WorksPhysics and Astronomy
Equilibrium propagation in non-equilibrium quantum systems
Abstract
Equilibrium propagation (EP) is a powerful training algorithm for self-learning physical systems that is widely used for a variety of machine learning inspired tasks. It relies on the existence of an energy function that depends on tunable training parameters, and can be extended to apply to many-body quantum systems in equilibrium. A key challenge however is the need to repeatedly relax to the system ground state after each parameter update. We show here a route for circumventing this problem by extending EP to many-body driven-dissipative quantum systems, where the dynamics naturally relaxes to a parameter-dependent non-equilibrium steady state. Despite the lack of a free energy, we find that for a class of open systems, a quantum detailed balance symmetry still allows for an efficient EP strategy. This relies crucially on a new class of non-equilibrium static susceptibility relations that we derive in detail. We analyze examples of our new non-equilibrium training paradigm for two experimentally-relevant models: a driven-dissipative transverse field Ising model, and a parametrically-driven bosonic lattice.
Introduction
Physical learning refers to the process by which a physical system learns some desired behavior by modifying internal degrees of freedom in response to some external stimuli [1]. These learning parameters are trained by a learning rule, which describes the dynamical process by which they are updated. Training often proceeds using gradient descent, where the physical dynamics use this learning rule to minimize some cost function. A central question in this emerging field is how to efficiently extract gradients from state preparations and measurements alone [2, 3]. Equilibrium propagation (EP) is a powerful training algorithm that allows one to extract gradients in this way for systems in thermal equilibrium [4]. This method has been proposed and successfully implemented in a variety of classical systems [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16], including non-linear resistor networks [17] and elastic networks [18].
It is interesting to ask if EP can be used for physical learning with quantum systems. Recent work shows this can be done for equilibrium quantum systems described by a Hamiltonian [19, 20, 21]. Despite its many advantages, a potential drawback of EP (especially relevant for quantum systems) is its requirement to constantly re-equilibrate the system after each nudge of model parameters. In the context of equilibrium quantum systems, EP can then only be performed through repeated preparation of ground or thermal states of many-body systems, for potentially thousands of training steps. Preparing ground states of many-body Hamiltonians is QMA-hard in the worst case [22, 23, 24], and while recent quantum Gibbs samplers (e.g. [25]) provably prepare thermal states efficiently in certain regimes (e.g. at high temperatures, or for 1D systems) [26, 27, 28, 29], efficient preparation for generic Hamiltonians at low temperature remains out of reach. This would seem to pose a serious challenge for large-scale quantum EP-based learning.
Figure 1. (a) A generic physical learning setup, where system parameters are trained so that the system reproduces a target input-output function from a set of parameters to measured quantities . (b) Implementation in a driven-dissipative system of interacting qubits. Internal parameters and inputs are drive amplitudes; the qubits interact with strength and are subject to decay at rate .
In this work, we explore an alternative approach that addresses this crucial state-preparation issue: we extend quantum EP to a setting where equilibration dynamics is not assumed, but is instead built-in to the description of the system. We study driven-dissipative quantum systems which relax to a non-equilibrium steady state that depends on system parameters, circumventing the problem of assuming an external means for repeated ground state cooling. The goal is to then use these parameter-dependent steady states for learning. Naively, it would seem that EP is now impossible: our new setting produces parameter-dependent non-thermal states, whereas EP requires the existence of an equilibrium free energy. The question then becomes whether it is ever possible to salvage EP learning in truly non-equilibrium quantum settings.
Surprisingly, the answer is yes. The key insight is that full thermal equilibrium is not necessary for ‘equilibrium’ propagation. Instead, EP just requires a specific symmetry of static susceptibilities in the steady state. While thermal equilibrium guarantees such symmetries, they can also occur robustly in more general non-equilibrium systems. A key contribution of our work is to derive susceptibility symmetries and demonstrate EP-based training in driven-dissipative quantum systems which satisfy a version of quantum detailed balance known as hidden time-reversal symmetry (hTRS) [30]. hTRS enables us to promote trivial gauge covariances in dissipative dynamics to a much more powerful set of relationships, which in turn guarantees susceptibility symmetries (see Figure 2 for a sketch of the key idea). Consider a driven-dissipative many-body system determined by a large set of drive parameters. In many cases, a simple unitary transformation relates the dynamics and steady state corresponding to one choice of parameters to those corresponding to another choice of parameters (e.g. with different drive phases). In Figure 2, models related in this way all lie on a circle in an abstract parameter space. The central physical result of our work, contained in Theorems theorem:tfim and theorem:bosonic below, is that hTRS allows us to non-trivially extend this gauge covariance to relate models that differ both in phases and magnitude of drive amplitudes (see orange arrow in Figure 2). As we show, this non-trivial mapping between different steady states directly allows us to prove the crucial susceptibility symmetries that enable EP.
Figure 2. Schematic of the space of drive parameters for a driven-dissipative quantum system. A circle of a fixed radius schematically represents the set of models whose steady states are related by a unitary transformation, i.e. for some unitary . When the system has hTRS, this gauge covariance is effectively ‘complexified’: we can write down an explicit relation between steady states that lie on circles of different radii, i.e. for drives anywhere in the parameter space of the model. This non-trivial extension of the unitary covariance is the main content of Theorems theorem:tfim and theorem:bosonic.
We demonstrate these new non-equilibrium susceptibility relations and EP based learning explicitly for two many-body dissipative systems that are experimentally relevant: a dissipative transverse-field Ising model (DTFIM) [31] (Theorem theorem:tfim), and an interacting bosonic lattice subject to two-photon driving [32] (Theorem theorem:bosonic). These examples are meant to be illustrative but not exhaustive: there are many other many-body open systems that also have hTRS and could be interesting for EP learning (e.g. dissipative spin chains [33, 34], and interacting fermionic models [35]).
The rest of the paper is organized as follows. Given our goal of bridging topics in physical learning theory with those from many-body open quantum systems, we discuss background topics from these fields relevant to our work in § 2.1 to § 2.3 (including EP and hidden time-reversal in open quantum systems). In § 3, we turn to the main setting of our work. We detail the EP training procedure in the disordered, dissipative transverse-field Ising model (DTFIM), and demonstrate a simple example of training. Then, in § 4, we do a similar analysis on an interacting bosonic lattice model, presenting the susceptibility symmetry and a simple training example.
We note that in classical systems, there has been some work to extend EP to general dynamical systems. These studies generally pursue one of two directions, each fundamentally distinct from our work. One strategy involves modifying the dynamics in each training step to make them more ‘equilibrium-like’ such that one can express an effective free energy [36, 37]. In contrast, we use hTRS to define a generating function for a subset of system observables we use for EP, without ever requiring global modifications of system dynamics. The other strategy, proposed in the context of driven-dissipative optical systems [38], develops an EP-style two phase training approach reliant on a susceptibility symmetry being approximately correct for slightly non-linear systems. In our work, the susceptibility symmetries are exact, and there is no linearity assumption: both models we study in this work are highly non-linear systems.
