15: Derivation of the Nakajima-Zwanzig and Mori-Zwanzig Formalisms

Fluctuations Near Equilibrium · LibreTexts

0%

Contents

    Text

    Size
    100%
    Theme
    Typeface
    Line length
    Spacing

    Read aloud

    How

    Speed: 1×
    Light up
    Highlight

    Symbols: Mulberry Symbols, © 2018–2026 Steve Lee, CC BY-SA 4.0 (licence, the set), the site's own copies.

    From LibreTexts, under CC BY-NC-SA 4.0, as the source states it; its equations are written in TeX, which the PDF shows typeset. This page at LibreTexts.

    15: Derivation of the Nakajima-Zwanzig and Mori-Zwanzig Formalisms

    The Nakajima-Zwanzig projection operator technique is a powerful formal method in non-equilibrium statistical mechanics that is used to derive the equation of motion for a "system of interest" that is interacting with a larger "reservoir" or "bath". At its core, the method allows you to "reduce" the full density matrix of the universe (system + bath) into a reduced density matrix for just the system by formally integrating out the bath degrees of freedom. The projection-operator formalism, developed independently by Nakajima (1958) and Zwanzig (1960) and later refined by Mori (1965), constitutes a cornerstone of modern non-equilibrium statistical mechanics. Starting from the exact Liouville equation, the formalism provides a systematic procedure to eliminate microscopic degrees of freedom while retaining exact, closed equations of motion for macroscopic observables of interest. The price paid is the appearance of memory kernels and fluctuating forces - objects that encode the influence of the discarded modes on the relevant ones.

    This chapter provides a self-contained, step-by-step derivation of (i) the Nakajima-Zwanzig (NZ) equation for the reduced density matrix \(\rho_{\mathrm{s}}(t)\), and (ii) the Mori-Zwanzig Generalized Langevin Equation (GLE) for dynamical variables \(\mathbf{A}(t)\). We follow the Zwanzig convention, in which the Liouville operator is defined through the classical Poisson bracket, so that no explicit factor of \(i\) appears. Derivations are given in full detail, with physical interpretation after each key step.

    15.1 Motivation and Physical Setting

    A generic physical system of interest - a spin coupled to a phonon bath, a colloidal particle suspended in a solvent, a molecule undergoing chemical reactions in a heat reservoir - can be divided conceptually into two parts:

    • A relevant (or "system") sector \(S\), described by a small number of slowly-varying degrees of freedom.
    • An irrelevant (or "bath") sector \(B\), comprising a large number of fast, microscopic degrees of freedom.

    The full microscopic dynamics of \(S+B\) is governed by a Hamiltonian

    \[ H=H_{S}+H_{B}+H_{S B}, \tag{15.1} \label{15.1} \]

    where \(H_{S B}\) is the coupling. In principle, Liouville's equation determines the time evolution of the full phase-space distribution (or density matrix) exactly. In practice, we neither know the precise initial state of \(B\) nor care about its detailed dynamics. The challenge is to derive exact, closed equations for the relevant sector alone.

    Naive perturbation theory in \(H_{S B}\) is inadequate for strong coupling or long times. The projection-operator approach [1, 2, 3] achieves the elimination exactly, at the cost of introducing an equation with a memory integral and an inhomogeneous noise term. These are not approximations - they are exact reformulations. Approximations are made only afterwards (e.g., Markov approximation, Born approximation), and the formalism keeps track of what is being discarded.

    15.2 The Liouville Operator and Zwanzig Convention

    15.2.1 Classical mechanics setting

    Let the full system have coordinates \(\left(q_{j}, p_{j}\right), j=1, \ldots, N\), and Hamiltonian \(H\). The time evolution of any observable \(A(q, p)\) along a Hamiltonian trajectory obeys

    \[ \dot{A}=\{H, A\} \equiv \sum_{j}\left(\frac{\partial H}{\partial p_{j}} \frac{\partial A}{\partial q_{j}}-\frac{\partial H}{\partial q_{j}} \frac{\partial A}{\partial p_{j}}\right) . \tag{15.2} \label{15.2} \]

    Following Zwanzig [3], we define the Liouville operator \(L\) by

    \[ L A \equiv\{H, A\}, \tag{15.3} \label{15.3} \]

    so that \(\dot{A}=L A\). The formal solution is \(A(t)=e^{L t} A(0)\).

