (αk ) ρB,k D (−αk )] (369a)

k
+ δx,0 δy,1 ∣0⟩⟨1∣ ⊗ ∏ Tr [D (αk ) ρB,k D (αk )] (369b)
k
+ δx,1 δy,0 ∣1⟩⟨0∣ ⊗ ∏ Tr [D (−αk ) ρB,k D (−αk )] (369c)
k
+ δx,1 δy,1 ∣1⟩⟨1∣ ⊗ ∏ Tr [D (−αk ) ρB,k D (αk )] . (369d)
k

Consider the Tr terms: for ∣0⟩⟨0∣ and ∣1⟩⟨1∣ by cycling in the trace the displacement operators cancel and Tr [ρB,k ] = 1. Thus, as
expected the diagonal terms do not change:

ρ0,0 (t) = ρ0,0 (0) , ρ1,1 (t) = ρ1,1 (0) . (370)

As for the off-diagonal terms:

1 ∣βk ∣2
Tr [D (±2αk ) ρB,k ] = ∫ d βk exp (−
2
) ⟨βk ∣D (±2αk ) ∣βk ⟩ . (371)
π⟨Nk ⟩ ⟨Nk ⟩

Now:

⟨β∣D (±2α) ∣β⟩ = exp [± (αβ ∗ − α∗ β)] ⟨β∣ ± 2α + β⟩ (372a)

1
= exp [± (αβ ∗ − α∗ β)] exp [β ∗ (±2α + β) − (∣β∣2 + ∣ ± 2α + β∣2 )] (372b)
2
= exp (−2∣α∣2 ± 2 (αβ ∗ − α∗ β)) . (372c)
 60

Thus:

1 ∣βk ∣2
Tr [D (±2αk ) ρB,k ] = exp (−2∣αk ∣2 ) ∫ d βk exp (−
2
± 2 (αk βk∗ − αk∗ βk )) (373a)
π⟨Nk ⟩ ⟨Nk ⟩
exp (−2∣αk ∣2 )
= [π⟨Nk ⟩ exp (−4∣αk ∣2 ⟨Nk ⟩)] (373b)
π⟨Nk ⟩
1
= exp [−4∣αk ∣2 (⟨Nk ⟩ + )] (373c)
2
⎡ 2 ⎤
⎢ ∗
λk (e iω k t
− 1) 1 1 ⎥⎥
⎢
= exp ⎢−4 ∣ ∣ ( βω + )
2 ⎥⎥
(373d)
⎢ ω e k −1
⎣ k
⎦
1 − cos(ωk t) βωk
= exp [−4∣λk ∣2 2
coth ] . (373e)
ωk 2

Thus decay of the off-diagonal terms goes as e−2γ(t)t , with

βωk 1 − cos(ωk t) βωk ωk t

γ(t) = 2 ∑ ∣λk ∣2 coth = ∑ ∣λk ∣2 coth tsinc2 , (374)
k 2 ωk2 t k 2 2

which coincides with the exact result quoted in Sec. X C, specifically Eq. (333) with τ replaced by t.

XII. QUANTUM TRAJECTORIES AND UNRAVELLING THE LINDBLAD EQUATION

Solving the Lindblad equation numerically is demanding. For a d-dimensional system Hilbert space, the density matrix is
d × d, involving d2 − 1 real numbers that one must store and update at each time-step. Is there a more space-efficient alternative?
It turns out that instead one can propagate a wavefunction (only 2d − 1 real numbers, so a quadratic savings), at the expense
of introducing statistical averaging over many runs. A very interesting side-benefit of this so-called unravelling procedure is
that each wavefunction undergoes a “quantum trajectory”, that can be correlated to an individual sequence of quantum events,
whereas the density matrix instead corresponds to an ensemble of such events.
Let us write down the Lindblad equation [Eq. (265)] in the following form:
2
d
1
ρ̇ = −i[H, ρ] + ∑ γk (Lk ρL†k − {L†k Lk , ρ}) (375)
k=1 2

Here Lk are the Lindblad operators and γk are scalars. As is clear from the derivation presented in Sec. X, the number of
non-zero terms in the sum is at most d2 . If one sets ∥Lk ∥ = 1 then the scalars γk can be understood as rates of the corresponding
relaxation process.9
There are multiple ways we can proceed to study this equation:

1. Derive γk , Lk given the description of open system;

2. Find equivalent dynamics of the wavefunction ∣ψ(t)⟩ (in the closed system case the wavefunction is a d-dimensional
vector over C such that ⟨ψ∣ψ⟩ = 1; this time we will let its norm be arbitrary);

3. Suppose that measurements are performed repeatedly on the system, and derive the equation for dynamics given a string
of measurement outcomes.

Here we will address points 2 and 3. In a very narrow sense we will address 1, if the closed system + measurement apparatus
are thought of as an open system.

9 Here we use the operator norm ∥O∥:

√
∥O∥ = max∣v⟩∶⟨v∣v⟩=1 ⟨v∣O† O∣v⟩
√
This norm is the largest eigenvalue of O† O. For Hermitian O, it reduces to the largest absolute value of eigenvalues of O.
 61

A. Method summary

To begin, we rewrite the Lindblad equation, Eq. (265), in the form

ρ̇ = −i (HC ρ(t) − ρ(t)HC† ) + ∑ γα Lα ρ(t)L†α , (376)

α

where
i
HC = H − ∑ γα Lα Lα
†
(377)
2 α

is called the “conditional Hamiltonian”. Note that it is non-Hermitian. Consider the evolution of a pure state ∣ψ(0)⟩ subject to
HC :
HC
∣ψ(0)⟩ z→ e−iHC t ∣ψ(0)⟩ = ∣ψ̃(t)⟩ . (378)

Since HC is non-Hermitian, the norm of ∣ψ̃(t)⟩ decreases over time (hence the tilde):
d
∥ ∣ψ̃(t)⟩ ∥2 = ⟨ψ(0)∣ eiHC t (iHC† )e−iHC t + eiHC t (−iHC )e−iHC t ∣ψ(0)⟩
† †
(379a)
dt
= i ⟨ψ(0)∣ eiHC t (HC† − HC )e−iHC t ∣ψ(0)⟩
†
(379b)
= − ∑ γα ⟨ψ(0)∣ e iHC† t
L†α Lα e−iHC t ∣ψ(0)⟩ (379c)
α

= − ∑ γα ∥Lα e−iHC t ∣ψ(0)⟩ ∥2 ≤ 0 . (379d)

α

The action of the other term in Eq. (376) can be viewed as inducing a “quantum jump”:

Lα ∣ψ̃(t)⟩ γα ∥Lα ∣ψ̃(t)⟩ ∥2

∣ψ̃(t)⟩ z→ = ∣ψ(t)⟩ with probability pα = , (380)
∥Lα ∣ψ̃(t)⟩ ∥ ∑α γα ∥Lα ∣ψ̃(t)⟩ ∥2
where the post-jump state ∣ψ(t)⟩ is normalized, and pα tells us the probability that the particular jump Lα was realized.
If a jump took place at time t, then the probability that the next jump takes place in the interval (t, t + τ ] is given by

Pr(jump in (t, t + τ ] ∣ jump at t) = 1 − ∥e−iHC τ ∣ψ̃(t)⟩ ∥2 . (381)

In this way, the probability of a second jump at τ = 0 is zero, but the probability increases exponentially as τ grows.
Putting these steps together one arrives at the following algorithm for evolution from t = 0 to tf :
1. Initialize the state as ∣ψ(0)⟩, set j = 1

2. Evolve under the conditional Hamiltonian: ∣ψ̃(tj )⟩ = e−iHC t ∣ψ(tj )⟩

Lα ∣ψ̃(tj +τ )⟩
3. Perform a jump at tj + τ with probability given by Eq. (381): ∣ψ̃(tj + τ )⟩ z→ ∥Lα ∣ψ̃(tj +τ )⟩∥
, with the index α chosen with
γα ∥Lα ∣ψ̃(tj +τ )⟩∥
2
probability pα =
∑α γα ∥Lα ∣ψ̃(tj +τ )⟩∥2

4. If a jump took place, advance j to j + 1: call the new (normalized state) ∣ψ(tj+1 )⟩ and set tj+1 = tj + τ
5. Return to step 2, unless tj+1 ≥ tf

6. Repeat K times from step 1, calling the output from the k th round ψk (tf ), and construct ρ(tf ) = 1
K ∑k=1 ∣ψk (tf )⟩⟨ψk (tf )∣,
K

stop when ρ(tf ) has converged

It turns out that this algorithm converges to the solution ρ(tf ) of the Lindblad equation at t = tf (see, e.g., Section 7.1 of
Ref. [1], and also the proof below). Its major advantage is that, as mentioned above, it propagates wavefunctions rather √ than
density matrices, thus resulting in a quadratic space savings. The error in the approximation of ρ(tf ) decreases as 1/ K. By
the “no-free lunch theorem” it should be the case that it is sufficient to use K on the order of the Hilbert space dimension, so that
the total cost is conserved. However, in practice fewer repetitions may suffice, so that the quantum trajectories algorithm may in
fact be more efficient than brute force solution of the Lindblad equation.
 62

Each sequence {ψk (0), ψ̃k (t1 ), ψk (t1 ), ψ̃k (t2 ), ψk (t2 ), . . . , ψk (tj ), ψ̃k (tj ), . . . } is a “quantum trajectory”. It describes a
series of norm-decreasing evolutions interrupted by quantum jumps. This provides an interesting and insightful interpretation of
what actually takes place during open quantum system evolution. Consider, e.g., generalized amplitude damping (Sec. VII G).
An atom undergoes spontaneous emission to its ground state, but due to thermal excitation it can repopulate its excited state. As
we saw in Sec. IX D 3, the probability of a transition from the excited state to the ground state increases exponentially with time,
which is in accordance with Eq. (381). But now we see that the actual emission event is a “jump”, whereby the atom suddenly
and discontinuously finds itself in the ground state. The process can also work in the opposite direction, and by absorbing energy
from the bath, the atom can find itself in an excited state, etc. The downward transition event is accompanied by the emission
of a photon (by energy conservation), or phonon, or some other elementary excitation, which can be detected. And indeed, such
quantum trajectories have been measured in quantum optics experiments (see, e.g., Ref. [16] and references therein).
We now proceed to give a more careful and detailed description and analysis.

B. Equivalent dynamics of the wavefunction

1. Naive form

Starting over, we note that we can rewrite the Lindblad equation as follows, in the limit dt → 0:
1
ρ(t + dt) = ρ(t) − i[H, ρ(t)]dt − ∑ γk {L†k Lk , ρ(t)}dt + ∑ γk Lk ρ(t)L†k dt (382a)
k=1 2 k=1

= e−iHC dt ρ(t)eiHC dt + ∑ Mk ρ(t)Mk†

†
where (382b)
k=1
i √
HC = H − †
∑ γk Lk Lk , Mk = γk dtLk . (382c)
2 k=1
†
Here again HC is the non-Hermitian conditional Hamiltonian. If we define M0 = eiHC dt then this is the standard channel
decomposition that we started with:

ρ(t + dt) = ∑ Mk ρ(t)Mk† (383)

k=0

We note that instead of using a differential equation solver to obtain ρ(t + dt), using the non-selective measurement formalism
of Sec. VI A we can instead mathematically “simulate” the above formula in the following way:
1. choose k ≥ 0 with probability pk = Tr[Mk ρ(t)Mk† ];

2. set ρ(t + dt) = 1

N
Mk ρ(t)Mk† ;
3. repeat for the next time step dt.
This simulation uses random numbers {k}. It is easy to see that the expectation value of the density matrix at some later time T
is exactly the same as the solution of the master equation:

limdt→0 Av{k} ρ(T, {k}) = ρ(T ) (384)

Now we note that the whole process was linear with respect to ρ(t) = ∑i pi ∣ψi (t)⟩⟨ψi (t)∣. So we can work with the states
instead! Generate a random number i with probability pi given by initial conditions, so as to choose ∣ψi (0)⟩ as the initial state
(a pure state). Then follow these instructions with normalized ∣ψ(t)⟩ at each step to produce ∣ψ(t + dt)⟩:
1. choose k ≥ 0 with probability pk = ⟨ψ(t)∣Mk† Mk ∣ψ(t)⟩

2. set ∣ψ(t + dt)⟩ = 1

N
Mk ∣ψ(t)⟩
3. repeat for the next time step dt
We have derived the equivalence, so we know that

limdt→0 Av{k},i ∣ψ(T, {k}, i)⟩⟨ψ(T, {k}, i)∣ = ρ(T ) . (385)

Here we average over random numbers {k}, i to obtain the same density matrix as the solution of the master equation. For a
small range of k and simple operators Lk this method already leads to substantial savings, as one never needs to store d × d
 63

matrices during the simulation, only d-dimensional vectors. However, note that to obtain the average in practice one needs to
sample from {k}, i some number of times K, repeating the whole simulation. In principle K can be as large as d, thus defeating
the purpose of the method, but in practice one can observe convergence of the average with increasing K, e.g. by studying the
dispersion of some observable

D(O) = Av{k},i (⟨ψ(T, {k}, i)∣O∣ψ(T, {k}, i)⟩)2 . (386)

Convergence is often achieved for K ≪ d.

2. Telegraph noise form

We consider a slightly different perspective that is essentially the same as above, but we note that k = 0 corresponding to
†
M0 = eiHC dt dominates the probability distribution for k in the limit dt → 0:

⟨ψ(t)∣M0† M0 ∣ψ(t)⟩ = 1 − O(dt) = 1 − ∑ γk ⟨ψ(t)∣L†k Lk ∣ψ(t)⟩dt + O(dt2 ) (387)

k=1

this means that one does not need to calculate ⟨ψ(t)∣Mk† Mk ∣ψ(t)⟩ every dt. One only calculates

pno-jump = ⟨ψ(t)∣M0† M0 ∣ψ(t)⟩ , (388)

and generates an auxilliary random variable JUMP= 0, 1 with probability pno-jump , 1 − pno-jump respectively. Only if JUMP= 1 we
ask which k actually happened.
Looking at Eq. (387) we see that at first the probability of a jump happening within an interval [t, t + τ ] increases from 0
linearly with τ , and at large τ it approaches 1 exponentially. The coefficient in front of the linear dependence is ∣ψ⟩-dependent,
but weakly so. There is a well-known stochastic process given by

pjump = rdt . (389)

In other words, independent jumps occur with rate r per unit of time. This process is called telegraph noise. The simula-
tion method described above is a quantum evolution interrupted by essentially independent jumps following a telegraph noise
distribution. Below we will study a different type of noise.

3. Stochastic Schrödinger equation approach

What we did above was produce a map from a wavefunction ∣ψ(t)⟩ plus a random variable ξ to the wavefunction at the next
time step ∣ψ(t + dt)⟩. The way we proved that this map is equivalent to the original master equation is by observing that

Avξ ∣ψ(t + ∆t, ξ)⟩⟨ψ(t + ∆t, ξ)∣ = ∣ψ(t)⟩⟨ψ(t)∣ + L(∣ψ(t)⟩⟨ψ(t)∣)∆t + O(∆t) . (390)

Here L is the generator of the Lindblad equation we are trying to simulate.

Let us now demonstrate that the Lindblad equation can also be derived from a stochastic Schrödinger equation approach. For
simplicity, let us consider a generator with just one Hermitian term:

1 1
L = AρA − A2 ρ − ρA2 . (391)
2 2
Let the random variable ξ actually be a stochastic function of time ξ(t) on the interval [t, t + ∆t]. Define the time-step for our
trajectory as:
t+∆t
ξ(τ )dτ
∣ψ(t + ∆t)⟩ = eiA ∫t ∣ψ(t)⟩ . (392)

This is the solution of the differential equation:

d
∣ψ(t)⟩ = iAξ(t)∣ψ(t)⟩ . (393)
dt
 64

We can do a second order Taylor series expansion of Eq. (408). The average of the density matrix after our time-step is then
given by:

Avξ ∣ψ(t + ∆t, ξ)⟩⟨ψ(t + ∆t, ξ)∣ = ∣ψ(t)⟩⟨ψ(t)∣ (394a)

t+∆t t+∆t
+ Avξ (iA ∫ ξ(τ )dτ ∣ψ(t)⟩⟨ψ(t)∣ − i∣ψ(t)⟩⟨ψ(t)∣A ∫ ξ(τ )dτ ) (394b)
t t
t+∆t t+∆t A2 A2
+ Avξ [∫ ∫ ξ(τ )ξ(τ ′ )dτ dτ ′ (A∣ψ(t)⟩⟨ψ(t)∣A − ∣ψ(t)⟩⟨ψ(t)∣ − ∣ψ(t)⟩⟨ψ(t)∣ )] + O(∆t2 ) . (394c)
t t 2 2
Note that before the choice of ξ is made, we don’t really know what the smallness of the next order in Taylor series is. First of
all we set

Avξ ξ(τ ) = 0 , (395)

to get rid of the first order in A. We also define the correlation function

C(τ, τ ′ ) = Avξ ξ(τ )ξ(τ ′ ) ≡ C(τ − τ ′ ) (396)

to be translation-invariant in time (i.e., to depend only the difference τ − τ ′ ). Together these two conditions define the first two
moments of Gaussian stochastic random variable. We then have:

Avξ ∣ψ(t + ∆t, ξ)⟩⟨ψ(t + ∆t, ξ)∣ = ∣ψ(t)⟩⟨ψ(t)∣ (397a)

t+∆t t+∆t 2 2
A A
+∫ ∫ C(τ − τ ′ )dτ dτ ′ (A∣ψ(t)⟩⟨ψ(t)∣A − ∣ψ(t)⟩⟨ψ(t)∣ − ∣ψ(t)⟩⟨ψ(t)∣ ) + O(∆t2 ) . (397b)
t t 2 2
We would like
t+∆t t+∆t
∫ ∫ C(τ − τ ′ )dτ dτ ′ ∼ ∆t . (398)
t t

We note that this will be the case if C(t) is peaked at 0 with width w ≪ ∆t and height C0 :
t+∆t t+∆t
∫ ∫ C(τ − τ ′ )dτ dτ ′ ≈ wC0 ∆t (399)
t t

Setting wC0 = 1 will recover the desired Lindblad generator L given in Eq. (391). Since w is the smallest timescale in the
problem we can just choose

C(t) = δ(t) (400)

where δ(t) is the Dirac delta function. We have proven:

Avξ ∣ψ(t + ∆t, ξ)⟩⟨ψ(t + ∆t, ξ)∣ = ∣ψ(t)⟩⟨ψ(t)∣ (401a)

2 2
A A
+ (A∣ψ(t)⟩⟨ψ(t)∣A − ∣ψ(t)⟩⟨ψ(t)∣ − ∣ψ(t)⟩⟨ψ(t)∣ ) ∆t + O(∆t2 ) . (401b)
2 2

Now the smallness of the remaining terms can be guaranteed as O(∆t2 ), and we have indeed recovered the Lindblad generator
L given in Eq. (391).
Using the same idea for the derivation, we can prove the equivalence between the original Lindblad equation (375) and the
following differential equation on ∣ψ(t)⟩:
d √
∣ψ(t)⟩ = −i(H − ∑ Lk ξk (t))∣ψ(t)⟩, Avξ ξk (t)ξm (t′ ) = δkm δ(t − t′ ) γk . (402)
dt k

Here δkm is the Kronecker delta function. The equivalence states that

Avξ ∣ψ(T, ξ)⟩⟨ψ(T, ξ)∣ = ρ(T ) (403)

the limit is included in δ-function and the definition of the differential equation, so no additional limit needs to be taken here. In
t+∆t
practice, though, some discretization scheme needs to be applied and the numerical simulation uses ∫t ξ(τ )dτ instead of the
raw ξ(t).
 65

4. Comparison between the telegraph noise and Stochastic Schrödinger equation approaches

If we compare the Stochastic Schrödinger equation approach to the telegraph noise method, we find that ξ tel (t) is a sequence
of randomly spaced peaks with 0 in between. It is possible to arrange for the correlation function of that signal to be C(t) = δ(t),
however the higher order correlation functions will be vastly different from the Gaussian noise that is usually used for stochastic
differential equations. The defining characteristic of the Gaussian noise is that higher order correlations (or moments) are
T
expressed via C(t) according to Wick’s theorem. Another way to think about it is that the Fourier transforms ∫0 ξ(t)eikt dt are
i.i.d. random variables for each k for Gaussian noise, but not for telegraph noise.
Let us discuss the properties of individual ∣ψ(t, ξ)⟩ or ∣ψ(t, {k})⟩ for a given realization of random variables, under the two
approaches. One way to look at this is to take an observable O (s.t. ∥O∥ = 1) and follow its average:

⟨ψ(t, ξ)∣O∣ψ(t, ξ)⟩ or ⟨ψ(t, {k})∣O∣ψ(t, {k})⟩ . (404)

If the closed system evolution of the observable O(t) has a characteristic frequency ω ∼ ∥[O, H]∥ and the relaxation has the
characteristic rates r ∼maxk γk , then there are two possible regimes: ω ≪ r and ω ≫ r. The qualitative picture that we will see
is as follows:

ω≪r ω≫r
telegraph smooth curves interrupted rapid sine wave interrupted
by discontinuities by discontinuities
stochastic noisy diffusive behaviur noisy almost periodic behavior

Even though the two methods are both equivalent to the same master equation, other characteristics such as the dispersion D(O)
or the diffusion coefficient of individual trajectories vary between the two methods. Thus, we find very different visual behavior
of individual trajectories. It is possible to interpolate between the two by chossing a non-Gaussian ξ(t). We note that the results
for a single trajectory are reminiscent of experimental measurements. We next make this analogy more precise.

C. Weak measurements

One way is to choose the distribution of the random process ξ(t) in such a way that an individual trajectory ⟨ψ(t, ξ)∣O∣ψ(t, ξ)⟩
matches the measurement output M (t) of some repeated measurement. However this is an unphysical approach. What we should
be doing is to come up with a mapping M (ξ) since ξ contain the information about random choices made outside of the system,
while ⟨ψ(t, ξ)∣O∣ψ(t, ξ)⟩ contains information “private” to the system, something that has not been measured yet.
We note that the first method with the decomposition given in Eq. (383) can be directly interpreted as a measurement where
k is an answer. The stochastic one requires some transformations before this can be done, as the width of the δ-function is
the smallest time-scale that is faster than the supposed data collection timescale. We do not know of any research that makes
this connection. There is a lot of research connecting weak measurements with trajectories, which could be seen as such an
interpretation of stochastic equations. The difference with telegraph noise is that every Mk is close to identity I with a small
probability in front.

XIII. ANALYTICAL SOLUTION OF THE GENERAL LINDBLAD EQUATION

In this section we discuss the analytical solution of the Lindblad equation in arbitrary dimensional (but finite) Hilbert spaces.

A. The coherence vector

Let us first introduce a “nice” operator basis for B(HS ), where d = dim(HS ). Let F0 = IS and choose M other traceless,
Hermitian operators {Fj }Mj=1 , where M = d − 1, such that
2

Tr(Fj ) = 0 , Tr(Fj Fk ) = δjk , Fj† = Fj . (405)

A common choice is the generators of su(d) (just as in the single-qubit case we chose the Pauli matrices), but for our purposes
the explicit form of the operator basis won’t matter. Note that this is similar to what we did in Sec. X A, except that for later
convenience we make our basis choice somewhat more explicit here.
 66

We can now expand any operator in this basis, including the density matrix:
M
1 1
ρ= F0 + ∑ vj Fj = I + F⃗ ⋅ v⃗ , (406)
d j=1 d

where v⃗ = (v1 , . . . , vM )T ∈ RM is called the “coherence vector” (a generalized Bloch vector), and F⃗ = (F1 , . . . , FM ) collects
the operator basis into a vector. Thus the components of the coherence vector are

vj = Tr(ρFj ) . (407)

In analogy to Eq. (168) for the single qubit case, we shall see that as a consequence of the Lindblad equation ρ̇ = Lρ, the
coherence vector satisfies the first order, inhomogeneous differential equation

v⃗˙ = G⃗
v + c⃗ . (408)

Moreover, the decomposition of L as

L = LH + LD , (409)

with

LH [⋅] = −i[H, ⋅] (410a)

1
LD [⋅] = ∑ aij (Fi ⋅ Fj − {Fj Fi , ⋅}) , (410b)
ij 2

induces the decomposition of G into G = Q + R, where LH [ρ] ↝ Q⃗ v and LD [ρ] ↝ R⃗ v + c⃗.

To explain the form of the dissipative term given in Eq. (410b), recall the original form given in Eq. (266). Combine this with
the unitary transformation between the operator basis and the Lindblad operators given in Eq. (295), to see that we can always
transform between the non-diagonal and diagonal forms of the Lindblad equation. This transformation preserves positivity, i.e.,
we know that the coefficient matrix a ≡ (aij ) is positive semi-definite.
Note that the normalization convention we have chosen for the coherence vector is slightly different from the Bloch vector,
since we did not divide v⃗ ⋅ F⃗ by d in Eq. (406). As a result, the coherence vector is confined to a sphere with a radius less than
one. Recall that the purity P = Tr(ρ2 ) [Eq. (64)] satisfies P ≤ 1. Thus
2
1 1 1
1 ≥ Tr(ρ2 ) = Tr [( I + F⃗ ⋅ v⃗) ] = + ∑ Tr(Fi Fj )vi vj = + ∥⃗
v ∥2 , (411)
d d ij d

i.e.,

1 1/2
0 ≤ ∥⃗
v ∥ ≤ (1 − ) . (412)
d
1/2
The upper bound is saturated for pure states, which thus live on the surface of an d2 −1-dimensional sphere with radius (1 − d1 ) .

B. Just the non-dissipative part

Let us assume that LD = 0. In this case we have, starting from Eq. (407):

v̇k = Tr (ρ̇Fk ) = −iTr ([H, F⃗ ⋅ v⃗]Fk ) (413a)

⎛ ⎞
= −iTr ∑(HFj Fk − Fj HFk )vj = i ∑ Tr (H[Fk , Fj ]) vj (413b)
⎝j ⎠ j

= (Q⃗
v )k , (413c)

i.e.,

v⃗˙ = Q⃗
v, (414)
 67

where

Qjk ≡ iTr (H[Fj , Fk ]) . (415)

Note that the appearance of the commutator [Fj , Fk ] is a good reason to use as an operator basis the generators of a Lie algebra,
for which the commutator can be expressed in terms of the algebra’s structure constants.
The matrix M × M dimensional Q is clearly skew symmetric: Qjk = −Qkj , i.e.,

Q = −QT . (416)

The solution of Eq. (414) is

v⃗(t) = eQt v⃗(0) ≡ Ω(t)⃗

v (0) . (417)

The evolution operator Ω is orthogonal:

ΩT Ω = eQ t eQt = e−Qt eQt = I ,

T
(418)

where we used the skew-symmetry of Q. This immediately implies that the norm of the coherence vector is preserved: ∥⃗ v (t)∥2 =
v⃗ (0)Ω Ω⃗
T T
v (0) = ∥⃗
v (0)∥ .
2

Thus, the evolution of the coherence vector in the absence of the dissipative part LD = 0 is a rotation in RM , generated by Q.

C. Full Lindblad equation for the coherence vector

Let us now assume that both LH , LD ≠ 0. Starting again from Eq. (407), and using Eq. (410b), we have:

1 1 1
v )k + ∑ aij Tr [Fi ( I + F⃗ ⋅ v⃗) Fj Fk − {Fj Fi , I + F⃗ ⋅ v⃗}Fk ]
v̇k = (Q⃗ (419a)
ij d 2 d
1 1
v )k + ∑ ∑ aij Tr [(Fj Fk Fi − (Fk Fj Fi + Fj Fi Fk )) Fl ] vl + ∑ aij Tr([Fi , Fj ] Fk )
= (Q⃗ (419b)
l ij 2 d ij
= [(Q + R)⃗
v ]k + ck , (419c)

where
1
Rkl ≡ ∑ aij Tr [(Fi Fl Fj − {Fj Fi , Fl }) Fk ] (420a)
ij 2
1
ck ≡ ∑ aij Tr([Fi , Fj ] Fk ) . (420b)
d ij

Thus, we have established that Eq. (408) holds, with G = Q + R, and with Q, R, and c⃗ as given in Eqs. (415) and (420),
respectively.

D. Solution for diagonalizable and invertible G

Equation (408) is a linear, first order, inhomogeneous differential equation. Solving it is a standard exercise in linear algebra.
For simplicity, let us assume that G is diagonalizable over RM and also invertible. Neither of these assumptions holds in general,
and we deal with the general case in the next subsection.
We look for a solution in the form

v⃗(t) = v⃗(0) (t) + v⃗(∞) , (421)

where v⃗(0) (t) is the homogeneous part and v⃗(∞) is the inhomogeneous part. Let x
⃗(k) and λk represent the eigenvectors and
(possibly degenerate and complex) eigenvalues of G, i.e.,

x(k) = λk x
G⃗ ⃗(k) , k = 1, . . . , M . (422)
 68

It is then straightforward to check by direct differentiation and substitution that

M
v⃗(0) (t) = ∑ sk eλk t x
⃗(k) (423a)
k=1

v⃗(∞) = −G−1 c⃗ (423b)

in the solution of Eq. (408). Indeed:

M
v⃗˙ = G⃗ v (0) (t) + G⃗
v (t) + c⃗ = G⃗ v (∞) + c⃗ = ∑ sk eλk t λk x
⃗(k) − GG−1 c⃗ + c⃗ = v⃗˙ (0) , (424)
k=1

as required. The coefficients sk are determined by the initial condition v⃗(0):

M
v⃗(0) (0) = ∑ sk x
⃗(k) = X s⃗ , ⃗(k) ,
colk (X) = x (425)
k=1

i.e., X is the matrix whose columns are the eigenvectors of G. Also, v⃗(0) (0) = v⃗(0) − v⃗(∞) . Thus

s⃗ = X −1 (⃗
v (0) + G−1 c⃗) . (426)

Now, since the eigenvalues are in general complex numbers, they can be decomposed as λk = R(λk )+iI(λk ). The imaginary
part describes a rotation of the coherence vector (though we can be sure that since this vector lives in RM , such rotations
are ultimately described by an orthogonal (purely real) matrix). The real part is constrained by complete positivity and trace
preservation to be non-positive, or else the norm of the coherence vector would not be bounded [recall Eq. (412)]. Thus,
the overall behavior of the coherence vector is described by rotations at frequencies given by {I(λk )}, some of which are
exponentially damped on a timescale given by the set of non-zero {R(λk )}.

