![Arxiv:2001.01572V1 [Cond-Mat.Stat-Mech] 29 Dec 2019 H Neligdtriitcdnmc Sas Concep- a Also from Question](http://data.docslib.org/img/c67828452097219356998b7300c9a015-1.webp)
arXiv:2001.01572v1 [cond-mat.stat-mech] 29 Dec 2019 h neligdtriitcdnmc sas concep- a also from question. is interesting origin dynamics tually whose deterministic terms, and underlying noise sometimes arise the Finally, contain not equations do. do these they usually circumstances they which interesting why under is both It understand terms. memory to contain not do equations otoa otefntoa eiaieo reenergy free a of derivative the functional where [4–6] the part, pro- to irreversible systems, close-to-equilibrium portional an for con- and is, They part reversible irreversible form. a general of a nonequilibrium sist share Moreover, all [3]. equations larger transport much is degrees freedom microscopic of of en- number and the momentum, although density, of mass, ergy by number are characterized fluids small well example, a For already only [2]. variables using “relevant” of described class so-called large be very can a of systems dynamics nonequilibrium the that derived. be can they which and in equations both way transport it of the nonequilibrium field, understanding of this an structure have in the to working important For Hamil- therefore (e.g., is motion). of motion equations of un- ton’s equations the phenomenological from microscopic starting from derlying derivations either from driven obtained or of are considerations study type the this in of e.g., nonequilib- arise, active on and These focused is field systems. this rium in part research large a modern equilibrium, of thermodynamic on focus to tend ∗ 2 1 orsodn uhr raphael.wittkowski@uni-muenster author: Corresponding hsi o osbefratv ytm [7]. systems active [1]. for microorganisms dir possible swimming not into are is energy example This convert good that A particles motion. are particles Active ti eyrmral n yn en biu fact obvious means no by and remarkable very a is It mechanics statistical to introductions basic Although 2 nte on sta oto hs transport these of most that is point Another . 1 rjcinoeaosi ttsia ehnc:apedagogi a mechanics: statistical in operators Projection otmte.Tasoteutosfrsystems for equations Transport matter. soft ytm.Apiain,i atclrt pnrlxto an relaxation spin int discussed. to general also particular a c give in to gene method Applications, way this more a systems. as how to mechanics explain statistical formalism we on the Thereby, course relate i mechanics. we This quantum Moreover, research. and current theory. in used the is wh of rare it formalism, form extremely the Mori-Zwanzig are in the – methodology forms to advanced introduction ac more systematic However, its in mechanics. particular statistical Ther in in derivations. courses structure microscopic advanced general of and the procedure general into the insights and many gives equ procedure. and coarse-graining macroscopic research systematic derive a through to dynamics allowing mechanics, statistical h oiZazgpoeto prtrfraimi n fth of one is formalism operator projection Mori-Zwanzig The .INTRODUCTION I. et¨ lsh ihlsUiesta ¨ntr -84 M Westf¨alische Wilhelms-Universit¨at M¨unster, D-48149 1 ntttfu hoeicePyi,Cne o otNanosci Soft for Center Physik, f¨ur Theoretische Institut ihe eVrugt te Michael .de 1 ected n ahe Wittkowski Raphael and fcluaint e hs(e es 1,1]frsc cal- such for 13] [11, Refs. amount significant (see a this requires see it to calculation cases, of both in (and be used can are) approaches ap- both both and macro- equivalent, Although are by proaches [15]. determined information density available approxi- probability means scopically by to relevant system way a a of a of density as Instead, probability actual discussed frequently. Hilbert the primarily less mate time- a is used of of method is case the product idea the scalar the In or space operators, 22]. projection [12, formed variables dependent subspace by relevant certain formed a the scalar onto space by appropriate project Hilbert can an one time- through a product, which, For of in idea observables, all op- the ways. projection through the different erator introduces one in methods, derived typically independent are and methods these presented since time-dependent understanding, is dents’ it where methods 27–30]. time-independent and 13, is [11, 22] operator 12, meth- projection general [10, distinguished: the same be where the can forms share ods tech- main many all Two in they structure. differ although forms details, different nical Moreover, method Kawasaki-Gunton 27]. or [14, 18], [17, Mori-Zwanzig- technique different Forster [26], under [25], known method formalism operator is Nakajima-Mori-Zwanzig projection and Zwanzig forms as such of exists names, variety [10], large Mori Hajime a and in [9], Zwanzig Robert [8], 24]. jima [23, physics the- particle and relaxation fluid 22], spin 21, as [20], theory [13, physics such ory functional matter density fields, active dynamical of [14–19], 11], number has [3, large mechanics formalism a mi- Mori-Zwanzig underlying in The the applications from way dynamics. systematic croscopic transport a nonequilibrium in derive equations to allows which cedure, formalism Zwanzig necletwyt td hs set sthe is aspects these study to way excellent An hsdffrnecnfr ao btcefrstu- for obstacle major a form can difference This Naka- Sadao by developed originally formalism, The yaia est ucinlter,are theory, functional density dynamical d outo otesuyo nonequilibrium of study the to roduction toso oinfo h microscopic the from motion of ations fr,i savlal nrdeto basic of ingredient valuable a is it efore, ti motn samto nphysical in method a as important is It esbeitoutost hsmto – method this to introductions cessible cue ohbscadmdr versions modern and basic both ncludes fnnqiiru rnpr equations transport nonequilibrium of c lossuet oudrtn the understand to students allows ich nti ril,w ieasml and simple a give we article, this In . a set fsaitclmechanics statistical of aspects ral eta ol fnonequilibrium of tools central e nb noprtdit lecture a into incorporated be an ¨unster, Germany 1, 81] ti oregann pro- coarse-graining a is It [8–13]. ∗ ence, a approach cal Mori- 2 culations). ten derived heuristically). It therefore describes re- However, as we shall see, it is possible to turn this versible classical mechanics. The second part with into an advantage. The simple form of the Mori-Zwanzig the prefactor ν describes irreversible dissipative dy- formalism, with time-independent projection operators, namics, which increases the entropy. can be motivated and derived based solely on ideas that the students are familiar with from quantum mechan- 3. The equation is memory-less, i.e., the rate of change ics. This gives important insights into many aspects of of ~u(~r, t) depends only on the current value ~u(~r, t), nonequilibrium statistical mechanics. Therefore, a simple but not on values ~u(~r, s) at previous times s