    15.2.2 Evolution of the phase-space density

    The phase-space distribution function \(\rho(q, p, t)\) satisfies the Liouville equation, which expresses conservation of probability along Hamiltonian trajectories:

    \[ \frac{\partial \rho}{\partial t}=-L \rho . \tag{15.4} \label{15.4} \]

    The minus sign arises because \(\rho\) transforms contravariantly: if observables are pushed forward in time by \(e^{L t}\), then the distribution is pulled back by \(e^{-L t}\). Equation \ref{15.4} is the starting point for everything that follows.

    15.3 Projection Operators: Setup and Algebra

    15.3.1 Definition and basic properties

    The key idea is to split any phase-space distribution \(\rho\) into two orthogonal parts: a relevant part \(\mathcal{P} \rho\) and an irrelevant part \(\mathcal{Q} \rho\). This is accomplished by a projection operator \(\mathcal{P}\) satisfying

    \[ \mathcal{P}^{2}=\mathcal{P} \quad \text { (idempotency). } \tag{15.5} \label{15.5} \]

    Its complement is

    \[ \mathcal{Q} \equiv \mathcal{I}-\mathcal{P}, \quad \mathcal{Q}^{2}=\mathcal{Q}, \quad \mathcal{P} \mathcal{Q}=\mathcal{Q} \mathcal{P}=0 . \tag{15.6} \label{15.6} \]

    Every distribution decomposes uniquely as \(\rho=\mathcal{P} \rho+\mathcal{Q} \rho\).

    15.3.2 A concrete choice of \(\mathcal{P}\)

    A standard and physically transparent choice is

    \[ \mathcal{P} \rho(t) \equiv \rho_{B}^{\mathrm{eq}} \operatorname{Tr}_{B}[\rho(t)]=\rho_{B}^{\mathrm{eq}} \otimes \rho_{S}(t), \tag{15.7} \label{15.7} \]

    where \(\rho_{B}^{\mathrm{eq}}\) is the equilibrium distribution of the bath and \(\rho_{S}(t)=\operatorname{Tr}_{B}[\rho(t)]\) is the reduced density of the system. One readily verifies \(\mathcal{P}^{2}=\mathcal{P}\) : applying \(\mathcal{P}\) twice simply traces out the bath again, recovering the same \(\rho_{S}(t)\).

    Physical picture. The projection \(\mathcal{P} \rho\) retains only the information about the system degrees of freedom, discarding all correlations between system and bath. The irrelevant part \(\mathcal{Q} \rho\) encodes precisely those system-bath correlations that build up over time due to the interaction \(H_{S B}\). The NZ equation will show how \(\mathcal{Q} \rho\) feeds back into the evolution of \(\mathcal{P} \rho\) through a memory integral.

    15.4 The Nakajima-Zwanzig Equation

    15.4.1 Splitting the Liouville equation

    Apply \(\mathcal{P}\) and \(\mathcal{Q}\) to both sides of Equation \ref{15.4}. Since \(\rho=\mathcal{P} \rho+\mathcal{Q} \rho\) and \(\mathcal{P}+\mathcal{Q}=\mathcal{I}\), we obtain two coupled equations:

    \[ \begin{align*} & \frac{\partial}{\partial t} \mathcal{P} \rho(t)=-\mathcal{P} L \mathcal{P} \rho(t)-\mathcal{P} L \mathcal{Q} \rho(t), \tag{15.8} \label{15.8}\\[4pt] & \frac{\partial}{\partial t} \mathcal{Q} \rho(t)=-\mathcal{Q} L \mathcal{P} \rho(t)-\mathcal{Q} L \mathcal{Q} \rho(t) . \tag{15.9} \end{align*} \]

    These are exact. Equation \ref{15.8} governs the relevant part but is not closed: it is driven by the irrelevant part \(\mathcal{Q} \rho\) through the coupling term \(-\mathcal{P} L \mathcal{Q} \rho\). Our strategy is to solve Equation \ref{15.9} for \(\mathcal{Q} \rho(t)\) in terms of \(\mathcal{P} \rho\) and substitute back.