E. Solution for general G

The general case is where G is not diagonalizable over RM , and may not be invertible. In this case we can still use a similarity
transformation S to transform G into Jordan canonical form:

⎛ J1 ⎞
GJ = SGS −1
=⎜
⎜ ⋱ ⎟,
⎟ (427)
⎝ Jq ⎠

where the q Jordan blocks have the form

⎛ µj 1 ⎞
⎜ µj ⋱ ⎟
⎜ ⎟
⎜ ⎟
Jj = ⎜ ⋱ ⋱ ⎟ = µj I + Kj . (428)
⎜ ⎟
⎜ ⎟
⎜ µj 1 ⎟
⎝ µj ⎠

d
The µj ’s are the (possibly degenerate, complex) eigenvalues and Kj are nilpotent matrices: Kj j = 0, where dj is the dimension
of Jj . When all dj = 1, G is diagonalizable and GJ reduces to the diagonalized form of G.
Applying S from the left to Eq. (408) yields

S v⃗˙ = SGS −1 S⃗
v + S⃗
c Ô⇒ w ⃗ + c⃗′ ,
⃗˙ = GJ w (429)

where w⃗ = S⃗v and we defined c⃗′ = S⃗c. This is still a linear, first order, inhomogeneous differential equation. The different Jordan
blocks don’t couple, so we can solve this as a set of q independent problems, and take the direct sum of all the sub-solutions.
Consider first the case of a 2 × 2 Jordan block, i.e., dj = 2. The homogeneous part becomes:

(0) µ 1 (0)
⃗˙ j = ( j
w ⃗j ,
)w (430a)
µj
 69

i.e.,
(0) (0) (0)
ẇj,1 = µj wj,1 + wj,2 (431a)
(0) (0)
ẇj,2 = µj wj,2 . (431b)

(0) (0) (0)

Solving the second of these yields wj,2 (t) = eµj t wj,2 (0), which can be substituted into the first, and solved to yield wj,1 (t) =
(0) (0)
eµj t (wj,1 (0) + wj,2 (0)t).
Similarly, the dj = 3 case yields:
(0) (0) (0)
ẇj,1 = µj wj,1 + wj,2 (432a)
(0) (0) (0)
ẇj,2 = µj wj,2 + wj,3 (432b)
(0) (0)
ẇj,3 = µj wj,3 , (432c)

which is easily solved in the same manner, and gives:

(0) (0)
wj,3 (t) = eµj t wj,3 (0) (433a)
(0) (0) (0)
wj,2 (t) =e µj t
(wj,2 (0) + wj,3 (0)t) (433b)
(0) (0) (0) (0) t2
wj,1 (t) = eµj t (wj,1 (0) + wj,2 (0)t + wj,3 (0) ). (433c)
2!
(0)
⃗j
The general pattern can now be inferred. The solution for a general dj dimensional Jordan block is a vector w =
(0) (0)
⃗j,1 , . . . , w
(w ⃗j,dj )T with components:

dj
(0) (0) tn−k
w ⃗j,n (0)
⃗j,k (t) = eµj t ∑ w , k = 1, . . . , dj . (434)
n=k (n − k)!

The general solution of the homogenous part is then

q
(0)
⃗ (0) (t) = ⊕ w
w ⃗j (t) , (435)
j=1

where the direct sum notation means that the summands need to be joined into a single column vector. The new aspect of the
general G case is thus the appearance of the degree dj − 1 polynomials in t. These polynomials induce an additional non-trivial
time-dependence in addition to the rotations and exponential decay we found for the case of diagonalizable G. Note that we
can be certain that for all dj > 1 [when the degree of the polynomial in Eq. (434) is ≥ 1], the corresponding R(µj ) < 0, since a
positive or zero real part would violate the general norm upper bound (412).
As for the inhomogeneous part, we can write the solution of Eq. (429) as

⃗
w(t) ⃗ (0) (t) + w
=w ⃗ (∞) , (436)

and find the particular solution that satisfies

⃗ (∞) = −⃗
GJ w c′ . (437)

Depending on the rank r(G) of G, this equation has either zero [r(G) = 1], one [r(G) = M ], or infinitely many [0 < r(G) < M ]
solutions. The first case is unphysical, the second is unproblematic, and for the third every initial condition still determines a
corresponding final state in a unique way.

F. Phase Damping Example

As a simple example meant to illustrate how we construct and solve the differential equation for the coherence vector, assume
that a single qubit is subject to a magnetic field along the z direction along with dephasing:

ρ̇ = −i[ωZ, ρ] + γ(ZρZ − ρ) . (438)

 70

As a fixed operator basis satisfying the conditions in Eq. (405), we choose the Pauli matrices:
√
Fj = σ j / 2 , (439)

with the normalization due to the requirement that Tr(Fi Fj ) = δij . The Q matrix elements [Eq. (415)] are then
1
Qjk = iω Tr (Z[σj , σk ]) , (440)
2
and are non-vanishing only when [σj , σk ] ∝ Z, i.e., [X, Y ] = 2iZ and [Y, X] = −2iZ. Therefore Q12 = 2ω = −Q12 , and all
other Q matrix elements are zero.
Next, we need to calculate the R matrix and the c⃗ vector, using Eq. (420). Note that, in this case, only a33 = 2γ is non-zero in
the a-matrix of the Lindblad equation (the factor of 2 is due to the normalization of the F ’s). Therefore ∑ij reduces to just the
term with i = j = 3:
1
ck = γTr([Z, Z] σk ) = 0 (441a)
23/2 d
1 1 1
Rkl = Tr (Zσl Zσk − {Z 2 , σl } σk ) = γTr (Zσl Zσk − σl σk ) . (441b)
4 2 4
Clearly, σk must equal σl in order for the trace to be non-zero. When σk = σl = X, or when σk = σl = Y , we get 14 γTr(−I − I) =
−γ, whereas when σk = σl = Z we get 0. Thus R = diag(−γ, −γ, 0). Combining with the result for Q, we have:

⎛ −γ 2ω 0 ⎞
G=⎜ ⎟
⎜ −2ω −γ 0 ⎟ . (442)
⎝ 0 0 0⎠

This G matrix is diagonalizable but not invertible (its rank is 2), so we are in a scenario that is in between that of Secs. XIII D
and XIII E. Non-invertibility only affects the existence of the limit of v⃗(t) as t → ∞. Since G is diagonalizable, all its Jordan
blocks have dimension dj = 1, i.e., they are simply the eigenvalues. The eigenvalues are −γ ± 2iω and 0. This corresponds to
a coherence vector rotating at angular frequency 2ω in the X − Y plane, while exponentially decaying towards the Z axis with
rate γ. This means that the entire Z axis is the limit as t → ∞, hence there is no unique final state. However, every initial state
decays to a unique final state (its projection onto the Z axis).

XIV. DERIVATION OF THE LINDBLAD EQUATION FROM THE CUMULANT EXPANSION AND COARSE GRAINING

We now present a derivation of the Lindblad equation (LE) from first principles, following Ref. [17]. This derivation avoids
the so-called rotating wave approximation (RWA), which is the most commonly used approach to deriving the LE. We shall
return to an RWA-based approach later.

A. Cumulant expansion

Let λ be a small, dimensionless parameter, and consider the Hamiltonian

H = HS + HB + λHSB (443)

with

HSB = A ⊗ B (444)

where A is a Hermitian system operator and B is a Hermitian bath operator. We have restricted ourself to a single term to
simplify the notation, but the more general case with multiple terms follows in an analogous fashion.
Define:

H0 ≡ HS ⊗ IB + IS ⊗ HB , (445a)
−itHS −itHB
U0 (t) ≡ exp (−itH0 ) = US (t) ⊗ UB (t) = e ⊗e , (445b)
ρ̃SB (t) ≡ U0† (t)ρSB (0)U0 (t) , (445c)
 71

where ρ̃SB (t) is the state in the interaction picture (recall Sec. X B). We have the interaction picture Hamiltonian

H̃(t) = U0† (t)HSB U0 (t) = US† (t)AUS (t) ⊗ UB† (t)BUB (t) ≡ A(t) ⊗ B(t) . (446)
The density matrix in the interaction picture satisfies
d
ρ̃SB (t) = −i [λH̃(t), ρ̃SB (t)] , (447)
dt
which we can solve formally by integration followed by substitution and iteration:
t
ρ̃SB (t) = ρSB (0) − i ∫ ds [λH̃(s), ρ̃SB (s)] (448a)
0
t t s
= ρSB (0) − iλ ∫ ds [H̃(s), ρSB (0)] + (−iλ)2 ∫ ds ∫ ds′ [H̃(s), [H̃(s′ ), ρSB (0)]] + ⋯ , (448b)
0 0 0

and it is clear how this continues. A simple norm estimate (see Sec. ??) shows that the norm of the nth order term is
O[(∥HSB ∥t)n ]. Therefore a sufficient convergence condition is λ∥HSB ∥t < 1. Terms of third order and above can be ne-
glected provided λ∥HSB ∥t ≪ 1. This is known as the Born approximation.
We are interested in the reduced density matrix:
ρ̃(t) = TrB [ρ̃SB (t)] ≡ Λλ (t)ρ(0) . (449)

The cumulant expansion is given by introducing unknown, to be determined operators K (n) in the exponent:
∞
Λλ (t) = exp ( ∑ λn K (n) (t)) (450a)
n=1
1 2
= I + λK (1) (t) + λ2 (K (2) (t) + (K (1) (t)) ) + O(λ3 ) , (450b)
2
where in the second line we used a Taylor expansion of the exponential. We solve for K (n) by matching powers of λ with
Eq. (448). We get:
t
K (1) (t)ρ(0) = −i ∫ ds TrB ([H̃(s), ρSB (0)]) . (451)
0

We will see later that, without loss of generality, this can always be made to vanish (for a stationary bath) by shifting the operator
B, i.e:
K (1) (t)ρ(0) = 0 . (452)
The next order in λ gives:
t s
K (2) (t)ρ̃(0) = − ∫ ds ∫ ds′ TrB ([H̃(s), [H̃(s′ ), ρSB (0)]]) . (453)
0 0

Expanding the double commutator gives:

TrB ([H̃(s), [H̃(s′ ), ρSB (0)]]) = [A(s)A(s′ )ρ(0) − A(s′ )ρ(0)A(s)] Tr [B(s)B(s′ )ρB ] + h.c. (454a)
′ ′ ′
= [A (s)A(s )ρ(0) − A(s )ρ(0)A (s)] Tr [B (s)B(s )ρB ] + h.c.
† † †
(454b)
′ ′ ′
= [A (s)A(s )ρ(0) − A(s )ρ(0)A (s)] B(s, s ) + h.c.
† †
(454c)
where
B(s, s′ ) ≡ ⟨B † (s)B(s′ )⟩ = B(s′ , s)∗ , (455)
and
⟨X⟩B ≡ Tr[ρB X] , (456)
and ρB is, e.g., the thermal (Gibbs) state of the bath [Eq. (237)]. Equation (455) holds since:
⟨B † (s)B(s′ )⟩ = Tr[ρB (0)B † (s)B(s′ )] = (Tr[B(s′ )† B(s)ρB (0)]† )∗ = (Tr[ρB (0)B(s′ )† B(s)]† )∗ = ⟨B † (s′ )B(s)⟩∗ .
(457)
 72

B. The second order cumulant

It turns out to be convenient to express the interaction picture system operator A(t) in the frequency domain. To do so, let us
first expand HS in its eigenbasis:

HS = ∑ εa ∣εa ⟩⟨εa ∣ , (458)

a

where {εa } are the eigenenergies of HS . Thus

A(t) = US† (t)AUS (t) = ∑ e−i(εb −εa )t ∣εa ⟩⟨εa ∣A∣εb ⟩⟨εb ∣ = ∑ Aω e−iωt , (459)
a,b ω

where ω ≡ εb − εa is a Bohr frequency, and

Aω ≡ ∑ ⟨εa ∣A∣εb ⟩∣εa ⟩⟨εb ∣ . (460)

εb −εa =ω

To clarify, the sum over εb − εa = ω in Eq. (460) is over all pairs of eigenenergies {εb , εa } whose difference gives the same Bohr
frequency ω. The sum over ω in Eq. (459) is a sum over all Bohr frequencies (negative, zero, and positive). This then gives the
following map from time 0 to t:

K (2) (t)ρ(0) = ∑ Bωω′ (t) (Aω ρ(0)A†ω′ − A†ω′ Aω ρ(0)) + h.c., (461)
ω,ω ′

where
t s
Bωω′ (t) ≡ ∫ ds′ ei(ω s−ωs ) B(s, s′ ) .
′ ′
ds ∫ (462)
0 0

We will see that Eq. (461) can be rewritten in the form of a Lindblad generator:
1
K (2) (t)ρ(0) = −i [Q(t), ρ(0)] + ∑ bωω′ (t) [Aω ρ(0)A†ω′ − {A†ω′ Aω , ρ(0)}] , (463)
ω,ω ′ 2

where the elements of the matrix b(t) are given by

t t
bωω′ (t) ≡ ∫ ds′ ei(ω s−ωs ) B(s, s′ ) = b∗ω′ ω (t) ,
′ ′
ds ∫ (464)
0 0

and we will show that b(t) is positive semi-definite.

The “Lamb shift” term is

Q(t) = ∑ Qωω′ (t)A†ω′ Aω , (465)

ω,ω ′

where

i
Qωω′ (t) = − (Bωω′ − Bω∗ ′ ω ) (466a)
2
i t s
= − ∫ ds ∫ ds′ (ei(ω s−ωs ) B(s, s′ ) − e−i(ωs−ω s ) B(s′ , s)) .
′ ′ ′ ′
(466b)
2 0 0
∗
Note that (Qωω′ ) = Qω′ ω , so that Q† = Q, as required for the interpretation of Q as a Hamiltonian.

C. Why the first order cumulant can be made to vanish

We argued [Eq. (452)] that we can shift the bath operator B such that K (1) (t)ρ(0) = 0. Here we show why.
Let ρB (0) = ∑µ λµ ∣µ⟩⟨µ∣ and

Bd (t) ≡ diag(B(t)) = ∑ Bµµ (t)∣µ⟩⟨µ∣ , (467)

µ
 73

i.e., the diagonal part of B in the eigenbasis of ρB (0). Here Bµµ (t) = ⟨µ∣ B(t) ∣µ⟩. Let us define a new bath operator

B ′ (t) ≡ B(t) − Bd (t) . (468)

Then

⟨B ′ (t)⟩ = ⟨B(t)⟩ − ⟨Bd (t)⟩ = ∑ λµ ⟨µ∣ B(t) ∣µ⟩ − ∑ λµ ⟨µ∣ [∑ Bνν (t)∣ν⟩⟨ν∣] ∣µ⟩ = 0 . (469)
µ µ ν

′
Let HSB = A ⊗ B ′ , so that H̃ ′ (t) = U0† (t)HSB
′
U0 (t). Then

TrB ([H̃ ′ (t), ρSB (0)]) = TrB ([A(t) ⊗ B ′ (t), ρS (0) ⊗ ρB (0)]) = ⟨B ′ (t)⟩[A(t), ρS (0)] = 0 . (470)

Therefore, K ′(1) (t)ρ(0) = 0, with K ′(1) defined with the modified system-bath interaction HSB ′
. The price we have to pay
for this is the shift of B to B . This shift manifests itself only through the bath correlation function B(s, s′ ) [Eq. (455)]. The
′

shifted correlation function becomes B ′ (s, s′ ) = ⟨B † (s)B ′ (s′ )⟩, and nothing else changes, since the bath operators only appear
′

through the bath correlation function.

D. Derivation of the Lindblad equation

We will now prove that Eq. (461) can be transformed into Eq. (463). It turns out that the unequal upper integration limits
in Bωω′ [Eq. (462)] are problematic, while the equal upper integration limits in bωω′ [Eq. (464)] are what allows us to prove
complete positivity, as we show in Sec. XIV E directly below. To replace the unequal upper limits by equal limits we note the
following relations for the integral, where for notational simplicity we suppress the t-dependence for now:
t s t t t t
Bωω′ ≡ ∫ ds′ ei(ω s−ωs ) B(s, s′ ) , = [∫ ds′ − ∫ ds′ ] ei(ω s−ωs ) B(s, s′ ) ,
′ ′ ′ ′
ds ∫ ds ∫ ds ∫ (471a)
0 0 0 0 0 s
t t t s′
= [∫ ds′ − ∫ ds′ ∫ ds] ei(ω s−ωs ) B(s, s′ ) ,
′ ′
ds ∫ (471b)
0 0 0 0
t t t s
=∫ ds′ ei(ω s−ωs ) B(s, s′ ) − ∫ ds′ ei(ω s −ωs) B(s′ , s) ,
′ ′ ′ ′
ds ∫ ds ∫ (471c)
0 0 0 0
= bωω′ − Bω∗ ′ ω , (471d)

where bωω′ [Eq. (464)] has the desired equal upper integration limits. It follows immediately that
∗
Bωω ′ = bω ′ ω − B ω ′ ω . (472)

Therefore, the first summand +h.c. in Eq. (461) yields:

∗ ∗
∑ [Bωω′ Aω ρAω′ + Bωω′ Aω′ ρAω ] = ∑ [bωω′ Aω ρAω′ + bω′ ω Aω′ ρAω − (Bω′ ω Aω ρAω′ + Bω′ ω Aω′ ρAω )] ,
† † † † † †
(473a)
ω,ω ′ ω,ω ′
∗
= ∑ [bωω′ Aω ρA†ω′ + bω′ ω Aω′ ρA†ω − (Bωω ′ Aω ′ ρAω + Bωω ′ Aω ρA ′ )] ,
† †
ω (473b)
ω,ω ′

where in the second term on the RHS we have switched ω ↔ ω ′ , which is permissible since we are summing over all ω and ω ′ .
Furthermore, this second term is now exactly in the form of the original term, so we have the result:

∗ 1
∑ [Bωω′ Aω ρAω′ + Bωω′ Aω′ ρAω ] = ∑ [bωω′ Aω ρAω′ + bω′ ω Aω′ ρAω ] = ∑ bωω′ Aω ρAω′ .
† † † † †
(474)
ω,ω ′ 2 ω,ω′ ω,ω ′

The second summand +h.c. in Eq. (461) is of the form −A†ω′ Aω ρ(0), which reminds us of the anti-commutator term in the
Lindblad equation, except that it doesn’t have the factor of 1/2. However, note that since b∗ω′ ω = bωω′ = Bωω
∗
′ + Bω ′ ω , where we
∗
used Eq. (472). Therefore by writing bωω′ = 2 (bωω′ + bω′ ω ), and again using Eq. (472), we have:
1

1 1
Bωω′ = bωω′ + (Bωω′ − Bω∗ ′ ω ) . (475)
2 2
 74

This allows us to write the second summand +h.c. in Eq. (461) as:
∗
− ∑ [Bωω′ A†ω′ Aω ρ + Bωω ′ ρAω Aω ′ ]
†
(476a)
ω,ω ′
1 1 ∗ ∗
=− ∑ (bωω′ Aω′ Aω ρ + bω′ ω ρAω Aω′ ) − ∑ [(Bωω′ − Bω′ ω ) Aω′ Aω ρ + (Bω′ ω − Bω′ ω ) ρAω Aω′ ]
† † † †
(476b)
2 ω,ω′ 2 ω,ω′
1 1 ∗
=− ∑ bωω′ (Aω′ Aω ρ + ρAω′ Aω ) − ∑ (Bωω′ − Bω′ ω ) [Aω′ Aω ρ − ρAω′ Aω ]
† † † †
(476c)
2 ω,ω′ 2 ω,ω′
1 1 ∗
=− ∑ bωω′ {Aω′ Aω , ρ} − ∑ (Bωω′ − Bω′ ω ) [Aω′ Aω , ρ] .
† †
(476d)
2 ω,ω′ 2 ω,ω′
We can now write the RHS of Eq. (461) as:
∑ Bωω′ (t) (Aω ρ(0)Aω′ − Aω′ Aω ρ(0)) + h.c.
† †
(477a)
ω,ω ′
1 1 ∗
= ∑ bωω′ Aω ρ(0)A†ω′ − ∑ bωω′ {Aω′ Aω , ρ(0)} − ∑ (Bωω′ − Bω′ ω ) [Aω′ Aω , ρ(0)]
† †
(477b)
ω,ω ′ 2 ω,ω′ 2 ω,ω′
−i 1
= −i[ ∑ (Bωω′ − Bω∗ ′ ω ) [A†ω′ Aω , ρ(0)]] + ∑ bωω′ (t)[Aω ρ(0)A†ω′ − {A†ω′ Aω , ρ(0)} ] , (477c)
ω,ω ′ 2 ω,ω ′ 2
which is Eq. (463), together with the identification of the term in the commutator as the Lamb shift Q(t) as defined in Eq. (465).

E. Complete positivity

Clearly, the dissipative (second) term on the RHS of Eq. (463) appears to be in Lindblad form, but we must still prove the
positivity of the matrix b(t). To this end we again expand the bath density matrix in its eigenbasis, and use this to write the
correlation function B(s, s′ ) = ⟨B † (s)B(s′ )⟩B explicitly. Let v⃗ be some arbitrary vector; then positivity amounts to showing
that v⃗b(t)⃗
v † ≥ 0 for all v⃗. Indeed:
t t
v † = ∑ vω bωω′ (t)vω∗ ′ = ∫
v⃗b(t)⃗ ds′ ∑(vω e−iωs ) ∑(vω′ e−iω s )∗ Tr[∑ λµ ∣µ⟩ ⟨µ∣ B † (s)B(s′ )]
′ ′
ds ∫ (478a)
ωω ′ 0 0 ω ω′ µ

= ∑ λµ ⟨µ∣ F † (t)F (t) ∣µ⟩ = ∑ λµ ∥F (t) ∣µ⟩ ∥2 ≥ 0, (478b)

µ µ

where F (t) ≡ ∫0 dsB(s) ∑ω vω∗ e−iωs . Note how it was crucial in this proof that the upper limits of the integrals are the same,
t

since otherwise the factorization would have failed.

Therefore, our quantum map is given by:
2
K (2) (t)
ρ̃(t) = eλ ρ(0) . (479)
2
The only approximation we have introduced so far is the truncation at order λ , i.e., the Born approximation. The CP map (479)
is in principle already sufficient, and one can use it to compute Kraus operators. However, in order to find the time-dependent
system state ρ̃(t) one has to compute eλ K (t) for each t, which is laborious. In order to arrive at a master equation, with the
2 (2)

associated advantages (e.g., a quantum trajectories unravelling) we need to introduce an additional, Markovian approximation.

F. LE from the cumulant expansion and coarse-graining

Let us show how to obtain the LE from the results above. Expanding the exponential in Eq. (479) to second order in λ, we
have:
1
ρ̃(t) − ρ̃(0) = −i [λ2 Q(t), ρ(0)] + ∑ λ2 bωω′ (t) [Aω ρ(0)A†ω′ − {A†ω′ Aω , ρ(0)}] . (480)
ω,ω ′ 2
It is straightforward to check that Q(0) = bωω′ (0) = 0 (due to the upper integration limit being 0). Therefore, dividing both sides
of Eq. (480) by τ , and setting t = τ , we have:

˙ 0 = −i [λ2 ⟨Q̇⟩0 , ρ(0)] + ∑ λ2 ⟨ḃωω′ ⟩0 [Aω ρ(0)A† ′ − 1

⟨ρ̃⟩ ω {A†ω′ Aω , ρ(0)}] , (481)
ω,ω ′ 2
 75

where we used the coarse-graining definition, Eq. (310).

Similarly to Sec. X, the path to the Lindblad equation is to now introduce a Markovian assumption in terms of the coarse-
graining timescale τ . The Markovian assumption amounts to assuming that both ⟨Q̇⟩0 and ⟨ḃωω′ ⟩0 are constant for all t, i.e., that
⟨Q̇⟩j = ⟨Q̇⟩0 and ⟨ḃωω′ ⟩j = ⟨ḃωω′ ⟩0 for all j. This can be rigorously justified by first assuming that the bath correlation function
is translationally invariant, i.e., B(s, s′ ) = B(s − s′ ). This is true for stationary baths. A bath is stationary if

[HB , ρB (0)] = 0 , (482)

which implies that ρB (t) = UB (t)ρB (0)UB† (t) = ρB (0). This is the case, e.g., if ρB (0) = e−βHB /Z, i.e., is a Gibbs state. In
addition we assume that the bath correlation function decays over a timescale τB , i.e.,

B(t) ∼ e−t/τB , (483)

while the coarse graining is done over a much longer timescale, so that the integrand in Eq. (464) has already decayed. The RHS
of Eq. (518) is then valid for all times, allowing us to also shift the time argument of ρ to arbitrary jτ . Let us now define the
Lamb-shift and the Lindblad rates as:

HLS ≡ λ2 ⟨Q̇⟩0 , (484a)

γωω′ ≡ λ ⟨ḃωω′ ⟩0 .
2
(484b)

Moreover, we assume that τ is very small on the timescale τS over which ρ(t) changes, so that ⟨ρ̃⟩
˙ j = [ρ̃((j + 1)τ ) − ρ̃(jτ )]/τ
can be replaced by ρ̇(t). These assumptions can be summarized as

τB ≪ τ ≪ τS . (485)

We can thus write the interaction picture Lindblad equation in the final form:
1
˙
ρ̃(t) = −i [HLS , ρ(t)] + ∑ γωω′ [Aω ρ(t)A†ω′ − {A†ω′ Aω , ρ(t)}] . (486)
ω,ω ′ 2

The RHS contains the free parameter τ , which can be determined using Eq. (485). Everything else is determined in terms of the
given specification of the Hamiltonian H = HS + HB + HSB and the initial state of the bath ρB (0). In particular,
• The Bohr frequencies ω are determined by HS ;
• The Lindblad operators are determined by the system operator A in HSB and the Bohr frequencies (i.e., HS );
• The bath correlation function B(s, s′ ) is determined by the bath operator B in HSB , the bath Hamiltonian HB (which
determines the time-dependence of B(t)), and the initial bath state ρB ;
• The Lamb shift is determined by the bath correlation function and the Bohr frequencies.

G. Illustration using the spin-boson model for phase damping

Consider once more the spin-boson model defined in Sec. X C. Let us denote the eigenvalues of HS = −(g/2)Z by ε± = ±g/2
and their respective eigenvectors by ∣ε− ⟩ = ∣0⟩ (ground state) and ∣ε+ ⟩ = ∣1⟩ (excited state). Using Eq. (460), the Lindblad
operators are then given by:
A−g = ∣ε+ ⟩⟨ε+ ∣Z∣ε− ⟩⟨ε− ∣ = 0 (487)
A0 = ∣ε+ ⟩⟨ε+ ∣Z∣ε+ ⟩⟨ε+ ∣ + ∣ε− ⟩⟨ε− ∣Z∣ε− ⟩⟨ε− ∣ = Z (488)
Ag = ∣ε− ⟩⟨ε− ∣Z∣ε+ ⟩⟨ε+ ∣ = 0 . (489)
Thus, only the (elastic, or on-shell) ω = 0 term contributes to the sums over ω. This means that the Lamb shift is given by:

λ2 λ2
HLS = Q(τ ) = Q00 (t)A†0 A0 ∝ I , (490)
τ τ
so that [HLS , ρ̃(t)] = 0. The dissipative part of the LE [Eq. (486)] is given by:
1 1
∑ γω,ω′ [Aω ρAω′ −
†
{A†ω′ Aω , ρ}] = γ00 (Z ρ̃Z − {I, ρ̃}) = γ (Z ρ̃Z − ρ̃) , (491)
ω,ω ′ 2 2
 76

where
λ2 λ2 τ τ
′ ′
γ ≡ γ00 = b00 (τ ) = ∫ ds ∫ ds B(s, s ) , (492)
τ τ 0 0

and where we used Eq. (464). We already computed this decay rate when we solved the spin-boson model analytically, and
found it in Eq. (374). The result after coarse graining is given in Eq. (333).
While we already saw the solution of the corresponding LE in Sec. IX D 2, let us solve it again using a nice and useful
“vectorization” trick. Let us define:

⎛ col1 (ρ) ⎞
⎜ ⋮ ⎟
vec(ρ) ≡ ⎜
⎜
⎟
⎟ (493)
⎜ colj (ρ) ⎟
⎝ ⋮ ⎠

i.e., vec(ρ) corresponds to stacking the columns of ρ (in some basis). We now use the identity [18]:

vec (ABC) = (C T ⊗ A) vec (B) (494)

where (A, B, C) are arbitrary matrices of appropriate dimensions allowing their multiplication. Using this, we can write the LE
ρ̃˙ = γ(Z ρ̃Z − I ρ̃I) as
vec (ρ̃)
˙ = γ (Z ⊗ Z − I ⊗ I) vec(ρ̃) ≡ Lvec(ρ̃) . (495)
Conveniently, L is diagonal with entries (0, −2γ, −2γ, 0), so we can immediately write:

⎛1 ⎞ ⎛ ρ00 (0) ⎞
⎜ exp(−2γt) ⎟ ⎜ ρ10 (0) ⎟
vec(ρ̃(t)) = exp(L)vec(ρ̃(0)) = ⎜
⎜
⎟⎜
⎟⎜
⎟
⎟ (496)
⎜ exp(−2γt) ⎟ ⎜ ρ01 (0) ⎟
⎝ 1 ⎠ ⎝ ρ11 (0) ⎠
Therefore, we find as before:

ρ00 (0) exp(−2γt)ρ01 (0)

ρ̃(t) = ( ) (497)
exp(−2γt)ρ10 (0) ρ11 (0)
Transforming back to the Schrödinger picture, the result is adjusted to

ρ00 (0) exp(−2γt − igt)ρ01 (0)

ρ̃(t) = ( ). (498)
exp(−2γt + igt)ρ10 (0) ρ11 (0)

XV. FIRST-PRINCIPLES DERIVATION OF THE LINDBLAD EQUATION FROM THE BORN, MARKOV, AND ROTATING
WAVE APPROXIMATIONS

We now present our last derivation of the Lindblad equation. This is the standard approach found in textbooks such as [1], but
we will add some clarifications concerning the limitations of the validity of this approach. We will also discuss the differences
between this and the cumulant-based approach.