iPLX ∝ (iLX,A)= −(X, iLA), (16) C. Mori-Zwanzig equation in a general context such that iPLX is of order iLA and can be dropped. Equation (14), which is local in time, is much easier If we take a look at the equation of motion (14), we to solve than Eq. (9). Physically, the main difference is see that we have arrived at the general form we were that we have replaced the memory integral by the simple interested in: term D A (t). The fact that Eq. (14) only depends on ij j 1. The equations of motion depend only on the rele- the present state explains the name “Markovian approx- vant variables {A }. imation”: A Markov process is, in stochastics, a random i process that has no memory. 2. They consist of a part Ωij Aj that describes re- The other task we had to address is the scalar product. versible organized motion and of a dissipative con- 8 For this, we first require the notion of a phase-space dis- tribution Dij Aj . tribution function ρ(~q, ~p, t). In the Gibbsian framework of statistical mechanics7, where a many-particle system 3. They are memoryless, i.e., they only depend on the is described using an ensemble, which is a hypothetical current state of the system. set of infinitely many copies of the system with different Thus, Eq. (14) is a paradigmatic case for a transport initial conditions, this function specifies the probability equation in statistical mechanics. In fact, all relevant that a system that is chosen at random from this ensem- transport equations can be derived from this formalism. ble is in the microstate specified by the phase-space co- Moreover, we now have found a way to understand where ordinates ~q and ~p [33]. In the quantum-mechanical case, the general structure described above comes from and where the state of a system is specified by a wave func- under which conditions it holds. tion |ψi, the many-particle system is instead described Up to Eq. (9), our derivation was formally exact. If we by the statistical operatorρ ˆ = p |ψi hψ| , where p i i i i i describe a system not based on all microscopic degrees is the probability that the system is in the quantum state P of freedom, but only with a reduced set of variables, the |ψi . i price we have to pay is that we need to know about the A standard choice is to use a generalized correlation state of the system at previous times to fully determine function. For close-to-equilibrium quantum systems, this its temporal evolution. There are, however, certain cir- is given by the Mori product [11] cumstances under which this is not necessary, namely if 1 the relevant variables are varying very slowly compared (X, Y )= dα Tr(ˆρXe−αβHY †eαβH ) (17) to other degrees of freedom. In this case, an approximate Z0 equation can be derived that gives a closed dynamics for the relevant variables without memory effects. This requires that the set of relevant variables, which the formalism itself cannot determine, contains all slow 5 Strictly speaking, the integral is extended to a time τ , which c macroscopic variables. For example, in the description is much longer than the relaxation time scale τR, but smaller than the time scale of recurrence. “Recurrence” is the fact that of fluids, a reasonable choice is to use mass density, mo- a Hamiltonian system will, after a sufficiently long time, always mentum density, and energy density [11]. If we restrict return arbitrarily close to its initial state. Since recurrence oc- ourselves to one variable only – say, the mass density – we curs, for macroscopic systems, on time scales vastly longer than the age of the universe, this is no practical problem [25] and we can simply set the upper integration boundary to infinity. 6 This matrix is also referred to as “diffusion matrix” [13]. 7 8 There is an alternative conceptual framework known as Boltz- From the general structure (14), it is not obvious that Ωij Aj is mannian statistical mechanics, which is not based on ensembles reversible, while Dij Aj is not. To see this, considerations on the [33]. This framework is of historical and conceptual interest, but symmetries of the coefficients Ωij and Dij under time reversal not frequently used in practical calculations. are required, which can be found in Ref. [11]. 5 make the strong assumption that all other variables relax Nevertheless, most transport equations relevant for on the time scale on which the chosen variable changes out-of-equilibrium systems, such as the Navier-Stokes [15]. equation (1), are nonlinear. These nonlinearities are very The formalism does also allow to handle the case in important for the dynamics of these systems, e.g., when which memory effects are relevant, since the memoryless studying pattern formation. Hence, it is a significant case is just a particular approximation of the more gen- advantage if we are able to derive nonlinear equations. eral transport equation (9) in which memory effects are This is possible in two ways. One option is to project present. Thus, memory kernel can be calculated system- onto a larger set of relevant variables that also includes atically based on the formalism [29, 35–37]. nonlinear functions of the {Ai}. This is the basic idea It is interesting that the Mori-Zwanzig formalism al- behind the derivation of Fokker-Planck equations within lows to derive irreversible macroscopic equations of mo- the Mori-Zwanzig formalism [11, 12]. The other alterna- tion from the reversible microscopic laws. As shown in tive, which we will present here, is to use time-dependent Ref. [11], the dissipative terms lead to a monotonous in- projection operators [11]. crease of entropy and thus allow to prove a H-theorem, Another limitation of the standard method is that it as long as the Markovian approximation holds (and only is restricted to time-independent Hamiltonians. This is then). Thus, the assumption of Markovian behavior of not the most general case, since time-dependent exter- the macroscopic variables is central to the emergence of nal fields can have an important influence on the micro- macroscopic irreversibility. Moreover, as shown in Ref. scopic structure and macroscopic dynamics of a system. [25], one also requires an assumption about initial condi- Therefore, time-dependent Hamiltonians, discussed fur- tions, namely that the irrelevant part of the probability ther below in Section IIIC, have become an active field of density (see below) vanishes in the distant past. If we research in the context of Mori-Zwanzig theory, both for had made the same assumption for the future, we could time-independent [22] and time-dependent [13, 30] pro- have