    15.4.2 Formal solution for the irrelevant part

    Equation \ref{15.9} is a first-order linear ODE in \(\mathcal{Q} \rho\) :

    \[ \frac{\partial}{\partial t} \mathcal{Q} \rho(t)=-\mathcal{Q} L \mathcal{Q} \rho(t)-\mathcal{Q} L \mathcal{P} \rho(t) \tag{15.10} \label{15.10} \]

    The homogeneous propagator is \(e^{-\mathcal{Q} L t}\), which generates dynamics entirely within the irrelevant subspace (since \(\mathcal{Q} L \mathcal{Q}\) appears, not the full \(L\) ). Using the method of integrating factors (variation of constants), the solution is

    \[ \mathcal{Q} \rho(t)=e^{-\mathcal{Q} L t} \mathcal{Q} \rho(0)-\int_{0}^{t} d \tau e^{-\mathcal{Q} L(t-\tau)} \mathcal{Q} L \mathcal{P} \rho(\tau) \tag{15.11} \label{15.11} \]

    Let us examine each term:

    • First term \(e^{-\mathcal{Q} L t} \mathcal{Q} \rho(0)\) : the irrelevant part of the initial distribution, propagated forward by the projected dynamics. If \(\mathcal{Q} \rho(0)=0\) (the common choice of initially uncorrelated system and bath), this term vanishes.
    • Second term: the irrelevant correlations generated at all previous times \(\tau<t\) by the coupling of \(\mathcal{P} \rho(\tau)\) to \(\mathcal{Q}\), each subsequently propagated to time \(t\) by \(e^{-\mathcal{Q} L(t-\tau)}\).

    15.4.3 Derivation of the closed equation for \(\mathcal{P} \rho\)

    Substitute Equation \ref{15.11} into Equation \ref{15.8}:

    \[ \begin{aligned} \frac{\partial}{\partial t} \mathcal{P} \rho(t)= & -\mathcal{P} L \mathcal{P} \rho(t)-\mathcal{P} L\left[e^{-\mathcal{Q} L t} \mathcal{Q} \rho(0)\right] \\[4pt] & +\int_{0}^{t} d \tau \underbrace{\mathcal{P} L e^{-\mathcal{Q} L(t-\tau)} \mathcal{Q} L \mathcal{P}}_{\mathcal{K}(t-\tau)} \rho(\tau) \end{aligned} \nonumber \]

    Collecting terms, the Nakajima-Zwanzig equation reads:

    \[ \frac{\partial}{\partial t} \mathcal{P} \rho(t)=-\mathcal{P} L \mathcal{P} \rho(t)+\int_{0}^{t} d \tau \mathcal{K}(t-\tau) \mathcal{P} \rho(\tau)+\mathcal{I}(t), \tag{15.12} \label{15.12} \]

    with

    \[ \begin{align*} & \mathcal{K}(t) \equiv \mathcal{P} L e^{-\mathcal{Q} L t} \mathcal{Q} L \mathcal{P} \quad \text { (memory kernel) }, \tag{15.13} \label{15.13}\\[4pt] & \mathcal{I}(t) \equiv-\mathcal{P} L e^{-\mathcal{Q} L t} \mathcal{Q} \rho(0) \quad \text { (inhomogeneous/noise term) } \tag{15.14} \end{align*} \]

    15.4.4 Physical interpretation of each term

    • 1.Streaming term \(-\mathcal{P} L \mathcal{P} \rho(t)\) : the instantaneous, Markovian evolution of the relevant sector due to the projected Liouvillian. In the spin-boson picture, this term would describe the precession of the spin under its own field, ignoring the bath.
    • 2.Memory integral \(\int_{0}^{t} d \tau \mathcal{K}(t-\tau) \mathcal{P} \rho(\tau)\) : the cumulative non-Markovian feedback of the bath on the system. The kernel \(\mathcal{K}(t-\tau)\) encodes how correlations generated at time \(\tau\) influence the relevant sector at time \(t\). The memory decays on a timescale \(\tau_{B}\) characteristic of the bath. In the limit \(\tau_{B} \rightarrow 0\), the kernel collapses to a delta function and one recovers a Markovian (GKSL-type) master equation.
    • 3.Inhomogeneous term \(\mathcal{I}(t)\) : a noise contribution from the initial system-bath correlations. Under the assumption \(\mathcal{Q} \rho(0)=0\) (the system and bath are initially uncorrelated), this term vanishes exactly, which is the standard initial condition in open quantum systems.