A. Setting up

Our starting point is identical to the one we used in the cumulant expansion approach (Sec. XIV A). The only difference is
that we now consider the more general system-bath interaction
HSB = g ∑ Aα ⊗ Bα , (499)
α

where g has units of energy. Thus, in the interaction picture:

H̃(t) = g ∑ Aα (t) ⊗ Bα (t) , (500)
α
 77

and

Aα (t) = US† (t)Aα US (t) , US (t) = e−iHS t (501a)

−iHB t
Bα (t) = UB† (t)Bα UB (t) , UB (t) = e . (501b)

Formally integrating the Liouville-von Neumann equation

d
ρ̃SB (t) = −i [H̃(t), ρ̃SB (t)] , (502)
dt
we have:
t
ρ̃SB (t) = ρSB (0) − i ∫ ds [H̃(s), ρ̃SB (s)] . (503)
0

Let us now substitute this solution back into Eq. (502) and take the partial trace:
d d t
ρ̃(t) = TrB { ρ̃SB (t)} = −iTrB {[H̃(t), ρSB (0)]} + (−i)2 TrB {[H̃(t), ∫ ds [H̃(s), ρ̃SB (s)]]} . (504)
dt dt 0

Just as we argued in Sec. XIV C, the first order term can again be made to vanish provided we shift the bath operators. We are
thus left with
d t
ρ̃(t) = −TrB {[H̃(t), ∫ ds [H̃(s), ρ̃SB (s)]]} . (505)
dt 0

t 0 t
Let us change variables to τ = t − s, so that ∫0 ds = − ∫t (−dτ ) = ∫0 dτ , and:
d t
ρ̃(t) = −TrB {[H̃(t), ∫ dτ [H̃(t − τ ), ρ̃SB (t − τ )]]} . (506)
dt 0

B. Born approximation

To proceed we now make our first approximation. For a sufficiently large bath that is in particular much larger than the system,
it is reasonable to assume that while the system undergoes non-trivial evolution, the bath remains unaffected, and hence that the
state of the composite system at time t is

ρ̃SB (t) = ρ̃(t) ⊗ ρB (0) + χ(t) ≈ ρ̃(t) ⊗ ρB , (507)

where ρB is the time-independent, stationary bath state, and the correlations χ(t) can be neglected. This is (again) called the
Born approximation.
Using this and Eq. (500), we have:
dρ̃ t
= −g 2 ∑ TrB {[Aα (t) ⊗ Bα (t), ∫ dτ [Aβ (t − τ ) ⊗ Bβ (t − τ ), ρ̃(t − τ ) ⊗ ρB ]]} . (508)
dt α,β 0

Let’s expand the double commutator:

TrB [Aα (t) ⊗ Bα (t), [Aβ (t − τ ) ⊗ Bβ (t − τ ), ρ̃(t − τ ) ⊗ ρB ]] (509a)

= Aα (t)Aβ (t − τ )ρ̃(t − τ )Tr[Bα (t)Bβ (t − τ )ρB ] (509b)
− Aβ (t − τ )ρ̃(t − τ )Aα (t)Tr[Bβ (t − τ )ρB Bα (t)] (509c)
− Aα (t)ρ̃(t − τ )Aβ (t − τ )Tr[ρB Bβ (t − τ )Bα (t)] (509d)
+ ρ̃(t − τ )Aβ (t − τ )Aα (t)Tr[Bα (t)ρB Bβ (t − τ )] . (509e)

We now assume again that the bath is stationary (i.e., [ρB , HB ] = 0). As in Eq. (456), let ⟨X⟩B ≡ Tr[ρB X]. Similarly to
Eq. (457), we define the bath two-point correlation function:

Bαβ (t, t − τ ) ≡ ⟨Bα (t)Bβ (t − τ )⟩B = Tr(eiHB t Bα e−iHB t eiHB (t−τ ) Bβ e−iHB (t−τ ) ρB ) (510a)
−iHB (t−τ ) iHB t −iHB t iHB (t−τ ) −iHB τ
= Tr(e e Bα e e Bβ ρB ) = Tr(e iHB τ
Bα e Bβ ρB ) = ⟨Bα (τ )Bβ ⟩B (510b)
= Bαβ (τ, 0) ≡ Bαβ (τ ) , (510c)
 78

where we used the bath stationarity assumption to go the second line, and in the third line we denoted Bαβ (τ, 0) by Bαβ (τ )
for simplicity, since only the time shift τ matters, so we can measure everything from t = 0. Thus, Bαβ (τ ) measures the
autocorrelation of the bath after time τ . Note that in the Bαβ (τ ) notation we implicitly associate t = τ with the first index (in
this case α), whereas the second index is associated with t = 0. Also,
∗
Bβα (τ ) = Tr[(ρB Bβ (τ )Bα )† ] = Tr[Bα (UB† (τ )Bβ UB (τ ))† ρB ] = Tr[ρB Bα UB† (τ )Bβ UB (τ )] = Bαβ (0, τ ) (511a)
= Tr[ρB UB (τ )Bα UB† (τ )Bβ ] = Bαβ (−τ ) . (511b)
Then, noting that the terms in lines (509b) and (509e) are Hermitian conjugates, as are the terms in lines (509c) and (509d),
we have:
dρ̃ t
= −g 2 ∑ ∫ dτ {Bαβ (τ ) [Aα (t), Aβ (t − τ )ρ̃(t − τ )] + h.c.} . (512)
dt αβ 0

C. Markov approximation and Redfield equation

Note that the RHS of Eq. (512) depends on the entire history of the system state, since the argument of ρ̃(t − τ ) ranges from t
to 0 as τ increases from the lower to the upper limit of the integral. Thus, Eq. (512) is time-nonlocal. We would like to arrive at
a time-local differential equation for the system state, which depends only on t, but not on the state’s history.
To attain this, at this point we need to introduce our second approximation, the Markov approximation. Informally, it states that
the bath has a very short correlation time τB , i.e., that the correlation function Bαβ (τ ) decays rapidly with some characteristic
timescale τB , e.g., ∣Bαβ (τ )∣ ∼ e−τ /τB . We also assume that
g ≪ 1/τB , t ≫ τB . (513)
The first of these is a weak-coupling limit (g is small), and the second states that we do not expect our approximation to be
accurate for times t that are comparable to the bath correlation time (instead, we only consider times much larger than the latter).
Now, since the correlation function Bαβ (τ ) is essentially zero for τ ≫ τB , and since we assume that t ≫ τB , we can replace
ρ̃(t − τ ) by ρ̃(t), since the short “memory” of the bath correlation function causes it to keep track of events only within the short
period [0, τB ]. Under this approximations, Eq. (512) becomes:
dρ̃ t
= −g 2 ∑ ∫ dτ {Bαβ (τ )[Aα (t), Aβ (t − τ )ρ̃(t)] + h.c.} , (514)
dt α,β 0

which is known as the Redfield equation. It is notoriously non-CP, which means that the density matrix can be become non-
positive (though various fixes have been proposed [19, 20]).
Moreover, for the same reason (correlation function negligible for τ ≫ τB ) we can extend the upper limit of the integral to
infinity without changing the value of the integral.
dρ̃ ∞
= −g 2 ∑ ∫ dτ {Bαβ (τ )[Aα (t), Aβ (t − τ )ρ̃(t)] + h.c.} + O(g 4 τB3 ) , (515)
dt α,β 0

That Eq. (512) can be replaced by Eq. (515) can be proven rigorously under the following sufficient condition [21], as we will
show in Sec. XVIII D:
∞
∫ τ n ∣Bαβ (τ )∣dτ ∼ τBn+1 , n ∈ {0, 1, 2} . (516)
0

This is satisfied, e.g., by an exponentially decaying correlation function. Indeed:

∞ dn ∞ dn ∞ dn
∫ τ n e−τ /τB dτ = ∫ e−τ /τB dτ = ( −τB e−τ /τB ∣0 ) = τB = n!τBn+1 . (517)
0 d(−1/τB ) 0
n d(−1/τB ) n d(−1/τB )n

More generally, if ∣Bαβ (τ )∣ ∼ e−(τ /τB ) where k > 0, we have:

k

∞ 1 n + 1 n+1
∫ τ n ∣Bαβ (τ )∣dτ = Γ( ) τB , (518)
0 k k
where Γ(x) is the gamma function [recall that Γ(n + 1) = n! for n ∈ N]. Thus, in fact even a subexponential (k < 1) decay will
suffice.
Note that thanks to Eq. (516), the integral in Eq. (515) is of order τB . Thus the ratio between the leading order correction and
the integral is (g 4 τB3 )/(g 2 τB ) = (gτB )2 ≪ 1, by our assumption that gτB ≪ 1.
 79

D. Going to the frequency domain

After dropping the correction term, Eq. (515) is now a differential equation for ρ̃(t), but is not yet in Lindblad form. To
convert it into this form we once again convert the system operators A(t) to the frequency domain. The procedure is essentially
the same as in Sec. XIV B, except that we need to keep track of the system operator index as well. Thus, after expanding HS in
its eigenbasis as HS = ∑a εa ∣εa ⟩⟨εa ∣, we have

Aα (t) = US† (t)Aα US (t) = ∑ e−i(εb −εa )t ∣εa ⟩⟨εa ∣Aα ∣εb ⟩⟨εb ∣ = ∑ Aα (ω)e−iωt , (519)
a,b ω

where ω ≡ εb − εa is a Bohr frequency, and

Aα (ω) ≡ ∑ ⟨εa ∣Aα ∣εb ⟩∣εa ⟩⟨εb ∣ = Aα (−ω) ,

†
(520)
εb −εa =ω

where the last equality follows since Hermitian conjugation interchanges εa and εb . Also, note that since Aα (t) is Hermitian,
−iωt
∑ Aα (ω)e = ∑ A†α (ω)eiωt . (521)
ω ω

Returning to Eq. (515), consider the two terms in the commutator [Aα (t), Aβ (t − τ )ρ̃(t)]:

Aα (t)Aβ (t − τ )ρ̃(t) = ∑ eiω t e−iω(t−τ ) A†α (ω ′ )Aβ (ω)ρ̃(t) = ∑ eiωτ ei(ω −ω)t A†α (ω ′ )Aβ (ω)ρ̃(t)
′ ′
(522a)
ωω ′ ωω ′

Aβ (t − τ )ρ̃(t)Aα (t) = ∑ e−iω(t−τ ) eiω t Aβ (ω)ρ̃(t)A†α (ω ′ ) = ∑ eiωτ ei(ω −ω)t Aβ (ω)ρ̃(t)A†α (ω ′ ) .

′ ′
(522b)
ωω ′ ωω ′

The entire τ -dependence is thus in the factor eiωτ , which motivates collecting everything that is τ -dependent in Eq. (515) into
one function:
∞
Γαβ (ω) ≡ ∫ dτ eiωτ Bαβ (τ ) , (523)
0

which is the one-sided Fourier transform of the bath correlation function. This allows us to rewrite Eq. (515) as
dρ̃
= −g 2 ∑ ∑ {Γαβ (ω)ei(ω −ω)t [A†α (ω ′ ), Aβ (ω)ρ̃(t)]} + h.c.
′
(524)
dt α,β ω,ω ′

Note that Γ as defined here has dimensions of time, and g 2 Γ has units of frequency.

E. Rotating Wave Approximation

Alas, Eq. (524) is still not in Lindblad form. The problem is the “non-secular” (off-diagonal) terms with ω ≠ ω ′ . While
these did not present a problem in the cumulant derivation (recall that we proved complete positivity in Sec. XIV E), they do
now. Therefore we next introduce the final approximation, known as the rotating wave approximation (RWA), sometimes also
called the secular approximation. This approximation is based on the idea that the terms with ω ≠ ω ′ in Eq. (524) are rapidly
oscillating if t ≫ ∣ω − ω ′ ∣−1 , which thus (roughly) average to zero. Since we already assumed that t ≫ τB , the former assumption
is consistent provided we also assume that the Bohr frequency differences satisfy

min ∣ω − ω ′ ∣ > 1/τB . (525)

ω≠ω ′

Note that this means that also the Bohr frequencies themselves (by setting ω ′ = 0) must be large compared to the inverse of the
bath correlation time, and this therefore excludes the treatment of systems with gaps that are small relative to 1/τB (this has
implications for the applicability to systems that are typically of interest in adiabatic quantum computing, for example). Also
note that, combining this with the previous assumption [Eq. (513)], we get:

g ≪ 1/τB < min′ ∣ω − ω ′ ∣ (526)

ω≠ω

This shows that the coupling also lower bounds the Bohr frequencies.
 80

Let
∞
γαβ (ω) = ∫ eiωτ Bαβ (τ )dτ , (527)
−∞

i.e., the full Fourier transform of the bath correlation function. Using Eq. (511):
∞ ∞
∗
γαβ (ω) = ∫ e−iωτ Bβα (−τ )dτ = ∫ eiωτ Bβα (τ )dτ = γβα (ω) , (528)
−∞ −∞

i.e., γ(ω) is a Hermitian matrix. The inverse Fourier transform is

1 ∞
Bαβ (τ ) = e−iω τ γαβ (ω ′ )dω ′ .
′
∫ (529)
2π −∞
Then
∞ ∞ 1 ∞ 1 ∞ ∞
Γαβ (ω) = ∫ eiωτ Bαβ (τ )dτ = ∫ e−iω τ γαβ (ω ′ )dω ′ = dω ′ γαβ (ω ′ ) ∫ dτ ei(ω−ω )τ .
′ ′
eiωτ dτ ∫ ∫ (530)
0 0 2π −∞ 2π −∞ 0
∞
Now recall that the Dirac δ function can be represented as δ(x) = 1
2π ∫−∞ dτ e
ixτ
. When the integration lower limit is 0 instead
of −∞, we have the identity
∞ 1
∫ dτ eixτ = πδ(x) + iP ( ) , (531)
0 x
where the Cauchy principal value is defined as

1  f (x)
P ( ) [f ] = lim ∫ dx , (532)
x →0 − x
for smooth functions f with compact support on the real line R. Substituting Eq. (531) into Eq. (530), we can thus write
1
Γαβ (ω) = γαβ (ω) + iSαβ (ω) , (533)
2
where
1 ∞ 1
Sαβ (ω) = ∫ γαβ (ω ′ )P ( ) dω ′ = Sβα
∗
(ω) , (534)
2π −∞ ω − ω′
and we used the fact that γ is Hermitian in the last equality. Therefore:
1
γαβ (ω) = Γαβ (ω) + Γ∗βα (ω) , Sαβ (ω) = (Γαβ (ω) − Γ∗βα (ω)) . (535)
2i
Finally, we will show in Sec. XV F 1 that by introducing Eq. (533) and the RWA into Eq. (524), we arrive at the interaction
picture Lindblad equation:
dρ̃ 1
= −i[HLS , ρ̃(t)] + g 2 ∑ ∑ γαβ (ω)(Aβ (ω)ρ̃(t)A†α (ω) − {A†α (ω)Aβ (ω), ρ̃(t)}) , (536)
dt ω αβ 2

where the Lamb shift Hamiltonian is given by

HLS ≡ g 2 ∑ ∑ Sαβ (ω)A†α (ω)Aβ (ω) . (537)

ω αβ

To justify calling HLS a Hamiltonian we should show that it is Hermitian:

∗
†
HLS = g 2 ∑ Sαβ (ω)A†β (ω)Aα (ω) = g 2 ∑ Sβα (ω)A†β (ω)Aα (ω) = g 2 ∑ Sαβ (ω)A†α (ω)Aβ (ω) = HLS . (538)
αβω αβω αβω

We will show in Sec. XV F 2 that

[HLS , HS ] = 0 . (539)
 81

Note that γαβ (ω) and Sαβ (ω) as defined in Eqs. (527) and (534) have dimensions of time, while g 2 γαβ (ω) and g 2 Sαβ (ω) have
units of frequency. The factor g 2 can always be reabsorbed into the definition of γαβ (ω) and Sαβ (ω).10
We will show in Sec. XV F 3 that we can transform back to Schrödinger picture via ρ(t) = US (t)ρ̃(t)US† (t) and thus finally
obtain the RWA-LE:
dρ 1
= −i[HS + HLS , ρ] + g 2 ∑ ∑ γαβ (ω)(Aβ (ω)ρA†α (ω) − {A†α (ω)Aβ (ω), ρ}) . (540)
dt ω αβ 2

We will show in Sec. XV F 4 that:

γ(ω) > 0 , (541)
as required for complete positivity.
Let us now provide all the missing steps indicated above.

F. The missing steps

1. From Born-Markov [Eq. (524)] to the RWA-LE [Eq. (536)]

Let us start by expanding the commutator and the Hermitian conjugate term in Eq. (524), relabelling indices, and combining
terms. This gives us for the summands:

ei(ω −ω)t (Γαβ (ω)A†α (ω ′ )Aβ (ω)ρ̃(t) + Γ∗βα (ω ′ )ρ̃(t)A†α (ω ′ )Aβ (ω)) − ei(ω −ω)t (Γαβ (ω) + Γ∗βα (ω ′ )) Aβ (ω)ρ̃(t)A†α (ω ′ ) .
′ ′

(542)
Applying the RWA (i.e., setting ω = ω ′ ) and substituting Γαβ (ω) = 21 γαβ (ω) + iSαβ (ω), this becomes:
∗
1
γ (ω)A†α (ω)Aβ (ω)ρ̃(t) + 12 γβα
2 αβ
(ω)ρ̃(t)A†α (ω)Aβ (ω) (543a)
∗
+ iSαβ (ω)A†α (ω)Aβ (ω)ρ̃(t) − iSβα (ω)ρ̃(t)A†α (ω)Aβ (ω) − γαβ (ω)Aβ (ω)ρ̃(t)A†α (ω) . (543b)
Since γ(ω) and S(ω) are Hermitian this becomes:
1
γ (ω) (A†α (ω)Aβ (ω)ρ̃(t) + ρ̃(t)A†α (ω)Aβ (ω))
2 αβ
(544a)
+ iSαβ (ω) (A†α (ω)Aβ (ω)ρ̃(t) − ρ̃(t)A†α (ω)Aβ (ω)) − γαβ (ω)Aβ (ω)ρ̃(t)A†α (ω) (544b)
= 21 γαβ (ω) {A†α (ω)Aβ (ω), ρ̃(t)} + iSαβ (ω) [A†α (ω)Aβ (ω), ρ̃(t)] − γαβ (ω)Aβ (ω)ρ̃(t)A†α (ω) . (544c)
Putting this back into the original sum in Eq. (524) then gives us our desired result:
d ρ̃
= −ig 2 ∑ Sαβ (ω) [A†α (ω)Aβ (ω), ρ̃(t)] + g 2 ∑ γαβ (ω) (Aβ (ω)ρ̃(t)A†α (ω) − 12 {A†α (ω)Aβ (ω), ρ̃(t)}) (545a)
dt α,β,ω α,β,ω

= −i [HLS , ρ̃(t)] + g 2 ∑ γαβ (ω) (Aβ (ω)ρ̃(t)A†α (ω) − 21 {A†α (ω)Aβ (ω), ρ̃(t)}) . (545b)
α,β,ω

2. Proof of Eq. (539)

Let us write the system operators [Eq. (520)] as

Aα (ω) = ∑ Π(εa )Aα Π(εb ) = A†α (−ω) , (546)
εb −εa =ω

where the projectors Π(εa ) = ∣εa ⟩⟨εa ∣ are the eigenprojectors of HS , i.e.,
HS = ∑ εa Π(εa ) , (547)
a

10 Also note that in our derivation of the LE using coarse graining (Sec. XIV) we did not include the coupling strength g. Instead we used a dimensionless
parameter λ when we wrote down the system-bath interaction as λHSB , where HSB has dimensions of energy. As a result, γωω′ in the CG-LE has units of
̵ = 1).
frequency, while as noted above, in the RWA-LE γ(ω) has units of time, and g 2 γ(ω) has units of frequency (or energy, since we’re using units where h
 82

and hence HS Π(εa ) = Π(εa )HS = εa Π(εa ). Then:

HS A†α (ω)Aβ (ω) = ∑ εa Π(εa ) ∑ Π(εi )A†α Π(εj ) ∑ Π(εl )Aβ Π(εk ) (548a)
a εi −εj =ω εk −εl =ω

= ∑ εi Π(εi )A†α Π(εj ) ∑ Π(εl )Aβ Π(εk ) , (548b)

εi −εj =ω εk −εl =ω

and similarly:

A†α (ω)Aβ (ω)HS = ∑ Π(εi )A†α Π(εj ) ∑ εk Π(εl )Aβ Π(εk ) . (549)
εi −εj =ω εk −εl =ω

It follows that

[HS , A†α (ω)Aβ (ω)] = ∑ (εi − εk ) Π(εi )A†α Π(εj )Π(εl )Aβ Π(εk ) (550a)
εi −εj =ω
εk −εl =ω

= ∑ (εi − εk ) Π(εi )A†α Π(εj )Aβ Π(εk ) (550b)

εi −εj =ω
εk −εj =ω

=0, (550c)

where the second line follows from the product of the two inner projection operators, and the third line from the summation
conditions, which set εi = εk . Consequently:

[HS , HLS ] = g 2 ∑ Sαβ (ω) [HS , A†α (ω)Aβ (ω)] = 0 . (551)

αβω

3. Transformation back to the Schrödinger picture

Recall that ρ̃(t) = eiHS t ρ(t)e−iHS t , so d ρ̃

dt
= i [HS , ρ] + eiHS t dρ
dt
e−iHS t , and hence:

dρ d ρ̃
= −i [HS , ρ] + e−iHS t eiHS t . (552)
dt dt
Also, using Eq. (547) again:

Aα (ω)eiHS t = ∑ Π(ε)Aα Π(ε′ )eiε t .

′
(553)
ε′ −ε=ω

Thus:

e−iHS t Aβ (ω)ρ̃A†α (ω)eiHS t = ∑ e−iHS t Π(εj )Aβ Π(εi )eiHS t ρ(t)e−iHS t Π(εk )A†α Π(εl )eiHS t (554a)
εi −εj =ω
εk −εl =ω

= ∑ ei(−εj +εi −εk +εl )t Π(εj )Aβ Π(εi )ρ(t)Π(εk )A†α Π(εl ) (554b)
εi −εj =ω
εk −εl =ω

= ∑ Π(εj )Aβ Π(εi )ρ(t)Π(εk )A†α Π(εl ) (554c)

εi −εj =ω
εk −εl =ω

= Aβ (ω)ρA†α (ω) , (554d)

and

e−iHS t A†α (ω)Aβ (ω)ρ̃(t)eiHS t = ∑ e−iHS t Π(εi )A†α Π(εj )Π(εl )Aβ Π(εk )eiHS t ρ(t)e−iHS t eiHS t (555a)
εi −εj =ω
εk −εl =ω

= ∑ ei(−εi +εk )t Π(εi )A†α Π(εj )Π(εl )Aβ Π(εk )ρ(t) (555b)
εi −εj =ω
εk −εl =ω

= ∑ ei(−εj +εl )t Π(εi )A†α Π(εj )Π(εl )Aβ Π(εk )ρ(t) = A†α (ω)Aβ (ω)ρ(t) , (555c)
εi −εj =ω
εk −εl =ω
 83

FIG. 6. Left: original integration region. Right: new integration region.

and similarly for the second term in the anti-commutator. This shows that
1 1
e−iHS t (Aβ (ω)ρ̃(t)A†α (ω) − {A†α (ω), Aβ (ω), ρ̃(t)})eiHS t = (Aβ (ω)ρ(t)A†α (ω) − {A†α (ω), Aβ (ω), ρ(t)}) . (556)
2 2
Now, since we showed that HS and HLS commute:

e−iHS t [HLS , ρ̃(t)]eiHS t = e−iHS t HLS eiHS t e−iHS t ρ̃(t)eiHS t − e−iHS t ρ̃(t)eiHS t e−iHS t HLS eiHS t (557a)
= [HLS , ρ(t)] . (557b)

Hence, using Eqs. (536) and (552) we obtain Eq. (540) as required.

4. Proof that γ(ω) > 0

We’ll give two different proofs.

a. First proof The idea is to establish the following identity:
Lemma 2.
+∞ 1 T T
γαβ (ω) = ∫ eiωu Bαβ (u)du = lim ∫ dt ∫ eiω(t−s) Bαβ (t − s)ds . (558)
−∞ T →∞ T 0 0

Proof. Consider the following integral:

1 T T
I(ω, T ) ≡ ∫ dt ∫ e
iω(t−s)
Bαβ (t − s)ds . (559)
T 0 0

First, we change the variables from (t, s) to (u, s) with u = t − s. For every value of s, sweeping t from 0 to T will yield a
horizontal line of length T in the (u, s) plane. The new integration region is therefore a parallelogram in the variables (u, s),
as illustrated in Fig. 6. We can split this region into u ∈ [−T, 0] and u ∈ [0, T ], and perform the integration over s first. As is
clear from the figure, s varies from −u to T in the u ∈ [−T, 0] region, and from 0 to T − u in the u ∈ [0, T ] region. The area is
preserved so the Jacobian yields 1. Consequently,
T T 0 T T T −u
∫ ds ∫ dt = ∫ du ∫ ds + ∫ du ∫ ds . (560)
0 0 −T −u 0 0

When we integrate over some function independent of s,

0 T T T −u 0 T T
∫ du ∫ dsf (u) + ∫ du ∫ dsf (u) = ∫ duf (u)(T + u) + ∫ duf (u)(T − u) = ∫ f (u)(T − ∣u∣) . (561)
−T −u 0 0 −T 0 −T

Therefore, after the change of variables we get

1 T
I(ω, T ) = ∫ e Bαβ (u)(T − ∣u∣)du
iωu
(562a)
T −T
T 1 T
= ∫ eiωu Bαβ (u)du − ∫ eiωu Bαβ (u)∣u∣du . (562b)
−T T −T
 84

∞
Now recall that in the Markov approximation we assumed [Eq. (516)] that ∫0 un ∣Bαβ (u)∣du ∼ τBn+1 , where τB < ∞ is the bath
correlation time. Therefore, using Eq. (510c):
T T 0
∫ eiωu ∣u∣Bαβ (u)du = ∫ eiωu uBαβ (u)du − ∫ eiωu uBαβ (u)du (563a)
−T 0 −T
T T
=∫ eiωu uBαβ (u)du + ∫ e−iωu uBβα
∗
(u)du (563b)
0 0
∞ ∞
∗
≤∫ u∣Bαβ (u)∣du + ∫ u∣Bβα (u)∣du ∼ 2τB2 . (563c)
0 0
T
∫−T e ∣u∣Bαβ (u)du = 0, and
1 iωu
Consequently limT →∞ T

lim I(ω, T ) = γαβ (ω) (564)

T →∞

as claimed.
t
Now, for any vector v = (v1 , v2 , . . . ) we have
+∞
v † γ(ω)v = ∑ vα∗ γαβ (ω)vβ = ∑ vα∗ vβ ∫ eiωu Bαβ (u)du (565a)
αβ αβ −∞

1 T T
∗
= lim ∑ vα vβ ∫ dt ∫ e
iω(t−s)
Bαβ (t − s)ds (565b)
T →∞ T αβ 0 0

1 T T
∗ iωt −iωs
= lim ∑ λµ ⟨µ∣ ∫ vα e Bα (t)dt ∫ vβ e Bβ (s)ds∣µ⟩ (565c)
T →∞ T αβ,µ 0 0
2
1 T
−iωs
= lim ∑ λµ ∣∑ ∫ vα Bα (s)e ds∣µ⟩∣ ≥ 0 . (565d)
T →∞ T µ α 0

Therefore γ(ω) ≥ 0.
b. Second proof The following proof uses Bochner’s theorem as suggested, e.g., in the textbook [1].
Since γ(ω) is Hermitian [Eq. (528)] we can diagonalize it using a unitary transformation:
∗
D ≡ U γU † ⇒ Dαβ = ∑ Uαi γij Uβj . (566)
i,j
∞
D is diagonal so we need only consider the diagonal elements (i.e., the eigenvalues of γ). Plugging in γij = ∫−∞ eiωs Bij (s)ds
gives
∞ ⎛ ∗ ⎞
Dα = ∫ eiωs ∑ Uαi Bij (s)Uαj ds . (567)
−∞ ⎝ i,j ⎠
We wish to show that Dα is non-negative for each α. To do this we must consider the function in parenthesis. Dα is the Fourier
transform of this function so if we can show that it is of positive type then Dα must be positive by Bochner’s theorem [22].
Define the following function with {ti } an arbitrary time partition:
∗
α
fmn ≡ ∑ Uαi Bij (tm − tn )Uαj . (568)
i,j

Now use the property ⟨Bα (s)Bβ (0)⟩ = ⟨Bα (t)Bβ (t − s)⟩ [Eq. (510c)] to write fmn
α
as

∗ ⎛ ∗ ⎞
α
fmn = ∑ Uαi Tr [ρB Bi (tm )Bj (tn )] Uαj = Tr ρB ∑ Uαi Bi (tm ) ∑ Uαj Bj (tn ) . (569)
i,j ⎝ i j ⎠
We need to show that f α is a positive matrix. For arbitrary ∣v⟩ we have
⎡ ⎤
⎢⎛ √ ⎞⎛ √ ⎞⎥⎥
∗ ⎢ ∗ ∗
⟨v∣ f ∣v⟩ = ∑ vm vn fmn = Tr ⎢ ∑ vm Uαi ρB Bi (tm ) ∑ vn Uαj Bj (tn ) ρB ⎥
α α
(570a)
⎢⎝i,m ⎠ ⎝j,n ⎠⎥
m,n
⎣ ⎦
⎡ ⎤
⎢⎛ √ ⎞⎛ √ ⎞ ⎥
= Tr ⎢⎢ ∑ vm ∗
Uαi ρB Bi (tm ) ∑ vm Uαi ∗
Bi (tm ) ρB ⎥⎥ (570b)
⎢⎝i,m ⎠ ⎝i,m ⎠⎥
⎣ ⎦
≡ Tr (Mα† Mα ) ≥ 0 , (570c)
 85

where the final inequality follows from the fact that Mα† Mα is non-negative which follows immediately from right polar decom-
posing Mα (then Mα† Mα = RU † U R = R2 ≥ 0).
We have established that ⟨v∣ f α ∣v⟩ ≥ 0 for any time partition {ti }. Therefore Dα is positive by Bochner’s theorem. Conse-
quently, γ is a positive matrix since all its eigenvalues are non-negative.

XVI. THE KUBO-MARTIN-SCHWINGER (KMS) CONDITION AND THE GIBBS STATE AS A STATIONARY STATE
LINDBLAD EQUATION

In this section we formalize the folklore notion that “systems like to relax into lower energy states”, and that systems “tend to
equilibrate”.

A. The KMS condition

Consider a general system-bath Hamiltonian of the form HSB = ∑a Aa ⊗ Ba (we’re using a and b since we’ll reserve β
for the inverse temperature in this subsection). Let us assume again that the bath state is stationary [Eq. (482)], which as we
saw implies that ρB (t) = UB (t)ρB (0)UB† (t) = ρB (0) ≡ ρB . We also saw that this means that the bath correlation function is
time-translation-invariant:

⟨Ba (t + τ )Bb (t)⟩ = ⟨Ba (τ )Bb (0)⟩ , (571)

where for notational simplicity we dropped the B subscript we used before in ⟨X⟩B = Tr[ρB X].
If we assume not only that the bath state is stationary, but that it is also in thermal equilibrium at inverse temperature β, i.e.,
ρB = e−βHB /Z, then it follows that the correlation function satisfies the Kubo-Martin-Schwinger (KMS) condition [1]:

⟨Ba (τ )Bb (0)⟩ = ⟨Bb (0)Ba (τ + iβ)⟩ . (572)

The proof is the following calculation:

1
⟨Ba (τ )Bb ⟩ = Tr[ρB UB† (τ )Ba UB (τ )Bb ] = Tr[Bb e−(β−iτ )HB Ba e−iτ HB ] (573a)
Z
1
= Tr[Bb ei(τ +iβ)HB Ba e−i(τ +iβ)HB e−βHB ] = Tr[ρB Bb UB† (τ + iβ)Ba UB (τ + iβ)] (573b)
Z
= ⟨Bb Ba (τ + iβ)⟩ . (573c)

Note that using the same technique it also follows that

⟨Ba (τ )Bb ⟩ = ⟨Bb (−τ − iβ)Ba ⟩ . (574)

If in addition the correlation function is analytic in the strip between τ = −iβ and τ = 0, then it follows that the Fourier
transform of the bath correlation function satisfies the frequency domain KMS condition:

γab (−ω) = e−βω γba (ω) . (575)

τ
0

−iβ

FIG. 7. Contour used in proof of the KMS condition.

 86

This is an extremely important condition, which is used in proving “detailed balance”, as we shall see when we discuss the Pauli
master equation, in Sec. XVII.
To prove this let us use the time-domain KMS condition, Eq. (574):
∞ ∞
γab (ω) = ∫ dτ eiωτ ⟨Ba (τ )Bb (0)⟩ = ∫ dτ eiωτ ⟨Bb (−τ − iβ)Ba (0)⟩ (576)
−∞ −∞

To perform this integral we replace it with a contour integral in the complex τ plane, ∮C dτ eiωτ ⟨Bb (−τ − iβ)Ba (0)⟩, with the
contour C as shown in Fig. 7. This contour integral vanishes by the Cauchy-Goursat theorem [23] since the closed contour
encloses no poles (by assumption, the correlation function ⟨Bb (τ )Ba (0)⟩ is analytic in the open strip (0, −iβ) and is continuous
at the boundary of the strip [24]), so that

∮ (. . . ) = 0 = ∫ (. . . ) + ∫ (. . . ) + ∫ (. . . ) + ∫ (. . . ) (577)
C ↑ ↓ → ←

where (. . . ) is the integrand of Eq. (576), and the integral ∫→ is the same as in Eq. (576). After making the variable transformation
τ = −x − iβ, where x is real, we have
∞
∫ (. . . ) = −e ∫
βω
dx e−iωx ⟨Bb (x)Ba ⟩ = −eβω γba (−ω) . (578)
← −∞

Assuming that ⟨Ba (±∞ − iβ)Bb (0)⟩ = 0 (i.e., the correlation function vanishes at infinite time), we further have ∫↑ (. . . ) =
∫↓ (. . . ) = 0, and hence we find the result:

0 = γab (ω) + 0 + 0 − eβω γba (−ω) (579)

which proves Eq. (575).