Finally, before proceeding, we wish to clarify the focus and scope of our work. When studying physical learning, there are two essential yet distinct questions that one can ask: that of trainability (i.e. how efficiently does a training algorithm perform gradient descent?) and that of expressivity (i.e. how good is a specific system at learning complex data?). Our focus here is on a new training algorithm based on EP in non-equilibrium quantum systems; as such, our focus will be on the first question of trainability. The question of expressivity of generic non-equilibrium systems with hTRS is of course interesting and important, but we leave a more thorough study of this question for future work.
Background
Physical learning using quantum ground states
To set the context for our work, we first review the basics of how closed quantum systems can be used for learning, and how equilibrium propagation can be implemented in this context [19]. The overarching goal is to use an input-output relation defined by a physical system with many internal parameters to learn a given target function using a set of training data. Our system will be an -site many-body quantum system where a subset of lattice sites will be used to encode input data via local Hermitian operators . The remaining lattice sites will be used as trainable parameters, and will each have an associated Hermitian operator . The Hamiltonian is:
where and are tuneable local fields, and describes fixed interaction terms.
As we now detail, expectation values in the parameter-dependent ground state of the system will define the relevant input-output relation. We partition the non-input sites into a set of sites and for some integer . We denote the control fields and observables on the first set (the ‘output’ set) of sites as
Similarly, we denote the control fields on the second set of sites (the ‘internal’ sites) as
We use () to refer to the vector of all control fields (observables) on sites , () to denote control fields (observables) for only the output sites , and () for control fields (observables) for only the internal sites . The output of the system can then be defined in terms of the ground state expectation values
These outputs are a function of the parameter-dependent ground state, hence defining our trainable input-output relation .
We next need to specify how training data generated by sampling from the target function will be used to train the internal parameters . This can be achieved by defining an appropriate cost function that is then minimized using gradient descent. We will use throughout this work the mean squared error (MSE), , where is the size of the output vector . Training occurs iteratively, where one first computes the gradient for the current choice of parameters; as we discuss, there are different ways to do this. This gradient is then used to update the internal training parameters in each training step via
where is a tunable hyperparameter referred to as the learning rate.
The gradient of the cost function can be written as
where we define
This is an error vector that encodes all the information about the cost function, and is defined only on the output sites. The question now becomes how to efficiently compute the needed gradients of observables .
Parameter-shift method
The most direct approach to computing the needed gradients is known as the parameter-shift method. It involves using a finite difference approximation for the derivatives in Eq. (6). To this end, we take measurements of the output observables in two ground states. First, we prepare a ‘free’ state, i.e. the ground state corresponding to our current choice of system parameters:
The second state is the ‘nudged’ ground state, i.e. the ground state corresponding to a new set of system parameters where just one of the training parameters has been changed from the “current" parameter values by a small amount . We can write the new set of parameters as where is a unit vector defined as
The nudged ground state is then defined as
We can then approximate the gradient of the cost function as
Each of the needed gradients is computed using a three-step procedure:
Measure output observables in present state to obtain .
Perturb a single training parameter .
Prepare the system in a new ground state , and then again measure the output observables to obtain .
The gradient is then computed as in Eq. (11). It is important to note that we must repeat this procedure to get the gradient for each of the training parameters: a new nudged ground state must be prepared for each . Therefore, for a single gradient update, we need to prepare independent ground states (one free state and a nudged state for each training parameter), and in each of them perform measurements. This results in a total of measurements per training update. One of course needs to repeat each of these measurements several times to faithfully estimate the expectation values, as is standard in any quantum experiment. In many cases, the limiting aspect of this procedure will be the large number of state preparations required of different ground states, across potentially tens of thousands of training steps. Preparing many-body systems in low-temperature states is generically difficult, both from a practical experimental perspective, but also on fundamental grounds. All these reasons motivate the search for better methods for extracting gradients.
Equilibrium propagation
Equilibrium propagation (EP) [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16] is an alternative to the parameter shift method that offers a remarkable advantage: by exploiting the fact that the system state is governed by a parameterized energy (or free energy), it extracts gradients with a dramatically smaller number of independent state preparations.
In the first phase of the EP training procedure (the “free" phase) we prepare the ground state corresponding to the current set of system parameters, and then measure all observables in this state (i.e. both on the internal sites and output sites) obtaining . This is in contrast to the parameter-shift method, where we must only measure the output observables in both phases. In the second “nudged" phase, we prepare a single new perturbed ground state. This new ground state corresponds to simultaneously nudging all the local fields on the output sites, with a nudge that is proportional to the size of the error vector: . This yields the new ground state
We once again measure all observables in this new state, obtaining . The above measurement results are enough to calculate all the needed gradients via (see App. Figure 5 for a quick proof):
The form of Eq. (13) directly shows the advantage of EP over the parameter shift method: all training parameters are updated simultaneously (i.e. using a single nudged state ), and one does not require a new nudged state for each training parameter. Hence, in EP, only two states are needed per gradient update (), independent of the number of parameters being trained. Further, we perform only total measurements per training update. In this way, EP scales more favorably than the parameter shift method in terms of total state preparations, and when also in terms of total measurements.
EP and Eq. (13) crucially rely on the fact that we are always using states that minimize an energy function (or free energy). Specifically, EP requires that parameter-dependent expectation values obey a susceptibility symmetry, namely:
In App. Figure 5, we show how this susceptibility symmetry is sufficient to prove Eq. (13). Note that the expectation values here each depend on the parameters , as they are taken in -dependent ground states.
We emphasize that this relation is formally distinct from Onsager reciprocity (see e.g. Ref. ), as it does not rely on any notion of time-reversal symmetry. Instead, it is a direct consequence of the existence of an energy function. To be explicit, the parameter-dependent ground state energy of our system can be written as a ground-state expectation value of the system Hamiltonian:
Given the form of our Hamiltonian in Eq. (1), it thus follows that we can express expectation values of an observable for some parameters in terms of a derivative with respect to its conjugate field:
where both sides should be evaluated at . This is a direct consequence of the well-known Hellmann-Feynman theorem [39]. The susceptibility symmetry Eq. (14) is then a simple consequence of the symmetry of second derivatives (sometimes referred to as Schwarz’s theorem):
We thus see that the EP training protocol will work in any setting where we are working with parameter-dependent states that minimize a parameter-dependent energy or free energy; this is the standard setting for EP [4]. This might give the impression that once we move away from equilibrium systems, EP will not be possible, as one is no longer guaranteed to have a meaningful notion of a free energy. This is overly pessimistic. We note that all EP requires is the effective symmetry of susceptibilities expressed in Eq. (14). We will show that such relations can hold even in non-equilibrium quantum settings without a well-defined free energy. Instead, these relations can result as the consequence of generalized notions of Onsager symmetry and quantum detailed balance [30].