    15.5 The Mori-Zwanzig Generalized Langevin Equation

    15.5.1 Change of perspective: observables instead of distributions

    While the NZ equation acts on density matrices (distributions over phase space), Mori's formulation [2] focuses on the Heisenberg-picture evolution of a chosen set of slow variables (observables) \(\mathbf{A}=\left(A_{1}, A_{2}, \ldots, A_{n}\right)\). These might be the total momentum, magnetization, or energy of a subsystem - whatever quantities are expected to evolve slowly relative to the bath.

    The time evolution is:

    \[ \mathbf{A}(t)=e^{L t} \mathbf{A}, \quad \frac{d \mathbf{A}(t)}{d t}=L \mathbf{A}(t)=e^{L t} L \mathbf{A} . \tag{15.15} \label{15.15} \]

    The projection operator now acts on observables: given an inner product \((\cdot, \cdot)\) on the space of observables (see below), we define

    \[ \mathcal{P} B \equiv(B, \mathbf{A})(\mathbf{A}, \mathbf{A})^{-1} \mathbf{A}, \tag{15.16} \label{15.16} \]

    which projects any observable \(B\) onto the subspace spanned by \(\mathbf{A}\).

    15.5.2 Key operator identity

    The derivation of the GLE rests on a single operator identity known as the Dyson decomposition (or Dyson identity):

    \[ e^{L t}=e^{\mathcal{Q} L t}+\int_{0}^{t} d \tau e^{L(t-\tau)} \mathcal{P} L e^{\mathcal{Q} L \tau} \tag{15.17} \label{15.17} \]

    Proof. Differentiate the right-hand side with respect to \(t\) and verify it satisfies the same ODE \(\frac{d}{d t} e^{L t}=L e^{L t}\) with the same initial condition \(e^{L \cdot 0}=\mathcal{I}\). Alternatively, write \(e^{L t}-e^{Q L t}\) and differentiate to obtain the desired integral equation.

    15.5.3 Derivation of the GLE

    We apply the Dyson identity Equation \ref{15.17} to act on \(L \mathbf{A}\) :

    \[ e^{L t} L \mathbf{A}=e^{\mathcal{Q} L t} L \mathbf{A}+\int_{0}^{t} d \tau e^{L(t-\tau)} \mathcal{P} L e^{\mathcal{Q} L \tau} L \mathbf{A} \tag{15.18} \label{15.18} \]

    Now decompose \(L \mathbf{A}\) using the projection:

    \[ L \mathbf{A}=\mathcal{P} L \mathbf{A}+\mathcal{Q} L \mathbf{A} . \tag{15.19} \label{15.19} \]

    Substituting the definitions of \(\Omega, \mathbf{F}(t)\), and \(\Gamma(t)\) (given below), one arrives at the Mori-Zwanzig Generalized Langevin Equation:

    \[ \frac{d \mathbf{A}(t)}{d t}=\Omega \mathbf{A}(t)-\int_{0}^{t} d \tau \Gamma(\tau) \mathbf{A}(t-\tau)+\mathbf{F}(t), \tag{15.20} \label{15.20} \]

    with

    \[ \begin{align*} \Omega & \equiv(L \mathbf{A}, \mathbf{A})(\mathbf{A}, \mathbf{A})^{-1} \quad \text { (frequency/drift matrix) } \tag{15.21} \label{15.21}\\[4pt] \mathbf{F}(t) & \equiv e^{\mathcal{Q} L t} \mathcal{Q} L \mathbf{A} \quad(\text { fluctuating force }), \tag{15.22}\\[4pt] \Gamma(t) & \equiv(\mathbf{F}(t), \mathbf{F}(0))(\mathbf{A}, \mathbf{A})^{-1} \quad \text { (memory/friction kernel). } \tag{15.23} \label{15.23} \end{align*} \]