The KMS condition (575) is important, since it tells us that transitions involving negative Bohr frequencies are exponentially
suppressed, as e−βω , compared to the opposite transitions involving positive Bohr frequencies. I.e., when a system is coupled to
a thermal bath, an excitation in the system is exponentially suppressed relative to a relaxation event at the same frequency.11

B. The Gibbs state is a stationary state of the RWA-LE

Consider a bath at inverse temperature β. We would like to show that the system Gibbs state

⎛ e−βε0 ⎞
1 −βHS 1 ⎜ ⎟,
ρG = e = ⎜ e−βε1 ⎟ Z = Tr[e−βHS ] (580)
Z Z
⎝ ⋱ ⎠

is always a stationary state, in the sense that ρ̇G = 0. Here the energies are listed in increasing order, starting from ground state
energy 0 . We will show this here directly from the RWA-LE, and given an alternative derivation from the Pauli master equation
in Sec. XVII.
In the Schrödinger picture the RWA-LE has the form:

ρ̇ = −i [HS + HLS , ρ] + D(ρ) , (581)

where the dissipator is

1
D(ρ) = g 2 ∑ ∑ γαβ (ω) (Aβ (ω)ρA†α (ω) − {A†α (ω)Aβ (ω), ρ}) . (582)
αβ ω 2

To show that ρ̇G = 0, consider first the Hamiltonian part. That [HS , ρG ] = 0 follow immediately from Eq. (580). Now recall
that [HS , HLS ] = 0 [Eq. (539)]. Thus HS and HLS are diagonalizable in the same basis, i.e., there exists a unitary V such that
V HS V † = D1 and V HLS V † = D2 , where D1 and D2 are both diagonal (and of course commute). Therefore
1
V [HLS , ρG ]V † = V HLS V † V ρG V † − V ρG V † V HLS V † = [D2 , e−βD1 ] = 0 , (583)
Z

11 Recall Eq. (460): ω = εb − εa < 0 corresponds to a transition from ∣εb ⟩ to ∣εa ⟩, i.e., from energy εb to a higher energy εa .
 87

which means that [HLS , ρG ] = 0.

Next let us consider the dissipative part. This requires us to calculate ρG A†α (ω) and Aβ (ω)ρG . Now, for any pair of operators
A and B it is easy to prove (e.g., by Taylor expansion) that:
∞ n
(−α)
e−αA BeαA = ∑ [A, B]n , (584)
n=0 n!
where the nested commutator is defined recursively via
[A, B]n = [A, [A, B]n−1 ] , [A, B]0 = B . (585)
Simplifying our notation via ∣a⟩ ≡ ∣εa ⟩, let us write the system operators [Eq. (520)] as
Aα (ω) = ∑ Πa Aα Πb = A†α (−ω) , (586)
b−a=ω

where the projectors Πa = ∣a⟩⟨a∣ are in the energy basis, i.e., HS Πa = Πa HS = aΠa , where HS = ∑a aΠa . Using the property
Πa Πb = δab Πa , note that:
[HS , Aα (ω)] = ∑ aΠa ∑ Πa′ Aα Πb − ∑ Πa Aα Πb ∑ a′ Πa′ (587a)
a b−a′ =ω b−a=ω a′
= ∑ aΠa Aα Πb − ∑ bΠa Aα Πb = ∑ (a − b)Πa Aα Πb (587b)
b−a=ω b−a=ω b−a=ω
= −ωAα (ω) (587c)
[HS , A†α (ω)] = ωA†α (ω) , (587d)
where Eq. (587d) follows by taking the Hermitian conjugate of Eq. (587c).
Therefore:
[HS , Aα (ω)]n = (−ω)n Aα (ω) (588a)
[HS , A†α (ω)]n = ω n Aα (ω) . (588b)
Hence, using Eq. (584):
∞
(−β)n (−ω)n
e−βHS Aα (ω)eβHS = ∑ Aα (ω) = eβω Aα (ω) , (589)
n=0 n!
which tells us that
Aα (ω)ρG = e−βω ρG Aα (ω) . (590)
It follows by Hermitian conjugation that:
ρG A†α (ω) = e−βω A†α (ω)ρG . (591)
We are now ready to consider the terms in the dissipator, Eq. (582). Commuting ρG to the right we find:
Aβ (ω)ρG A†α (ω) = e−βω Aβ (ω)A†α (ω)ρG (592a)
−βω
ρG A†α (ω)Aβ (ω) =e A†α (ω)ρG Aα (ω) = A†α (ω)Aβ (ω)ρG , (592b)
The action of the dissipator thus becomes:
D(ρG ) = g 2 ∑ ∑ γαβ (ω) (e−βω Aβ (ω)A†α (ω) − A†α (ω)Aβ (ω)) ρG . (593)
αβ ω

Let us now separate the sum over ω as ∑ω<0 +(ω = 0) + ∑ω>0 . Recall that KMS result [Eq. (575)]: γαβ (−ω) = e−βω γβα (ω).
We know from Eq. (586) that Aα (0) = A†α (0), so that the ω = 0 cancels since the remaining sum is over all α and β, and by
KMS, γαβ (0) = γβα (0). As for the sum over negative frequencies, using KMS and Eq. (586) again we have:

∑= γαβ (−ω ′ ) (eβω Aβ (−ω ′ )A†α (−ω ′ ) − A†α (−ω ′ )Aβ (−ω ′ )) ρG
′
∑ (594a)
ω<0 ω ′ =−ω>0

= ∑ γβα (ω ′ )e−βω (eβω A†β (ω ′ )Aα (ω ′ ) − Aα (ω ′ )A†β (ω ′ )) ρG ,

′ ′
(594b)
ω ′ >0
 88

so that
−βω
∑ ∑ = γβα (ω) (Aβ (ω)Aα (ω) − e Aα (ω)A†β (ω)) ρG
†
(595a)
αβ ω<0

= −∑ ∑ , (595b)
αβ ω>0

and hence ∑ω<0 + ∑ω>0 = 0.

So, the dissipator is also zero, and the Gibbs state is indeed stationary:

ρ̇G = 0 . (596)

C. Return to equilibrium, quantum detailed balance, and ergodicity under the RWA-LE

A natural next question is under which conditions the Gibbs state is actually reached. To answer this we need to define the
concept of ergodicity. A system is ergodic if it holds that for any arbitrary system operator X

[X, Aα (ω)] = [X, A†α (ω)] = 0, ∀α, ω (597)

if and only if X is proportional to the identity operator.

It is possible to prove that if a system is ergodic and in addition L = −i[H, ⋅] + LD satisfied the quantum detailed balance
condition with respect to the stationary state ρ̃ (the state for which Lρ̃ = 0)

[H, ρ̃] = 0 (598a)

(L†D A, B) = (A, L†D B) (598b)

for (A, B) ≡ Tr[ρ̃A† B] and all A, B ∈ domain(L† ), then for any initial state ρ(0) the stationary state is the Gibbs state. I.e., the
t→∞
Gibbs state is an attractor for the dynamics: ρ(t) = eLt ρ(0) ÐÐ→ ρG . This is a fundamental result, as it tells us the conditions
under which a system is guaranteed to become thermally equilibrated. The proof is given in Sec. 1.3.4 of Ref. [2] (see also
Ref. [30]).
However, not all systems are ergodic [25]. For example, consider a system of N qubits coupled to a bath such that
N
Aα = ∑ σjα , α ∈ {x, y, z} . (599)
j=1

Clearly, all Aα are invariant under permutations, so that they commute with the elements of the permutation group. This means
that Eq. (597) is satisfied for operators X that are not proportional to the identity (e.g., the SWAP operator between any pair of
qubits), and hence such a system is not ergodic. Indeed, Eq. (599) describes “collective decoherence”, under which there exist
subspaces that are invariant under the action of the Aα operators, and undergo unitary dynamics [26, 27]. Initial states in such
subspaces do not converge to the Gibbs state, and do not equilibrate.
More generally, if the system-bath interaction possesses some symmetry (e.g., a permutational symmetry as above), then
ergodicity does not hold and the system need not equilibrate [28].

XVII. PAULI MASTER EQUATION

Sometimes we are particularly interested in finding out the evolution of just the populations (diagonal elements) in the energy
eigenbasis. For example, this is the case in adiabatic quantum computing and quantum annealing, where the answer to a
computation is encoded in the ground state [29]. In other applications we are interested in finding out the Gibbs distribution
ρG [Eq. (580)] in order to compute various thermodynamic averages ⟨X⟩ = Tr(XρG ), where X could be any observable
of interest; the Gibbs state is an example of a state that is diagonal in the energy eigenbasis, i.e., the eigenbasis {∣a ⟩} of
HS = ∑a a ∣a ⟩⟨a ∣ = ∑a α Πa .
Recall that the RWA-LE in the Schrödinger picture is

dρ 1
= −i [HS + HLS , ρ] + g 2 ∑ ∑ γαβ (ω) [Aβ (ω)ρA†α (ω) − {A†α (ω)Aβ (ω), ρ}] .
dt αβ ω 2
 89

The population in the ath energy eigenbasis state is:

pa (t) = ⟨a ∣ ρ(t) ∣a ⟩ = ρaa (t) = Tr[Πa ρ] . (600)
Our goal is to derive a master equation for the evolution of these populations, known as the Pauli master equation. We will
see that the populations in the energy eigenbasis are decoupled from the coherences (off diagonal elements) in the same basis.
Consider then, the time-derivative of the populations, while using the fact that HS is time-independent (and hence so are its
eigenvalues and eigenvectors):
ṗa = ⟨a ∣ρ̇∣a ⟩ = Tr[Πa ρ̇] (601a)
= −i⟨a ∣[HS , ρ]∣a ⟩ − i⟨a ∣[HLS , ρ]∣a ⟩ (601b)
1
+ g 2 ∑ ∑ γαβ (ω)⟨a ∣ [Aβ (ω)ρA†α (ω) − {A†α (ω)Aβ (ω), ρ}] ∣a ⟩ . (601c)
αβ ω 2

The first term in Eq. (601b) is:

⟨a ∣[HS , ρ]∣a ⟩ = ⟨a ∣HS ρ∣a ⟩ − ⟨a ∣ρHS ∣a ⟩ = a ⟨a ∣ρ∣a ⟩ − a ⟨a ∣ρ∣a ⟩ = 0 . (602)
As for ⟨a ∣[HLS , ρ]∣a ⟩, recall that [HS , HLS ] = 0, which means that HS and HLS share a common eigenbasis, i.e., the en-
ergy eigenbasis {∣a ⟩}; hence HLS is diagonal in the same basis and the same calculation as in Eq. (602) also implies that
⟨a ∣[HLS , ρ]∣a ⟩ = 0. Therefore there is no contribution from the unitary part to the evolution of the populations in the energy
eigenbasis.
Now consider the dissipative part, i.e., line (601c). Recall that
Aβ (ω) = ∑ ∣a ⟩⟨a ∣Aβ ∣b ⟩⟨b ∣ = ∑ ∣a⟩Aab,β ⟨b∣ , A†α (ω) = ∑ ∣b⟩Aba,α ⟨a∣ (603)
b −a =ω b−a=ω b−a=ω

where we again used the simplified notation a ↦ a. We have for the first term in line (601c):
⟨a ∣ Aβ (ω)ρA†α (ω) ∣a ⟩ = ⟨a∣ ∑ Aa′ b′ ,β ∣a′ ⟩⟨b′ ∣ρ ∑ Ab′′ a′′ ,α ∣b′′ ⟩⟨a′′ ∣a⟩ (604a)
ω=b′ −a′ ω=b′′ −a′′
= ∑ Aab′ ,β ρb′ b′′ Ab′′ a,α (604b)
ω=b′ −a
ω=b′′ −a
= ∑ Aab′ ,β pb′ Ab′ a,α = ∑ Aaa′ ,β pa′ Aa′ a,α , (604c)
ω=b′ −a ω=a′ −a

where to go the second line we used ⟨a∣a′ ⟩ = δaa′ and ⟨a′′ ∣a⟩ = δa′′ a , and to go to the third line we used the fact that b′ must
equal b′′ due to the summation constraints.
Similarly,
′ ′
⟨a∣ A†α (ω)Aβ (ω)ρ ∣a⟩ = ∑ ⟨a∣ Ab′ a′ ,α ∣b ⟩⟨a ∣ ∑ Aa′′ b′′ ,β ∣a′′ ⟩⟨b′′ ∣ρ ∣a⟩ (605a)
ω=b′ −a′ ω=b′′ −a′′
= ∑ Aaa′ ,α Aa′ b′′ ,β ρb′′ a (605b)
ω=a−a′
ω=b′′ −a′
= ∑ Aaa′ ,α Aa′ a,β pa , (605c)
ω=a−a′

and
′ ′
⟨a∣ ρA†α (ω)Aβ (ω) ∣a⟩ = ∑ ⟨a∣ ρAb′ a′ ,α ∣b ⟩⟨a ∣ ∑ Aa′′ b′′ ,β ∣a′′ ⟩⟨b′′ ∣a⟩ (606a)
ω=b′ −a′ ω=b′′ −a′′
= ∑ ρab′ Ab′ a′ ,α Aa′ a,β (606b)
ω=b′ −a′
ω=a−a′
= ∑ pa Aaa′ ,α Aa′ a,β , (606c)
ω=a−a′

which is the same result as in Eq. (605).

Combining Eqs. (604)-(606), we have:

ṗa = ∑ ( ∑ Aa′ a,α Aaa′ ,β pa′ − ∑ Aaa′ ,α Aa′ a,β pa ) γαβ (ω) . (607)
αβ ω=a′ −a ω=a−a′
 90

Since the index a is fixed, the sum over ω really only involves varying a′ . Thus:

ṗa = ∑ ∑ γαβ (a′ − a)Aa′ a,α Aaa′ ,β pa′ − γαβ (a − a′ )Aaa′ ,α Aa′ a,β pa . (608)
αβ a′

Now define a transition matrix W via

W (a∣a′ ) ≡ ∑ γαβ (a′ − a)Aa′ a,α Aaa′ ,β . (609)

αβ

Note that W (a∣a′ ) ≥ 0. To prove this, let u be the unitary matrix that diagonalizes γ: γαβ = ∑α′ uαα′ γα′ u∗βα′ . Then:

⎛ ⎞
W (a∣a′ ) = ∑ uαα′ γα′ (a′ − a)u∗βα′ Aa′ a,α Aaa′ ,β = ∑ γα′ (a′ − a) (∑ uαα′ Aa′ a,α ) ∑ u∗βα′ Aaa′ ,β (610a)
α′ αβ α′ α ⎝β ⎠
2
= ∑ γα′ (a′ − a) ∣Ãa′ a,α′ ∣ ≥ 0 , (610b)
α′

where Ãa′ a,α′ = ∑α uαα′ Aa′ a,α , and we used the Hermiticity of Aβ to write Aaa′ ,β = A∗a′ a,β . Eq. (608) can thus be simplified
as:

ṗa = ∑ W (a∣a′ )pa′ − W (a′ ∣a)pa . (611)

a′

This represents a closed set of rate equations for the populations {pa }.
If we assume that the KMS condition γαβ (−ω) = e−βω γβα (ω) (for ω > 0) holds, then this allows us to write, for a > a′ :

γαβ (a′ − a) = e−β(a−a ) γβα (a − a′ ).

′
(612)

Then W (a∣a′ ) can be rewritten as:

W (a∣a′ ) = ∑ e−β(a−a ) γβα (a − a′ )Aa′ a,α Aaa′ ,β = e−β(a−a ) ∑ γαβ (a − a′ )Aa′ a,β Aaa′ ,α = e−β(a−a ) W (a′ ∣a) .
′ ′ ′
(613)
αβ αβ

This is the detailed balance condition:

“ ↑ ” W (a∣a′ )
= = e−β(a−a ) .
′
(614)
“ ↓ ” W (a′ ∣a)

It says that the rate for an “up” transition, from the low energy state ∣a′ ⟩ to the high energy state ∣a⟩, is exponentially less likely
than the reverse, “down” transition, with the exponent given by the energy difference in units of the bath temperature. This is
an extremely important result, since it establishes rigorously the intuition that at very low temperatures (relative to the smallest
energy gap) systems tend to relax towards their ground states. This is a special case of the quantum detailed balance condition
we mentioned in Sec. XVI C.
Finally, we can also reestablish that the Gibbs state is the stationary state (recall that we showed this in Sec. XVI B). For a
stationary state ṗa = 0. It follows from Eq. (611) that in this case:

W (a∣a′ ) pa e−βa
= ′ = e−β(a−a ) = −βa′
′
(615a)
W (a ∣a) pa
′ e
e−βa
⇒ pa = ; Z = ∑ e−βa , (615b)
Z a

which is the Gibbs distribution.

XVIII. LINDBLAD EQUATION IN THE SINGULAR COUPLING LIMIT (SCL)

All our derivations of the LE so far have assumed the weak coupling limit of system-bath coupling. Somewhat surprisingly,
the opposite limit of strong coupling also allows us to derive the Lindblad equation, while avoiding the use of the RWA.
 91

A. Derivation

Assume that the Hamiltonian takes the form

1 1
H = HS + HSB + 2 HB , (616)
 
where HSB = g ∑α Aα ⊗ Bα as in Eq. (499), so that the Aα , Bα operators are dimensionless. Since we are interested in the
limit of small , this is called the singular coupling limit (SCL). In this limit the bath Hamiltonian dominates over the system
and system-bath Hamiltonians.
Note that in order for the Gibbs state of the bath to remain invariant (ρB = e−βHB /Z), the bath must be in thermal equilibrium
with respect to HB /2 at the temperature T /2 → ∞. Thus, we can also interpret the SCL as a high temperature limit. For a
more detailed discussion see Ref. [31].
Our starting point is the interaction picture Born approximation [Eq. (512)], which we write here with  included:
dρ̃ 1 t
= −g 2 2 ∑ ∫ dτ {Bαβ (τ ) [Aα (t), Aβ (t − τ )ρ̃(t − τ )] + h.c.} . (617)
dt  αβ 0

Let us transform this to the Schrödinger picture via Eq. (552):

dρ 1 t
= −i[HS , ρ(t)] + g 2 ∑ 2 ∫ dτ US (t) ([Aβ (t − τ )US† (t − τ )ρ(t − τ )US (t − τ )Aα (t)−
dt αβ  0

Aα (t)Aβ (t − τ )US† (t − τ )ρ(t − τ )US (t − τ )] Bαβ (τ ) + h.c.) US† (t) . (618)

We can perform a change of variables to τ = 2 τ ′ , and take the limit  → 0, so that τ → 0. Then, recalling Eq. (501), the various
terms in Eq. (618) transform as follows:

US (t)Aβ (t − τ )US† (t − τ )ρ(t − τ )US (t − τ )Aα (t)US† (t) = US† (−τ )Aβ ρ(t − τ )US† (τ )Aα
→ Aβ ρ(t)Aα = Aβ ρ(t)A†α (619a)
US (t)Aα (t)Aβ (t − τ )US† (t − τ )ρ(t − τ )US (t − τ )US† (t) = Aα US† (−τ )Aβ ρ(t − τ )US† (τ )
→ Aα Aβ ρ(t) = A†α Aβ ρ(t) (619b)
i2 τ ′ HB /2 −i2 τ ′ HB /2
Bαβ (τ ) = Tr (e Bα e Bβ ρB ) = Tr (UB† (τ ′ )Bα UB (τ ′ )Bβ ρB ) ′
= Bαβ (τ ) (619c)
1 t t −2
∞
′
∫ dτ = ∫ dτ → ∫ dτ ′ . (619d)
2 0 0 0

Thus the  → ∞ strong coupling and bath limit, is essentially a Markovian limit, as it allows us to extend the integration limit to
∞ and make ρ time-local. It also removes the time dependence from the Aα system operators.
Applying the transformations in Eq. (619) to Eq. (618) gives:
dρ ∞
= −i[HS , ρ(t)] + g 2 ∑(Aβ ρ(t)A†α − A†α Aβ ρ(t)) ∫ dτ Bαβ (τ ) + h.c. . (620)
dt αβ 0

Now recall Eqs. (523) and (533), which tell us that

∞ 1
∫ dτ Bαβ (τ ) = Γαβ (0) = γαβ (0) + iSαβ (0) . (621)
0 2
Thus
dρ 1
= −i[HS + HLS , ρ(t)] + g 2 ∑ γαβ (0) (Aβ ρ(t)A†α − {A†α Aβ , ρ(t)}) (622a)
dt αβ 2
∞
HLS = ∑ Sαβ (0)A†α Aβ , γαβ (0) = ∫ dτ Bαβ (τ ) . (622b)
αβ −∞

Note that the SCL keeps only the ω = 0 component out of all the Bohr frequencies, so it is clearly a more “extreme” limit than
the WCL. We can understand this as a consequence of the fact that the SCL is designed to accelerate the internal evolution of
the bath by rescaling the bath Hamiltonian via HB ↦ HB /2 ; this means that all system frequencies are effectively zero relative
to the very high effective bath evolution frequency, and only the static component ω = 0 survives.
 92

B. Examples contrasting the WCL and SCL

Let us consider a single qubit.

1. Phase damping when [HS , HSB ] = 0

We assume that
1
HS = − ωz σ z , HSB = gσ z ⊗ B. (623)
2
For the interaction Hamiltonian in Eq. (623), there is only a single system operator Az = σ z = ∣0⟩⟨0∣ − ∣1⟩⟨1∣. The eigenstates are
∣ε0 ⟩ = ∣0⟩ and ∣ε1 ⟩ = ∣1⟩. Considering the RWA-LE (the weak coupling limit case) Eq. (536) and ⟨εa ∣ Az ∣εb ⟩ ∝ δab , there is only
a single Lindblad operator that is non-zero:

Az (0) = σ z , (624)

as given by Eq. (520). This follows since [HS , HSB ] = 0. Therefore, the RWA-LE [Eq. (540)] takes the simple form

d 1
ρ(t) = −i [HS , ρ(t)] + g 2 γ(0) (σ z ρ(t)(σ z )† − {(σ z )† σ z , ρ(t)}) , (625)
dt 2
where we have also used the fact that HLS ∝ I. This form is the same as what is predicted in the SCL, since only the ω =
0 component appears. We have encountered this equation several times before [e.g., Eq. (277)]. After expanding ρ(t) =
∑i,j∈{0,1} ρij ∣i⟩ ⟨j∣, and taking matrix elements in the computational basis (which here is equivalent to the energy eigenbasis)
we obtain:

ρ00 (t) = ρ00 (0) = 1 − ρ11 (t) , (626a)

(c)
ρ01 (t) = exp(−t/T2 + iωz t)ρ01 (0) = ρ∗10 (t) , (626b)

where
(c) 1
T2 = , (627)
2g 2 γ(0)

where the ‘c’ superscript denotes the computational basis (we shall shortly see a second T2 associated with the energy eigenba-
sis). This is the familiar phase damping channel, where only the off-diagonals elements (transverse magnetization) decay with a
(c)
characteristic timescale T2 . The stronger the coupling to the bath g, the shorter the qubit coherence time. Note that the qubit
(c) (c)
energy gap ωz plays no role in the result for T2 , and T2 here is entirely determined by the spectrum of the bath correlation
function at zero frequency. In this example there is no thermal relaxation (the T1 time is infinite), since the population of the
energy states remains fixed, as a consequence of [HS , HSB ] = 0.

2. Phase damping when [HS , HSB ] ≠ 0

Let us now replace the system Hamiltonian so that [HS , HSB ] ≠ 0. Specifically, consider

1
HS = − ωx σ x , HSB = gσ z ⊗ B . (628)
2
We shall see that there is a sharp contrast between the WCL and SCL, with the WCL resulting in decoherence in the energy
eigenbasis, while the SCL results in decoherence in the computational basis, just as in the previous subsection, when HS and
HSB were commuting.
a. WCL The energy eigenstates of HS are ∣ε0 ⟩ = ∣+⟩ with eigenvalue − 21 ωx (ground state) and ∣ε1 ⟩ = ∣−⟩ with eigenvalues
1
ω (excited state), where ∣±⟩ = √12 (∣0⟩ ± ∣1⟩). Therefore the possible Bohr frequencies are ω ∈ {0, ±ωx }. Since σ z ∣±⟩ = ∣∓⟩,
2 x
we find Az (0) = 0, and the non-zero Lindblad operators are:

Az (ωx ) = ∣+⟩⟨+∣σ z ∣−⟩⟨−∣ = ∣+⟩⟨−∣ , Az (−ωx ) = ∣−⟩⟨−∣σ z ∣+⟩⟨+∣ = ∣−⟩⟨+∣ . (629)

 93

Note that we now have a non-trivial Lamb shift term:

HLS = S(ωx )∣−⟩⟨−∣ + S(−ωx )∣+⟩⟨+∣ . (630)
Now we need to compute the terms in the RWA-LE [Eq. (540)]. It is most convenient to do so in the energy eigenbasis, i.e., the
basis that diagonalizes HS , namely the {∣±⟩} basis we used above. Note that:
1
HS + HLS = Ω+ ∣+⟩⟨+∣ + Ω− ∣−⟩⟨−∣ , Ω± = ωx + S(±ωx ) (631a)
2
1 1
g 2 ∑ ∑ γαβ (ω)⋯ = g 2 [γ(ωx ) (∣+⟩⟨−∣ρ∣−⟩⟨+∣ − (∣−⟩⟨−∣ρ + ρ∣−⟩⟨−∣)) + γ(−ωx ) (∣−⟩⟨+∣ρ∣+⟩⟨−∣ − (∣+⟩⟨+∣ρ + ρ∣+⟩⟨+∣))] .
ω αβ 2 2
(631b)

Writing ρ(t) = ∑i,j∈{+,−} ρij ∣i⟩⟨j∣, and taking matrix elements of Eq. (631), we find:

⟨−∣ ρ̇ ∣−⟩ = ρ̇−− = −i ⟨−∣ (HS + HLS )ρ − ρ(HS + HLS ) ∣−⟩ − g 2 γ(ωx )ρ−− + g 2 γ(−ωx )ρ++ , (632)
and the first (Hamiltonian) term is easily seen to vanish. Also, note that Tr[ρ(t)] = ρ++ (t) + ρ−− = 1 implies that ρ̇−− = −ρ̇++ .
After a similar calculation for the off-diagonal components, we find that the Lindblad equation for the density matrix components
is:
d d
−ρ++ = ρ−− = −g 2 γ(ωx )ρ−− (t) + g 2 γ(−ωx )ρ++ (t) (633a)
dt dt
d ∗ d 1
ρ (t) = ρ−+ (t) = Ωρ−+ (t) , Ω ≡ −i [Ω(−ωx ) − Ω(ωx )] − g 2 [γ(ωx ) + γ(−ωx )] . (633b)
dt +− dt 2
The solution for the off-diagonal elements [Eq. (633b)] is immediate: ρ−+ (t) = ρ−+ (0)e−iΩt , i.e.:

ρ∗+− (t) = ρ−+ (t) = ρ−+ (0)e−iωx t e−t/T2

′ (e)
, (634)
where
(e) 2
T2 = , ωx′ = ωx + S(ωx ) − S(−ωx ) , (635)
g 2 γ(ωx ) (1 + e−βωx )
where the ‘e’ superscript denotes the energy eigenbasis (as opposed to the computational basis) , and where we used the KMS
condition [Eq. (575)] to write γ(ωx ) + γ(−ωx ) = γ(ωx )(1 + e−βωx ). Contrast this result with Eq. (627), where the dephasing
rate depended only on γ(0) and did not exhibit a temperature dependence.
To solve for the populations, let us substitute ρ++ = 1 − ρ−− into Eq. (633a), so that we can write ρ̇−− = a − bρ−− , where
a = g 2 γ(−ωx ) and b = g 2 [γ(−ωx ) + γ(ωx )]. As a solution let us try the ansatz ρ−− (t) = ce−t/T1 + d, so that the initial condition
(e)

yields c = ρ−− (0) − d. Then

c
e−t/T1 = a − b (ce−t/T1 + d) = a − bce−t/T1 − bd ,
(e) (e) (e)
ρ̇−− = − (e)
(636)
T1
(e)
which tells us that d = a/b = γ(−ωx )/[γ(−ωx ) + γ(ωx )] and T1 = 1/b, i.e.:

(e) 1 (e)
T1 = T2 . (637)
2
Moreover, recall that the Gibbs state is
1 −βHS 1 12 βωx σx
ρG = e = e = pG (−)∣+⟩⟨+∣ + pG (+)∣−⟩⟨−∣ , (638)
Z Z
where
e±βωx /2
pG (±) = , Z = Tr(ρG ) = pG (−) + pG (+) = 2 cosh(βωx /2) . (639)
Z
Using this and the KMS condition, we have
γ(−ωx )
d= = PG (−) . (640)
γ(−ωx )(1 + eβωx )
 94

Using our ansatz we thus find for the populations, finally:

1 − ρ++ (t) = ρ−− (t) = pG (−) + [ρ−− (0) − pG (−)] e−t/T1

(e)
. (641)
We note several important facts about these results:
• The decoherence occurs in the energy eigenbasis, i.e., the off-diagonal components in the energy eigenbasis (not in the
(e)
computational basis) decay exponentially to zero with a timescale determined by T2 .
• The entire contribution of the Lamb shift is in shifting the rotation rate of the off-diagonal elements from ωx to ωx +
S(ωx ) − S(−ωx ) [Eq. (635)].
• The populations (ρ++ , ρ−− ) approach the Gibbs state associated with the Hamiltonian HS within a timescale determined
(e) βωx /2
by T1 [Eq. (641)]. In particular, for the ground state population: ρ++ → pG (+) = e Z .
(e) (e)
• The two timescales (T1 , T2 ) are strictly related (relaxation is twice as fast as dephasing) and have a non-trivial depen-
dence on the energy gap ωx .
(e) (e)
• Even in the zero temperature limit (β → ∞), the dephasing and relaxation times can be non-vanishing: T1 = T2 /2 =
1
g 2 γ(ωx )
> 0.
b. SCL Let us contrast this with what happens in the SCL case, Eq. (622). This simply becomes Eq. (625), with HS =
− 12 ωx σ x , i.e.:
ωx x
ρ̇ = i [σ , ρ] + g 2 γ(0) (σ z ρσ z − ρ) , (642)
2
In this case the evolution of the density matrix elements is most conveniently solved for in the computational basis. Taking
matrix elements in this basis yields:
d 1
ρ00 = −i ωx (ρ10 − ρ01 ) , (643a)
dt 2
d 1
ρ11 = −i ωx (ρ01 − ρ10 ) , (643b)
dt 2
d 1
ρ01 = i ωx (ρ11 − ρ00 ) − 2g 2 γ(0)ρ01 , (643c)
dt 2
d 1
ρ10 = i ωx (ρ00 − ρ11 ) − 2g 2 γ(0)ρ10 . (643d)
dt 2
This set of equations can be solved analytically for arbitrary initial conditions, but for brevity, let us consider the case where
the density matrix is initially in a uniform computational basis superposition (the ground state of the previous WCL case), i.e.,
ρ(0) = ∣+⟩⟨+∣. The solution is then given by:
1 1
ρ01 = ρ10 = e−t/T2 .
(c)
ρ00 = ρ11 = , (644)
2 2
(c)
In this case, the off-diagonal elements in the computational basis decay exponentially with a timescale determined by T2
[Eq. (627)], so we have decoherence in the computational basis regardless of the fact that the system Hamiltonian does not
commute with HSB . The predictions made under the WCL and SCL assumptions are thus starkly different.
c. Results for a bosonic bath So far we didn’t specify the bath, and hence γ(ω) was left unspecified as well. Let us now
assume that the bath is bosonic:
HB = ∑ ωk b†k bk , (645)
k

where bk is the annihilation operator associated with bosonic mode k, and the system-bath interaction is
HSB = gA ⊗ B , A = σz , B = ∑(gk /g)(bk + b†k ) . (646)
k

There is only a single bath correlation function, because there is only a single bath operator B. For a bath in a Gibbs state at
inverse temperature β it can be shown that the bath correlation function in this case is [21, Appendix H]:
(gk /g)2
⟨B(t)B⟩B = ∑ (e−iωk t + eiωk t−βωk ) . (647)
k 1 − e−βωk
 95

β=1, ω c =10 β=0.1, ω c =10

γ(ω) γ(ω)

10

3 8

6
2

1
2

ω ω
-10 10 20 30 40 50 60 -20 -10 10 20 30 40 50

FIG. 8. The Lindblad rate γ(ω) for an Ohmic spectral density [Eq. (650c)], for η = 1/(2π), ωc = 20, and low temperature β = 10 (left) or high
temperature β = 0.1 (right). It can be checked numerically that the peak is always at ω ≈ ωc for sufficiently large β, or at ω = 0 for sufficiently
small β. Note that γ(0) = 1/β.