Physical learning using non-equilibrium steady states of driven-dissipative quantum systems
We now consider the driven-dissipative analogue of the closed-system learning problem introduced in § 2.1. Consider a quantum many-body open system whose dynamics are Markovian and governed by a time-independent generator. The system is described by a density matrix , whose dynamics follows a Gorini-Kossakowski-Sudarshan-Lindblad (GKSL) form master equation [40, 41] ( throughout):
Here the Lindblad dissipative superoperator is defined as . The first term on the RHS of Eq. (18) describes purely Hamiltonian dynamics (which by itself would generate unitary evolution), while the second term describes both dissipation and incoherent driving. Each dissipative term is parameterized by a “jump operator" ; these operators describe how the system is coupled to the dissipative environment. We will be interested in the steady-state of the dynamics (often termed a non-equilibrium steady state (NESS)), which we denote . Such a steady state satisfies:
In this work, we restrict attention to cases where for any fixed choice of parameters, there is a unique NESS.
For the open-system version of the learning problem in § 2.1, we consider a setup governed by a parameter-dependent master equation of the form of Eq. (18). For simplicity, we assume that the jump operators are parameter-independent, whereas the system Hamiltonian is a function of both input variables and training parameters , i.e. . This then yields a parameter dependent NESS:
Outputs now correspond to the expectation value of output-site observables in this steady state, and we have an analogous input-output relation to the closed system. Critically, in this case, these observables need not be the ones conjugate to the input drives ; indeed, as we will show, hTRS requires that these observables in fact do not appear in the Hamiltonian at all.
This general setup gives us a parameter-dependent state and a parameter dependent input-output relation, just like the Hamiltonian ground state case. We could thus directly employ the parameter-shift method to perform learning in this setting. Note that even at this level, there is in principle an advantage over the ground state learning protocol of § 2.1. There, one had to assume some mechanism for cooling to the new ground state every time parameters were modified. In contrast, by specifying a full dissipative master equation with a unique steady state, we are defining both, a family of parameter-dependent steady states, and also a means for preparing these steady states (i.e. the dissipative dynamics directly leads to the NESS at long times, no other external state preparation machinery or “fridge" is required). Given the general difficulty of preparing ground states of many-body Hamiltonians, this already represents a potentially large advantage.
While the parameter-shift method works in this open-systems setting without modification, at first glance there is no way to employ an EP strategy: in general, does not correspond to the minimum of a well-defined free energy functional (see e.g. [42]). There is thus no built-in structure that guarantees that the crucial susceptibility symmetry required for EP (c.f. Eq. (14)) will hold. A key result of our work is to show that all hope is not lost: there is a class of open quantum systems where Eq. (14) holds despite not having any well-defined notion of a free energy. The crucial ingredient, as we now explain, is a suitable notion of quantum detailed balance.
Quantum detailed balance in dissipative systems
To employ an EP strategy for physical learning with open quantum systems, the simplest approach would be to restrict oneself to Lindblad master equations whose steady states always correspond to thermal states of the system Hamiltonian (e.g. through the use of Davies generators [43, 44, 45, 46]). While formally possible, constructing the required dissipative dynamics requires highly non-local, parameter-dependent jump operators (see e.g. [47]), leading to an experimentally infeasible setup. Further, in this case there is nothing truly new in the open systems setting, as one is again just using a thermal equilibrium state with a well-defined free energy for learning.
We instead take a different approach: we work with experimentally-feasible dissipative dynamics (i.e. based on simple local dissipators) that lead to NESS that are not simply Gibbs states of the Hamiltonian. To maintain this richness while still being able to employ the remarkable advantages of an EP training approach, we focus on dissipative dynamics that obey a specific kind of quantum detailed balance, so-called hidden time-reversal symmetry (hTRS) [30]. Unlike simpler versions of quantum detailed balance, this symmetry can hold even when the NESS is very far from being a Gibbs state. We will show below that hTRS can enable exactly the kind of susceptibility symmetry in Eq. (14) needed for EP-based learning. This hTRS can be viewed as a modified version of the more familiar KMS detailed balance, sometimes referred to as the SQDB- condition (where the modification introduces a non-trivial anti-unitary operator that plays the role of time-reversal) [48, 49]. A detailed and formal discussion is given in [30, 50]; here we just state the basic definition and its immediate consequences.
Recall that the simplest kind of classical detailed balance implies that the steady state is constrained by more than just the requirement of global stationarity: there are additional “local" constraints on the dynamics, namely that the probability current between any pair of states must vanish in the steady-state. In a similar fashion, hTRS is a form of quantum detailed balance that places independent constraints on each element of the dynamical generator. An open quantum system described by a Lindbladian Eq. (18) has its dynamics generated both by dissipative terms corresponding to the jump operators , and also via the non-Hermitian Hamiltonian . For the system to have hTRS, each of these is constrained: there must exist an anti-linear operator 1 such that
If these conditions hold, it immediately follows that is a steady state of the Lindbladian.
The existence of hTRS has some immediately relevant consequences. In particular, any open quantum system with hTRS obeys a kind of Onsager symmetry that relates certain steady-state correlation functions. Any Lindblad system with hTRS necessarily satisfies:
Here, all expectation values are taken in the NESS. We stress that only this subset of correlation functions obey the above Onsager-like symmetry, as opposed to all correlation functions. As discussed extensively in Ref. [30], this means that open systems with hTRS can still have highly non-thermal NESS, e.g. NESS that do not commute with the system Hamiltonian. Surprisingly, there are a number of many-body open systems that exhibit hTRS, including dissipative collective spin models [31], interacting bosonic models [32], and driven-dissipative spin chain systems [34, 33]. In what follows, we study two of these models as paradigmatic examples, and show that they can indeed be used for EP-based learning. As we will see, hTRS in these models directly enables the crucial susceptibility symmetry Eq. (14) needed for EP-based learning.
EP-based learning in a dissipative transverse field Ising model (DTFIM)
We now turn to the main focus of this work: a detailed study of a specific many-body open quantum system whose parameterized NESS can be used for EP-based learning despite the lack of a meaningful free energy. We will explore a dissipative, disordered transverse-field Ising model (DTFIM). Ref. [31] showed that this system has hTRS, and used this symmetry to derive an exact analytic solution for the steady state (something that will allow us to simulate large systems, see App. section:mps). We now introduce the model and show how it can be used to enable EP-based learning.