proven the unphysical statement that the entropy jection operators. always decreases. For this reason, projection operator methods have also attracted the attention of philosophers of physics who A. Relevant probability density are interested in where, given the reversible microscopic physics, irreversibility comes from. Philosophers study In the following, we focus on the time-dependent pro- projection operators due to the general insights they pro- jection operator method presented in the classical text- vide into the origin of irreversible transport equations book by Grabert [11]. It is a very general method appli- [38, 39]. A matter of debate here is, roughly speaking, cable to nonlinear far-from-equilibrium dynamics, allows whether the fact that irreversible laws arise from coarse- to describe also the dynamics of fluctuations [28], and has graining methods, which are often justified by the dis- recently been extended to incorporate time-dependent interest in microscopic details, lead to some undesirable Hamiltonians [13]. amount of subjectivity. After all, the fact that entropy al- The time-dependent projection operator method is ways increases seems to be a physical matter of fact which closely related to the usual methodology of statistical is not related to our ignorance of microscopic details of mechanics and extends it towards nonequilibrium sys- a statistical-mechanical system. A recent philosophical tems. In statistical mechanics, the configuration of a discussion of this problem based on the Mori-Zwanzig many-particle system is described through a probability projection operator method can be found in Ref. [40]. distribution ρ, which in classical mechanics is a function Moreover, it is discussed what precisely justifies the as- on phase space and in quantum mechanics a Hermitian sumption of Markovian behavior [2]. For a more detailed operator (“density matrix”). For equilibrium systems, presentation of the debate in philosophy of physics, with ρ is constructed by optimizing a thermodynamic poten- an emphasis on how it relates to the one in statistical tial or the entropy. When choosing the microcanonical mechanics, see Ref. [41]. ensemble, the entropy, defined as S = −kB Tr(ρ ln(ρ)), is maximized with the constraint that the total energy of the system is fixed (or, more precisely, that the total III. ADVANCED FORMS energy is between E and E + ∆E with ∆E/E ≪ 1). In the information-theoretic approach to statistical Although the theory based on Eqs. (9) and (14) is al- mechanics, pioneered by E. T. Jaynes [42], this method ready a very powerful and, in principle, formally exact of maximizing the entropy is given an epistemic justifi- method, it has certain drawbacks. The most important cation: The probability distribution is introduced as a one is that Eq. (14) is always a linear equation. As dis- way incorporating what we know (and what we do not cussed in Ref. [12], the variables A and A2, although they know) about the microscopic configuration of a system. are obviously related, are different variables in a Hilbert It is, as he argues, rational to assign to all microscopic space of dynamical functions. Therefore, if we project configurations that are, as far as we know, possible, the onto A, the dependence on A2 belongs to the orthogonal same probability. This is formalized by demanding that dynamics contained in memory and noise terms. the probability distribution is chosen in such a way that 6 it maximizes the Shannon entropy From this form of the projection operator, it is not im- mediately clear how it is related to the time-independent σ = − pi ln(pi) (18) definition. It can be shown through a longer calculation Xi (see Refs. [11, 13]) that the operator (22) contains the op- erator (3) as a limiting case and that it can be rewritten with the probabilities pi of the possible microscopic con- using a generalized scalar product. However, the general figurations. For example, if we consider a many-particle case can be more easily understood from the relevant- system about which we know nothing but the fact that density point of view than from the scalar-product point it has total energy E, particle number N, and volume of view. V , then we can maximize the information entropy, i.e., a We first make clear that P (t) defined by Eq. (22) is measure for how indifferent we are with respect to miss- still a projection operator. It has the property ing information, and arrive at the microcanonical distri- bution. Likewise, if our macroscopic information is that P (t)P (t′)X = P (t′)X. (23) we have an average energy E, we recover, by the same procedure, the canonical distribution with the temper- This is a generalization of the usual projection operator ature T arising from the Lagrange multiplier fixing the property P 2X = PX. (Note that the property (23) de- average energy [42]. pends on the specific definition (22) of the projection op- While this method thus gives the desired results for the erator and can be different for other definitions [13, 23].) equilibrium case, it can also be applied to nonequilibrium Moreover, if we continue to think of the relevant variables systems [43]. Assume that we have, like above, a set of {Ai} as basis vectors in a Hilbert space of operators and relevant variables {Ai}. Let their mean values be {ai(t)}. add the identity to the set of relevant variables [11], then These mean values are the macroscopic information we P (t)X, thought of as an element of this Hilbert space, have about our system. Then, we choose our relevant still points in a direction spanned by the relevant vari- probability density as [28] ables (including the identity, which gives the first term). 1 Now, we can use the more general projection operator ρ¯(t)= e−λi(t)Ai . (19) (22) to derive the equations of motion for the relevant Z(t) variables. For this purpose, we use instead of the Dyson with the normalization Z(t) and the thermodynamic con- identity (8) the more general identity [11] jugates {λi(t)}. This density maximizes the informa- t tional entropy with respect to the constraint that our eiLt = eiLtP (t)+ dseiLs(P (s)iLQ(s) − P˙ (s))G(s,t) Z macroscopic information is given by the mean values 0 (24) {ai(t)} of the relevant variables [16]. The normalization + Q(0)G(0,t) Z(t) ensures that with the orthogonal dynamics propagator Tr(¯ρ(t)) = 1 (20) t and the thermodynamic conjugates {λ (t)} are chosen in G(s,t) = exp i dt′ LQ(t′) , (25) j R Z such a way that s