Let us introduce a spectral density J(ω) = ∑k (gk /g)2 δ(ω − ωk ) via

∞
∑(gk /g) ↦ ∫
2
dωJ(ω) , (648)
k 0

and let us further assume that it is Ohmic:

J(ω) = ηωe−ω/ωc , (649)
12
where ωc is a cut-off frequency and η is a dimensionless parameter.
With this model of the bath spectral density function, we can compute the rate γ(ω) as the Fourier transform of the bath
correlation function,
∞ ∞ ∞ J(ω ′ )
γ(ω) = ∫ dteiωt ⟨B(t)B(0)⟩ = ∫ dω ′ (e−iω t + eiω t−βω )
′ ′ ′
dteiωt ∫ (650a)
−∞ −∞ 0 1 − e−βω′
2πη ∣ω∣ e−∣ω∣/ωc
= (Θ(ω) + e−β∣ω∣ Θ(−ω)) (650b)
1 − eβ∣ω∣
2πηωe−∣ω∣/ωc
= . (650c)
1 − e−βω
where Θ(x) is the Heaviside step function (0 if x < 0 or 1 if x > 0). Note that the KMS condition is satisfied. The result is
shown in Fig. 8.
2πη(−ω)e−∣ω∣/ωc
γ(−ω) = = e−βω γ(ω) . (651)
eβω (e−βω − 1)
In the limit of large βω we can neglect e−βω in the denominator; differentiating we then get 2πηe−∣ω∣/ωc (1 − ω/ωc ), so that the
maximum is at ω = ωc . Also note that
2πη
lim γ(ω) = = 2πηkB T , (652)
ω→0 β
which tells that the transition rate in the limit of small gaps is linear in the temperature. This means that the SCL result for the
dephasing rate becomes
(c)
1/T2 = 2g 2 γ(0) = 4πg 2 ηkB T , (653)
meaning that the dephasing rate increasing in proportion to the temperature and the square of the coupling strength.
(e) (e)
For the WCL, recall that we found that the dephasing and relaxation rates in the energy eigenbasis are 1/T2 = 1/(2T1 ) =
1
2
[γ(ωx ) + γ(−ωx )] [Eq. (635)]. Considering Fig. 8, we see that γ(−ωx ) ≪ γ(ωx ), and that both rates are highly suppressed
when ωx ≫ ωc . For large β they are maximized when ωx ≈ ωc and become small for ωx < ωc , but are lower-bounded by
γ(0) = 1/β.

12 If J(ω) = ηω ζ e−ω/ωc then the ζ > 1 case is called super-Ohmic, and the 0 < ζ < 1 case is called sub-Ohmic.
 96

C. Example: collective vs independent phase damping

To close our discussion of the RWA-LE, let us revisit the phase damping model we considered in Sec. X C, but this time for n
qubits. Thus the system Hamiltonian is
n
HS = ∑ εα Zα . (654)
α=1

The eigenstates {∣a ⟩}2a=0−1 are just the computational basis states, i.e., all length-n bit strings.
n

We will consider two cases: collective and independent phase damping.

1. The collective case

In the collective phase damping case there is a qubit permutation symmetry and the qubits are all coupled to the same bosonic
modes. Thus the system-bath interaction is
n
HSB = ∑ gk Zα ⊗ (bk + b†k ) = gA ⊗ B , A = ∑ Zα , B = ∑(gk /g)(bk + b†k ) . (655)
k,α α=1 k

Since there is only one bath operator, the analysis starting from Eq. (647) holds without any change.

2. The independent case

Here each qubit is coupled to a separate bosonic bath. Thus the bath Hamiltonian is
n
HB = ∑ HB,α , HB,α = ∑ ωk,α b†k,α bk,α , (656)
α=1 k

where bk,α is the annihilation operator associated with bosonic mode k and qubit α, and the system-bath interaction is
n
HSB = ∑ gk,α Zα ⊗ (bk,α + b†k,α ) = g ∑ Aα ⊗ Bα , A α = Zα , Bα = ∑(gk,α /g)(bk,α + b†k,α ) . (657)
k,α α=1 k

The bath Gibbs state factors since operators belonging to different qubit indices commute:
1 1 −βHB,α
ρB = e−βHB = ⊗ ρB,α , ρB,α = e , (658)
Z α Z α

where Zα = Tre−βHB,α . In light of this case, consider the bath correlation functions, and recall that Tr(A ⊗ B) = TrA × TrB
for any pair of operators A and B:
⟨Bα (t)Bβ ⟩B = Tr (ρB eiHB,α t Bα e−iHB,α t Bβ ) (659a)
α≠β −iHB,α t
= Tr (ρB,α e iHB,α t
Bα e ) Tr (ρB,β Bβ ) (659b)
=0, (659c)
where the last equality follows since (as in Sec. XIV C) we can always ensure that Tr(ρB B) = 0. If α = β, we recover the
expression we obtained in the collective case but with the bath parameters corresponding to the α-th bath. Thus,
⟨Bα (t)Bβ ⟩B = δαβ ⟨Bα (t)Bα ⟩B . (660)
This, in turn, implies that
∞
γαβ (ω) = ∫ dteiωt ⟨Bα (t)Bβ ⟩B = δαβ γαα (ω) . (661)
−∞
If we again assume an Ohmic spectral density, now of the form
Jα (ω) = ηα ωe−ω/ωc,α , (662)
then the same calculation as in the collective case yields
2πηα ωe−∣ω∣/ωc,α
γαα (ω) = , (663)
1 − e−βω
where we have assumed that all baths are thermally equilibrated at the same inverse temperature β.
 97

3. Contrasting the dephasing rates in the collective and independent cases

We can now compare the predictions of the collective and independent dephasing models. Consider the time evolution of the
density matrix elements in the energy eigenbasis, i.e., ρ̇ab . Using the RWA-LE we have:
1
ρ̇ab = ⟨a ∣ρ̇∣b ⟩ = ⟨a ∣ ∑ γαβ (ω) (Aβ (ω)ρA†α (ω) − {A†α (ω)Aβ (ω), ρ}) ∣b ⟩ . (664)
αβ,ω 2
Evaluating this yields, after some algebra:
n
1
ind
1/τab = g 2 ∑ γαα (0)(Aaa,α − Abb,α )2
independent: (665a) ρ̇ab = −ρab /τab
ind
,
2 α=1
1
collective: ρ̇ab = −ρab /τab
col col
, 1/τab = g 2 γ(0)(Aaa − Abb )2 , (665b)
2
where we used the explicit form of the eigenstates of the system Hamiltonian in Eq. (654). We see that, as expected from
single-qubit dephasing case (recall, e.g., Sec. XIV G) that there is no change in the populations, i.e., ρ̇aa = 0. The solution to
these decoupled equations for the off-diagonal elements is of the form ρab (t) = ρab (0)e−t/τab , where τab is the dephasing time.
Let us compare the scaling of this time with the number of qubits n in the independent and collective dephasing settings.
• Independent-dephasing:
Aaa,α = ⟨a ∣Zα ∣a ⟩ = ±1 . (666)
Thus, (Aaa,α − Abb,α )2 = 4 for a ≠ b.
• Collective dephasing:
n
Aaa = ⟨a ∣ ∑ Zα ∣a ⟩ ∈ {−n, −n + 2, . . . , n − 2, n} (667)
α=1

Thus max(Aaa − Abb )2 = 4n2 and min(Aaa − Abb )2 = 0 for even n, or min(Aaa − Abb )2 = 4 for odd n.
There is thus a substantial difference between the two models. In the independent case, using Eq. (665a), we find 1/τab ind
=
O(n), or simply 1/τab = 2nγ(0) if all rates γαα (0) are equal [to γ(0)]. In the collective case, using Eq. (665b), we have a
ind

range of dephasing rates, varying from “superdecoherent” 1/τab col

= 2n2 γ(0), to “decoherence-free” 1/τab col
= 0 for even n or to
“subdecoherent” 1/τab = 2γ(0) for odd n. The decoherence-free case is of particular interest in quantum computing, and arises
col

for the zero-eigenvalue system eigenstates of the collective dephasing operator ∑nα=1 , i.e., states ∣a ⟩ that have an equal number
of 0’s and 1’s in the computational basis. Such states form a conserved subspace under the action of the RWA-LE, and hence are
called a decoherence-free subspace [26–28] (recall also our discussion of non-equilibration in Sec. XVI C). At the other extreme,
the states in the superdecoherent subspace dephase quadratically faster than in the independent dephasing case.

D. Bounding the Markov approximation error

Earlier we asserted that it is permissible to go from Eq. (512) to Eq. (515). Our goal is now to prove this, and in particular to
derive the associated error estimate, O(g 4 τB3 ).
Consider just one of the four (two due to the commutator, times two due to the h.c.) terms in Eqs. (512), and its Markov
approximation [as in Eq. (515)]:
t
true ≡ g 2 ∑ ∫ dτ Bαβ (τ )Aα (t)Aβ (t − τ )ρ̃(t − τ ) (668a)
αβ 0
∞
≈ g2 ∑ ∫ dτ Bαβ (τ )Aα (t)Aβ (t − τ )ρ̃(t) ≡ approx (668b)
αβ 0
∞ ∞
= g2 ∑ ∫ dτ Bαβ (τ )Aα (t)Aβ (t − τ ) (ρ̃(t) − ρ̃(t − τ )) +g 2 ∑ ∫ dτ Bαβ (τ )Aα (t)Aβ (t − τ )ρ̃(t − τ ) (668c)
αβ 0 αβ 0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∆1
∞
= g 2 ∑ ∆1 + true + g 2 ∑ ∫ dτ Bαβ (τ )Aα (t)Aβ (t − τ )ρ̃(t − τ ) . (668d)
αβ t
αβ
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∆2
 98

Thus, approx = true + ∆1 + ∆2 , or

error = ∥true − approx∥ = ∥∆1 + ∆2 ∥ ≤ ∥∆1 ∥ + ∥∆2 ∥ . (669)
This shows that in order to bound the error it suffices to bound ∥∆1 ∥ and ∥∆2 ∥ in a convenient norm, which we will take to
be the operator norm (see Appendix B for a discussion of the various norms we use here and their properties). The other three
terms in Eqs. (512) will obey exactly the same bound, since they are different from Eq. (668) only in the operator order, which
is removed once we take the norm. Thus, it suffices to concern ourselves with the term in Eq. (668).

1. Bound on ∥∆1 ∥

Using the triangle inequality and submultiplicativity of the operator norm ∥ ⋅ ∥∞ :

∞
∥∆1 ∥∞ ≤ ∫ dτ ∣Bαβ (τ )∣∥Aα (t)∥∞ ∥Aβ (t − τ )∥∞ ∥ρ̃(t) − ρ̃(t − τ )∥∞ (670a)
0
∞
=∫ dτ ∣Bαβ (τ )∣∥Aα ∥∞ ∥Aβ ∥∞ ∥ρ̃(t) − ρ̃(t − τ )∥∞ (670b)
0
∞
≤ η2 ∫ dτ ∣Bαβ (τ )∣∥ρ̃(t) − ρ̃(t − τ )∥∞ , (670c)
0

where in the second line we used unitary invariance, and where

η ≡ max ∥Aα ∥∞ . (671)
α

Now, by the mean value theorem of elementary calculus, there exists a point t′ ∈ [t − τ, t] such that
ρ̃(t) − ρ̃(t − τ ) ˙ ′
= ρ̃(t ) . (672)
τ
Therefore
∥ρ̃(t) − ρ̃(t − τ )∥∞ ≤ τ ˙ ′ )∥∞ ,
sup ∥ρ̃(t (673)
t′ ∈[t−τ,t]

and
∞
∥∆1 ∥∞ ≤ η 2 ∫ ˙ ′ )∥∞ .
dτ τ ∣Bαβ (τ )∣ sup ∥ρ̃(t (674a)
0 t′ ∈[t−τ,t]

˙ ′ )∥∞ we can return to Eq. (512):

To bound ∥ρ̃(t
t′
˙ ′ )∥∞ ≤ g 2 ∑ ∫
∥ρ̃(t dτ ∣Bαβ (τ )∣∥[Aα (t), Aβ (t − τ )ρ̃(t − τ )] + h.c.∥∞ (675a)
α,β 0

t′
≤ 4g 2 ∑ ∫ dτ ∣Bαβ (τ )∣∥Aα (t)Aβ (t − τ )ρ̃(t − τ )∥∞ (675b)
α,β 0

t′
≤ 4g 2 ∑ ∫ dτ ∣Bαβ (τ )∣∥Aα ∥∞ ∥Aβ ∥∞ ∥ρ̃(t − τ )∥1 (675c)
α,β 0

t′
≤ 4(ηg)2 M ∫ dτ ∣Bαβ (τ )∣ , (675d)
0

where in the second line we used the fact that all four terms in the first line (again, after the commutator and h.c.) have the same
operator norm, and where M ≡ ∑αβ 1 is the square of the number of summands in HSB = ∑α Aα ⊗ Bα . Now, since
t′ ∞
sup ∫ dτ ∣Bαβ (τ )∣ ≤ ∫ dτ ∣Bαβ (τ )∣ , (676)
t′ ∈[t−τ,t] 0 0

we have
∞ ∞
∥∆1 ∥∞ ≤ 4M η 4 g 2 ∫ dτ τ ∣Bαβ (τ )∣ ∫ dτ ∣Bαβ (τ )∣ ∼ 4M η 4 g 2 τB3 , (677)
0 0

where we used Eq. (516) once with n = 1, and once with n = 0.

 99

2. Bound on ∥∆2 ∥

Similarly,
∞
∥∆2 ∥∞ ≤ ∫ dτ ∣Bαβ (τ )∣∥Aα (t)∥∞ ∥Aβ (t − τ )∥∞ ∥ρ̃(t − τ )∥1 (678a)
t
∞
≤ η2 ∫ dτ ∣Bαβ (τ )∣ . (678b)
t
∞
Intuitively, we know that ∫t dτ ∣Bαβ (τ )∣ should be arbitrarily small as long as t ≫ τB , as we assumed in Eq. (513), since the
∞
correlation function decays over a timescale of τB . To formalize this, note that convergence of ∫t dτ ∣Bαβ (τ )∣ is guaranteed if

∣Bαβ (τ )∣ ∼ (τB /τ )x , x>1. (679)

Thus, we will assume that the correlation function decays no more slowly than this power-law dependence [this is even slower
than the subexponential decay we assumed to get Eq. (518)]. Under this assumption, we have
∞
∞ ∞ τB x τBx 1 τBx
∫ dτ ∣Bαβ (τ )∣ ∼ ∫ dτ ( ) = ∣ = . (680)
t t τ (1 − x)τ x−1 t x − 1 tx−1

Now, to use the assumption that t ≫ τB , let us write t = cτB , where c ≫ 1. Then:
∞ τB
∫ dτ ∣Bαβ (τ )∣ ∼ . (681)
t (x − 1)cx−1
Therefore, even with a power-law decaying correlation function, we have
τB
∥∆2 ∥∞ ≲ η 2 , (682)
(x − 1)cx−1
which can be made arbitrarily small by making c = t/τB large enough.

3. Putting the bounds together

We have seen that ∥∆1 ∥∞ ≲ 4M η 4 g 2 τB3 = O(g 2 τB3 ) and ∥∆2 ∥∞ can be made arbitrarily small. Thus the dominant contribu-
tion to the error comes from ∥∆1 ∥∞ , which is the error due to replacing all the intermediate-time states (at t − τ ) by the state at
the single time t. Moreover, we need t ≫ τB in order to ensure that ∥∆2 ∥∞ can be neglected.
When accounting for the additional g 2 prefactor in Eq. (668) (as well as ∑αβ , which just gives rise to another factor of M ),
we finally have from Eq. (669):

error = O(g 4 τB3 ) , (683)

as claimed.

E. The RWA-LE is the infinite coarse-graining time limit of the cumulant-LE

The RWA we used in Sec. XV E in order to derive the Lindblad equation leaves something to be desired. We simply dropped
terms with different Bohr frequencies, without a rigorous mathematical justification. We will now show that the RWA-LE can be
rigorously derived from the cumulant-LE, in the limit of an infinite coarse-graining timescale. This shows that the cumulant-LE
is truly more general than the (standard) RWA-LE.

1. Quick summary

For convenience, let us collect the main results of each of the two approaches. For simplicity we’ll set λ = g = 1 and also
assume that HSB = A ⊗ B (not a sum), so that we can drop the α index from Eq. (536). The RWA-LE is then:
1
˙
ρ̃(t) = −i [HLS , ρ̃(t)] + ∑ γ(ω) (Aω ρ̃(t)A†ω − {A†ω Aω , ρ̃(t)}) (684)
ω 2
 100

with
∞
γ(ω) = ∫ dseiωs B(s, 0) . (685)
−∞

The cumulant-LE is
′ 1
˙
ρ̃(t) = −i [HLS , ρ̃(t)] + ∑ γωω′ (τ ) (Aω ρ̃(t)A†ω′ − {A†ω′ Aω , ρ̃(t)}) , (686)
ω,ω ′ 2

where the rates γ keep a dependence on two different Bohr frequencies ω and ω ′ :
1 τ τ
γωω′ (τ ) = bωω′ (τ ) , bωω′ (τ ) = ∫ ds′ ei(ω s−ωs ) B(s, s′ ) .
′ ′
ds ∫ (687)
τ 0 0

Our goal is to show that in an appropriate sense the cumulant-LE tends to the RWA-LE in the limit as τ → ∞, where τ is the
coarse-graining timescale. More specifically, we will show that limτ →∞ γωω′ (τ ) = γ(ω)δωω′ [17]. We will assume stationarity,
i.e., B(s, s′ ) = B(s − s′ , 0).

2. A useful lemma

Lemma 3. The following equivalent form holds for γωω′ (τ ):

τ v
1 ω′ −ω ω′ − ω ω+ω ′
γ ωω ′ (τ ) = ei 2 τ ∫ dv cos ( (v − τ )) ∫ du ei 2 u B(u, 0) . (688)
τ 2
0 −v

Proof. In the RWA we dropped terms with ω ≠ ω ′ , so it makes sense to rewrite ω ′ s − ωs′ in terms of a sum and difference of
Bohr frequencies:
1 1
ω ′ s − ωs′ = (ω ′ − ω)v + (ω ′ + ω)u , (689)
2 2
where u = s − s′ and v = s + s′ . After this change of variables B(s − s′ , 0) = B(u, 0), and since s = (v + u)/2 and s′ = (v − u)/2,
1/2 1/2
the Jacobian of the transformation is ∣( )∣ = 1/2. In terms of the new variables the integration region is diamond
1/2 −1/2
shaped (a square rotated by π/4), bounded between the lines u = v and u = −v for v ∈ [0, τ ] and the lines u = 2τ − v and v − 2τ
for v ∈ [τ, 2τ ]. Thus:
τ v 2τ 2τ −v
1 ω ′ −ω ω+ω ′ 1 ω ′ −ω ω+ω ′
bωω ′ (τ ) = ∫ dv ei 2 v ∫ du ei 2 u B(u, 0) + ∫ dv ei 2 v ∫ du ei 2 u
B(u, 0) . (690)
2 2
0 −v τ −(2τ −v)

To get the integration limits to be the same we make a change of variables from v to 2τ − v in the second double integral:
τ v τ v
1 ω ′ −ω ω+ω ′ 1 ω ′ −ω ω+ω ′
bωω′ (τ ) = ∫ dv ei 2 [(v−τ )+τ ] ∫ du ei 2 u B(u, 0) + ∫ dv e−i 2 [(v−τ )−τ ] ∫ du ei 2 u B(u, 0) (691a)
2 2
0 −v 0 −v
τ v
′
i ω 2−ω τ ω′ − ω ω+ω ′
=e ∫ dv cos ( (v − τ )) ∫ du ei 2 u B(u, 0) . (691b)
2
0 −v

The claim now follows from Eq. (687).

3. The ω = ω ′ case

For ω = ω ′ we now have:

τ v
1
γωω (τ ) = ∫ dv ∫ du e B(u, 0) .
iωu
(692)
τ
0 −v
 101

v
Let U = ∫−v du eiωu B(u, 0). Recall the Leibnitz rule for differentiating a definite integral:
b(z) b(z)
∂z ∫ f (x, z)dx = ∫ ∂z f (x, z)dx + f (b(z), z)b′ − f (a(z), z)a′ . (693)
a(z) a(z)

Therefore dU = (eiωv B(v, 0) + e−iωv B(−v, 0)) dv. Then, integrating by parts (∫0 U dv = [U v]0 − ∫0 vdU ) gives:
τ τ τ

⎡ v ⎤τ τ
1 ⎢⎢ ⎥ 1
γωω (τ ) = ⎢v ∫ du e B(u, 0)⎥⎥ − ∫ dv v (eiωv B(v, 0) + e−iωv B(−v, 0)) .
iωu
(694)
τ⎢ ⎥ τ
⎣ −v ⎦0 0

Consider the second integral:

RRR τ RRR τ ∞
RRR 1 RRR 1 1
RRR ∫ dv ve B(v, 0)RRR ≤ ∫ dv v ∣B(v, 0)∣ ≤ ∫ dv v ∣B(v, 0)∣
iωv
(695a)
RRR τ 0 RRR τ 0 τ
0
1 τ →∞
∼ τB2 Ð→ 0 , (695b)
τ
where in the last step we used the assumption (516) that the bath correlation function decays with a finite timescale τB . Since
B(v, 0) = B ∗ (−v, 0) [recall Eq. (511b)], the third integral in Eq. (694) satisfies the same bound and limit. We are thus left with
∞
lim γωω (τ ) = ∫ du eiωu B(u, 0) = γ(ω) , (696)
τ →∞
−∞

where the last equality is due to Eq. (685).

4. The ω ≠ ω ′ case

For ω ≠ ω ′ we also perform integration by parts of Eq. (688), but we shall see that this time the boundary terms vanish. We
ω ′ −ω ω+ω ′
write γωω′ (τ ) = τ1 ei 2 τ ∫0 dV U (v), where now dV = cos ( ω 2−ω (v − τ )) dv and U (v) = ∫−v du ei 2 u B(u, 0). Then
τ ′ v

2 ω′ − ω
V (v) = sin ( (v − τ )) (697a)
ω′ − ω 2
ω+ω ′ ω+ω ′
dU /dv = ei 2 v
B(v, 0) + e−i 2 v
B(−v, 0) (697b)
τ
[U (v)V (v)]0 = U (τ )V (τ ) − U (0)V (0) = 0 . (697c)
Therefore:
ω ′ −ω τ
τ 2ei 2 τ (ω ′ − ω) ω+ω ′ ω+ω ′
γωω′ (τ ) = − ∫ V dU = − ′ ∫ dv sin ( (v − τ )) [ei 2 v B(v, 0) + e−i 2 v B(−v, 0)] . (698)
0 (ω − ω)τ 2
0

Changing from v to −v in the second term we get

⎡ τ 0 ⎤
2ei 2 τ ⎢⎢ ⎥
ω ′ −ω
(ω ′ − ω) (ω ′ − ω)
B(v, 0)⎥⎥
′ ′
i ω+ω i ω+ω
γωω′ (τ ) = − ′ ⎢ ∫ dv sin ( (v − τ )) e 2 v
B(v, 0) + ∫ dv sin ( (−v − τ )) e 2 v
(ω − ω)τ ⎢ 2 2 ⎥
⎣0 −τ ⎦
(699a)
ω ′ −ω τ
ei 2 τ ω′ − ω sgn(v) ω′ − ω
= ′ ∫ dv [sin ( τ ) (eiωv + eiω v ) + cos ( τ ) (eiωv − eiω v )] B(v, 0) ,
′ ′
(699b)
(ω − ω)τ 2 i 2
−τ

where we used the angle sum identity for the sine in the last equality. Thus:
ω ′ −ω
ei 2 τ ω′ − ω ′ ω′ − ω
lim γ ωω ′ (τ ) = lim [sin ( τ ) (γ(ω) + γ(ω )) + 2 cos ( τ ) (S(ω) − S(ω ′ ))] . (700)
τ →∞ τ →∞ (ω ′ − ω)τ 2 2
∞
where we have used that for Γ(ω) = ∫0 ds eiωs B(s, 0) [recall Eq. (523)], we have γ(ω) = Γ(ω) + Γ∗ (ω) and 2iS(ω) =
Γ(ω) − Γ∗ (ω) [recall Eq. (535)]. Since nothing cancels with the overall τ −1 , we find that the ω ≠ ω ′ term vanishes.
A similar calculation could be done for the Lamb shift term (466). Therefore, the RWA results can be understood as the
τ → ∞ limit of the coarse-graining timescale.
 102

XIX. THE NAKAJIMA-ZWANZIG EQUATION

The master equations we have developed so far are approximations to the true dynamics. In this section we take a step back
and derive an exact master equation. Since it is exact, it will naturally be non-Markovian.
Consider the total Hamiltonian

H = H0 + αHSB , H0 = HS + HB , (701)

where 0 < α < 1 is a dimensionless parameter. Let us work in the interaction picture, so that the total system-bath state ρ̃ satisfies

∂t ρ̃ = −iα[H̃(t), ρ̃(t)] ≡ αLρ̃(t) , (702)

where as usual H̃(t) = U0† (t)HSB U0 (t), with U0 (t) = e−iH0 t , and HSB = ∑ Sα ⊗ Bα . We abbreviate ∂t ≡ ∂t ∂
. For the rest
of this section we drop the tilde decoration on states to simplify the notation, so that, e.g., ρ denotes the interaction-picture
system-bath state.

A. Feshbach P-Q partitioning

Consider a fixed bath state ρB . As usual, ρS = TrB ρ is the system state of interest. Consider the projection superoperator P
defined via

Pρ = TrB (ρ) ⊗ ρB . (703)

That P is a projection follows from applying it twice:

P 2 ρ = P[TrB (ρ) ⊗ ρB ] = TrB [TrB (ρ) ⊗ ρB ] ⊗ ρB = TrB (ρ) ⊗ ρB = Pρ . (704)

Define the orthogonal projection Q via

Q=I −P . (705)

We call Pρ the “relevant” part, and Qρ the “irrelevant part”. This procedure is sometimes called Feshbach P-Q partitioning,
after a method introduced in nuclear scattering theory [32].
We are interested in deriving a master equation for ∂t (Pρ). Now, note that

∂t (Pρ) = ∂t [TrB (ρ) ⊗ ρB ] = TrB (∂t ρ) ⊗ ρB = P(∂t ρ) , (706)

i.e., [P, ∂t ] = 0. Therefore, using Eq. (702):

∂t (Pρ) = αPLρ . (707)

Likewise:

∂t (Qρ) = ∂t [(I − P)ρ] = αLρ − αPLρ = α(I − P)Lρ = αQLρ . (708)

Let us now insert I = P + Q into the last two equations:

∂t (Pρ) = αPL(P + Q)ρ = αPLPρ + αPLQρ (709a)

∂t (Qρ) = αQL(P + Q)ρ = αQLPρ + αQLQρ . (709b)

These are coupled differential equations for the relevant (Pρ) and irrelevant (Qρ) parts. To solve them, let us eliminate the
irrelevant part.
Define

X̂ ≡ PX , X̄ ≡ QX (710)

for any operator X. then Eq. (709) can be rewritten more compactly as:

∂t ρ̂ = αL̂ρ̂ + αL̂ρ̄ (711a)

∂t ρ̄ = αL̄ρ̂ + αL̄ρ̄ . (711b)
 103

B. Derivation

We can formally solve the second of these equations and substitute the solution into the first. Consider first ∂t ρ̄ = αL̄ρ̄. This
has the immediate solution ρ̄(t) = T+ exp (α ∫t0 L̄(t′ )dt′ ) ρ̄(t0 ), where T+ denotes the usual forward Dyson time-ordering. We
t

thus define
α ∫tt L̄(t′ )dt′
G(t, t0 ) ≡ T+ e 0 . (712)
Eq. (711b) contains another term, and we can easily guess that the solution integrates over this term, but first applies G, i.e.:
t
ρ̄(t) = G(t, t0 )ρ̄(t0 ) + α ∫ G(t, t′ )L̄(t′ )ρ̂(t′ )dt′ . (713)
t0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
⊛

To verify that this is the formal solution of Eq. (711b), we apply the Leibnitz rule (693) to get ∂t ∫t0 f (t, t′ )dt′ = f (t, t) +
t

∂t f (t, t′ )dt′ , and also note that G(t, t0 ) has the property G(t, t) = I, ∂t G(t, t′ ) = L̄(t)G(t, t′ ). Using all of the above we
t
∫t0
have:
t
∂t ⊛ = αG(t, t)L̄(t)ρ̂(t) + α ∫ ∂t G(t, t′ )L̄(t′ )ρ̂(t′ )dt′ = αL̄(t)ρ̂(t) + αL̄(t)⊛ = αL̄(t) (ρ̂(t) + ⊛) . (714)
t0

Therefore, if we differentiate Eq. (713) we find:

∂t ρ̄(t) = αL̄(t)G(t, t0 )ρ̄(t0 ) + αL̄(t)(ρ̂(t) + ⊛) = αL̄(t)ρ̂(t) + αL̄(t) (G(t, t0 )ρ̄(t0 ) + ⊛) , (715)
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
ρ̄(t)

which agrees with Eq. (711b) as required.

