betaFeedback

WorksPhysics and Astronomy

Optimality of the Petz-Fuchs Smoothed State

Abstract

Quantum state smoothing is an estimation technique that uses past and future measurement information obtained from a continuous-in-time measurement record to reconstruct the state of an open quantum system. Recently, there has been a new approach to quantum state smoothing based on quantum retrodiction known as the Petz-Fuchs smoothed state. Here I prove that this Petz-Fuchs smoothed state majorizes all other smoothed states that only depend on the filtered state (the operator conditioned on the past measurement record) and the retrofiltered effect (the operator describing the future measurement record). To do this, I characterize the set of all possible smoothed states satisfying this condition.

Introduction

Smoothing is a classical estimation technique that uses a continuous-in-time measurement record gathered both prior and posterior to a time tt to estimate the state of a classical, Markovian open system [1, 2]. As this technique uses information after time tt one cannot implement this technique in real time, unlike its counterpart, filtering [3, 4], which only utilizes the measurement record prior to the estimation time (the past record). In the case of Markovian open quantum systems [5], one can consider a similar scenario, where one performs a continuous-in-time weak measurement of the environment to gain more information about the state of the system. The concept of filtering (also known as quantum trajectory theory) [6, 7, 8, 9] follows analogously from the classical theory, yielding a conditional quantum state based on the past measurement record. However, directly adapting the classical smoothing theory to the quantum scenario posed problems due to the non-commutative nature of the filtered state ρF(t)\rho_{\mathrm{F}}(t) and the retrofiltered effect E^R(t)\hat{E}_{\mathrm{R}}(t), an effect operator describing the measurement record posterior to time tt (the future record).

While there were many approaches to utilizing past and future measurement information in quantum systems [10, 11, 12, 13, 14] (see Ref. [15] for a more comprehensive discussion of these techniques), recently [16], there was a new approach to quantum state smoothing proposed. This approach to smoothing viewed classical smoothing from the perspective of Bayesian retrodiction, whereby the smoothed state is dynamically constructed backwards-in-time from the filtered state at the final time, conditioning on the future record as it evolves. Applying the quantum analogue of classical Bayesian retrodiction, the Petz recovery map [17, 18], to the filtered quantum state, one obtains the Petz-Fuchs smoothed state,

ρSPF(t)∝ρF(t)E^R(t)ρF(t).(1)\rho_{\mathrm{S}}^{\mathrm{PF}}(t) \propto\sqrt{\rho_{\mathrm{F}}(t)}\hat{E}_{\mathrm{R}}(t)\sqrt{\rho_{\mathrm{F}}(t)}. \tag*{(1)}

While this state is a valid smoothed quantum state, it is unclear in what sense, if any, this state is optimal. As quantum retrodiction is a currently active research field [19, 20, 21, 22], with different approach to retrodiction existing, the Petz-Fuchs smoothed state is just one of many possible valid smoothed quantum state, some of which may be a better in terms of, say, purity or entropy. In this paper I will show that the Petz-Fuchs smoothed state majorizes all other valid smoothed quantum states that only depend on the filtered state and retrofiltered effect, and hence is the purest possible smoothed state that one can assign to the system. To do this, I will first characterize the entire set of smoothed quantum states that depend only on the filtered state and retrofiltered effect, from which one can show that the Petz-Fuchs state majorizes all other elements of this set.

The structure of this paper is as follows. First, in Sec. II, I will begin by briefly reviewing both classical and quantum state filtering and smoothing, as well as the four criteria that a state must satisfy in order to be consider a valid smoothed quantum state. Next, in Sec. III, I will characterize the set of valid smoothed quantum states that only depend on the filtered state and retrofiltered effect. Finally, in Sec. IV, I will show that the Petz-Fuchs smoothed state majorizes the set of valid smoothed quantum states.

Brief Review: Filtering and Smoothing

This section will only provide a brief review of the necessary concepts required in this paper. For a more comprehensive review of classical state filtering and smoothing, see Refs. [2, 3], and for quantum state filtering and smoothing, see Refs. [9, 16]. Note, here I will only review Petz-Fuchs smoothing; for an in depth review of an alternative formulation, known as Guevara-Wiseman smoothing, see Ref. [15, 23].

Classical Systems

Let us consider an Markovian open classical system described by a vector of parameters x\mathbf{x}, which, for simplicity, we will assume can only take discrete values in the set X\mathsf{X}. The state of this system at time tt is defined as the probability distribution of the parameters, ℘(x;t)\wp(\mathbf{x};t). As we are considering a discrete Markovian system, the dynamics are given by

℘(x;t+dt)=∑x′∈Xφ(x;t+dt∣x′;t)℘(x′;t),(2)\wp(\mathbf{x};t+\mathrm{d}t)=\sum_{\mathbf{x}'\in\mathcal{X}}\varphi(\mathbf{x};t+\mathrm{d}t|\mathbf{x}';t)\wp(\mathbf{x}';t), \tag*{(2)}

with initial condition ℘(x;t0)\wp(\mathbf{x};t_0), where φ(x;t+dt∣x′;t)\varphi(\mathbf{x};t+\mathrm{d}t|\mathbf{x}';t) describing the dynamical transition probabilities for the system. Here, the transition probabilities must satisfy the completeness condition ∑x∈Xφ(x;t+dt∣x′;t)=1\sum_{\mathbf{x}\in\mathcal{X}}\varphi(\mathbf{x};t+\mathrm{d}t|\mathbf{x}';t)=1 to preserve total probability. This state is generally referred to as the unconditioned state, and, in the absence of additional measurement information, is the most accurate description of the system.