Tr(¯ρ(t)A )= a (t), (21) where expR(·) denotes a right-time ordered exponential i i (see Section IIIC). What might be confusing here is that which is called the “macroequivalence condition”, stating the argument of the exponential is now LQ(t) rather than that the relevant density gives the correct mean values QL as in the time-independent case. The reason is that, for the relevant variables. for a time-independent projection operator, we have [13] Although Eq. (19) is a standard choice, a relevant den- iLQ(t−s) sity can be any function of the mean values that satis- QG(s,t)= Qe fies Eqs. (20) and (21) [11]. It is helpful, in general, to 2 i 2 choose the relevant density in such a way that it is a = Q 1 + (t − s)iLQ + (t − s) LQLQ + ··· 2 good approximation for the actual microscopic density, i2 in particular if they coincide for t = 0. = Q + i(t − s)QLQ + (t − s)2QLQLQ + ··· (26) 2 i2 = 1 + i(t − s)QL + (t − s)2QLQL + ··· Q B. Time-dependent projection operators 2 = eiQL(t−s)Q. For the time-dependent case, Grabert [11, 28] defines the projection operator by As we can see, if we compare this to Eq. (8), we have an additional operator Q at the end. For this reason, the ∂ρ¯(t) identity (24) is applied to iLA, while the identity (8) is P (t)X = Tr(¯ρ(t)X) + (Aj − aj (t)) Tr X . (22) ∂aj (t) applied to QiLA. 7
Applying Eq. (24) to iLA gives the equation of motion equation. In the close-to-equilibrium case, one can show that mean values and fluctuations follow similar equa- A˙i(t)= vi(t)+Ωij (t)δAj (t) tions of motion, such that Eq. (9) can be recovered by t adding them [11, 13]. Equation (33), on the other hand, + ds Ki(t,s)+ φij (t,s)δAj (s) (27) is a transport equation for the mean values of the observ- Z 0 ables/operators rather than for the operators themselves, + Fi(t, 0), as it was the case in Eq. (9). For this reason, it is possi- ble that the transport equations are nonlinear. Nonlinear where we have introduced the organized drift equations do not always make sense for the microscopic observables, which are frequently defined microscopically vi(t) = Tr(¯ρ(t)iLAi), (28) as sums over delta functions. Moreover, the microscopic the fluctuations δAi(t)= Ai(t) − ai(t), the collective fre- observables always follow a linear equation, namely the quencies Liouville equation. No such restrictions hold, in general, for the mean values. ∂ρ¯(s) Of course, obtaining a transport equation for the mean Ωij (t) = Tr iLAi , (29) ∂aj (t) values {ai(t)} is also possible in the time-independent case by averaging over Eq. (9). This transport equa- the after-effect functions tion will then be a linear equation, since the prefac- tors are time-independent and have no dependence on Ki(t,s) = Tr ρ¯(s)iLQ(s)G(s,t)iLAi , (30) the macroscopic state determined by the {ai(t)}. Al- the memory functions though it is always formally possible to apply the time- independent projection operator, it is therefore most use- ∂ρ¯(s) ful if one is close to equilibrium, such that thermody- φij (t,s) = Tr iLQ(s)G(s,t)iLAi namic nonlinearities are not important [13]. The case of ∂aj(s) (31) time-independent projection operators can be recovered ∂2ρ¯(s) − a˙ k(s) Tr G(s,t)iLAi , from the time-dependent case, if one linearizes the rele- ∂aj (s)∂ak(s) vant density in the thermodynamic conjugates, i.e., if one assumes deviations from equilibrium to be small [11, 13]. and the random forces
F (t, 0) = Q(0)G(0,t)iLA . (32) i i C. Time-dependent Hamiltonians In the more general formalism, we get separate equa- tions of motion for the mean values and the fluctuations. An even more general case is one in which the Hamil- From averaging Eq. (27) we obtain [11]9 tonian is time-dependent, which, in general, also leads to a time-dependent Liouvillian [13]. For time-independent t projection operators, this case has been considered in a˙ i(t)= vi(t)+ dsKi(s,t)+ fi(t) (33) Z0 Refs. [22, 44–47]. Generalizations to time-dependent pro- jection operators are derived in Refs. [13, 30]. We here with the mean random force fi(t) = Tr(ρ(0)Fi(t, 0)). If present the method derived in Ref. [13], which is appli- we assume ρ(0) =ρ ¯(0), we have fi(t)=0. Subtracting cable also to quantum systems. Eq. (33) from Eq. (27) gives for the fluctuations [11] This topic requires familiarity with time-ordered ex- t ponentials, which students will typically learn about in an advanced quantum mechanics or a quantum field the- δA˙ i(t)=Ωij (t)δAj (t)+ ds φij (t,s)δAj (s) Z0 (34) ory course. However, when it comes to the Mori-Zwanzig + δFi(t, 0) formalism, certain subtleties become important that are typically ignored in the treatment of time-dependent with δFi(t, 0) = Fi(t, 0) − fi(t, 0). Hamiltonians. The considerations relevant here thus can Taking a look at Eqs. (33) and (34), we can notice that be a very valuable ingredient of a quantum mechanics Eq. (34), which describes the fluctuations, has a struc- course on time-dependent Hamiltonians, even if one is ture that is relatively similar to that of Eq. (9), which not interested in their use in statistical mechanics. Here, we know from the time-independent case. The reason is we explain how this can be done. that, for the fluctuations, we have written down a linear The starting point is, again, the Liouville equation
A˙(t) = iLH(t)A(t), (35) where the Liouvillian is now time-dependent10. For t> 0, 9 A more general equation is obtained by considering an initial time t = u rather than t = 0. The physical significance of u is, in this case, that information about the history of the system from u onwards is taken into account [28]. 10 See below for a discussion of the subscript H, which denotes the 8 we can integrate Eq. (35), giving As the Liouvillian is defined as the commutator with the Hamiltonian, one therefore needs to distinguish t ′ ′ ′ between a Heisenberg-picture Liouvillian LH(t) and a A(t)= A0 + i dt LH(t )A(t ) (36) Z0 Schr¨odinger-picture Liouvillian LS(t), corresponding to the commutators with Schr¨odinger- and Heisenberg- with A0 = A(0). This equation can be solved iteratively: picture Hamiltonians, respectively. Above, we have t shown that in terms of Heisenberg-picture Liouvillians, ′ ′ A(t)= A0 + i dt LH(t )A0 the time evolution can be written as Z0 t t t′ (37) ′ ′ 2 ′ ′′ ′ ′′ A(t) = expL i dt LH(t ) A0. (42) + i dt dt LH(t )LH(t )A0 + ··· . Z0 Z0 Z0 As is shown in Ref. [13], the time evolution can also be If we perform this integration infinitely often, we will written as get an infinite number of terms with increasingly high t numbers of Liouvillians. Looking at the second-order ex- ′ ′ A(t) = expR i dt LS(t ) A0. (43) pression (37), one can already see that, because of the Z0 integration boundaries, these are ordered in such a way that the Liouvillian that depends on the latest time is Remarkably, a right-time-ordered exponential of standing on the left. Thus, the Liouvillians are said to Schr¨odinger-picture Liouvillians is equivalent to a be in left-time order. Performing infinitely many integra- left-time-ordered exponential of Heisenberg-picture tions, we get Liouvillians. A direct proof of this is sketched in Appendix B. For the right-time-ordered