Substituting the solution for ρ̄(t) into Eq. (711a), we have:
t
∂t ρ̂(t) = αL̂(t)ρ̂(t) + αL̂(t)G(t, t0 )ρ̄(t0 ) + α2 ∫ L̂(t)G(t, t′ )L̄(t′ )ρ̂(t′ )dt′ . (716)
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ t0
(a) (b) ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
(c)

• We can show that term (a) can always be made to vanish in a similar way to what we did in Sec. XIV C. To see this, note
that
L̂(t)ρ̂(t) = PL(t)Pρ(t) = PL(t)TrB [ρ(t)] ⊗ ρB = −iP [H̃(t), ρS (t) ⊗ ρB ] (717a)
= −i ∑ TrB (Aα (t)ρS (t) ⊗ Bα (t)ρB ) ⊗ ρB − TrB (ρS (t)Aα (t) ⊗ ρB Bα (t)) ⊗ ρB (717b)
α
= −i ∑ [A(t), ρS (t)] ⟨Bα (t)⟩ ⊗ ρB = 0 (717c)
α

since ⟨Bα (t)⟩ can be made zero in the same way as in Eq. (469), i.e., [HB , ρB (0)] = 0.
• Term (b) is an inhomogeneity that depends on the initial condition and measures how much correlation there is in the
initial state:
ρ̄(0) = (I − P)ρ(0) = ρ(0) − TrB [ρ(0)] ⊗ ρB . (718)
It vanishes for a factorized initial state, i.e., if ρ(0) = ρS (0) ⊗ ρB (the same fixed initial state we chose for the bath at the
beginning of the derivation).
Thus, assuming a factorized initial state Eq. (716) becomes:
t
∂t ρ̂(t) = ∫ K(t, t′ )ρ̂(t′ )dt′ (719a)
t0

K(t, t′ ) ≡ α L̂(t)G(t, t′ )L̄(t′ )P . 2

(719b)
Equation (719) is called the (homogeneous) Nakajima-Zwanzig master equation (NZ-ME), and the superoperator K is called the
memory kernel (note that we multiplied it from the right by P, which we can do since it acts on ρ̂ = Pρ). If we include the (b)
term αL̂(t)G(t, t0 )ρ̄(t0 ) from Eq. (716) on the RHS we have the inhomogeneous NZ-ME.
The NZ-ME is exact, non-perturbative, and in the inhomogeneous case it can even describe non-factorized initial conditions.
It is clearly non-local in time, in the sense that the RHS retains a memory of the entire history of the state evolution, weighted
via the memory kernel. The Nakajima-Zwanzig equation is an integro-differential equation, and solving it is essentially as hard
as solving the original Liouville-von Neumann equation (702). Nevertheless, it provides an important and convenient starting
point for perturbative expansions, as we shall see shortly.
 104

C. From the Nakajima-Zwanzig equation to the Born master equation

Consider a perturbative expansion in α. To lowest order we have from Eq. (712):

G(t, t0 ) = I + O(α) , (720)

so at the same order the memory kernel becomes

K(t, t′ ) = α2 L̂(t)[I + O(α)]L̄(t′ )P = α2 PL(t)QL(t′ )P + O(α3 ) , (721)

and hence:
t
∂t [Pρ(t)] = ∂t ρS (t) ⊗ ρB = α2 ∫ PL(t)QL(t′ )Pρ(t′ )dt′ (722a)
t0
t
= α2 ∫ PL(t)L(t′ )Pρ(t′ )dt′ (722b)
t0
t
= −α2 ∫ P [H̃(t), [H̃(t′ ), ρS (t′ ) ⊗ ρB ]] dt′ (722c)
t0
t
= −α2 ∫ TrB [H̃(t), [H̃(t′ ), ρS (t′ ) ⊗ ρB ]] ⊗ ρB dt′ (722d)
t0

where in the second line we used PL(t)P = 0 [Eq. (717)]. Applying one final TrB to both sides finally gives
t
∂t ρS (t) = −α2 ∫ TrB [H̃(t), [H̃(t′ ), ρS (t′ ) ⊗ ρB ]] dt′ , (723)
t0

which we recognize as the Born master equation [Eq. (512)] discussed in Sec. XV B.

D. The O(α3 ) term of the Nakajima-Zwanzig master equation

The O(α3 ) term comes from the α1 term in the propagator

G(t, t0 ) = I + g1 (t, t0 ) + O(α2 ) , (724)

where
t
g1 (t, t0 ) = α ∫ L̄(s)ds . (725)
t0

The O(α3 ) term is

t t t
α2 ∫ dt′ PL(t)g1 (t, t′ )QL(t′ )Pρ(t′ ) = α3 ∫ dt′ PL(t) ∫ dsQL(s)QL(t′ )P ρ(t′ ) (726a)
t0 t0 t′
t t
= α3 ∫ ∫ dt′ dsPL(t)QL(s)QL(t′ )Pρ(t′ ) , (726b)
t0 t′

where

PL(t)QL(s)QL(t′ )P = PL(t)(I − P)L(s)(I − P)L(t′ )P (727a)

= PL(t)L(s)L(t′ )P − PL(t)L(s)PL(t′ )P − PL(t)PL(s)L(t′ )P + PL(t)PL(s)PL(t′ )P .
(727b)

It turns out that we can always ensure that

PL(t1 )⋯L(tn )P = 0 (728)

for any odd n and any ordering of the time argument, by appropriately shifting the bath operators. Therefore the order O(α3 )
term vanishes, and the Nakajima-Zwanzig master equation is unchanged at this order, namely:
t
∂t ρS (t) = −α2 ∫ TrB [H̃(t), [H̃(t′ ), ρS (t′ ) ⊗ ρB ]] dt′ + O(α4 ) . (729)
t0
 105

XX. THE TIME CONVOLUTIONLESS (TCL) MASTER EQUATION

The Nakajima-Zwanzig equation (719) contains a convolution with a complicated memory kernel [Eq. (719b)]: ∂t ρ̂(t) =
K(t, t′ )ρ̂(t′ )dt′ . It seems that this is an unavoidable feature of an exact, non-Markovian master equation. In this section we
t
∫t0
will see that it is possible to remove the memory kernel by making a type of short-time approximation, and arrive at a fully time-
local, convolutionless master equation. The main insight we’ll need to achieve this, is that the memory kernel can be removed
by formally back-propagating the system state.

A. Derivation

1. Back-propagation

Let us start again from the Liouville-von-Neumann equation ∂t ρ̃ = −iα[H̃(t), ρ̃(t)] ≡ αLρ̃(t) [Eq. (702)]. Its formal solution
is

ρ̃(t) = T+ eα ∫t′ L(s)ds ρ̃(t′ ) = U+ (t, t′ )ρ̃(t′ ) ,

t
(730)

where U+ (t, t′ ) is a forward time-ordered superoperator. This can be inverted so that

ρ̃(t′ ) = T− e−α ∫t′ L(s)ds ρ̃(t) = U− (t, t′ )ρ̃(t) ,

t
(731)

which defines the backward time-ordered superoperator U− (t, t′ ). To get an explicitly representation note first that by substituting
ρ̃(t′ ) from Eq. (731) into Eq. (730) we get U+ (t, t′ )U− (t, t′ ) = I. Now, since
t − t′
U+ (t, t′ ) = T+ eα ∫t′ L(s)ds = lim eα∆tL(t−∆t) ⋯eα∆tL(t +∆t) eα∆tL(t )
t
(∆t = lim ),
′ ′
(732)
∆t→0 N →∞ N

in order to have U+ (t, t′ ) and U− (t, t′ ) be each other’s inverse, it must be that U− (t, t′ ) has the opposite order and α is replaced
by −α, so that when multiplied the two products cancel equal and opposite terms. I.e.,
t − t′
U− (t, t′ ) = T− e−α ∫t′ L(s)ds = lim e−α∆tL(t ) e−α∆tL(t +∆t) ⋯e−α∆tL(t−∆t)
t
(∆t = lim ).
′ ′
(733)
∆t→0 N →∞ N

Applying P to both sides of Eq. (731), and again dropping the tilde decoration to simplify the notation (though we continue
to work in the interaction picture) we have ρ̂(t′ ) = Û− (t, t′ )ρ(t), so that Eq. (713) becomes:

ρ̄(t) = G(t, t0 )ρ̄(t0 ) + Σ(t)ρ(t) (734a)

t
Σ(t) ≡ α ∫ G(t, t′ )L̄(t′ )Û− (t, t′ )dt′ . (734b)
t0

Note that the superoperator Σ(t) is not chronologically ordered since it contains both forward [via G(t, t′ ); recall Eq. (712)] and
backward time propagation. For this reason we do not write Σ(t, t0 ), despite the dependence of Σ(t) on t0 , since that notation
is reserved for propagation from t0 to t.13 Equation (734a) has removed the memory kernel and replaced it by (the even more
complicated object) Σ(t). However, in terms of the time-dependence of ρ, it is time-local, i.e., depends only on t (apart from
the initial condition t0 ). Next we solve this equation.

2. Solving for the relevant part

Let us insert I = P + Q into Eq. (734a):

ρ̄(t) = G(t, t0 )ρ̄(t0 ) + Σ(t)(P + Q)ρ(t) (735a)

Ô⇒ ρ̄(t) = G(t, t0 )ρ̄(t0 ) + Σ(t)ρ̂(t) + Σ(t)ρ̄(t) (735b)
Ô⇒ (I − Σ(t)) ρ̄(t) = G(t, t0 )ρ̄(t0 ) + Σ(t)ρ̂(t) . (735c)

13 We could write Σt0 (t) without danger of confusion, but this more cumbersome notation won’t turn out to be particularly helpful.
 106

We can solve this for ρ̄(t) provided I − Σ(t) is invertible, i.e., provided Σ(t) is not too close from identity. Since Σ(t0 ) = 0,
we can conclude that I − Σ(t) is invertible for sufficiently short evolution times. In addition, Σ(t) = O(α), so invertibility
should also hold provided the system-bath coupling is sufficiently weak. Thus, from now we shall assume that I − Σ(t) is indeed
invertible, which is the only assumption we shall make to arrive at the TCL master equation. Then:
−1 −1
ρ̄(t) = (I − Σ(t)) G(t, t0 )ρ̄(t0 ) + (I − Σ(t)) Σ(t)ρ̂(t) , (736)

and substituting this into Eq. (711a) we find:

−1 −1
∂t ρ̂(t) = αL̂(t)ρ̂(t) + αL̂ (I − Σ(t)) G(t, t0 )ρ̄(t0 ) + αL̂ (I − Σ(t)) Σ(t)ρ̂(t) (737a)
−1 −1
= αL̂ (I − Σ(t)) G(t, t0 )Qρ̄(t0 ) + αL̂(t) [I + (I − Σ(t)) Σ(t)] P ρ̂(t) , (737b)

where in the second line we used the freedom to insert a P and Q in front of ρ̂ and ρ̄, respectively. Note that
−1
I + (I − Σ) Σ = (I − Σ)−1 (I − Σ) + (I − Σ)−1 Σ = (I − Σ)−1 (I − Σ + Σ) = (I − Σ)−1 . (738)

We have thus arrived at the time-convolutionless master equation (TCL-ME):

∂t ρ̂(t) = J (t)ρ̄(t0 ) + K(t)ρ̂(t) , (739)

where
−1
J (t) ≡ αL̂ (I − Σ(t)) G(t, t0 )Q inhomogeneity , (740a)
−1
K(t) ≡ αL̂(t)(I − Σ(t)) P TCL generator . (740b)

The most salient feature of the TCL-ME is that (when the inhomogeneity vanishes, e.g., for factorized initial conditions) it is
purely time-local, in stark contrast to the NZ-ME [Eq. (719)].

B. Perturbation theory

Despite the formal appearance of the result we have found so far, it is a convenient starting point for perturbation theory.

1. Matching powers of α

n
Let us write (I − Σ(t))−1 = ∑∞
t
n=0 Σ (t), i.e., as a geometric series. It follows from Eq. (734b) that Σ (t) = α (∫t0 ⋯) , so
n n n

that after substitution into K(t) [Eq. (740b)] we have a series expansion in powers of α:
∞
K(t) = αL̂(t) ( ∑ Σn (t)) P (741a)
n=0
∞
= ∑ αn Kn (t) , (741b)
n=1

where we need to determine the operators Kn (t). To do so we need to first expand Σ(t) in powers of α. It also follows from
α t L̄(t′ )dt′
Eq. (734b) that the expansion must start from α1 , since G(t, t0 ) = T+ e ∫t0 = I + O(α) [Eq. (712)]:
∞
Σ(t) = ∑ αm Σm (t) . (742)
m=1

Substituting this expansion into Eq. (741a) yields a cumulant expansion:

⎡ ⎤
⎢ ⎥
∞ ∞ n ∞ ∞
m ⎢
K(t) = αL̂(t) [ ∑ ( ∑ α Σm (t)) ] P = L̂(t) ⎢αI + ∑ α m+1
Σm (t) + ∑ α m+m′ +1
Σm (t)Σm (t) + ⋯⎥⎥ P
′ (743)
⎢ m,m′ =1 ⎥
n=0 m=1
⎣ m=1
⎦
 107

Matching terms of equal power of α with Eq. (741b) yields, for the lowest four orders:
α1 ∶ K1 (t) = αL̂(t)P = 0 (744a)
α ∶ K2 (t) = αL̂(t)Σ1 (t)P
2
Redfield equation (744b)
α ∶ K3 (t) =
3
αL̂(t) (Σ21 (t) + Σ2 (t)) P =0 (744c)
α4 ∶ K4 (t) = αL̂(t) (Σ31 (t) + {Σ1 (t), Σ2 (t)} + Σ3 (t)) P lowest order non-Markovian (744d)
The vanishing of K1 (t) is for the same reason as in Eq. (717); that of K3 (t) is explained below. First we need to explicitly find
the lowest order Σm (t)’s. The expansions of G(t, s) = T+ eα ∫s L̄(t )dt and U− (t, t′ ) = T− e−α ∫t′ L(s)ds [Eq. (733)] yield:
t ′ ′ t

2
α2 t t
G(t, t′ ) = I + α ∫ L̄(s)ds +
T+ (α ∫ L̄(s)ds) + ⋯ (745a)
t′ 2! t′
t 2 t 2
α
Û− (t, t′ ) = P [I − α ∫ L(s)ds + T− (∫ L(s)ds) + ⋯] . (745b)
t′ 2! t′

We can now collect equal powers of α in Σ(t) = α ∫t0 G(t, t′ )L̄(t′ )Û− (t, t′ )dt′ = ∑∞
t
m=1 α Σm (t):
m

t
α1 ∶ Σ1 (t) = ∫ L̄(t′ )Pdt′ (746a)
t0
t t t t
α2 ∶ Σ2 (t) = − ∫ dt′ L̄(t′ )P ∫ L(s)ds + ∫ dt′ (∫ L̄(s)ds) L̄(t′ )P (746b)
t0 t′ t0 t′
t s
=∫ ds ∫ dt′ [L̄(s)L̄(t′ )P − L̄(t′ )PL(s)] , (746c)
t0 t0

where in the last line we switched the order of integration via ∫t0 dt′ ∫t′ ds = ∫t0 ds ∫t0 dt′ .
t t t s

Therefore, using Eq. (744b):

t t t
K2 (t) = L̂(t) ∫ L̄(t′ )dt′ P = L̂(t) ∫ (I − P)L(t′ )dt′ P = L̂(t) ∫ L(t′ )dt′ P , (747)
t0 t0 t0

where we again used PL(t)P = 0, which we also use repeatedly below.

To calculate K3 (t), first note that Σ21 (t) = ∫t0 ∫t0 dt′ dt′′ QL(t′ )PQL(t′′ )P = 0, since PQ = 0. Second, note that K3 (t)
t t

contains the term L̂(t)L̄(t′ )[PL(s)P] = 0. The final term it contains is L̂(t)L̄(s)L̄(t′ )P = PL(t)(I −P)L(s)(I −P)L(t′ )P =
PL(t)L(s)L(t′ )P = 0, by Eq. (728). Therefore K3 (t) = 0.

2. The TCL-ME at second order yields the Redfield equation

Let us consider the lowest non-vanishing order of the TCL-ME, Eq. (739). At this order:
∂t ρ̂(t) = α2 K2 (t)ρ̂(t) , (748)
where we have assumed a factorized initial condition, so that the inhomogeneity vanishes. We already found K2 (t) in Eq. (747),
so what remains is to make it explicit using the definition of the projection to the relevant part:
t
K2 (t)ρ̂(t) = ∫ PL(t)L(t′ )dt′ Pρ(t) (749a)
t0
t
=∫ dt′ TrB [−iH̃(t), [−iH̃(t′ ), (TrB ρ(t)) ⊗ ρB ]] ⊗ ρB . (749b)
t0

Thus, after applying TrB to both sides:

t
∂t ρS (t) = −α2 ∫ dt′ TrB [H̃(t), [H̃(t′ ), ρS (t) ⊗ ρB ]] . (750)
t0

This is the Redfield equation, Eq. (514). It is identical to the Born-Markov approximation [Eq. (512)], except for the finite upper
limit of the integral. It is also nearly identical to the second order NZ-ME [Eq. (723)], the only difference being the fact that,
by construction, Eq. (750) is time-local, in the sense that the argument of ρS is t rather than t′ . This is an important difference:
whereas when we derived the RWA-LE we had to just assume that we can replace t′ by t [in going from Eq. (512) to Eq. (514)],
here this is a systematic result of our derivation.
 108

C. Example: spin-boson model of a qubit in a cavity

As an application of the TCL-ME we now consider a qubit in a cavity. This is an analytically solvable model subject to a
simplifying assumption about the initial condition. As such, it will allow us to compare the predictions of the TCL to an exact
result.
Consider as usual the total Hamiltonian H = H0 + HSB , where H0 = HS + HB , with
HS = ω0 ∣1⟩⟨1∣ = ω0 σ+ σ− , HB = ∑ ωk b†k bk = ∑ ωk nk , (751a)
k k

HSB = σ+ ⊗ B + σ− ⊗ B † , B = ∑ gk bk . (751b)
k

Here σ+ = ∣1⟩⟨0∣ and σ− = ∣0⟩⟨1∣ are the qubit raising and lowering operators, while bk and b†k are the bosonic lowering and raising
operators for mode k, satisfying the canonical bosonic commutation relations [bk , b†k′ ] = δkk′ . The gk are coupling constants
with dimensions of energy, and nk is the number operator for mode k. This Hamiltonian describes a qubit (the system) with
ground state ∣0⟩ of energy 0 and excited state ∣1⟩ with energy ω0 coupled to a QHO bath. The coupling either excites the qubit
and removes excitations from the bath, or v.v. It will be useful to think of the bath in this case as electromagnetic modes of cavity.
As usual, let us transform to the interaction picture wrt H0 , so that
H̃(t) = U0† (t)HSB U0 (t) = σ+ (t) ⊗ B(t) + h.c. (752a)
−iωk t
σ+ (t) = e iωt
σ+ , B(t) = ∑ e gk bk . (752b)
k

Then the joint system-bath state ∣φ(t)⟩ (assume it is pure) in the interaction picture is given by ∣φ(t)⟩ = U (t) ∣φ(0)⟩, where
U (t) = T+ exp (−i ∫0 H̃(t′ )dt′ ).
t

This model is not analytically solvable in general. However, we shall assume that the cavity supports at most one photon.
Under this assumption the model becomes analytically solvable, as we shall see.

1. Analytical solution in the 1-excitation subspace

a. The 1-excitation subspace is conserved Let ∣0⟩B denote the vacuum state of the bath and consider the following joint
system-bath states:
∣ψ0 ⟩ = ∣0⟩S ⊗ ∣v⟩B , (753a)
∣ψ1 ⟩ = ∣1⟩S ⊗ ∣v⟩B , (753b)
∣ψk ⟩ = ∣0⟩S ⊗ ∣k⟩B , (753c)
where ∣k⟩ = b†k ∣v⟩B = ∣01 , . . . , 0k−1 , 1k , 0k+1 , . . .⟩ denotes the state with one photon in mode k (∣k⟩ is not to be confused with
the usual labels for the computational basis of a qubit). Assume that the initial joint system-bath state contains at most a single
excitation, i.e.:
∣φ(0)⟩ = c0 ∣ψ0 ⟩ + c1 (0) ∣ψ1 ⟩ + ∑ ck (0) ∣ψk ⟩ . (754)
k

We wish to show that under the Hamiltonian above this remains true for all times, i.e., for all t:
∣φ(t)⟩ = c0 (t) ∣ψ0 ⟩ + c1 (t) ∣ψ1 ⟩ + ∑ ck (t) ∣ψk ⟩ . (755)
k

This is intuitively clear, since the system-bath coupling either excites the qubit while removing a photon, or v.v., and H0 creates
no new excitations. Nevertheless, let us give a formal argument for completeness.
Define the excitation number operator by
N = σ+ σ− ⊗ I + I ⊗ ∑ b†k bk . (756)
k

The name is well deserved since:

N ∣ψ0 ⟩ = (σ + σ− ∣0⟩) ⊗ ∣v⟩ + ∣0⟩ ⊗ ∑ b†k bk ∣v⟩ = 0 ⋅ ∣ψ0 ⟩ (757a)
k

N ∣ψ1 ⟩ = (σ + σ− ∣1⟩) ⊗ ∣v⟩ + ∣1⟩ ⊗ ∑ b†k bk ∣v⟩ = ∣1⟩ ⊗ ∣v⟩ = 1 ⋅ ∣ψ1 ⟩ (757b)
k

N ∣ψk ⟩ = (σ + σ− ∣0⟩) ⊗ ∣v⟩ + ∣0⟩ ⊗ ∑ b†k′ bk′ ∣k⟩ = ∣0⟩ ⊗ ∑ δkk′ ∣k⟩ = 1 ⋅ ∣ψk ⟩ , (757c)
k′ k′
 109

where we used σ− ∣0⟩ = bk ∣v⟩ = 0. I.e., N counts the number of excitations.

Next, note that the excitation number operator commutes with the total Hamiltonian H. That [N, H0 ] = 0 is obvious. As for
HSB , note first that [σ+ σ− , σ± ] = ±σ± , and [nk′ , bk ] = −bk δkk′ , [nk′ , b†k ] = b†k δkk′ . Therefore:

[N, HSB ] = [σ+ σ− , σ+ ] ⊗ B + [σ+ σ− , σ− ] ⊗ B † + σ+ ⊗ [∑ b†k bk , B] + σ− ⊗ [∑ b†k bk , B † ] (758a)

k k

= σ+ ⊗ B − σ− ⊗ B † + σ+ ⊗ [∑ nk , ∑ gk′ bk′ ] + σ− ⊗ [∑ nk , ∑ gk′ b†k′ ] (758b)

k k′ k k′

= σ+ ⊗ B − σ− ⊗ B † − σ+ ⊗ B + σ− ⊗ B † = 0 . (758c)

This means that N is a conserved quantity, i.e., its eigenvalues are conserved under the evolution generated by H, or by H̃(t)
in the interaction picture. It also means that H and N share a common set of eigenvectors, which can be indexed using the
eigenvalues of both H and N . Eigenvectors with different eigenvalues of N don’t mix under the dynamics generated by H
or H̃(t). This explains why, assuming the initial state is Eq. (754), the state subsequently must be as in Eq. (755): the state
∣ψ0 ⟩ has eigenvalue 0 under N [Eq. (757a)] and evolves as a separate one-dimensional subspace, and the states ∣ψ1 ⟩ and ∣ψk ⟩
have eigenvalue 1 under N [Eqs. (757b), (757c)], and also evolve as a separate two-dimensional subspace. U (t) evolves each
subspace separately and does not couple different subspaces labeled by different eigenvalues of N .
Note that i∂t ∣ψ0 ⟩ = H̃(t) ∣ψ0 ⟩ = 0, which means, since ∣ψ0 ⟩ evolves separately, that ∣ψ0 (t)⟩ = ∣ψ0 (0)⟩. Therefore c0 (t) =
c0 (0).
Even though the subspace spanned by {∣ψ0 ⟩ , ∣ψ1 ⟩ , ∣ψk ⟩} contains both 0 and 1 excitations, we loosely refer to it as the
1-excitation subspace.
b. Schrödinger dynamics in the 1-excitation subspace Substituting Eq. (755) into the Schrödinger equation, we have:

i∂t ∣φ(t)⟩ = ċ1 (t) ∣ψ1 ⟩ + ∑ ċk (t) ∣ψk ⟩ (759a)

k

= H̃(t) ∣φ(t)⟩ = (σ+ (t) ⊗ B(t) + σ− (t) ⊗ B † (t)) (c0 (0) ∣ψ0 ⟩ + c1 (t) ∣ψ1 ⟩ + ∑ ck (t) ∣ψk ⟩) (759b)
k

= [σ+ (t) ⊗ B(t)] ∑ ck (t) ∣0⟩ ⊗ ∣k⟩ + c1 (t)[σ− (t) ⊗ B † (t)](∣1⟩ ⊗ ∣v⟩) (759c)
k
= ∣1⟩ ⊗ ∑ gk ∣v⟩ ck (t)eiω0 t−iωk t + c1 (t) ∣0⟩ ⊗ ∑ gk∗ ∣k⟩ e−iω0 t+iωk t (759d)
k k

= ∑ gk ck (t)ei(ω0 −ωk )t ∣ψ1 ⟩ + ∑ gk∗ c1 (t)e −i(ω0 −ωk )t

∣ψk ⟩ . (759e)
k k

Multiplying by ⟨ψ1 ∣ and ⟨ψk ∣ gives us two coupled differential equations for the amplitudes c1 and ck :

c˙1 (t) = −i ∑ gk ck (t)ei(ω0 −ωk )t (760a)

k

c˙k (t) = −igk∗ c1 (t)e−i(ω0 −ωk )t . (760b)

Integrating Eq. (760b) gives:

t
ck (t) − ck (0) = −i ∫ dt′ gk∗ c1 (t′ )e−i(ω0 −ωk )t .
′
(761)
0

For simplicity, let us assume that the cavity starts in the vacuum state, i.e., ck (0) = 0. Then, after substituting the above into
Eq. (760a) we obtain:
t
c˙1 (t) = − ∫ dt′ f (t − t′ )c1 (t′ ) , (762)
0

where the “memory function” f is:

∞
f (t) = ∑ ∣gk ∣ ei(ω0 −ωk )t = ∫ dωJ(ω)ei(ω0 −ω)t ,
2
(763)
k 0

where J(ω) is the bath spectral density, formally given as usual by J(ω) = ∑k ∣gk ∣2 δ(ω − ωk ).
 110

Since Eq. (762) is a convolution, it can be solved by means of a Laplace transform,

∞
Lap[f ] ≡ fˆ(s) ≡ ∫ dt e−st f (t) , (764)
0

since the Laplace transform of a convolution of two functions is the product of their Laplace transforms:
t
Lap[∫ dt′ f (t − t′ )c1 (t′ )] = fˆ(s)ĉ1 (s) . (765)
0

Also, the Laplace transform of a derivative of a function g(t) is

∂g
L[ ] = sg̃(s) − g(0) . (766)
∂t
Therefore
c1 (0)
ĉ1 (s) = . (767)
s + fˆ(s)
This completes the analytical solution, since given the spectral density we can compute the excited state amplitude c1 (t) by
inverse Laplace transform of ĉ1 (s), and from there the ck (t) amplitudes via Eq. (761). Finally, recall that c0 (t) = c0 (0).
Eq. (755) then gives us the joint system-bath state in the 1-excitation subspace.
c. System-only state With the analytical solution in hand for the joint system-bath state ∣φ(t)⟩, we can find the system-only
state:
1 − ∣c1 ∣ c0 c∗1 (t)
2
ρ (t) ρ01 (t)
ρ(t) = TrB (∣φ(t)⟩⟨φ(t)∣) = ( 00 ) = ( 2 ) . (768)
ρ∗01 (t) ρ11 (t) c∗0 c1 (t) ∣c1 ∣
2 2 2 2
Note that normalization implies that ∣c0 ∣ + ∣c1 (t)∣ + ∑k ∣ck (t)∣ = 1, so that 1 − ∣c1 ∣ ≠ ∣c0 ∣2 (indeed, c0 is constant), which is
why ρ00 (t) ≠ ∣c0 ∣2 . To verify Eq. (768), let us explicitly calculate the partial trace, recalling that ∣φ(t)⟩ = [c0 ∣0⟩ + c1 (t) ∣1⟩] ∣v⟩ +
∣0⟩ ∑k ck (t) ∣k⟩:
TrB (∣φ(t)⟩⟨φ(t)∣) = ⟨v∣φ(t)⟩⟨φ(t)∣v⟩ + ∑⟨k∣φ(t)⟩⟨φ(t)∣k⟩ (769a)
k
= [c0 ∣0⟩ + c1 (t) ∣1⟩][c∗0 ⟨0∣ + c∗1 (t) ⟨1∣] + ∣0⟩⟨0∣ ∗
∑ ck′ δkk′ (t)ck′′ (t)δk′′ k (769b)
k′ ,k′′

= [∣c0 ∣2 + ∑ ∣ck (t)∣2 ]∣0⟩⟨0∣ + c0 c∗1 (t)∣0⟩⟨1∣ + c∗0 c1 (t)∣1⟩⟨0∣ + ∣c1 (t)∣2 ∣1⟩⟨1∣ . (769c)
k

d. Exact master equation To connect the analytical solution to the master equation framework, let us now find the exact
master equation satisfied by ρ(t). To do so, we differentiate Eq. (768), to find:

−∂ ∣c ∣ c0 ċ∗1 (t)
2
ρ̇ = ( ∗ t 1 2) . (770)
c0 ċ1 (t) ∂t ∣c1 ∣
The system-bath Hamiltonian describes an excitation and relaxation process. Therefore, recalling Eq. (282), a reasonable ansatz
for the exact master equation in the interaction picture is of the form
i 1
ρ̇ = − S(t)[σ+ σ− , ρ(t)] + γ(t) (σ− ρ(t)σ+ − {σ+ σ−, ρ(t)}) , (771)
2 2
where the first term represents the Lamb shift and the second term represents relaxation. We will shortly verify this ansatz.
Meanwhile, note that unlike Eq. (282), the relaxation rate γ is now time-dependent. This is an important difference, since there
is now no guarantee that the rate is always positive and finite.
Let us now check and confirm the ansatz. Note that
ρ11 0 ∣c (t)∣2 0
σ− ρσ+ = ( )=( 1 ) (772a)
0 0 0 0
0 −ρ01 0 −c0 c∗1 (t)
[σ+ σ− , ρ] = ( )=( ∗ ) (772b)
ρ10 0 c0 c1 (t) 0
0 ρ01 0 c0 c∗1 (t)
{σ+ σ− , ρ} = ( )=( ∗ ), (772c)
ρ10 2ρ11 c0 c1 (t) 2∣c1 (t)∣2
 111

where we used Eq. (768) for the second equality in each line. If Eq. (771) holds then it must be true, using the first equality in
each line of Eq. (772), that:

⎛ γ(t)∣c1 ∣2 ( 2i S(t) − 12 γ(t)) c0 c∗1 (t)⎞

ρ̇ = . (773)
⎝(− 2i S(t) − 12 γ(t)) c∗0 c1 (t) −γ(t)∣c1 ∣2 ⎠

Comparing the off-diagonal elements of Eqs. (770) and Eq. (773) we find that they agree provided ċ1 = − 12 c1 (t)[γ(t) + iS(t)],
i.e.:
ċ1 (t)
S(t) = −2I ( ) (774a)
c1 (t)
ċ1 (t)
γ(t) = −2R ( ). (774b)
c1 (t)
We have thus identified the Lamb shift rate and relaxation rate from the exact master equation (771).
But, to ensure that the ansatz is correct we still need to confirm that this identification also works for the diagonal elements.
Let c1 (t) = r(t)eiθ(t) . Then ∂t ∣c1 ∣2 = 2ṙr, and also ċ1 = ṙeiθ(t) + iθ̇c1 , which implies ċ1 /c1 = ṙ/r + iθ̇, i.e., R(ċ1 /c1 ) = ṙ/r.
Therefore, if Eq. (774b) holds then:
γ(t)∣c1 (t)∣2 = −2(ṙ/r)r2 = −2ṙr = −∂t ∣c1 (t)∣2 , (775)
as required if Eqs. (770) and Eq. (773) are to agree.
e. Connection with the TCL formalism Note that Eq. (771) is in the form of the TCL-ME, since it is time-local. Namely,
we can introduce a time-local generator and rewrite it as
ρ̇ = KS (t)ρ(t) = TrB [K(t)ρ(t) ⊗ ρB ] , (776)
where K(t) is the TCL generator [Eq. (739)], which can be computed directly from the time-local generator KS (t), which we
identify here as KS (t) = − 2i S(t)[σ+ σ− , ⋅] + γ(t) (σ⋅ σ+ − 12 {σ+ σ−, ⋅}). Next, recall that K(t) = ∑∞
n=1 α Kn (t) [Eq. (741b),
2n
∞
where we have shifted the bath operators so all odd orders vanish]. Correspondingly, KS (t) = ∑n=1 α Kn (t), and therefore
2n

also
∞ ∞
γ(t) = ∑ α2n γ2n (t) , S(t) = ∑ α2n S2n (t) . (777)
n=1 n=1

To make the connection between the exact solution of the qubit-in-cavity model and this perturbative expansion of the TCL-
ME, recall that we started from the Liouville-von-Neumann equation in the form ∂t ρ̃ = −iα[H̃(t), ρ̃(t)] ≡ αLρ̃(t) [Eq. (730)].
This means that if we were to introduce the dimensionless parameter α into the formulation of the qubit-in-cavity model, it
would multiply the coupling constants gk , and hence we would need to replace f (t) with α2 f (t) in Eq. (763). Then Eq. (762)
is replaced by
t
c˙1 (t) = −α2 ∫ dt′ f (t − t′ )c1 (t′ ) . (778)
0

If we consider the Laplace transform solution for c1 (t), given by the inverse Laplace transform of Eq. (767), then to lowest
order in α we simply have c1 (t) = c1 (0). The reason is that the inverse Laplace transform of c1 (0)/s [where have taken α → 0
in Eq. (767)] is c1 (0). Therefore to lowest order in α, Eq. (778) yields c˙1 (t) = −α2 c1 (0) ∫0 dt′ f (t − t′ ) + O(α3 ), and it follows
t

from Eq. (774) that

t
S2 (t) = 2I (∫ dt′ f (t − t′ )) (779a)
0
t
γ2 (t) = 2R (∫ dt′ f (t − t′ )) . (779b)
0

D. Jaynes-Cummings model on resonance

Having derived the exact master equation for a qubit in a cavity, we can now apply it to compare the predictions of various
master equations to the exact solution. To do so we need to specify the bath spectral density J(ω). We will consider the Jaynes-
Cummings model on resonance, a model in which the cavity supports a single mode with a frequency ω0 equal to that of the
qubit. First we consider the case where the cavity is completely isolated from the external world, then we consider the case
where the cavity is coupled to the external electromagnetic field.
 112

1. Isolated cavity

Assume that the cavity has opaque walls that act as infinitely tall barriers, so that no radiation can leak into or out of the cavity.
In this case, with ω0 being the qubit transition frequency, since the cavity only has a single mode, at this frequency, the spectral
density becomes

J(ω) = ∑ ∣gk ∣2 δ(ω − ωk ) ↦ ∣g∣2 δ(ω − ω0 ) . (780)

k

Therefore the memory function f (t) [Eq. (763)] is

∞
f (t) = ∫ dωJ(ω)ei(ω0 −ω)t = ∣g∣2 , (781)
0

and the amplitude of the qubit excited state, c1 (t), then satisfies
t t
c˙1 (t) = − ∫ dsf (t − t′ )c1 (s) = −∣g∣2 ∫ dt′ c1 (t′ ) . (782)
0 0

Rather than using the Laplace transform solution, it is simpler to differentiate both sides to get

c¨1 (t) = −∣g∣2 c1 (t) . (783)

The solution of this differential equation is

c1 (t) = A cos(∣g∣t) + B sin(∣g∣t) , (784)

where A and B are constants. Thus, the population of the excited state is ρ11 (t) = ∣c1 (t)∣2 , which oscillates with a period given
by π/∣g∣, as expected from a qubit coupled to an oscillator resonant with it.