    15.5.4 Physical interpretation of each term

    • 1.Frequency matrix \(\Omega \mathbf{A}(t)\) : the instantaneous "precession" of the slow variables due to the projected streaming. For a harmonic oscillator it becomes \(-\omega_{0}^{2} x\); for a spin it gives the Larmor precession term.
    • 2.Memory integral \(\int_{0}^{t} d \tau \Gamma(\tau) \mathbf{A}(t-\tau)\) : the dissipative (friction) contribution. The kernel \(\Gamma(\tau)\) is the autocorrelation of the fluctuating force, evaluated in the projected subspace. It encodes how much "drag" the bath exerts on the slow variables at lag \(\tau\).
    • 3.Fluctuating force \(\mathbf{F}(t)=e^{\mathcal{Q} L t} \mathcal{Q} L \mathbf{A}\) : the noise driving the slow variables. It lives entirely in the irrelevant subspace (since \(\mathcal{Q} L \mathbf{A}\) is acted on by \(e^{\mathcal{Q} L t}\), which never leaves \(\operatorname{ran} \mathcal{Q}\) ). Consequently, \(\mathbf{F}(t)\) is orthogonal to \(\mathbf{A}\left(t^{\prime}\right)\) for all \(t^{\prime} \leq t\) :

      \[ (\mathbf{F}(t), \mathbf{A})=0 \quad \forall t \geq 0 . \tag{15.24} \label{15.24} \]

      This orthogonality is the algebraic statement of the fluctuation-dissipation theorem.

    15.5.5 The fluctuation-dissipation theorem

    A key consequence of the Mori-Zwanzig construction is the second fluctuation-dissipation theorem:

    \[ \Gamma(t)=(\mathbf{F}(t), \mathbf{F}(0))(\mathbf{A}, \mathbf{A})^{-1} . \tag{15.25} \label{15.25} \]

    This is not an approximation but an exact identity within the formalism. It states that the friction kernel \(\Gamma(t)\) governing dissipation is completely determined by the time correlations of the fluctuating force \(\mathbf{F}(t)\). In the classical scalar case, with \(\Gamma(t)=\gamma \delta(t)\) (white noise, Markov limit), one recovers the familiar Einstein relation \(\gamma=k_{B} T / D\).

    15.6 Connections and Limiting Cases

    15.6.1 The Markov approximation

    If the bath correlation time \(\tau_{B}\) is much shorter than the system relaxation time \(\tau_{S}\) (i.e., \(\tau_{B} \ll \tau_{S}\) ), the memory kernel decays rapidly and we may replace \(\mathbf{A}(t-\tau)\) by \(\mathbf{A}(t)\) in the GLE:

    \[ \int_{0}^{t} d \tau \Gamma(\tau) \mathbf{A}(t-\tau) \approx\left(\int_{0}^{\infty} d \tau \Gamma(\tau)\right) \mathbf{A}(t) \equiv \gamma \mathbf{A}(t) . \tag{15.26} \label{15.26} \]

    The GLE reduces to an ordinary Langevin equation:

    \[ \dot{\mathbf{A}}(t)=(\Omega-\gamma) \mathbf{A}(t)+\mathbf{F}(t), \tag{15.27} \label{15.27} \]

    with white noise. The Markov approximation is thus clearly identified as the assumption of a vanishing bath correlation time, and its validity can be assessed by examining the decay of \(\Gamma(\tau)\). Together, NZ and MZ provide a complete, exact framework for coarse-graining microscopic dynamics. The key physical content - memory, fluctuations, and dissipation - emerges naturally and is related by exact identities that transcend any perturbative approximation.

    Bibliography

    • [1] S. Nakajima, On Quantum Theory of Transport Phenomena, Prog. Theor. Phys. 20, 948 (1958).
    • [2] R. Zwanzig, Ensemble Method in the Theory of Irreversibility, J. Chem. Phys. 33, 1338 (1960).
    • [3] H. Mori, Transport, Collective Motion, and Brownian Motion, Prog. Theor. Phys. 33, 423 (1965).
    • [4] R. Zwanzig, Nonequilibrium Statistical Mechanics, Oxford University Press (2001).

    This page titled 15: Derivation of the Nakajima-Zwanzig and Mori-Zwanzig Formalisms was last modified on Mon, 02 Nov 2026 06:09:27 GMT and is shared under a CC BY-NC-SA 4.0 license and was authored, remixed, and/or curated by Pabitra N. Sen.