exponentials, ∞ t tn−1 n one can prove the identity [13] A(t)=A0 + i dt1 ··· dtn LH(t1) ··· LH(tn)A0 Z0 Z0 t nX=1 ′ ′ t (38) expR i dt LS(t ) Z =exp i dt′ L (t′) A , 0 L H 0 t Z0 ′ ′ = expR i dt LS(t ) P (t) Z0 which defines the left-time-ordered exponential expL(·). t s (44) In analogy, one can also define a right-time-ordered ex- ′ ′ ponential + ds expR i dt LS(t ) expR(·), where later times are standing on the Z0 Z0 right. P (s)iL (s)Q(s) − P˙ (s) G(s,t) When working with time-dependent Hamiltonians, it is S important to distinguish between the Heisenberg picture + Q(0)G (0,t), and the Schr¨odinger picture of quantum mechanics. In the Schr¨odinger picture, the wave functions or statistical which is a generalization of Eq. (24). Applying this to operators are time-dependent, while in the Heisenberg iLS(t)Ai again gives the general equation of motion (27). picture, the observables or operators are time-dependent. The operator A(t) in the Heisenberg picture is related to IV. APPLICATIONS the corresponding Schr¨odinger-picture expression via
A(t)= U †(t)AU(t) (39) In this section, we present two typical applications of the Mori-Zwanzig formalism to illustrate how it can be with the unitary time-evolution operator used. The Bloch equations are derived using a time- t independent projection operator, while dynamical den- i ′ ′ sity functional theory is derived with a time-dependent U(t) = expL − ~ dt HS(t ) , (40) Z0 projection operator. where HS(t) denotes the Schr¨odinger-picture Hamilto- nian. Since the Hamiltonian is time-dependent, the A. Spin relaxation and the Bloch equations Hamiltonians at different points in time do not nec- essarily commute, such that there is a difference be- The treatment of spin relaxation is a standard applica- tween the Schr¨odinger-picture Hamiltonian HS(t) and the Heisenberg-picture Hamiltonian tion of the Mori-Zwanzig formalism [13, 21, 22]. Here, we present a derivation that is a strongly simplified form of † HH(t)= U (t)HS(t)U(t). (41) the one that can be found in Refs. [13, 21], which we will follow closely. Details on the spin algebra can be found in Ref. [13]. We consider N spins in a time-independent magnetic Heisenberg picture. field B~ = B0~ez, where B0 denotes the modulus of the 9
∞ T field and ~ez = (0, 0, 1) its orientation. The system has τ2,− = dsK33(s), (55) a total spin S~ given by the sum over the individual spin Z0 ~ operators {Si}. We choose the relevant variables as we can, after a Markovian approximation11 and an aver- aging removing the random force, find the Bloch equa- S + tions A~ = ∆Sz (45) S− ˙ Sz − Sz,eq Sz = − , (56) τ1 with ∆Sz = Sz − hSiz,eq (the subscript eq denotes an ˙ −1 S± = (∓γB0 − τ2,±)S±. (57) equilibrium average) and the spin ladder operators S± = Sx ± iSy. Our Hamiltonian reads The terms ∝ γB0 describe the precession in a magnetic field. Relaxation towards the equilibrium values is de- H = H + H − γB S . (46) l sl 0 z scribed by the terms ∝ Sz/τ1 and ∝ S±/τ2,± with the relaxation times τ1 and τ2,±. Hl describes lattice interactions commuting with the spin operator S~, Hsl describes interactions of spin and lattice, and the last term accounts for the interaction with the B. Dynamical density functional theory magnetic field, where γ is the gyromagnetic ratio. If we work in the high-temperature limit, the scalar As an example for the application of the time- product of two observables is the expectation value of dependent projection operator technique, we use the the product of the two observables. In this case, we can derivation of classical dynamical density functional the- easily calculate the normalization matrix ory (DDFT) [14, 15]. DDFT is a theory for the time 2 0 0 evolution of the one-body density n(~r) in a colloidal or † 1 2 atomic fluid which is based on a free-energy functional hA~A~ ieq = N~ 0 1 0 (47) 4 0 0 2 F (t). While the original derivations have started from Langevin [48, 49] or Smoluchowski [50] equations, projec- and the frequency matrix tion operators have become an important tool in DDFT. In particular, since they can be applied to arbitrary vari- 10 0 ables, they can (and have been) used to derive extensions Ω= −iγB0 00 0 . (48) of DDFT towards additional variables, such as energy 0 0 −1 density [17], entropy density [18], and momentum den- sity [19]. For the random force, we find We suggest DDFT as an example for three reasons. First, it is relatively simple. Second, it is an extremely i ~ iQLt ~ iQLt ~ important theory in soft matter physics, such that stu- F (t)= e iQLA = e ~[Hsl, A]. (49) dents, in addition to learning about projection operators, All that is left now is to calculate the memory matrix, also learn another method that is of more general impor- the integral over which will – after a Markovian approx- tance. Third, DDFT is a good example of a rather gen- imation – give the dissipative matrix. Due to the typical eral class of nonequilibrium theories known as “gradient symmetries of the Hamiltonian, off-diagonal terms of the dynamics theories” [4, 5], where the rate of change of a memory matrix vanish, leading to a decoupling of the variable or set of variables is proportional to the gradient equations for the three variables. We obtain, after some of the functional derivative of a free energy. Seeing the calculation, the memory-matrix diagonal elements microscopic derivation of DDFT can thus further con- tribute to the general understanding of the microscopic 2 1 origins of irreversible transport equations. K (t)= eiQLt[H ,S ] [S ,H ] , 11 ~4 sl + − sl (50) N eq We follow closely the derivation by Espa˜nol and L¨owen presented in Ref. [15].12 The considered system consists 4 1 K (t)= eiQLt[H ,S ] [S ,H ] , 22 ~4 sl z z sl (51) N eq 2 1 K (t)= eiQLt[H ,S ] [S ,H ] . 11 To make the presentation simpler, we have ignored one aspect 33 ~4 sl − + sl (52) N eq that is discussed in Refs. [13, 21]: For a Markovian approxima- tion to be allowed in the case of spin relaxation, one should first Making the definitions make a transformation to the rotating frame in order to remove fast precession effects, such that one can actually assume the ∞ variables to be slow. 12 τ1 = dsK22(s), (53) A difference in our presentation is that we use, for simplicity, the Z0 relevant density (19). In Ref. [15], a more general density is used ∞ that, in addition, has a factor accounting for the equilibrium τ2,+ = dsK11(s), (54) configuration. Z0 10
of N classical particles of mass m. As a relevant variable, Inserting Eq. (63) into Eq. (33) and performing a Marko- we choose the number density13 vian approximation gives the equation of motion