2. Leaky cavity

Next we consider the case where, instead of opaque walls, the cavity allows photons to leak out or in. It can be shown that in
this case the memory function is
1
f (t) = e−t/τB (785)
2τM τB
where τM is a Markovian timescale whose exact meaning will become apparent below, and τB is the usual bath correlation time
(decay time of ⟨B(t)B(0)⟩B ). Moreover, it can be shown that α2 = τB /τM , where α is the dimensionless system-bath coupling
strength we have used as a dimensionless prefactor for HSB in the TCL-ME.
The excited state amplitude c1 (t) then satisfies
t 1 t
c˙1 = − ∫ dt′ f (t − t′ )c1 (t′ ) = − dt′ e−(t−t )/τB c1 (t′ ) .
′
∫ (786)
0 2τM τB 0

It is again simpler to differentiate once more rather than use the Laplace transform:
1 1
c̈1 + ċ1 + c1 = 0 , (787)
τB 2τM τB
a simple second order differential equation. Its solution is:

− 2τt tδ 1 tδ
c1 (t) = c1 (0)e B [cosh ( )+ sinh ( )] , (788)
2 τB δ 2
where
√
1 2 1 √
δ= − = 1 − 2α2 . (789)
τB2 τM τB τB

The excited state population is ρ11 (t) = ∣c1 (t)∣2 . We thus have two distinct cases:
 113

a. Weak coupling This is the case when α2 = τB /τM ≤ 1/2, so that δ ∈ R. Then S(t) = 0 [Eq. (774a)] and Eqs. (774b)
and (788) yield:
2
τM τB
cosh ( tδ
2
)
γ(t) = (790a)
δ cosh ( 2 ) + τB sinh ( tδ
tδ 1
2
)
− τt tδ 1 tδ 2
ρ11 (t) = ρ11 (0)e B ∣[cosh ( )+ sinh ( )]∣ . (790b)
2 τB δ 2
In this case the population decays, i.e., the dynamics is Markovian-like.
b. Strong coupling This is the case when α2 = τB /τM > 1/2, so that δ ∈ I. Then S(t) ≠ 0 [Eq. (774a)] and Eq. (774b)
and (788) yield:
2
τM τ B
cos ( t∣δ∣
2
)
γ(t) = (791a)
δ cos ( t∣δ∣
2
)+ 1
τB
sin ( t∣δ∣
2
)
2
− τt t∣δ∣ 1 t∣δ∣
ρ11 (t) = ρ11 (0)e B ∣[cos ( )+ sin ( )]∣ . (791b)
2 τB ∣δ∣ 2
In this case the population exhibits damped oscillations, i.e., the dynamics is non-Markovian.
With this analytical solution in hand, we are ready to compare to the predictions of the TCL-ME.

3. Comparison to TCL-ME, Markov limit, and NZ-ME

Recall that the TCL-ME expansion is, in the present case, equivalent to an expansion of γ(t) and S(t) in powers of α, as in
Eq (777). We can thus obtain the γ2n (t) terms for the weak coupling case by expanding γ(t) of Eq. (790b) in powers of α, and
similarly for the strong coupling case.
We can also use Eq. (779b), so that:
t t 1 1
γ2 (t) = 2R (∫ dt′ f (t − t′ )) = R (∫ dt′ e−(t−t )/τB ) = (1 − e−t/τB ) ,
′
(792)
0 0 τM τB τM
which is clearly an example of the weak coupling case (as expected for a low-order-in-α expansion) since the rate exhibits no
oscillations. Note that γ2 (t) has a rise time of τB to its asymptotic value of 1/τM .
Recall that the TCL-2 result is exactly the Redfield equation, as we showed in Sec. XX B 2. Moreover, if we take the upper
limit of the integral to infinity we have the Markov limit. Therefore:

γ2 (∞) = 1/τM ≡ γ0 , (793)

which explains the subscript M notation we used all along in this example. We already know the solution in the Markovian
limit: ρ11 (t) = ρ11 (0)e−t/τM .
By doing the expansion to fourth order in α we find:14
1 τM
γ4 (t) = (1 − e−t/τB + [sinh(t/τB ) − t/τB ]e−t/τB ) , (794)
τM τB

which has the limiting behavior γ4 (∞) = τ1M + 2τ1B > γ2 (∞).
What about the NZ-ME? It can be shown that to second order in α, the NZ-ME yields exactly the same result as TCL-2,
except that two changes are needed: (1) e−t/τB is replaced by e−t/(2τB ) in the results for ρ11 (t), and (2) δ is replaced by
√
′ 1 4 1 √
δ = − = 1 − 4α2 . (795)
τB2 τM τB τB
Figure 9 shows these various results in terms of the deviation of the excited state population from the exact result. Focusing
on panels (a)-(c), which report results for the weak coupling case, it illustrates a number of points:

14 Note that the result given in the book [1] differs from Ref. [33][Eq. (69)]; the latter is the correct one.
 114

• All approximations, except Markov, are good for very short times (shorter than τB ).
• The Markov approximation initially overestimates the depopulation of the excited state, the underestimates it for longer
times. It is a particularly poor approximation for times shorter than τB , which is the rise-time of the curves in panel (a).
• TCL-2 (Redfield) underestimates the depopulation of the excited state for intermediate times.
• TCL-2 converges to Markov in the long-time limit.
• NZ-2 overestimates the depopulation of the excited state for intermediate times.
• TCL-4 is a better approximation than both the TCL-2 and the Markov approximation. Its rate γ4 (t) goes above the Markov
rate, as expected since γ4 (∞) − =1 2τ1B .

4. Breakdown of the NZ-ME and TCL-ME expansions for strong coupling

What about the strong coupling case? The exact result is shown in Fig. 9(d), and exhibits damped oscillations. The second
order NZ-ME also exhibits damped oscillations, but the excited state population becomes negative! This result is physically
non-sensical and is a clear example of violation of complete positivity of the evolution map. The TCL-4 approximation is good
for short times but fails to capture the oscillations. To understand this let us take a step back and recall that the TCL-ME requires

FIG. 9. Damped Jaynes-Cummings model on resonance. Exact solution (exact), TCL-ME to second (TCL 2) and fourth order (TCL 4),
NZ-ME to second order (GME 2), and the RWA-LE (Markov). (a) Decay rate of the excited state population, (b) the population of the excited
state, including a stochastic simulation of the TCL-ME with 105 realizations (diamonds for TCL 2and stars for TCL 4), and (c) deviation of the
approximate solutions from the exact result, for 1/γ0 ≡ τM = 5τB (weak coupling). (d) Population of the excited state for 1/γ0 ≡ τM = 0.2τB
(strong coupling). Source: Ref. [33].
 115

the invertibility of the operator I − Σ. The present example serves to illustrate how this invertibility condition can be violated,
and how therefore the TCL can break down.
(1) (2) (3)
Assume that for different initial conditions {ρ11 (0), ρ11 (0), ρ11 (0), . . . } there is a common time t0 at which the exact
11 (t0 ) = 0. This is indeed the case shown in Fig. 9(d), as is easy to verify from Eq. (791b): solving for its roots
solution gives ρexact
we have:
2
tan(∣δ∣tn /2) = −∣δ∣τB Ô⇒ tn = (arctan(∣δ∣τB ) + nπ) Ô⇒ t0 = min tn , (796)
∣δ∣ n

where n runs over the integers. Now, since the TCL-ME is time-local, i.e., it only “knows” about the current time t, this means
that for t ≥ t0 it is impossible to invert the evolution back to the initial condition, as this information is lost in a time-local
description. We therefore expect the TCL-ME to give unreliable results when the exact solution predicts a vanishing population.
This is precisely what is seen in Fig. 9(d).
Mathematically, we can see this another way. Eq. (791a) tells us that γ(t) diverges at the same times t = tn where ρexact
11 (t) = 0.
More fundamentally, this is because c1 (t) = 0 implies via Eq. (774b) that γ(t) diverges (unless ċ1 (t) = 0 at the same time). But
if γ(t) diverges then it does not have a Taylor series, so the various γ2n (t) are undefined, and the TCL-ME expansion does not
exist.

XXI. POST MARKOVIAN MASTER EQUATION

We have seen a variety of approaches to describing the reduced system dynamics via master equations, ranging from the
exact Nakajima-Zwanzig equation, via the time-convolutionless, to the Markovian limit. In this section we will review a master
equation approach that naturally interpolates between the Markovian limit and the limit of exact dynamics, as expressed in
terms of CP maps via the Kraus OSR [34]. The key idea will be to understand both limits as arising from a non-selective
measurement process of the bath state. The exact dynamics corresponds to a single measurement at the final time, whereas
Markovian dynamics corresponds to the limit of infinitely many measurements. The interpolation will thus limit the number of
measurements in order to arrive at an non-Markovian approximation.

A. Measurement interpretation of the Kraus OSR and the Lindblad equation

Consider the usual setup of open system evolution, with the initial state ρ(0) = ρS (0) ⊗ ρB evolving under a joint unitary U
to the final state ρ(t) = U (t)ρ(0)U † (t). The reduced system state at the final time is ρS (t) = TrB [ρ(t)]. We wish to show
that this can be understood equivalently as a projective measurement of the bath at the final time, as depicted schematically in
Fig. 10.
Suppose that we measure the bath at the final time t via the complete set of projection operators {Pk = ∣k⟩⟨k∣}. Thus, if
outcome k was observed, then the joint state transforms as
Pk (IS ⊗ Pk )ρ(t)(IS ⊗ Pk )
ρ(t) z→ ≡ ρ(k) (t) (797)
pk
with probability pk = Tr[(IS ⊗ Pk )ρ(t)]. The reduced system state for this outcome is

(k) ⟨k∣ ρ(t) ∣k⟩

ρS (t) = TrB [ρ(k) (t)] = ∑ ⟨k ′ ∣ ρ(k) (t) ∣k ′ ⟩ = . (798)
k′ pk

Assuming we do not keep track of the measurement outcome, i.e., the measurement is non-selective, the final system state is the
(k)
mixed state ensemble [recall Eq. (68)] {pk , ρS (t)}, i.e.,
(k)
ρS (t) = ∑ pk ρS (t) = ∑ ⟨k∣ ρ(t) ∣k⟩ = TrB [ρ(t)] , (799)
k k

i.e., exactly the Kraus OSR result. Thus we can indeed understand the Kraus OSR as joint unitary evolution followed by a single
non-selective measurement of the bath at the final time t.
In other words, we have shown that the following two evolutions are equivalent:
U (t) TrB
ρ(0) z→ ρ(t) z→ ρS (t) (800a)
U (t) PB TrB (k) non-selective
ρ(0) z→ ρ(t) z→ ρ(k) (t) z→ ρS (t) z→ ρS (t) , (800b)
 116

P M (a)
0 t
P M M M M (b)
0 τ 2τ t −τ t
FIG. 10. Measurement approach to open system dynamics. P=preparation, M= measurement, time proceeds from left to right. (a) Exact Kraus
operator sum representation. (b) Markovian approximation.

where PB denotes a projective measurement of the bath with projectors {Pk }.

For the Lindblad equation, we have already shown in Sec. IX A that it can be understood as arising from a sequence of
infinitesimal CP maps. More specifically, we showed that the LE
1
ρ̇S = −i[H, ρS ] + ∑ Lα ρL†α − {L†α Lα , ρS } (801)
α≥1 2
is equivalent to the sequence of CP maps
ρS (t + τ ) = ∑ Kα ρS (t)Kα† , (802)
α≥0

where τ → 0 and
1
K0 = I + (−iH − ∑ L Lα )τ
†
conditional evolution (803a)
2 α≥1 α
√
Kα = Lα τ , (α ≥ 1) jumps . (803b)
Since we have just shown that each CP map can be understood as a projective measurement of the bath, we see that the LE can
also be understood as representing an infinite sequence of such measurements, taking place in intervals of length τ . Since each
such measurement disentangles the system and bath state, it can be viewed as a preparation step of a new product state between
the system and bath; see Fig. 10.

B. Interpolating between the two limits: derivation of the PMME

Having seen that the exact Kraus OSR and the fully Markovian LE are two measurement limits, it is natural to consider an
intermediate scenario, of a finite number of intermediate measurements between the initial and final times. Consider the simplest
case, of a single projective measure of the bath at a random time t′ ∈ (0, t), and note that the more measurements we introduce,
the more Markovian the evolution becomes. We assume that the same CPTP map Λ governs the evolution in the period [0, t′ )
and (t′ , t], as shown in Fig. 11. The measurement produces a random system state ρ(t′ ) (where we from here on we drop the
subscript S since we are interested only in the system dynamics), which is then propagated to ρ(t), i.e., ρ(t) = Λ(t − t′ )ρ(t′ ).
But since we do not know the outcome, nor the time t′ , we introduce a weighting function k(t − t′ , t) (the choice to make the
argument depend on the remaining time interval t − t′ rather than t′ is for later convenience). The final state ρ(t) can then be
represented in the following form:
t
ρ(t) = ∫ k(t − t′ , t) Λ(t − t′ )ρ(t′ )dt′ (804)
0 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
weight (kernel)

Λ (t ′) ρ (0) Λ (t − t ′) ρ (t ′)
P M M
0 t′ t
FIG. 11. A single projective measurement of the bath is preceded and followed by a CPTP map Λ. For that specific outcome ρ(t′ ) = Λ(t′ )ρ(0)
and ρ(t) = Λ(t − t′ )ρ(t′ ). To account for all possible outcomes each such trajectory is weighted as in Eq. (804).
 117

It is convenient to change variables to s = t − t′ , so that:

t
ρ(t) = ∫ k(s, t)Λ(s)ρ(t − s)ds . (805)
0

Our purpose is to arrive at a master equation, so let us differentiate Eq. (805) with respect to t:
∂ρ ∂ t
= ∫ k(s, t)Λ(s)ρ(t − s)ds (806a)
∂t ∂t 0
t ∂k(s, t) ∂ρ(t − s)
= k(t, t)Λ(t)ρ(0) + ∫ ( Λ(s)ρ(t − s) + k(s, t)Λ(s) ) ds . (806b)
0 ∂t ∂t

The first term corresponds to performing the bath measurement at t = 0 and then evolving from ρ(0) via Λ(t). This term can thus
be dropped [formally, by setting k(t′ = 0, t) = k(s = t, t) = 0] since we assumed that the intermediate measurement weighted
by k occurs in the open interval (0, t). To make further progress let us specify the form of the CP map Λ. For simplicity, let us
assume that the intermediate evolutions are themselves Markovian:

Λ(t) = eLt , (807)

where L is a Lindbladian, since this is the unique way to ensure that Λ is CPTP in the Markovian case. Then

∂ρ(t − s) ∂eL(t−s)
= ρ(0) = LeL(t−s) ρ(0) = Lρ(t − s) , (808)
∂t ∂t
so that Eq. (806b) simplifies to:

∂ρ t ∂k(s, t)
=∫ ( + k(s, t)L) eLs ρ(t − s)ds . (809)
∂t 0 ∂t
We now seek to ensure that this evolution is trace-preserving. This requires the RHS to be traceless, since then 0 = Tr∂t ρ =
∂t Trρ = 0, so that Trρ(t) = const. It is sufficient to this end to demand that ∂t k(s, t) = 0, since the second term is already
traceless:
t t
Tr [∫ ds k(s, t)LeLs ρ(t − s)] = ∫ ds k(s, t)Tr [LeLs ρ(t − s)] = 0 , (810)
0 0

since for a Lindbladian L acting on any operator X

1 1 1
Tr[LX] = Tr[∑ Lα XL†α − L†α Lα X − XL†α Lα ] = ∑ Tr[XL†α Lα ] − Tr[XL†α Lα ] − Tr[XL†α Lα ] = 0 . (811)
α 2 2 α 2

Now, since ∂t k(s, t) = 0, it follows that k(s, t) = ck(s), where c is a constant we can choose to be 1. Therefore

k(s, t) ≡ k(s) . (812)

Then Eq. (809) reduces to:

∂ρ t
= L ∫ k(s)eLs ρ(t − s)ds (813a)
∂t 0
= Lk(t)eLt ∗ ρ(t) , (813b)

where in the second line ∗ denotes a convolution. Equation (813) is the PMME.
Now consider two special cases of Eq. (813):
• k(s) = δ(s): In this case the PMME reduces to ∂ρ
∂t
= Lρ(t), which is the standard Lindblad equation. Therefore the
PMME includes the LE as a special case.
t
• Expanding the exponential to zeroth order in L (assuming ∥Lt∥ ≪ 1), the PMME reduces to ∂ρ ∂t
= L ∫0 k(s)ρ(t − s)ds,
which is a form that has been proposed heuristically in the literature on non-Markovian master equations.
• Since the PMME involves a convolution, it can be viewed as a special case of the NZ-ME. Namely, we can write the
PMME in the NZ-ME form ∂t ρ̂(t) = ∫0 K(t, t′ )ρ̂(t′ )dt′ [recall Eq. (719b)], where K(t, t′ ) is directly obtainable from
t

Eq. (813a).
 118

C. Solution of the PMME

To solve the PMME (813) we can use the Laplace transform (764). Recall that the Laplace transform of the convolution of
two functions is the product of their Laplace transforms: Eq. (765), and also recall the result for the Laplace transform of a
derivative in Eq. (766). Therefore, upon taking the Laplace transform of the PMME we find:

sρ̃(s) − ρ(0) = LLap[k(t)eLt ]ρ̃(s). (814)

The Laplace transform satisfies the following shifting property:

Lap[f (t)eat ] = f˜(s − a), (815)

but to use it requires a few extra steps, since it is not immediately clear how to deal with eLt in this context. Thus, we find that
it is most convenient to work in the eigenbasis of L. Since L is not normal ([L, L† ] ≠ 0 in general), it can have distinct right
and left eigenvectors, i.e., we can find a set of operators {Ri } and {Li } such that LRi = λi Ri and Li L = λi Li . Both sets are
complete, and they are mutually orthonormal in the sense that after normalization Tr[Li Rj ] = δij .
We can therefore expand ρ in this so-called “damping basis” (the basis of right eigenvectors of L), to get:

ρ(t) = ∑ µi (t)Ri , (816)

i

where the expansion functions are given by

µj (t) = ∑ µi (t)Tr(Lj Ri ) = Tr[Lj ρ(t)] . (817)

i

Substituting into the PMME Eq. (813a) we obtain

∂µi t
Ri = ∑ L ∫ k(s)eLs µi (t − s)Ri ds (818a)
∂t i 0
t
= ∑ λi ∫ k(s)eλi s µi (t − s)Ri ds , (818b)
i 0

where we used eLs Ri = eλi s Ri . Multiplying both sides by Li and taking the trace yields:
∂µi t
= λi ∫ k(s)eλi s µi (t − s)ds. (819)
∂t 0

At this point we can take the Laplace transform of both sides and use the shifting property (815), to get:

sµ̃i (s) − µi (0) = λi Lap[k(t)eλi t]µ̃i (s) (820a)

= λi k̃(s − λi )µ̃i (s) (820b)

Therefore:
1
µ̃i (s) = µi (0). (821)
s − λi k̃(s − λi )
Finally, taking the inverse Laplace transform:

µi (t) = ξi (t)µi (0) , (822)

where:
1
ξi (t) ≡ Lap−1 [ ], µi (0) = Tr[Li ρ(0)] . (823)
s − λi k̃(s − λi )
This completes the exact solution of the PMME.
To summarize, given L we need to compute its eigenvalues λi and associated left and right eigenvectors, and given the kernel
k(t) we need to compute its Laplace transform. Using the initial condition ρ(0), we can then compute ξi (t) and µi (t), from
which we obtain ρ(t) using Eq. (816). The kernel k(t) was assumed to satisfy the condition k(0) = 0.
 119

D. The PMME as a map, and its relation to the TCL-ME

The solution of the PMME can be viewed as a map Φ:

ρ(t) = ∑ µi (t)Ri = ∑ ξi (t)µi (0)Ri = ∑ ξi (t)Tr[Li ρ(0)]Ri = Φ[ρ(0)] , (824)

i i i

where

Φ[X] ≡ ∑ ξi (t)Tr[Li X]Ri . (825)

i

Let us assume that ξi (t) ≠ 0 ∀t. If this is the case then Φ is invertible, i.e., if we let

Φ−1 [X] = ∑ ξi−1 (t)Tr[Li X]Ri (826)

i

then

Φ−1 ○ Φ[X] = ∑ ξi−1 (t)Tr[Li Φ(X)]Ri = ∑ ξi−1 (t)ξj (t)Tr[Lj X]Tr[Li Rj ]Ri = ∑ Tr[Li X]Ri = X , (827)
i ij i

as required. Therefore, using ρ(t) = Φ[ρ(0)] we can write ρ(t − t′ ) = Φ(t − t′ )[ρ(0)] = Φ(t − t′ )Φ−1 (t)[ρ(t)], and so we have:

∂ρ t
= [L ∫ k(t′ )eLt Φ(t − t′ )Φ−1 (t)dt′ ] ρ(t) ≡ K(t)ρ(t) ,
′
(828)
∂t 0

where K(t) is now a convolutionless generator, and Eq. (828) is time-local. Therefore, despite the appearance of the convolution
in the PMME (813), it can be written in TCL-ME form. This is similar to what we did to transform the NZ-ME into the TCL-ME,
where an invertibility assumptions was likewise assumed (recall Sec. XX A 2). It is an interesting open problem to identify the
conditions under which the TCL-ME reduces to the PMME.

E. Complete Positivity of the PMME

Due to the freedom in choosing the kernel k(t), complete positivity is not a guaranteed feature of the PMME. The following
theorem provides us with a way to construct a complete positivity test.
Consider a linear map Φ ∶ Cd×d ↦ Cd×d , i.e., Φ acts on operators represented by d × d matrices, acting on the Hilbert space
H = span{∣i⟩}di=1 . Let us pick ∣i⟩ as a column vector of zeroes, except for a single 1 in position i. Let C = {Φ[∣i⟩⟨j∣]}ij =
∑ij ∣i⟩⟨j∣ ⊗ Φ[∣i⟩⟨j∣]. I.e., C, known as the Choi matrix, is a d × d matrix of the d × d matrices Φ[∣i⟩⟨j∣], meaning that C is d2 × d2 .
Theorem 3 (Choi’s theorem [4]). Φ is completely positive if and only if C > 0.

Constructing the Choi matrix C for the PMME using Eq. (825) we have:

C = ∑ ∣i⟩⟨j∣ ⊗ ∑ ξk (t)Tr[Lk ∣i⟩⟨j∣]Rk = ∑ ξk (t) ∑ ∣i⟩⟨j∣ ⊗ ⟨j∣ Lk ∣i⟩ Rk = ∑ ξk (t) ∑ ∣i⟩⟨j∣(LTk )ij ⊗ Rk . (829a)
ij k k ij k ij

Hence:

C = ∑ ξk (t)LTk ⊗ Rk > 0 (830)

k

Eq. (830) is the complete positivity for the kernel k(t), for a given Lindbladian L and its set of left and right eigenvectors.

F. Example of the PMME: phase damping Lindbladian with an exponential kernel

To illustrate the solution of PMME, consider the phase damping Lindbladian:

γ
Lρ = (ZρZ − ρ) (831)
2
 120

To find the left and right eigenvectors of L, consider its action on the Pauli matrices {I, X, Y, Z}:
γ γ
(ZIZ − I) = 0 , LZ = (Z 3 − Z) = 0
LI = (832a)
2 2
γ γ
LX = (ZXZ − X) = −γX , LY = (ZY Z − Y ) = −γY . (832b)
2 2
Thus the Pauli matrices {Ri } = {I, X, Y, Z} are L’s right eigenvectors, with corresponding eigenvalues {λi } = {0, −γ, −γ, 0}.
Representing the Pauli matrices as vectors, i.e., as I = (1, 0, 0, 0)T , X = (0, 1, 0, 0)T , etc., we can write L as a diagonal matrix
with diagonal entries {0, −γ, −γ, 0}. It is then clear that the left eigenvectors are again the Pauli matrices, i.e., in this example
Li = Ri for i ∈ {I, X, Y, Z}, and the condition Tr(Li Rj ) = δij is automatically satisfied.
Let us express the density matrix in terms of the Bloch vector: ρ(t) = 12 (I + v⃗(t) ⋅ σ
⃗ ). The initial condition can then be written
as
1
µi (0) = Tr[Li ρ(0)] = vi (0) (833)
2
where vI (0) = 1.
Let us now assume that the kernel k(t) is:
k(t) = Ae−at . (834)
Recall that Lap(eat ) = 1/(s − a), so that after the Laplace transformation we have
A
k̃(s) = . (835)
s+a
Using Eq. (823) we thus find:
⎡ ⎤
⎢
−1 ⎢ 1 ⎥
ξi (t) = Lap ⎥.
⎢ ⎥ (836)
⎢ s − λi s−λi +a ⎥
A
⎣ ⎦
The L eigenvectors I and Z have the eigenvalue λ = 0, so that:
1
ξI,Z (t) = Lap−1 [ ] = e0t = 1 . (837)
s
The L eigenvectors X and Y have the eigenvalue λ = −γ, so that:
⎡ ⎤
⎢ ⎥
ξX,Y (t) = Lap−1 ⎢⎢
1 ⎥ = e− 21 (a+γ)t (cos ωt + a + γ sin ωt) ,
A ⎥
(838)
⎢ s + γ s+γ+a ⎥ 2ω
⎣ ⎦
√
where ω = 1
2
4γA − (γ + a)2 . Thus the density matrix is
1
ρ(t) = ∑ µi (0)ξi (t)Ri = [I + (vX (0)X + vY (0)Y )ξX,Y (t) + vZ (0)Z] . (839)
i 2
This describes a Bloch vector with fixed Z-component but with X and Y components undergoing damped oscillations with
frequency ω. This is clearly non-Markovian dynamics. The condition for oscillation is 4γA > (γ +a)2 ; otherwise the oscillations
become exponential decay.
Finally, we can use the complete positivity criterion we found above. The Choi matrix is:
C = ∑ ξk (t)LTk ⊗ Rk = ξI I T ⊗ I + ξX X T ⊗ X + ξY Y T ⊗ Y + ξZ Z T ⊗ Z (840a)
k

⎛ 1 0 0 ξX ⎞
⎜ 0 0 0 0 ⎟
= 2⎜
⎜
⎟.
⎟ (840b)
⎜ 0 0 0 0 ⎟
⎝ξ 0 0 1 ⎠
X

Its eigenvalues are easily found to be {0, 0, 2(1 + ξX ), 2(1 − ξX ). Therefore the PMME in this case corresponds to a CP map iff
∣ξX ∣ = ∣ξY ∣ < 1 , (841)
which is a condition on the problem parameters A, a, γ.
 121

G. Experimental determination of the Lindbladian L and kernel k(t)

Since both L and k(t) are phenomenological in the PMME, is there a way we can determine them experimentally? To do so,
we need to express the kernel in terms of measurable quantities. Let us assume that we ρ(t) can be determined via quantum
state tomography, let us guess L (perhaps based on physical intuition as to the prevalent noise). Then we know ρ(t), the initial
condition ρ(0), and the left and right eigenvector sets {Li , Ri }, so that we can compute ξi (t):

ρ(t) = ∑ µi (t)Ri = ∑ ξi (t)µi (0)Ri = ∑ ξi (t)Tr[Li ρ(0)]Ri (842a)

i
Ô⇒Tr[Lj ρ(t)] = Tr[Lj ρ(0)]ξj (t) (842b)
Tr[Li ρ(t)]
Ô⇒ξi (t) = , (842c)
Tr[Li ρ(0)]

which gives us way to compute ξi (t) from purely experimentally measurable quantities. But at the same time ξi (t) is related to
the kernel via Eq. (823). We can invert the latter for k(t) as follows:

˜ = 1 1 1
ξ(s) Ô⇒ k̃(s − λi ) = (s − ), (843)
s − λi k̃(s − λi ) λi ˜
ξi (s)

where we used the identity Lap−1 [k̃(s − λ)] = k(t)eλt , so that

e−λi t 1
k(t) = Lap−1 [s − ]. (844)
λi ξ˜i (s)

Note that in this expression only the RHS depends on the eigenvalue index i. This gives us an opportunity to optimize the choice
of the Lindbladian by minimizing the deviation for different i values, since they must all agree in order to give a unique result
for k(t). The experimental determination of L and k(t) is thus an iterative process involving this minimization.
 122

Appendix A: Linear algebra background and Dirac notation

Everything in this Appendix is about the finite dimensional case, unless explicitly noted otherwise.