DTFIM Lindbladian and susceptibility symmetry
The DTFIM we study describes Rabi-driven qubits (Pauli matrices ) that have spatially-varying drive amplitudes, and which interact via all-to-all Ising interactions. All qubits are driven at the same frequency, and assumed to have the same splitting frequency. Working in a rotating frame set by the drive, the system Hamiltonian is:
Here is the Ising interaction amplitude, is the drive detuning, and is the drive amplitude on qubit (all taken without loss of generality to be real). Each qubit is subject to local relaxation at a rate , with corresponding jump operators . The full dynamics of the system’s density matrix is then given by the Lindbladian:
This model features a number of tuneable parameters that grow linearly with the number of qubits , namely the real drive amplitudes (see Figure 1(b) for a schematic representation of the DTFIM). This allows our system to be used for a physical learning application.
As discussed extensively in Refs. [31, 51, 52, 53, 54, 55], the NESS of this model has a rich structure stemming from the competition between driving, interactions and loss. In the case of homogeneous driving () and for large , the model is known to exhibit phase transitions between a completely depolarized state (where the driving dominates) and a polarized state (where the dissipation dominates) [31, 50]. The obvious order parameter here is the average polarization, i.e. . We will instead focus on another set of observables that are also extremely sensitive to the dynamics: the steady state coherence of each qubit, as measured by . These are zero both in the phase where the dissipation dominates (i.e. all qubits unexcited), and when the drive dominates (i.e. all qubits depolarized). They are however non-zero and maximal when the system is tuned close to its phase transition.
Remarkably, the steady-state coherences satisfy exactly the kind of susceptibility symmetry that enables EP-style learning. Using hTRS, we can show that for any set of parameters, and for any pair of qubits , steady-state expectation values obey:
This susceptibility symmetry is a central result of our work, and the proof is given in App. section:dtfim-reciprocity-derivation. It arises from hTRS promoting a simple gauge covariance of the Lindbladian to a much more non-trivial symmetry that relates steady states of different instances of the DTFIM. Specifically, when we transform the Lindbladian using a unitary , it is mapped to a different DTFIM Lindbladian where drive phases are shifted; the steady state transforms in a corresponding way. This covariance of the Lindbladian under a gauge transformation is trivially true, and does not require hTRS. In the context of Figure 2, the ‘model space’ is precisely the complex plane of drive amplitudes, and the gauge covariance is simply the freedom to move along a circle of a fixed radius (i.e. relate steady states of DTFIMs with different drive phases). Critically however, as we show in App. section:dtfim-reciprocity-derivation, hTRS turns this into a much more powerful statement, relating steady states of DTFIM Lindbladians with drives differing in both phase and magnitude. This in turn enables us to prove Eq. (27).
We stress that while Eq. (27) appears analogous to Eq. (14) for a closed system, there are crucial differences. First, notice that the observable does not appear in the DTFIM Hamiltonian. It is in fact crucial for the symmetry to hold that this observable is orthogonal in the transverse plane to the driving term in the Hamiltonian. Second, this susceptibility symmetry does not hold for all possible choices of qubit observables. Both these features are just two ways the non-equilibrium nature of our system manifests itself. We now discuss how the above symmetry relation is used as the basis for an EP training protocol.
EP-based training using non-equilibrium steady states of the DTFIM
The training setup for EP in the DTFIM is analogous to the one described in § 2.1 for closed systems. We partition the drive amplitudes as:
The function input is encoded in the drives of the input spins . The output is defined as the expectation values
for some integer , giving us the required input-output relation. We refer to the spins as the output spins, and the spins as internal spins. As in the Hamiltonian case, we use to collectively denote the drives on all the output and internal spins, . These are the trainable parameters that are learned through EP.
Training using EP here again involves the same two-phase approach discussed in § 2.1, with each phase using a different NESS. In the first free phase, the system is allowed to relax to the NESS corresponding to the current set of system parameters, defined as
In this state, we measure the observables on all trainable spins. In the second nudged phase, we now simultaneously alter all of the drive amplitudes of the output spins, with the nudge of each drive taken to be proportional to the size of the error vector (c.f. Eq. (7)): . The system then relaxes to the new steady state corresponding to these parameters, yielding the ‘nudged’ NESS
We once again measure the observables on all internal and output spins. The EP gradient update is then given by
This follows from Eq. (27) via the same argument used for the closed system case, c.f. App. Figure 5.
We thus see that at a practical level, EP is implemented in our DTFIM model analogously to its implementation in a closed system (c.f. § 2.1). The key difference now however is that we are using a NESS which is not determined by a free energy. Just like the closed-system case, our quantum dissipative version of EP uses a nudge of all training parameters to calculate all gradient updates simultaneously.
In what follows, we use a ‘symmetric’ nudging procedure, as is now standard across implementations of EP (see e.g. [19]). This refers to performing two nudges of the output drives, of size , in opposite directions for better numerical convergence. Explicitly, the gradient update is then given by
We emphasize that in this case, symmetric EP needs three state preparations per gradient update; the first free state to determine the error vector in the current state, along with two nudged states and for the gradient update.
In the simulations that follow, we also use symmetric nudging for parameter-shift; the total number of state preparations per gradient step for this method is then . Each training parameter gets two nudges, using a finite difference to estimate (see Eq. (11)). Instead of a shared free state (as described in § 2.1.1), each parameter needs two independent states to estimate the gradient, yielding states per gradient update as opposed to in the one-sided nudging case.
Teacher-student learning
We now turn to a concrete demonstration of EP-based training of the DTFIM setup. As stated in the introduction, our goal is to show the basic efficacy of this training approach, and not conflate this issue with the distinct question of expressivity. To this end, we discuss teacher-student learning, a kind of supervised learning task where we know a priori that the target function is learnable. Teacher-student learning involves training one initialization of the system (the ‘student’) using training data generated from a different initialization of the system with different parameters and/or architecture (the ‘teacher’). It has been used extensively in the machine learning literature as a way to train small, simple neural networks from more complicated ones (see e.g. [56]).
We first construct the teacher system that will produce the target input-output relation. This will be an spin DTFIM where we just have a single input spin (with a drive amplitude ), and output spins. The drive amplitudes on the non-input spins are fixed at the values (see Table table:dtfim-1 for exact values), and we take . To produce the target input-output relation, for each value of in the input domain, we measure the outputs for in the corresponding NESS. This generates our training data that we will attempt to learn with a second system.
We then initialize a new -spin DTFIM (the ‘student’) with all the same system parameters as the teacher () except the training drives, which we set to some random values . The goal then is to train these student drives using the training data to learn the teacher input-output relation. Since the systems are identical, this is in some ways a trivial task, since we know that a solution always exists. This kind of setup is still useful, as it lets us benchmark a particular training strategy (in our case EP) both in terms of accuracy (i.e. the final converged error) and efficiency (i.e. how quickly is the student able to learn the internal parameters needed to mimic the teacher).