N a˙ i(t)= vi(t)+ Dij (t)λj (t) (65) nˆ(~r)= δ(~r − ~r ), (58) i 15 Xi=1 with the dissipative matrix ∞ where ~ri is the position of the i-th particle. It deserves iLs Dij (t)= ds Tr(¯ρ(t)(Q(t)iLAj )e (Q(t)iLAi)). (66) some comment, since it might not be clear to students, Z0 why we are able to derive a field theory for a variable The projected current is A(~r, t) even though all our considerations were based on variables {Ai(t)} that only depend on time. The basic Q(t)iLnˆ(~r)= −∇~ · J~(~r), (67) idea is that a field A(~r, t) is an infinite number of vari- ables indexed by the position ~r. Thus, we can reuse all which gives for the dissipative matrix the expression previous results for a set of variables {Ai}. Whenever ′ ′ D(~r, ~r ,t)= ∇~ ~r · (∇~ ~r′ · M(~r, ~r ,t)) (68) we encounter a sum over the variables’ index, we have to perform an integral over ~r, since this corresponds to with the mobility a sum over all relevant variables. For the same reason, ∞ ′ ~ ′ ~ a time derivative A˙ i becomes a partial time derivative M(~r, ~r ,t)= ds Tr(¯ρ(t)J(~r ) ⊗ J(~r, s)), (69) ∂A(~r, t)/∂t, since ~r is essentially an index. Z0 The mean value ofn ˆ(~r) is where ⊗ denotes a dyadic product. Assuming that the velocities {~vi = ~pi/m} of the individual particles are un- n(~r, t) = Tr(¯ρ(t)ˆn(~r)), (59) correlated and that the positions vary slowly, this sim- 16 where the trace is now an integral over phase space. First, plifies to we need to calculate the microscopic current, which in the ∞ ′ ′ classical case is done using the Poisson bracket. We find M(~r, ~r ,t)= ds Tr(¯ρ(s)J~(~r ) ⊗ J~(~r, s)) Z0 N ∞ N ~ ~ ~ ~ iLnˆ(~r)= (∇~pi H) · (∇~ri nˆ(~r)) = −∇ · J(~r) (60) ≈ ds Tr(¯ρ(s)~vi ⊗ ~vj (s)) Z Xi=1 0 i,jX=1 ′ with the current Tr(¯ρ(t)δ(~r − ~ri)δ(~r − ~rj (s))) (70) N N ~pi ′ J~(~r)= δ(~r − ~r ) (61) ≈ D01δij Tr(¯ρ(t)δ(~r − ~ri)δ(~r − ~rj )) m i Xi=1 i,jX=1 1 ′ and ~pi being the momentum of the i-th particle. = D0 n(~r, t)δ(~r − ~r ) Next, we calculate all terms in Eq. (33). We assume with the identity matrix 1 and the diffusion coefficient ρ(0)=ρ ¯(0), which allows to drop the random force f(t). ∞ The organized drift is 1 D0 = ds Tr(¯ρ(s)~vi · ~vj (s)). (71) 3Z v(~r, t)= −∇~ · Tr(¯ρ(t)J~(~r))=0. (62) 0 Replacing in Eq. (65) the sum over j by an integral over Here, the trace is a phase-space integral that, since the ~r′, this gives function J~ is odd in the momenta, vanishes. For the ∂n(~r, t) 3 ′ ′ ′ second term, we first use the fact that – as is shown in = d r (∇~ ~r · (∇~ ~r′ · M(~r, ~r ,t)))λ(~r ,t) Ref. [11] – the after-effect function (30) can be rewritten ∂t Z in the form 3 ′ ′ ′ = D0∇~ ~r · n(~r, t) d r (∇~ ~r′ δ(~r − ~r ))λ(~r ,t) Z Ki(t,s)= Rij (t,s)λj (s), (63) (72) = −D0∇~ · (n(~r, t)∇~ λ(~r, t)), where the retardation matrix Rij (t,s) is, for classical sys- tems, given by14
15 Rij (t,s) = Tr(¯ρ(s)(G(s,t)Q(t)iLAi)(Q(s)iLAj )). (64) In the Markovian approximation, one can replace ai(s) by ai(t), since variations of ai are of first order in iLAi. The relevant densityρ ¯(t) is a functional of the {ai(t)}, such that we can also replaceρ ¯(t) byρ ¯(s) [11]. The same argument applies for the pro- 13 Although ρ(~r) is a more common notation for the number density jection operators. Moreover, as in the case of time-independent than n(~r), we here use n(~r) in order to avoid confusion with the projection operators, we replace G(s, t) by exp(iL(t−s)). Finally, probability density. we substitute s → t − s and switch integration boundaries. 14 The corresponding quantum-mechanical expression can be found 16 Since we assume that the velocities vary quickly compared to the in Refs. [11, 13]. It is slightly more complicated and not required positions, we useρ ¯(t) for the expectation value of the positions here. The general structure remains the same. andρ ¯(s) for the expectation value of the velocities. 11 where in the last step we have integrated by parts. projection operators and time-dependent Hamiltonians. Finally, we can note that if we define a coarse-grained Particular attention has been paid on points that are, for entropy as someone who is new to the field, potentially difficult to understand, such as the relation between the scalar prod- S = −kB Tr(¯ρ(t) ln(¯ρ(t))), (73) uct and the relevant density approach. Relevant appli- cations from the literature to spin relaxation and DDFT the thermodynamic conjugates {λi} can be written as have also been presented. [28] We believe that the Mori-Zwanzig formalism, pre- 1 ∂S sented in this way, can form a valuable ingredient of λi(t)= . (74) basic and advanced courses in statistical mechanics and kB ∂ai(t) quantum mechanics. It is, in its simple forms, relatively For fields, the partial derivative becomes a functional easy to explain, and allows to introduce many impor- derivative. Introducing a free energy F by the Legen- tant general ideas of theoretical physics, such as the ori- dre transformation gin and general structure of dissipative transport equa- tions in statistical mechanics and the relation between F = U − TS (75) Schr¨odinger picture and Heisenberg picture. with the internal energy U and temperature T , we can rewrite Eq. (72) as ACKNOWLEDGEMENTS
∂n(~r, t) δF = βD0∇~ · n(~r, t)∇~ , (76) R.W. is funded by the Deutsche Forschungsgemein- ∂t δn(~r, t) schaft (DFG, German Research Foundation) – WI 4170/3-1. which is the traditional DDFT equation. Since D0 > 0, we can easily prove the H-theorem Appendix A: Proof of the Dyson identity dF δF ∂n(~r, t) = d3r dt Z δn(~r, t) ∂t We here present the standard derivation of the Dyson 3 δF ~ ~ δF = d r βD0∇ · n(~r, t)∇ (77) identity (8), which can be found, e.g., in Ref. [22]. Con- Z δn(~r, t) δn(~r, t) sider the quantity 2 3 δF −iLt iQLt = − d rβD0n(~r, t) ∇~ ≤ 0. W (t)= e e (A1) Z δn(~r, t) We have thus – by virtue of restricting ourselves to one and take the time derivative: relevant variable and the approximation that this vari- W˙ (t)= −e−iLtP iLeiQLt. (A2) able shows Markovian behavior – obtained an irreversible dissipative law, the DDFT equation, from the reversible Integrating this with respect to time from 0 to t gives Hamiltonian dynamics we started with. Moreover, the Mori-Zwanzig formalism provides a microscopic expres- t −iLs iQLs sion for the free energy. W (t)= W (0) − dse P iLe . (A3) Z The structure of Eq. (76), where the dynamics of the 0 relevant variable is driven by the gradient of a thermody- We use the initial condition W (0) = 1, insert Eq. (A1) namic conjugate, is a linear-response equation and a very into Eq. (A3), and multiply from the left by eiLt. This general property of close-to-equilibrium systems. This gives can be shown with the Mori-Zwanzig formalism in a more t general way [18]. General treatments also allow to show iQLt iLt iL(t−s) iQLs how generic features of nonequilibrium thermodynamics, e = e − dse P iLe , (A4) Z0 such as irreversibility or the Onsager relations, arise from the microscopic physics in the Mori-Zwanzig formalism which is the Dyson identity. [11].