1. Inner Product

The inner product of two vectors is a function operating on two copies of a vector space V that outputs a complex number,
f ∶ V × V ↦ C. By definition it must satisfy the following conditions:

f (∣v⟩, ∑ λi ∣wi ⟩) = ∑ λi f (∣v⟩, ∣wi ⟩) (A1a)

i i
∗
f (∣v⟩, ∣w⟩) = f (∣w⟩, ∣v⟩) (A1b)
f (∣v⟩, ∣v⟩) ≥ 0. (A1c)

It is easy to show that an immediate consequence is

∗
f (∣v⟩, ∑ λi ∣wi ⟩) = ∑ λ∗i f (∣wi ⟩, ∣v⟩) . (A2)
i i

We define the inner product between two Dirac kets as follows:

n ⎛ w1 ⎞
f (∣v⟩, ∣w⟩) ≡ ∑ vi∗ wi = (vi∗ , ..., vn∗ ) ⎜
⎜ ...
⎟ = ⟨v∣w⟩.
⎟ (A3)
i=1 ⎝ wn ⎠

2. Orthonormal Bases

Two vectors ∣v⟩ and ∣w⟩ are orthogonal if and only if their inner product is zero: ⟨v∣w⟩ = 0 ⇐⇒ ∣v⟩ ⊥ ∣w⟩. The norm of a
vector is
√
∥∣v⟩∥ ≡ ⟨v∣v⟩. (A4)

A unit vector is normalized: ∥∣v⟩∥ = 1. A set of vectors forms a basis if it spans the vector space and is linearly independent.
Using the previous definitions, we can then say that an orthonormal basis is a set of normalized orthogonal vectors that span the
vector space V and are linearly independent:
⎧
⎪
⎪1, if i = j
Orthonormal set: {∣vi ⟩}ni=1 , ⟨vi ∣vj ⟩ = δij , δij = ⎨
⎪
⎩0, if i ≠ j
⎪

3. Linear Operators

Another concept important to our formulation of quantum mechanics is that of linear operators. Consider an operator A that
maps one vector space to another:

A∶V ↦W (A5)

For A to be linear, it must be true that for a, b  C and ∣v⟩, ∣w⟩  V

A(a∣v⟩ + b∣w⟩) = aA∣v⟩ + bA∣w⟩  W (A6)

In words, the operator A acting on a linear combination of vectors in the space V produces a linear combination of the operator
acting on each vector individually, and this sum is an element of space W . A good example of a linear operator is the outer
product.
 123

4. Outer Product

If we consider vectors ∣v⟩ , ∣z⟩  V and ∣w⟩  W , the outer product of ∣v⟩ and ∣w⟩ is defined as follows:

A = ∣w⟩⟨v∣ ∶ (∣w⟩⟨v∣) ∣z⟩ ≡ ∣w⟩ (⟨v∣z⟩) = ⟨v∣z⟩∣w⟩ (A7)

² ´¹¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¶
outer product C

One important use of the outer product is in the case of expansion in an orthonormal basis. Consider a vector ∣v⟩  V and a set
of vectors {∣i⟩}ni=1 which forms an orthonormal basis set for V . We can equivalently write ∣v⟩ = ∑ni=1 ci ∣i⟩, in which ci is an
arbitrary constant. The inner product of some vector ∣j⟩ with ∣v⟩ produces the coefficient of ∣v⟩ in the given basis:

⟨j∣v⟩ = ∑ ci ⟨j∣i⟩ = ∑ ci δij = cj (A8)

i i

If we take the outer product of ∣v⟩ with itself, we generate an n × n identity matrix:

⎛ 1 0 ... 0 ⎞
n ⎜ 0 1 ... 0 ⎟
∑ ∣i⟩⟨i∣ = I = ⎜
⎜
⎟
⎟ (A9)
i=1 ⎜ ... ... ... ... ⎟
⎝ 0 0 ... 1 ⎠

We can confirm this is true by applying this inner product as an operator on a vector ∣v⟩:

n n n
(∑ ∣i⟩⟨i∣) ∣v⟩ = ∑ ∣i⟩ ⟨i∣v⟩ = ∑ ci ∣i⟩ = ∣v⟩ (A10)
i=1 i=1 ± i=1
ci

The operator acting on the vector returned the vector, and is known as the “resolution of the identity”. This special case of the
outer product is used to generate a matrix representation of an operator in the appropriate basis. If we consider an operator A
that preserves the space, A ∶ V ↦ V , multiplication of the operator by the identity matrix produces a matrix with elements that
perform the operation A in the following way:

A = IAI (A11)
n ⎛ ⎞ n
= (∑ ∣i⟩⟨i∣) A ∑ ∣j⟩⟨j∣ (A12)
i=1 ⎝j=1 ⎠
= ∑ ∣i⟩ ⟨i∣A∣j⟩⟨j∣ (A13)
i,j ´¹¹ ¹ ¹ ¸ ¹ ¹ ¹ ¶
aij

= ∑ aij ∣i⟩⟨j∣ (A14)

i,j

The scalar aij is known as a matrix element of the operator A. Recall that since the vectors ∣i⟩ and ∣j⟩ are members of an
orthonormal basis, aij ∣i⟩⟨j∣ is actually a matrix with all but the ijth element equal to zero and the (i, j)th element equal to aij :

⎛ 0 ... 0 ⎞
aij ∣i⟩⟨j∣ = ⎜ ⎟
⎜ .. aij ... ⎟ (A15)
⎝ 0 ... 0 ⎠

The sum over all combinations of i and j therefore produces a matrix with elements aij :

⎛ a11 ... a1n ⎞

∑ ija ∣i⟩⟨j∣ = ⎜ .. ... ... ⎟
⎜ ⎟ (A16)
i,j ⎝ an1 ... ann ⎠
 124

5. The Cauchy-Schwartz Inequality

The Cauchy-Schwartz inequality is

∣⟨v∣w⟩∣2 ≤ ⟨v∣v⟩⟨w∣w⟩. (A17)

It helps us make powerful statements about the properties of vectors in Hilbert space that define the domain of quantum mechan-
⃗ ⋅ ⃗b = ∥⃗
ics. In its elementary form it states that, from the definition of the inner product a a∥∥⃗b∥ cos θ, it follows that the magnitude
of the inner product of those vectors is less than or equal to the product of their norms: ∣⃗ a ⋅ ⃗b∣ ≤ ∥⃗
a∥∥⃗b∥.
We can prove this for Hilbert spaces while demonstrating the power of Dirac notation.

Proof. Pick an orthonormal basis whose first element is ∣1⟩ = ∣w⟩ /∥ ∣w⟩ ∥ (we can always do this using the Gram-Schmidt process
to complete the basis). Then, using the resolution of identity we have
n
⟨v∣w⟩ ⟨w∣v⟩ n
⟨v∣v⟩⟨w∣w⟩ = ⟨v∣I∣v⟩⟨w∣w⟩ = ∑⟨v∣i⟩⟨i∣v⟩⟨w∣w⟩ = ⟨w∣w⟩ + ∑ ∣⟨v∣i⟩∣2 ∥∣w⟩∥2
i=1 ∥∣w⟩∥ ∥∣w⟩∥ i=2 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ≥0
i=1
= ⟨v∣w⟩⟨w∣v⟩ + positive number (A18)

Therefore, since ⟨v∣w⟩⟨w∣v⟩ = ∣⟨v∣w⟩∣2 , we see that ∣⟨v∣w⟩∣2 ≤ ⟨v∣v⟩⟨w∣w⟩.

6. Trace equalities

The following are some useful equalities satisfied by the trace operation. They are easily provable by the rules of matrix
multiplication. A and B are arbitrary matrices of matching dimensions.

Tr(AB) = Tr(BA) (A19a)

Tr(A ⊗ B) = Tr(A)Tr(B) (A19b)
[Tr(AB)]∗ = Tr[B † A† ] . (A19c)

7. Positive operators

An operator is positive definite (or positive, for short) if all its eigenvalues are positive. An operator is positive semi-definite
if all its eigenvalues are non-negative. To test this for a given operator A, it suffices to prove that for all vectors ∣v⟩, the diagonal
matrix elements ⟨v∣ A ∣v⟩ are positive or non-negative, respectively. The reason is that this will obviously include the eigenvectors
of A.

8. Pauli matrices

The four Pauli matrices are:

1 0 0 1 0 −i 1 0
σ0 = I = ( ), σ 1 = σx = X = ( ), σ 2 = σy = Y = ( ), σ 3 = σz = Z = ( ). (A20)
0 1 1 0 i 0 0 −1

The last three are traceless by inspection. The Pauli matrices satisfy the identity

σk σl = δkl I + i ∑ εklm σm (A21)

m

where δkl is the Kronecker symbol (it is 1 if k = l, otherwise it is 0), and εklm is the completely anti-symmetric Levi-Civita
symbol [it is 1 if (klm) is an even permutation of (123), −1 if it is an odd permutation, and 0 if any index is repeated].
Since the Pauli matrices are traceless we also have the useful identity

Tr(σk σl ) = 2δkl . (A22)

 125

Appendix B: Unitarily invariant norms

√
Let V an inner product space equipped with the Euclidean norm ∥x∥ ≡ ∑i ∣xi ∣2 ⟨ei , ei ⟩, where x = ∑i xi ei ∈ V and V =
Span{ei }. Let A ∶ V ↦ V. Define
√
∣A∣ ≡ A† A . (B1)
Unitarily invariant norms are norms that satisfy, for all unitary U, V [35]:
∥U AV ∥ui = ∥A∥ui . (B2)
We list some important examples.
1. The trace norm:
∥A∥1 ≡ Tr∣A∣ = ∑ si (A) , (B3)
i

where si (A) are the singular values of A (i.e., the eigenvalues of ∣A∣). If A = ρ is a normalized quantum state, then
∥ρ∥1 = Trρ = 1.
2. The operator norm:
∥Ax∥
∥A∥∞ ≡ sup = max si (A) . (B4)
x∈V ∥x∥ i

Therefore ∥Ax∥ ≤ ∥A∥∞ ∥x∥. Also note that, by definition ∥A∥∞ ≤ ∥A∥1 , since the largest singular value is one of the
summands in ∥A∥1 .
3. The Hilbert-Schmidt norm:
√ √
∥A∥2 ≡ TrA† A = ∑ s2i (A) . (B5)
i
√ √
Again, by definition ∥A∥∞ ≤ ∥A∥2 , since ∑i s2i (A) ≥ maxi s2i (A) = ∥A∥∞ . In addition, ∥A∥21 = ∑i,j si (A)sj (A) ≥
∑i s2i (A) = ∥A∥22 .
We have thus established the ordering
∥A∥∞ ≤ ∥A∥2 ≤ ∥A∥1 . (B6)
All unitarily invariant norms satisfy the important property of submultiplicativity:
∥AB∥ui ≤ ∥A∥ui ∥B∥ui . (B7)
It follows that
∥AB∥∞ ≤ ∥A∥∞ ∥Bi ∥ , ∥B∥∞ ∥A∥i i = 1, 2, ∞ . (B8)
The norms of interest to us are also multiplicative over tensor products:
∥A ⊗ B∥i = ∥A∥i ∥B∥i i = 1, 2, ∞ . (B9)
As an application of unitarily invariant norms, let us revisit the convergence of the iterative expansion we saw in Eq. (448).
We have, for the nth order term:
t t1 tn−1
∥(−iλ)n ∫ dt1 ∫ dt2 ⋯ ∫ dtn [H̃(t1 ), [H̃(t2 ), . . . [H̃(tn ), ρSB (0)]] . . . ] ∥∞ (B10a)
0 0 0
t t1 tn−1
≤ λn ∫ dt1 ∫ dt2 ⋯ ∫ dtn ∥ [H̃(t1 ), [H̃(t2 ), . . . [H̃(tn ), ρSB (0)]] . . . ] ∥∞ (B10b)
0 0 0
t t1 tn−1
≤ λn 2n ∫ dt1 ∫ dt2 ⋯ ∫ dtn ∥H̃(t1 )∥∞ ∥H̃(t2 )∥∞ ⋯∥H̃(tn )∥∞ ∥ρSB (0)∥1 (B10c)
0 0 0
t t1 tn−1
= (2λ)n ∫ dt1 ∫ dt2 ⋯ ∫ dtn ∥HSB ∥n (B10d)
0 0 0
n
t
= (2λ∥HSB ∥)n . (B10e)
n!
To go from Eq. (B10a) to Eq. (B10b) we used the triangle inequality; to go from Eq. (B10b) to Eq. (B10c) we used the fact that
∥[A, B]∥ = ∥AB − BA∥ ≤ ∥AB∥ + ∥BA∥ ≤ 2∥A∥∥B∥ for any unitarily invariant norm; to go from Eq. (B10c) to Eq. (B10d) we
used the fact that ∥H̃(tj )∥ = ∥HSB ∥, since H̃(tj ) = U0† (t)HSB U0 (t) and U0 is unitary. Thus, the norm of the nth order term is
O[(∥HSB ∥t)n ].
 126

Appendix C: Distance and Fidelity between quantum states

Consider two quantum states represented by their density matrices ρ and σ. Suppose we perform a POVM measurement with
operators Ei , and obtain measurement outcome i with probability pi for state ρ, and qi for state σ:
pi = Tr(Ei ρ) (C1)
qi = Tr(Ei σ) (C2)
How close are the two outcomes, or equivalently, how close are the two distributions? We address this next.

1. Total variation distance and quantum distance

The total variation distance between two classical probability distributions p = {pi } and q = {qi } is defined as
1
D(p, q) ≡ ∑ ∣pi − qi ∣. (C3)
2 i

The total variation distance measure forms a metric on the space of classical probability distributions, as it satisfies all the three
properties of a metric, viz. the distance between the same variables is zero, it is symmetric, and it satisfies the triangle inequality:

D(x, x) = 0 (C4a)
D(x, y) = D(y, x) (C4b)
D(x, y) ≤ D(x, z) + D(z, y) (C4c)
The trace-norm distance can then be realized as a quantum analogue of the total variation distance.
1
D(ρ, σ) ≡ ∥ρ − σ∥1 (C5)
2
Here we have introduced the one-norm, also called the trace norm, which we define for an arbitrary matrix A:

∥A∥1 = ∑ σi (A) (C6)

i
√
where σi (A) are the singular values of A, i.e., the eigenvalues of ∣A∣ ≡= A† A. The name trace norm comes from

∥A∥1 ≡ Tr∣A∣. (C7)

While we’re at it, there is a useful inequality relating the trace norm and the operator norm [35]:

∥AB∥1 ≤ ∥A∥1 ∥B∥ (C8)

for any pair of operators A and B.

Some useful properties of the trace-norm distance are:
1. Bounded between 0 and 1: Clearly D(ρ, ρ) = 0 and D(ρ, σ) cannot be negative since it is the sum of non-negative
quantities (the singular values
√ are the absolute values of the eigenvalues). Also, by letting ρ = ∣ψ⟩⟨ψ∣ and σ = ∣φ⟩⟨φ∣ such
√
that ⟨ψ∣ φ⟩ = 0, we have Tr (ρ − σ)† (ρ − σ) = Tr ρ + σ = Tr(ρ + σ) = 2, where we used ρ + σ = (ρ + σ)2 and positivity.
Thus D(ρ, σ) = 1 in this case, and it’s not hard to see that D can’t be larger.
2. Invariance under a simultaneous unitary transformation of both ρ and σ:
1 1 1
D(U ρU † , U σU † ) = ∥U ρU † − U σU † ∥1 = ∥U (ρ − σ)U † ∥1 = ∥ρ − σ∥1 = D(ρ, σ) , (C9)
2 2 2
where we’ve used the fact that the trace norm is unitarily invariant [35].
3. If ρ and σ commute, the trace-norm distance reduces to the total variation distance between the set of paired eigenvalues
of ρ and σ. The pairing is done by their common eigenvectors (which they have by virtue of being commuting Hermitian
operators).
 127

4. It can be shown that if p and q are the probability distributions of ρ and σ for some POVM, D(ρ, σ) ≥ D(p, q). In other
words, the trace-norm distance is always an upper bound on the corresponding total variation distance. Moreover, there
always exists a POVM which saturates the bound.
Hence, an equivalent definition of the quantum distance measure is
D(ρ, σ) = sup D(p, q) (C10)
{POVM}

This is very useful since we’d like to find a measurement which makes the two states as distinguishable as possible. The
trace-norm distance automatically tells us how far apart the states would be if we could find such a measurement.

2. Fidelity Measures

A fidelity measure can be thought of as an overlap of two states, or the inner product between them. The classical fidelity is
defined as
√ √ √ √
F (p, q) = ∑ pi qi = ( ⃗p, ⃗q) , (C11)
i
√ √ √ √ √ √
i.e., it is the inner product between two vectors ⃗p = ( p1 , p2 , . . .) and ⃗q = ( q1 , q2 , . . .), whose elements are given
by square roots of the elements of classical probability distribution. The fidelity is not a metric since it doesn’t satisfy the
triangle inequality. However, arccos(F ) is a distance, also known as the Bures angle, or Bures length (related to the Bures or
Fubini-Study metric).

A quantum fidelity measure was first introduced by Uhlmann. The Uhlmann’s fidelity between two distribution ρ and σ is
clearly inspired by the classical fidelity, and is given by
√ √
F (ρ, σ) ≡ ∥ ρ σ∥1 . (C12)

3. The distance and fidelity inequality

Fidelity and distance both give us a sense of how close two states are. While the distance gives us the separation between
two states, fidelity measure the amount of overlap, or similarity of two states. We use two such measures, as while the distance
measure has a nice interpretation as resulting from the optimal POVM, the fidelity measure is often easier to calculate. The two
measures are related by the following inequality [36]:
√ √
1 − F ≤ D ≤ 1 − F 2 ⇐⇒ 1 − D ≤ F ≤ 1 − D2 . (C13)

4. Uhlman’s Theorem

Uhlman’s theorem gives a nice operational interpretation for the fidelity. Consider two states ρ and σ, acting on the same
Hilbert space H1 . Next consider the “doubled” Hilbert space given by H1 ⊗ H2 , where H2 = H1 .
One can always find two pure states ∣Ψ⟩ , ∣Φ⟩ ∈ H1 ⊗ H2 such that
ρ = TrH2 ∣Ψ⟩⟨Ψ∣ (C14a)
σ = TrH2 ∣Φ⟩⟨Φ∣. (C14b)
√ √
Indeed, if the spectral decomposition of ρ is ∑i ri ∣i⟩ ⟨i∣, then ∣Ψ⟩ = ∑i ri ∣i⟩ ⊗ ∣i⟩ yields TrH2 ∣Ψ⟩⟨Ψ∣ = TrH2 ∑ij ri rj ∣i⟩ ⟨j∣ ⊗
√
∣i⟩ ⟨j∣ = ∑ij ri rj ∣i⟩ ⟨j∣ Tr(∣i⟩ ⟨j∣) = ρ, and similarly for σ.
This procedure is called “purification”, and ∣Ψ⟩ is called a purification of ρ. While the purification of a state is not unique
√
(e.g., we could have picked ∣Ψ⟩ = ∑i eiθi ri ∣i⟩ ⊗ ∣i⟩ instead), it can clearly always be found. Uhlman’s theorem states that
F (ρ, σ) = sup ∣ ⟨Ψ∣Φ⟩ ∣ , (C15)
{∣Ψ⟩,∣Φ⟩}

i.e., the fidelity has the appealing interpretation of being the largest possible overlap among the purifications of the two states.
Thus it is also an inner product, just like the classical fidelity in Eq. (C11). Moreover, since ∣ ⟨Ψ∣Φ⟩ ∣ = ∣ ⟨Φ∣Ψ⟩ ∣, clearly
F (ρ, σ) = F (σ, ρ) . (C16)
 128

Using the definition of the trace norm and the positivity of ρ and σ, we have
√ √ √ √ √ √ √
∥ ρ σ∥1 = Tr ( ρ σ)† ( ρ σ) (C17a)
√√ √ √ √
= Tr σ ρ ρ σ (C17b)
√√ √
= Tr σρ σ (C17c)
√√ √ √ √
= Tr ρσ ρ = ∥ σ ρ∥1 , (C17d)
where the last line follows from Eq. (C16).
It turns out that, just like the trace distance is the maximum of the classical distance of the probability distributions from
arbitrary POVMs [Eq. (C10)], the quantum fidelity is the minimum of the classical fidelity of the probability distributions from
arbitrary POVMs [3][p.412]:
F (ρ, σ) = inf F (p, q). (C18)
{POVM}

5. Fidelity for a pure state passing through a noise channel

Suppose a pure state ∣ψ⟩ passes through a noise channel N , as depicted below, and we wish to compare the resultant mixed
state ρ = N (∣ψ⟩⟨ψ∣) with the original.

√
In this case we can simplify the expression for the fidelity (note that ∣ψ⟩⟨ψ∣ > 0, (∣ψ⟩⟨ψ∣)2 = ∣ψ⟩⟨ψ∣ ⇒ ∣ψ⟩⟨ψ∣ = ∣ψ⟩⟨ψ∣):
√√ √
F (ρ, ∣ψ⟩⟨ψ∣) = Tr ∣ψ⟩⟨ψ∣ρ ∣ψ⟩⟨ψ∣ (C19a)
√
= Tr ∣ψ⟩⟨ψ∣ρ∣ψ⟩⟨ψ∣ (C19b)
√
= ⟨ψ∣ρ∣ψ⟩Tr(∣ψ⟩⟨ψ∣) (C19c)
√
= ⟨ψ∣ρ∣ψ⟩. (C19d)
It turns out that in this case we can also obtain a tighter inequality than (C13),
1 − F 2 (ρ, ∣ψ⟩⟨ψ∣) ≤ D(ρ, ∣ψ⟩⟨ψ∣). (C20)

6. Fidelity is invariant under a joint unitary transformation

If we rotate ρ and σ by the same unitary transformation U , the Fidelity measure doesn’t change, i.e.
F (ρ, σ) = F (U ρU † , U σU † ) (C21)
To prove this, we note that
√the √
trace norm is a unitarily invariant norm, and hence is submultiplicative [recall Eq. (??)]. Also,
√ √ √ √
if A is positive, U AU = (U AU † )2 = U A AU † , so that
†

√ √
U AU † = U AU † . (C22)
Consequently,
√ √
F (U ρU † , U σU † ) = ∥ U ρU † U σU † ∥ (C23a)
1
√ √
= ∥U ρU † U σU † ∥1 (C23b)
√ √
= ∥U ρ σU † ∥1 (C23c)
√ √
= ∥ ρ σ∥1 (C23d)
= F (ρ, σ). (C23e)
 129

7. Fidelity of Noise channels

Consider a noise channel N that is completely positive and trace preserving (CPTP). Such maps can be represented by a set
of Kraus operators {Ki }, such that N (ρ) = ∑i Ki ρKi† and ∑i Ki† Ki = I. CPTP maps are contractive, i.e., they can only make
states become less distinguishable:
D(N (ρ), N (σ)) ≤ D(ρ, σ) (C24a)
F (N (ρ), N (σ)) ≥ F (ρ, σ) (C24b)
As a heuristic justification of these inequalities, consider a completely depolarizing noise channel which maps all states to
identity: N (ρ) = I ∀ρ. Then D(N (ρ), N (σ)) = 0 and F (N (ρ), N (σ)) = 1. At the other extreme, if N is a unitary rotation
(no decoherence), i.e., N (ρ) = U ρU † , then D(N (ρ), N (σ)) = D(ρ, σ) and F (N (ρ), N (σ)) = F (ρ, σ). Other CPTP maps lie
in between these two extremes.
Since the fidelity can only increase under a CPTP map it makes sense to define the fidelity of a noise channel by taking the
minimum over all input states ρ:
F (N ) ≡ inf F (ρ, N (ρ)). (C25)
ρ

Actually we can simplify this somewhat: we can show that the minimization doesn’t require general mixed states, but instead
pure states suffice. The reason that the fidelity satisfies “strong-concavity”, i.e., for any two convex combinations of mixed states
defined over the same index set,
√
F (∑ pi ρi , ∑ qi σi ) ≥ ∑ pi qi F (ρi , σi ) . (C26)
i i i

With this result, and the spectral decomposition ρ = ∑i λi ∣i⟩⟨i∣, we have from Eq. (C25)
F (N ) = inf F (∑ λi ∣i⟩⟨i∣, N (∑ λi ∣i⟩⟨i∣)) (C27a)
ρ
i i
√
≥ inf ∑ λi λi F (∣i⟩⟨i∣, N (∣i⟩⟨i∣)) (C27b)
ρ
i

≥ inf F (∣i⟩⟨i∣, N (∣i⟩⟨i∣) (∑ λi ) (C27c)

∣i⟩ i
= inf F (∣i⟩⟨i∣, N (∣i⟩⟨i∣) , (C27d)
∣i⟩

where in the penultimate line we used the fact that all terms of the form F (∣i⟩⟨i∣, N (∣i⟩⟨i∣) are non-negative, so eliminating all
but the smallest among them certainly makes the expression smaller.
Since every mixed state has a spectral decomposition, the infimum will be achieved for some pure state belonging to the
spectral decomposition of some mixed state. Hence the fidelity of a CPTP noise channel can be redefined as (∣ψ⟩ is a pure state)
√
F (N ) = inf F (∣ψ⟩⟨ψ∣, N (∣ψ⟩⟨ψ∣)) = inf ⟨ψ∣N (∣ψ⟩⟨ψ∣)∣ψ⟩ . (C28)
∣ψ⟩ ∣ψ⟩

8. Examples: fidelities of various noise channels

a. The pure-dephasing channel

Consider a channel that flips the phase of a qubit with probability p, and acts as identity otherwise.
NPD (ρ) = (1 − p)ρ + pZρZ (C29)
The fidelity of this channel can be calculated using Eq. (C28) as
F (NPD ) = inf F (∣ψ⟩⟨ψ∣, NPD (∣ψ⟩⟨ψ∣) (C30a)
∣ψ⟩
√
= inf ⟨ψ∣ NPD (∣ψ⟩⟨ψ∣) ∣ψ⟩ (C30b)
∣ψ⟩
√
= inf (1 − p) ⟨ψ∣ψ⟩ ⟨ψ∣ψ⟩ + p ⟨ψ∣Z∣ψ⟩ ⟨ψ∣Z∣ψ⟩ (C30c)
∣ψ⟩
√
2
= inf (1 − p) + p ⟨Z⟩ (C30d)
∣ψ⟩
 130

In this case the minimization is trivial, since, e.g., ⟨+∣Z∣+⟩ = 0. Therefore we have
√
F (NPD ) = 1 − p = 1 − p/2 + O(p2 ) (C31)

We see that the fidelity has been degraded by a term of order p. In other words, the pure-dephasing channel introduces an error
of order O(p) on the system.

b. The depolarizing channel

The depolarizing channel is represented by

p
NDep (ρ) = (1 − p)ρ + ∑ σ ρσ
α α
(C32)
3 α∈{x,y,z}

Proceeding as in in the previous example,

√
F (NDep ) = inf ⟨ψ∣ N (∣ψ⟩⟨ψ∣) ∣ψ⟩ (C33a)
∣ψ⟩
¿
Á p
= inf Á
2
À(1 − p) + ∑ ⟨ψ∣σ α ∣ψ⟩ (C33b)
∣ψ⟩ 3 α∈{x,y,z}

If ∣ψ⟩ = a ∣0⟩+b ∣1⟩, we obtain ⟨σ z ⟩ = ∣a∣2 −∣b∣2 , ⟨σ x ⟩ = 2 R(a∗ b) and ⟨σ y ⟩ = 2 I(a∗ b). The minimization over all a and b, subject
to ∣a∣2 + ∣b∣2 = 1, yields a = 1 and b = 0 as one possible solution (the easiest way to see this is to realize that the depolarizing
channel is completely symmetric, so any state will do, e.g., ∣0⟩). Thus,
√
p p
F (NDep ) = 1 − p + = 1 − + O(p2 ) (C34)
3 3
Thus, the error is again O(p).
 131

[1] H.-P. Breuer and F. Petruccione, The Theory of Open Quantum Systems (Oxford University Press, Oxford, 2002).
[2] R. Alicki and K. Lendi, Quantum Dynamical Semigroups and Applications (Springer Science & Business Media, 2007).
[3] M. A. Nielsen and I. L. Chuang, Quantum computation and quantum information (Cambridge University Press, 2010).
[4] M.-D. Choi, Linear Algebra and its Applications 10, 285 (1975).
[5] A. Peres, Physical Review Letters 77, 1413 (1996).
[6] M. Horodecki, P. Horodecki, and R. Horodecki, Physics Letters A 223, 1 (1996).
[7] P. Horodecki, M. Horodecki, and R. Horodecki, Physical Review Letters 82, 1056 (1999).
[8] H. Ollivier and W. H. Zurek, Physical Review Letters 88, 017901 (2001).
[9] C. A. Rodrı́guez-Rosario, K. Modi, A.-M. Kuah, A. Shaji, and E. C. G. Sudarshan, J. Phys. A 41, 205301 (2008).
[10] F. Buscemi, Phys. Rev. Lett. 113, 140502 (2014).
[11] J. M. Dominy, A. Shabani, and D. A. Lidar, Quant. Inf. Proc. 15, 1 (2015).
[12] J. M. Dominy and D. A. Lidar, Quant. Inf. Proc. 15, 1349 (2016).
[13] D. A. Lidar, Z. Bihary, and K. Whaley, Chem. Phys. 268, 35 (2001).
[14] L.-M. Duan and G.-C. Guo, quant-ph/9703036.
[15] C.W. Gardiner and P. Zoller, Quantum Noise, Springer Series in Synergetics, Vol. 56 (Springer, Berlin, 2000).
[16] T. A. Brun, Am. J. Phys. 70, 719 (2002).
[17] C. Majenz, T. Albash, H.-P. Breuer, and D. A. Lidar, Phys. Rev. A 88, 012103 (2013).
[18] R. Horn and C. Johnson, Topics in Matrix Analysis, Topics in Matrix Analysis (Cambridge University Press, 1994).
[19] P. Gaspard and M. Nagaoka, Journal of Chemical Physics 111, 5668 (1999).
[20] R. S. Whitney, Journal of Physics A: Mathematical and Theoretical 41, 175304 (2008).
[21] T. Albash, S. Boixo, D. A. Lidar, and P. Zanardi, New J. of Phys. 14, 123016 (2012).
[22] M. Reed and B. Simon, Methods of Modern Mathematical Physics: Fourier analysis, self-adjointness, Vol. 2 (Academic Press, 1975).
[23] J. H. Mathews and R. W. Howell, Complex Analysis: for Mathematics and Engineering, sixth ed. (Jones and Bartlett Pub. Inc., Sudbury,
MA, 2012).
[24] R. Haag, N. M. Hugenholtz, and M. Winnink, Comm. Math. Phys. 5, 215 (1967).
[25] R. Alicki, Physica A: Statistical Mechanics and its Applications 150, 455 (1988).
[26] P. Zanardi and M. Rasetti, Phys. Rev. Lett. 79, 3306 (1997).
[27] D. A. Lidar, I. L. Chuang, and K. B. Whaley, Phys. Rev. Lett. 81, 2594 (1998).
[28] D. A. Lidar and K. B. Whaley, Irreversible Quantum Dynamics, edited by F. Benatti and R. Floreanini, Lecture Notes in Physics, Vol.
622 (Springer Berlin / Heidelberg, 2003) pp. 83–120.
[29] T. Albash and D. A. Lidar, Reviews of Modern Physics 90, 015002 (2018).
[30] W. A. Majewski and R. F. Streater, Journal of Physics A: Mathematical and General 31, 7981 (1998).
[31] R. Alicki, D. A. Lidar, and P. Zanardi, Phys. Rev. A 73, 052311 (2006).
[32] H. Feshbach, Annals of Physics 5, 357 (1958).
[33] H.-P. Breuer, B. Kappler, and F. Petruccione, Physical Review A 59, 1633 (1999).
[34] A. Shabani and D. A. Lidar, Physical Review A 71, 020101 (2005).
[35] R. Bhatia, Matrix Analysis, Graduate Texts in Mathematics No. 169 (Springer-Verlag, New York, 1997).
[36] C. A. Fuchs and J. van de Graaf, IEEE Transactions on Information Theory, IEEE Transactions on Information Theory 45, 1216 (1999).