Figure 3. (a) Schematic depicting teacher-student learning, where a ‘student’ system with drives is used to learn the input-output relation generated by an identical ‘teacher’ system (with some initially unknown drives ). (b) Input-output relations for teacher-student learning with a spin DTFIM model (c.f. Eq. (25)) with one input spin and three output spins (). The input is the spin 11 drive amplitude , and the outputs are ; only two are plotted. Solid pink curves correspond to the target input-output relation defined by the teacher. The drive amplitudes (for spins 1 to 10) are trained either by EP (§ 2.1.2) or parameter-shift (§ 2.1.1). Black dotted ‘initial’ curves correspond to the student output pre-training. The blue dotted ‘learned’ curves correspond to the outputs of the student system after training has converged. Both methods converge to the same input-output relation. We use , and . We approximate finite differences with , and use a learning rate . Values for teacher and student drives (before and after training) are included in Table table:dtfim-1 in App. section:dtfim-teacher-student-examples. (c) The mean squared error (MSE) between the outputs and the targets as training progresses. We plot the evolution of the MSE for both training methods, EP and parameter shift, as a function of the number of DTFIM state preparations.
In Figure 3, we present an example of teacher-student learning for and (here ). We compare results for training based on EP, versus training based on the more naive parameter-shift method (see § 2.1.1), with all learning hyperparameters fixed across both methods. A further example (corresponding to different random teacher drives) is presented in App. section:dtfim-teacher-student-examples. All simulations in this model are performed using a formally exact matrix product state (MPS) representation of the exact steady-state solution (see App. section:mps for details). This allows efficient numerical evaluation of steady-state expectation values limited only by machine precision and does not require MPS simulation of the full master equation.
Figure 3(c) shows the evolution of the cost function being minimized, the MSE , as the training progresses for both EP and the parameter shift method applied to our DTFIM. It is plotted as a function of the total number of distinct DTFIM state preparations. We clearly see a relative speedup of EP versus parameter shift that scales as expected: the number of state preparations needed in EP for a converged minimal is smaller by a factor of compared to the number needed for parameter-shift. This is what would be expected from the general arguments in § 2.1 (for a symmetric nudging procedure, which we use).
It is also interesting to note that the learned student drives (in Table table:dtfim-1 in App. section:dtfim-teacher-student-examples) do not exactly match the teacher drives, despite the input-output relation being learned correctly. In particular, for this setup where we have a three dimensional function output, the converged student parameters only match those of the teacher on the three output spins. This is in fact true for all examples tested (see App. section:dtfim-teacher-student-examples), where the student drives match the teacher drives for only the output spins, but not necessarily the internal ones.
EP in a bosonic lattice with quadratic driving
We now show that our quantum non-equilibrium version of EP can apply for models with more structure and complexity than the DTFIM of the previous section. The model is a pair driven, dissipative and interacting bosonic lattice system (see Figure 4(a)), where again the NESS is not a simple Gibbs state. As discussed below, this model now has a non-trivial spatial structure. Despite this, hTRS still enables a static susceptibility symmetry in the NESS, which in turn makes EP possible and effective.
Model and susceptibility symmetry
Consider bosonic modes (with associated lowering operators ) with parametric (two-photon) drives associated with every pair of modes; the complex drive amplitude for a given pair is denoted . All drives are at the same frequency, and all modes have the same linear frequency (with a detuning from the parametric resonance). The system also has self-Kerr and cross-Kerr nonlinearities. We take these all to have magnitude , so that the interaction involves the total photon number of the system. Treating the driving within the rotating wave approximation and moving to a rotating frame, we obtain a time-independent Hamiltonian:
where is an arbitrary complex symmetric matrix, and is a photon number operator, i.e. . Finally, each mode is subject to local single-site photon loss at a dissipation rate (taken to be the same for each mode), yielding the Lindbladian
To use this system for physical learning, we will again view the drive amplitudes as trainable parameters. While we have some of the same ingredients as the DTFIM spin model (i.e. global interaction, driving, local dissipation), there are crucial differences. In addition to an unbounded local Hilbert space dimension, our system now also has a spatial structure (i.e. defines a connectivity graph). In contrast, the DTFIM is essentially zero-dimensional. These differences make the physics of this model far richer, and observables are generally more sensitive to variations of drive amplitudes (e.g. compare Figure 4(b) against Figure 3(b) for the DTFIM). Also, note that the number of tuneable parameters now grow quadratically in the system size , whereas it only grew linearly in the DTFIM.
The basic physics of this setup involves a competition between driving, dissipation and interactions. The drives populate the modes with photon pairs and generate squeezing correlations between them. This is opposed by the loss, detuning and the Kerr interactions. The competition between these two effects gives rise to dissipative phase transitions, non-trivial steady states and a host of rich quantum phenomena [32]. A number of related models have also been studied numerically in the literature [57, 58, 59]. Surprisingly, for all choices of parameters, the pair-driven bosonic model described by Eq. (34) and (35) has hTRS and is exactly solvable, something demonstrated in Ref. [32]. We make use of this structure in what follows.
We now move to the key new result of this section: the existence of a susceptibility symmetry in the NESS that enables EP. As shown in App. section:bosonic-proof and sketched below, hTRS again allows us to derive non-trivial constraints on static susceptibilities in the NESS, for any choice of model parameters. However, for these to be useful for EP, we need an additional constraint beyond simply having hTRS. We also need to restrict our model such that (in some fixed local gauge) all drive amplitudes are real. This is markedly different from the case of the DTFIM, where a useful susceptibility symmetry exists even if drive amplitudes are allowed to be complex (see App. section:dtfim-reciprocity-derivation for details). This additional constraint in our bosonic model is suggestive of a conventional time-reversal symmetry, i.e. there are no non-trivial Aharonov-Bohm fluxes associated with the connectivity between modes encoded in the matrix.
Similar to the DTFIM, the observables that turn out to be relevant for EP are associated with coherence in the NESS. We introduce for each pair of modes the Hermitian operators
and characterize coherent pair correlations between the modes and , whereas is the total photon number in these modes.
As derived in detail in App. section:bosonic-proof, if we restrict our model to always having real-valued drive amplitudes , hTRS implies a non-trivial susceptibility symmetry associated with the drive-dependent NESS:
where is the Kronecker delta. Here, as before, expectation values are taken in the NESS (which is itself a function of the real-valued ). Note that here and denote the same parameter since our model is only defined for a symmetric matrix. Therefore, the weights account for the fact that our Hamiltonian in Eq. (34) was defined to sum over all , and hence double counts the contribution associated with modes for which .
Note that there is a structural similarity to the corresponding relation for the DTFIM, Eq. (27). In the DTFIM, the relevant observables were the coherences , which did not explicitly appear in the Lindbladian, and which were conjugate to the drive operators . Similarly, in the bosonic model the do not appear directly in the Lindbladian; instead the real drive amplitudes couple directly to the conjugate quantities .