Appendix B: Relation between Heisenberg picture V. CONCLUSIONS and Schr¨odinger picture
In this article, we have provided a compact introduc- The identity [13] tion to the Mori-Zwanzig formalism, including both the t t standard, time-independent projection operator formal- ′ ′ ′ ′ expR dt iLS(t ) = expL dt iLH(t ) , (B1) ism and more advanced forms including time-dependent Z0 Z0 12 based on which a Mori-Zwanzig formalism for time- such that dependent Hamiltonians can be derived, can be proven in t two ways. An indirect proof based on expectation values ′ ′ expL dt iLH(t ) A that is valid to all orders is given in Ref. [13]. Here, we Z0 sketch a different proof that is not written down there. It t i ′ ′ is more lengthy, but conceptually easier, since it is based = A + ~ dt [HS(t ), A] Z0 on a direct calculation. ′ 2 t t We show the derivation up to second order, which is the i ′ ′′ ′′ ′ + ~ dt dt [HS(t ), [HS(t ), A]] + ··· first nontrivial one. The Heisenberg-picture Hamiltonian Z0 Z0 reads t (B6) = 1+ dt′ iL (t′) † S HH(t)= U (t)HS(t)U(t) Z0 ′ t t i t ′ ′ + dt′ dt′′ iL (t′′)iL (t′)+ ··· A = expR ~ dt HS(t ) HS(t) S S Z0 (B2) Z0 Z0 t i t ′ ′ = exp dt′ iL (t′) A. expL − ~ dt HS(t ) . R S Z0 Z0 Up to second order, we can thus write Since the operator A is arbitrary, we obtain from this Eq. (B1). We have thus shown, up to second order, that t i ′ ′ replacing the Heisenberg Liouvillian by the Schr¨odinger HH(t)= HS(t)+ dt HS(t )HS(t) ~Z Liouvillian corresponds to switching from left time order 0 (B3) t to right time order. i ′ ′ − ~ dt HS(t)HS(t )+ ··· Z0 Using Eq. (42), this gives for an arbitrary operator A the relation t ′ ′ expL dt iLH(t ) A Z0 t ′ ′ = 1+ dt iLH(t ) Z0 t t′ ′ ′′ ′ ′′ + dt dt iLH(t )iLH(t )+ ··· A Z0 Z0 t i ′ ′ = A + ~ dt [HH(t ), A] Z0 ′ 2 t t i ′ ′′ ′ ′′ + ~ dt dt [HH(t ), [HH(t ), A]] + ··· (B4) Z0 Z0 t i ′ ′ = A + ~ dt [HS(t ), A] Z0 ′ 2 t t i ′ ′′ ′′ ′ + ~ dt dt [HS(t )HS(t ), A] Z0 Z0 ′ 2 t t i ′ ′′ ′ ′′ − ~ dt dt [HS(t )HS(t ), A] Z0 Z0 ′ 2 t t i ′ ′′ ′ ′′ + ~ dt dt [HS(t ), [HS(t ), A]] + ··· . Z0 Z0
The terms of second order in HS can be rewritten as ′′ ′ ′ ′′ [HS(t )HS(t ), A] − [HS(t )HS(t ), A] ′ ′′ + [HS(t ), [HS(t ), A]] ′′ ′ ′ ′′ = HS(t )HS(t )A + AHS(t )HS(t ) (B5) ′ ′′ ′′ ′ − HS(t )AHS(t ) − HS(t )AHS(t ) ′′ ′ = [HS(t ), [HS(t ), A]], 13
[1] M. C. Marchetti, J. F. Joanny, S. Ramaswamy, T. B. Physics A: Mathematical and Theoretical 46, 355003 Liverpool, J. Prost, M. Rao, and R. A. Simha, “Hy- (2013). drodynamics of soft active matter,” Reviews of Modern [19] D. Camargo, J. de la Torre, D. Duque-Zumajo, Physics 85, 1143–1189 (2013). P. Espa˜nol, R. Delgado-Buscalioni, and F. Chejne, [2] J. Uffink, “Issues in the foundations of classical statisti- “Nanoscale hydrodynamics near solids,” Journal of cal physics,” in Philosophy of Physics, Handbook of the Chemical Physics 148, 064107 (2018). Philosophy of Science, edited by J. Butterfield and J. Ear- [20] A. Liluashvili, J. Onody,´ and T. Voigtmann, “Mode- man (Elsevier, Amsterdam, 2006) 1st ed. coupling theory for active Brownian particles,” Physical [3] D. Forster, Hydrodynamic Fluctuations, Broken Sym- Review E 96, 062608 (2017). metry, and Correlation Functions, 1st ed., Frontiers in [21] D. Kivelson and K. Ogan, “Advances in magnetic reso- Physics, Vol. 47 (Addison Wesley, Redwood City, 1989). nance,” (Academic Press, New York, 1974) Chap. Spin [4] U. Thiele, A. J. Archer, and M. Plapp, “Thermodynam- relaxation theory in terms of Mori’s formalism, pp. 71– ically consistent description of the hydrodynamics of free 155, 1st ed. surfaces covered by insoluble surfactants of high concen- [22] L. S. Bouchard, “Mori-Zwanzig equations with time- tration,” Physics of Fluids 24, 102107 (2012). dependent Liouvillian,” preprint, arXiv:0709.1358v2 [5] U. Thiele, “Recent advances in and future challenges (2007). for mesoscopic hydrodynamic modelling of complex wet- [23] T. Koide, “Derivation of transport equations using the ting,” Colloids and Surfaces A: Physicochemical and En- time-dependent projection operator method,” Progress gineering Aspects 553, 487–495 (2018). of Theoretical Physics 107, 525–541 (2002). [6] U. Thiele, T. Frohoff-H¨ulsmann, S. Engelnkemper, [24] X. Huang, T. Kodama, T. Koide, and D. H. Rischke, E. Knobloch, and A. J. Archer, “First order phase tran- “Bulk viscosity and relaxation time of causal dissipa- sitions and the thermodynamic limit,” New Journal of tive relativistic fluid dynamics,” Physical Review C 83, Physics 21, 123021 (2019). 024906 (2011). [7] R. Wittkowski, A. Tiribocchi, J. Stenhammar, R. J. [25] H.-D. Zeh, The physical basis of the direction of time 4 Allen, D. Marenduzzo, and M. E. Cates, “Scalar φ (Springer, Berlin, 1989). field theory for active-particle phase separation,” Nature [26] J. M. Dominy and D. Venturi, “Duality and condi- Communications 5, 4351 (2014). tional expectations in the Nakajima-Mori-Zwanzig for- [8] S. Nakajima, “On quantum theory of transport phenom- mulation,” Journal of Mathematical Physics 58, 082701 ena: steady diffusion,” Progress of Theoretical Physics (2017). 20, 948–959 (1958). [27] K. Kawasaki and J. D. Gunton, “Theory of nonlinear [9] R. Zwanzig, “Ensemble method in the theory of irre- transport processes: nonlinear shear viscosity and normal versibility,” Journal of Chemical Physics 33, 1338–1341 stress effects,” Physical Review A 8, 2048–2064 (1973). (1960). [28] H. Grabert, “Nonlinear transport and dynamics of fluc- [10] H. Mori, “Transport, collective motion, and Brownian tuations,” Journal of Statistical Physics 19, 479–497 motion,” Progress of Theoretical Physics 33, 423–455 (1978). (1965). [29] H. Meyer, T. Voigtmann, and T. Schilling, “On the [11] H. Grabert, Projection Operator Techniques in Nonequi- non-stationary generalized Langevin equation,” Journal librium Statistical Mechanics, 1st ed., Springer Tracts in of Chemical Physics 147, 214110 (2017). Modern Physics, Vol. 95 (Springer, Berlin, 1982). [30] H. Meyer, T. Voigtmann, and T. Schilling, “On the [12] R. Zwanzig, Nonequilibrium Statistical Mechanics, 3rd dynamics of reaction coordinates in classical, time- ed. (Oxford University Press, New York, 2001). dependent, many-body processes,” Journal of Chemical [13] M. te Vrugt and R. Wittkowski, “Mori-Zwanzig projec- Physics 150, 174118 (2019). tion operator formalism for far-from-equilibrium systems [31] R. Balian and M. V´en´eroni, “Time-dependent variational with time-dependent Hamiltonians,” Physical Review E principle for the expectation value of an observable: 99, 062118 (2019). mean-field applications,” Annals of Physics 164, 334–410 [14] A. Yoshimori, “Microscopic derivation of time-dependent (1985). density functional methods,” Physical Review E 71, [32] B. Holian and D. J. Evans, “Classical response theory in 031203 (2005). the Heisenberg picture,” Journal of Chemical Physics 83, [15] P. Espa˜nol and H. L¨owen, “Derivation of dynamical den- 3560–3566 (1985). sity functional theory using the projection operator tech- [33] R. Frigg, “A field guide to recent work on the foundations nique,” Journal of Chemical Physics 131, 244101 (2009). of statistical mechanics,” in The Ashgate Companion to [16] J. G. Anero, P. Espa˜nol, and P. Tarazona, “Functional Contemporary Philosophy of Physics, edited by D. Rick- thermo-dynamics: a generalization of dynamic density les (Ashgate, Aldershot, 2008) pp. 99–196. functional theory to non-isothermal situations,” Journal [34] J.-P. Hansen and I. R. McDonald, Theory of Simple Liq- of Chemical Physics 139, 034106 (2013). uids: with Applications to Soft Matter, 4th ed. (Elsevier [17] R. Wittkowski, H. L¨owen, and H. R. Brand, “Extended Academic Press, Oxford, 2009). dynamical density functional theory for colloidal mix- [35] H. Grabert, P. Talkner, and P. H¨anggi, “Microdynamics tures with temperature gradients,” Journal of Chemical and time-evolution of macroscopic non-Markovian sys- Physics 137, 224904 (2012). tems,” Zeitschrift f¨ur Physik B Condensed Matter 26, [18] R. Wittkowski, H. L¨owen, and H. R. Brand, “Mi- 389–395 (1977). croscopic approach to entropy production,” Journal of [36] H. Meyer, P. Pelagejcev, and T. Schilling, “Non- 14
Markovian out-of-equilibrium dynamics: a general nu- [44] S. Nordholm and R. Zwanzig, “A systematic derivation merical procedure to construct time-dependent mem- of exact generalized Brownian motion theory,” Journal ory kernels for coarse-grained observables,” preprint, of Statistical Physics 13, 347–371 (1975). arXiv:1905.11753 (2019). [45] F. Shibata, Y. Takahashi, and N. Hashitsume, “A gener- [37] A. J. Chorin, O. H. Hald, and R. Kupferman, “Optimal alized stochastic Liouville equation. Non-Markovian ver- prediction with memory,” Physica D: Nonlinear Phenom- sus memoryless master equations,” Journal of Statistical ena 166, 239–257 (2002). Physics 17, 171–187 (1977). [38] D. Wallace, “The arrow of time in physics,” in A Com- [46] T. Koide and M. Maruyama, “A new expansion of the panion to the Philosophy of Time, edited by H. Dyke Heisenberg equation of motion with projection operator,” and A. Bardon (John Wiley & Sons, Chichester, 2013) Progress of Theoretical Physics 104, 575–594 (2000). Chap. 16, pp. 262–281. [47] C. Uchiyama and F. Shibata, “Unified projection oper- [39] D. Wallace, “The quantitative content of statistical me- ator formalism in nonequilibrium statistical mechanics,” chanics,” Studies in History and Philosophy of Modern Physical Review E 60, 2636–2650 (1999). Physics 52, 285–293 (2015). [48] U. Marini Bettolo Marconi and P. Tarazona, “Dynamic [40] K. Robertson, “Asymmetry, abstraction, and auton- density functional theory of fluids,” Journal of Chemical omy: justifying coarse-graining in statistical mechanics,” Physics 110, 8032–8044 (1999). British Journal for the Philosophy of Science 69, axy020 [49] U. Marini Bettolo Marconi and P. Tarazona, “Dynamic (2018). density functional theory of fluids,” Journal of Physics: [41] M. te Vrugt, “The five problems of irreversibility,” in Condensed Matter 12, 413–418 (2000). preparation (2020). [50] A. J. Archer and R. Evans, “Dynamical density func- [42] E. T. Jaynes, “Information theory and statistical me- tional theory and its application to spinodal decompo- chanics,” Physical Review 106, 620–630 (1975). sition,” Journal of Chemical Physics 121, 4246–4254 [43] E. T. Jaynes, “Information theory and statistical me- (2004). chanics. II,” Physical Review 108, 171–190 (1975).