As in the case of the DTFIM, the proof of our bosonic susceptibility symmetry relies crucially on hTRS, and at a heuristic level directly connects to Figure 2. One starts with a trivial kind of gauge covariance under a transformation that mixes modes . This allows one to trivially relate the NESS for one choice of to another if the two matrices are unitarily related. hTRS turns this into something much more powerful: we can now relate the NESS for any choice of matrix to any other choice . This is the formal content of our Theorem theorem:bosonic (see App. section:bosonic-proof). This non-trivial relation between different NESS then lets us directly prove Eq. (39).
Figure 4. (a) Schematic depicting the encoding of an input-output relation in a dissipative, pair-driven bosonic lattice (here with modes). We show a setup with one input parameter and three output observables. The single input and training parameters (gray arrows) are driving amplitudes (matrix elements of ). The three outputs are the observables . There are 10 tuneable couplings in the system, and all but the input are trainable parameters, . (b) Input-output relations for teacher-student training in the four mode bosonic model depicted in (a). We plot two of the three outputs. Solid gray curves correspond to the target relation set by the teacher. Dashed ‘initial’ curves correspond to student outputs pre-training; dotted ‘learned’ curves correspond to these outputs post-training. Both EP and parameter-shift are used, and converge to the same relation. We use system parameters , . We approximate finite differences with , and use a learning rate . Values for teacher and student drives are included in Table table:bosonic in App. section:dtfim-teacher-student-examples. (c) Mean square error (MSE) as training progresses, for both EP and parameter-shift, as a function of the number of bosonic lattice steady state preparations.
EP-based training using NESS
In this section we describe how the NESS of the bosonic system can be used to perform EP-based training. The key difference from the training setup for the DTFIM described in § 3 is that we encode inputs and training parameters in a driving matrix that couples pairs of modes, as opposed to a vector of local single site drives . For a system of modes, the bosonic model has tunable real parameters, while the DTFIM only had parameters for spins.
It is useful to define sets , , which contain tuples corresponding (respectively) to the input, output and additional internal training drives. To be explicit, we partition the elements of our drive matrix as
The inputs are encoded in , while the (which refers to all couplings and ) are the trainable parameters. Note that here by definition we have that is the same pair as , and we therefore have one parameter and for each pair. The rescaling factor cancels the weights in Eq. (39), and the susceptibility symmetry takes the more natural form:
This is the form required for EP (see App. Figure 5), and is why the as opposed to are the natural training parameters.
Having specified how the drive elements in our model will be used either as inputs to our learning problem or as internal trainable parameters, we now turn to identifying the outputs. These are defined as
We again emphasize that only one output is defined for a given pair of modes or equivalently . Training proceeds similarly as in the DTFIM; the gradient update is given by
Here, the steady state () in the free (nudged) phase is defined identically as in Eq. (30) (Eq. (31)) where now is the steady state of the Lindbladian Eq. (35).
In Figure 4(b-c), we present results for teacher-student training (see § 3 for details on the training setup) on an mode system used to map a one dimensional input to a three dimensional output. To compute steady-state expectation values we constructed a formally exact matrix product state representation of the exact steady-state solution (see App. section:bosonic-mps for details). This MPS construction is very similar to the DTFIM with one significant difference: the local Hilbert space dimension of the DTFIM is 2, but it is unbounded for the bosonic model. While this requires us to introduce a local photon cutoff for the bosonic MPS, it also adds significantly more complexity to the state. We note that the local-photon-cutoff approximation can be controlled by increasing the photon cutoff.
Once again, we compare training using EP and parameter-shift (with all learning hyperparameters fixed across both training methods), and use a symmetric nudging procedure to compute gradients. One again sees (as expected) a significant learning speedup of EP over parameter shift. To be explicit, the number of state preparations needed for convergence of EP is improved by a factor of where denotes the cardinality of the set. What is more interesting in this example is the structure of the input-output relation in Figure 4(b): the outputs are distinct, non-monotonic functions of the input (in contrast to the DTFIM outputs in Figure 3(b)). One may then expect that this input-output relation could represent a much richer class of functions than the DTFIM even at smaller system sizes, potentially providing higher expressivity. At a heuristic level, this can be attributed to the additional complexity of the bosonic model that arises from it having an effective spatial structure. While a full analysis of the expressive power of this system is outside of the scope of the current work, it would be an interesting avenue for future exploration.
Outlook
In this work, we established that the highly effective equilibrium propagation strategy for physical learning can be extended to non-equilibrium, driven-dissipative quantum systems. This is possible despite the absence of any meaningful free energy in the non-equilibrium setting. We showed that the crucial ingredient was hidden time-reversal symmetry (a kind of quantum detailed balance). We derived fundamentally new general relations on how the NESS flows under a change of parameters using hTRS (Theorems theorem:tfim and theorem:bosonic) in two representative many-body models: a dissipative Ising model (§ 3) and a driven-dissipative bosonic lattice (§ 4), and showed how these directly enable EP. We benchmarked training in both models using a teacher-student learning method, confirming that EP provides a substantial scaling advantage over the direct parameter-shift method in these non-equilibrium settings.
Having established that EP is possible in driven-dissipative quantum systems, a natural next question is that of expressivity of non-equilibrium quantum systems with hTRS. The structure of the input-output relation of the bosonic model (see Figure 4(b)) is already encouraging. Exploring more systematic metrics to benchmark the kinds of functions these models can learn is an interesting avenue for future work. Furthermore, we need not limit ourselves to just these two models: there are a number of models that are known to have hTRS (e.g. driven spin chains [34, 33], interacting fermion models [35]). It would be interesting to explore our EP-based learning approach with these systems.
Another extremely promising direction for future work is to explore learning in systems that only approximately have hTRS, and thus only approximately satisfy our susceptibility symmetries (Eq. (27), Eq. (39)). This symmetry is used to approximate gradients needed to descend a loss landscape (gradient descent). Crucially, the gradients do not have to be exact: approximate gradients will still move training in the right direction, towards a minimum of the loss landscape, albeit training might take longer. This is the spirit of much related work in the field (see e.g. [38]). Using models that only approximately have hTRS could vastly expand the class of usable systems. This would both avoid issues of experimental fine-tuning, but could also potentially be used to enhance expressivity.
ACKNOWLEDGMENTS
We thank Arvind Murugan for helpful discussions. Research was sponsored by the Army Research Office and was accomplished under Grant Number W911NF-25-1-0261. AC also acknowledges support from the Simons Foundation through a Simons Investigator Award (Grant No. 669487).
AI statement. A large language model (Claude, Anthropic) was used to write and debug code for numerical simulations. It helped refine and generalize the key proofs of Theorems theorem:tfim and theorem:bosonic. All calculations presented were derived and verified independently by the authors. All text was written by the authors, although Claude helped catch typos and notational inconsistencies. The authors take full responsibility for the content of the work.
Figure 5. Output observables for the bosonic model with parameters corresponding to the teacher system (for ). For each photon cutoff, we use a bond dimension .
References
- [1]M. Stern and A. Murugan, ``Learning without neurons in physical systems,'' \em Annual Review of Condensed Matter Physics , vol. 14, no. 1, pp. 417--441, 2023.
- [2]A. Momeni, B. Rahmani, B. Scellier, L. G. Wright, P. L. McMahon, C. C. Wanjura, Y. Li, A. Skalli, N. G. Berloff, T. Onodera, \em et al. , ``Training of physical neural networks,'' \em Nature , vol. 645, no. 8079, pp. 53--61, 2025.
- [3]V. Lopez-Pastor and F. Marquardt, ``Self-learning machines based on hamiltonian echo backpropagation,'' \em Physical Review X , vol. 13, no. 3, p. 031020, 2023.
- [4]B. Scellier and Y. Bengio, ``Equilibrium propagation: Bridging the gap between energy-based models and backpropagation,'' \em Frontiers in computational neuroscience , vol. 11, p. 24, 2017.
- [5]B. Scellier, ``A deep learning theory for neural networks grounded in physics,'' \em arXiv preprint arXiv:2103.09985 , 2021.arxiv.org/abs/2103.09985
- [6]J. Kendall, R. Pantone, K. Manickavasagam, Y. Bengio, and B. Scellier, ``Training end-to-end analog neural networks with equilibrium propagation,'' \em arXiv preprint arXiv:2006.01981 , 2020.arxiv.org/abs/2006.01981
- [7]E. Martin, M. Ernoult, J. Laydevant, S. Li, D. Querlioz, T. Petrisor, and J. Grollier, ``Eqspike: Spike-driven equilibrium propagation for neuromorphic implementations,'' \em iScience , vol. 24, no. 3, p. 102222, 2021.
- [8]M. Ernoult, J. Grollier, D. Querlioz, Y. Bengio, and B. Scellier, ``Equilibrium propagation with continual weight updates,'' \em arXiv preprint arXiv:2005.04168 , 2020.arxiv.org/abs/2005.04168
- [9]B. Scellier, S. Mishra, Y. Bengio, and Y. Ollivier, ``Agnostic physics-driven deep learning,'' \em arXiv preprint arXiv:2205.15021 , 2022.arxiv.org/abs/2205.15021
- [10]M. J. Falk, A. T. Strupp, B. Scellier, and A. Murugan, ``Temporal contrastive learning through implicit non-equilibrium memory,'' \em Nature Communications , vol. 16, p. 2163, 2025.
- [11]Q. Wang, C. C. Wanjura, and F. Marquardt, ``Training coupled phase oscillators as a neuromorphic platform using equilibrium propagation,'' \em Neuromorphic Computing and Engineering , vol. 4, no. 3, p. 034014, 2024.
- [12]J. Laydevant, D. Markovi ć , and J. Grollier, ``Training an ising machine with equilibrium propagation,'' \em Nature Communications , vol. 15, no. 1, p. 3671, 2024.
- [13]M. Stern, D. Hexner, J. W. Rocks, and A. J. Liu, ``Supervised learning in physical networks: From machine learning to learning machines,'' \em Physical Review X , vol. 11, no. 2, p. 021045, 2021.
- [14]M. Stern, A. J. Liu, and V. Balasubramanian, ``Physical effects of learning,'' \em Physical Review E , vol. 109, p. 024311, 2024.
- [15]P. O'Connor, E. Gavves, and M. Welling, ``Training a spiking neural network with equilibrium propagation,'' in \em Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics (AISTATS) (K. Chaudhuri and M. Sugiyama, eds.), vol. 89 of \em Proceedings of Machine Learning Research , pp. 1516--1523, PMLR, 2019.
- [16]S. Dillavou, M. Stern, A. J. Liu, and D. J. Durian, ``Demonstration of decentralized physics-driven learning,'' \em Physical Review Applied , vol. 18, p. 014040, 2022.
- [17]S. Dillavou, B. D. Beyer, M. Stern, A. J. Liu, M. Z. Miskin, and D. J. Durian, ``Machine learning without a processor: Emergent learning in a nonlinear electronic metamaterial,'' \em arXiv preprint arXiv:2311.00537 , 2023.arxiv.org/abs/2311.00537
- [18]L. E. Altman, M. Stern, A. J. Liu, and D. J. Durian, ``Experimental demonstration of coupled learning in elastic networks,'' \em Physical Review Applied , vol. 22, p. 024053, 2024.
- [19]C. C. Wanjura and F. Marquardt, ``Quantum equilibrium propagation for efficient training of quantum systems based on onsager reciprocity,'' \em Nature Communications , vol. 16, no. 1, p. 6595, 2025.
- [20]S. Massar and B. M. Mognetti, ``Equilibrium propagation: the quantum and the thermal cases,'' \em Quantum Studies: Mathematics and Foundations , vol. 12, no. 1, p. 6, 2025.
- [21]B. Scellier, ``Quantum equilibrium propagation: Gradient-descent training of quantum systems,'' \em arXiv preprint arXiv:2406.00879 , 2024.arxiv.org/abs/2406.00879
- [22]A. Y. Kitaev, A. H. Shen, and M. N. Vyalyi, \em Classical and Quantum Computation , vol. 47 of \em Graduate Studies in Mathematics . \newblock Providence, RI: American Mathematical Society, 2002.
- [23]J. Kempe, A. Kitaev, and O. Regev, ``The complexity of the local Hamiltonian problem,'' \em SIAM Journal on Computing , vol. 35, no. 5, pp. 1070--1097, 2006.
- [24]R. Oliveira and B. M. Terhal, ``The complexity of quantum spin systems on a two-dimensional square lattice,'' \em Quantum Information \& Computation , vol. 8, no. 10, pp. 900--924, 2008.
- [25]C.-F. Chen, M. Kastoryano, F. G. Brand ã o, and A. Gily é n, ``Efficient quantum thermal simulation,'' \em Nature , vol. 646, no. 8085, pp. 561--566, 2025.
- [26]T. Bergamaschi and C.-F. Chen, ``Fast mixing of quantum spin chains at all temperatures,'' in \em Proceedings of the 58th Annual ACM Symposium on Theory of Computing , pp. 835--844, 2026.
- [27]C. Rouz é , D. Stilck Fran ç a, and Á . M. Alhambra, ``Efficient thermalization and universal quantum computing with quantum Gibbs samplers,'' \em Nature Physics , vol. 22, pp. 1123--1127, 2026.
- [28]C. Rouz é , D. Stilck Fran ç a, and Á . M. Alhambra, ``Optimal quantum algorithm for gibbs state preparation,'' \em Physical Review Letters , vol. 136, no. 6, p. 060601, 2026.
- [29]Y. Tong and Y. Zhan, ``Fast mixing of weakly interacting fermionic systems at any temperature,'' \em PRX Quantum , vol. 6, no. 3, p. 030301, 2025.
- [31]D. Roberts and A. A. Clerk, ``Exact solution of the infinite-range dissipative transverse-field ising model,'' \em Physical review letters , vol. 131, no. 19, p. 190403, 2023.
- [32]D. Roberts and A. Clerk, ``Competition between two-photon driving, dissipation, and interactions in bosonic lattice models: An exact solution,'' \em Physical Review Letters , vol. 130, no. 6, p. 063601, 2023.
- [34]A. Lingenfelter, M. Yao, A. Pocklington, Y.-X. Wang, A. Irfan, W. Pfaff, and A. A. Clerk, ``Exact results for a boundary-driven double spin chain and resource-efficient remote entanglement stabilization,'' \em Physical Review X , vol. 14, no. 2, p. 021028, 2024.
- [35]A. Lingenfelter and A. A. Clerk, ``Exact steady states of interacting driven dissipative fermionic systems with hidden time-reversal symmetry,'' \em arXiv preprint arXiv:2605.10846 , 2026.arxiv.org/abs/2605.10846
- [36]B. Scellier, A. Goyal, J. Binas, T. Mesnard, and Y. Bengio, ``Generalization of equilibrium propagation to vector field dynamics,'' \em arXiv preprint arXiv:1808.04873 , 2018.arxiv.org/abs/1808.04873
- [37]A. E. Scurria, D. V. Abeele, B. M. Mognetti, and S. Massar, ``Equilibrium propagation for non-conservative systems,'' \em arXiv preprint arXiv:2602.03670 , 2026.arxiv.org/abs/2602.03670
- [38]N. D. Cin, F. Marquardt, and C. C. Wanjura, ``Training nonlinear optical neural networks with scattering backpropagation,'' \em arXiv preprint arXiv:2508.11750 , 2025.arxiv.org/abs/2508.11750
- [39]R. P. Feynman, ``Forces in molecules,'' \em Physical review , vol. 56, no. 4, p. 340, 1939.
- [40]G. Lindblad, ``On the generators of quantum dynamical semigroups,'' \em Communications in mathematical physics , vol. 48, no. 2, pp. 119--130, 1976.
- [41]V. Gorini, A. Kossakowski, and E. C. G. Sudarshan, ``Completely positive dynamical semigroups of n-level systems,'' \em Journal of Mathematical Physics , vol. 17, no. 5, pp. 821--825, 1976.
- [42]R. Fazio, J. Keeling, L. Mazza, and M. Schir ò , ``Many-body open quantum systems,'' \em SciPost Physics Lecture Notes , p. 099, 2025.
- [43]E. B. Davies, ``Markovian master equations,'' \em Communications in Mathematical Physics , vol. 39, no. 2, pp. 91--110, 1974.
- [44]E. B. Davies, ``Markovian master equations. II ,'' \em Mathematische Annalen , vol. 219, no. 2, pp. 147--158, 1976.
- [45]R. Alicki, ``On the detailed balance condition for non- H amiltonian systems,'' \em Reports on Mathematical Physics , vol. 10, no. 3, pp. 249--258, 1976.
- [46]A. Kossakowski, A. Frigerio, V. Gorini, and M. Verri, ``Quantum detailed balance and KMS condition,'' \em Communications in Mathematical Physics , vol. 57, no. 2, pp. 97--110, 1977.
- [47]J. Guo, O. Hart, C.-F. Chen, A. J. Friedman, and A. Lucas, ``Designing open quantum systems with known steady states: Davies generators and beyond,'' \em Quantum , vol. 9, p. 1612, 2025.
- [48]F. Fagnola and V. Umanita, ``Generators of detailed balance quantum markov semigroups,'' \em Infinite Dimensional Analysis, Quantum Probability and Related Topics , vol. 10, no. 03, pp. 335--363, 2007.
- [49]F. Fagnola and V. Umanit à , ``Generators of kms symmetric markov semigroups on symmetry and quantum detailed balance,'' \em Communications in Mathematical Physics , vol. 298, no. 2, pp. 523--547, 2010.
- [50]D. D. Noachtar and A. A. Clerk, ``Exact metastability in a class of driven-dissipative quantum many-body systems,'' June 2026. \newblock arXiv:2606.07736 [quant-ph].arxiv.org/abs/2606.07736
- [51]M. Marcuzzi, E. Levi, S. Diehl, J. P. Garrahan, and I. Lesanovsky, ``Universal nonequilibrium properties of dissipative rydberg gases,'' \em Physical review letters , vol. 113, no. 21, p. 210401, 2014.
- [52]K. Tucker, D. Barberena, R. J. Lewis-Swan, J. K. Thompson, J. G. Restrepo, and A. M. Rey, ``Facilitating spin squeezing generated by collective dynamics with single-particle decoherence,'' \em Physical Review A , vol. 102, p. 051701, Nov. 2020.
- [53]E. Y. Song, D. Barberena, D. J. Young, E. Chaparro, A. Chu, S. Agarwal, Z. Niu, J. T. Young, A. M. Rey, and J. K. Thompson, ``A dissipation-induced superradiant transition in a strontium cavity- QED system,'' \em Science Advances , vol. 11, p. eadu5799, Apr. 2025.
- [54]Y.-X. Xiang, W. Li, Z. Bai, and Y.-Q. Ma, ``Switching Dynamics of Metastable Open Quantum Systems ,'' May 2025. \newblock arXiv:2505.05202 [quant-ph].arxiv.org/abs/2505.05202
- [55]N. Leppenen and E. Shahmoon, ``Quantum bistability at the interplay between collective and individual decay,'' Apr. 2024. \newblock arXiv:2404.02134 [quant-ph] version: 1.arxiv.org/abs/2404.02134
- [56]G. Hinton, O. Vinyals, and J. Dean, ``Distilling the knowledge in a neural network,'' \em arXiv preprint arXiv:1503.02531 , 2015.arxiv.org/abs/1503.02531
- [57]V. Savona, ``Spontaneous symmetry breaking in a quadratically driven nonlinear photonic lattice,'' \em Physical Review A , vol. 96, no. 3, p. 033826, 2017.
- [58]J. Lebreuilly, C. Aron, and C. Mora, ``Stabilizing arrays of photonic cat states via spontaneous symmetry breaking,'' \em Physical review letters , vol. 122, no. 12, p. 120402, 2019.
- [59]R. Rota, F. Minganti, C. Ciuti, and V. Savona, ``Quantum critical regime in a quadratically driven nonlinear photonic lattice,'' \em Physical review letters , vol. 122, no. 11, p. 110405, 2019.