To obtain more information about the system one can continuously measure the system obtaining a vector of noisy measurement signals yt\mathbf{y}_t. In the case of filtering, one makes use of the past measure record, defined as O←t={yτ:τ∈[t0,t)}\overleftarrow{\mathcal{O}}_t=\{\mathbf{y}_\tau:\tau\in[t_0,t)\} to obtain a more accurate description of the system at time t+dtt+\mathrm{d}t. Using Bayes theorem to update our knowledge of the state, the filtered state evolves as

℘~F(x;t+dt)=∑x′∈XFyt(x;t+dt∣x′;t)℘~F(x′;t),(3)\widetilde{\wp}_{\mathrm{F}}(\mathbf{x};t+\mathrm{d}t)=\sum_{\mathbf{x}'\in\mathcal{X}}\mathcal{F}_{\mathbf{y}_t}(\mathbf{x};t+\mathrm{d}t|\mathbf{x}';t)\widetilde{\wp}_{\mathrm{F}}(\mathbf{x}';t), \tag*{(3)}

where ℘F(x;t):=℘(x;t∣O←t)\wp_{\mathrm{F}}(\mathbf{x};t):=\wp(\mathbf{x};t|\overleftarrow{\mathcal{O}}_t) and the initial condition is ℘(x;t0)\wp(\mathbf{x};t_0). Here the tilde accent ∙~\widetilde{\bullet} denotes that this state is unnormalized and the conditional dynamical map is Fyt(x;t+dt∣x′;t)=φyt(x;t+dt∣x′;t)℘(yt∣x;t)\mathcal{F}_{\mathbf{y}_t}(\mathbf{x};t+\mathrm{d}t|\mathbf{x}';t)=\varphi_{\mathbf{y}_t}(\mathbf{x};t+\mathrm{d}t|\mathbf{x}';t)\wp(\mathbf{y}_t|\mathbf{x};t) and the unconditional transition probabilities have been modified to φyt(x;t+dt∣x′;t)\varphi_{\mathbf{y}_t}(\mathbf{x};t+\mathrm{d}t|\mathbf{x}';t) to account for possible measurement backaction, which can occur even in classical systems.

Finally, to obtain an even more accurate estimate of the system, one can use the past-future measurement record, O↔={yτ:τ∈[t0,T)}\overleftrightarrow{\mathcal{O}}=\{\mathbf{y}_\tau:\tau\in[t_0,T)\} where TT is the final time. Taking the Bayesian retrodiction approach, one can obtain the smoothed state via the dynamical equation

℘~S(x;t)=∑x′∈XRyt℘~F(x;t)(x;t∣x′;t+dt)℘~S(x′;t+dt),(4)\widetilde{\wp}_{\mathrm{S}}(\mathbf{x};t)=\sum_{\mathbf{x}'\in\mathcal{X}}\mathcal{R}^{\widetilde{\wp}_{\mathrm{F}}(\mathbf{x};t)}_{\mathbf{y}_t}(\mathbf{x};t|\mathbf{x}';t+\mathrm{d}t)\widetilde{\wp}_{\mathrm{S}}(\mathbf{x}';t+\mathrm{d}t), \tag*{(4)}

where ℘S(x;t):=℘(x;t∣O↔)\wp_{\mathrm{S}}(\mathbf{x};t):=\wp(\mathbf{x};t|\overleftrightarrow{\mathcal{O}}) and the final condition is ℘~S(x;T):=℘~F(x;T)\widetilde{\wp}_{\mathrm{S}}(\mathbf{x};T):=\widetilde{\wp}_{\mathrm{F}}(\mathbf{x};T). The Bayesian retrodictive map [24] is defined as

Ryt℘~F(x;t)(x;t∣x′;t+dt)=Fyt(x′;t+dt∣x;t)℘~F(x;t)∑x′′∈XFyt(x′;t+dt∣x′′;t)℘~F(x′′;t).(5)\mathcal{R}^{\widetilde{\wp}_{\mathrm{F}}(\mathbf{x};t)}_{\mathbf{y}_t}(\mathbf{x};t|\mathbf{x}';t+\mathrm{d}t)=\frac{\mathcal{F}_{\mathbf{y}_t}(\mathbf{x}';t+\mathrm{d}t|\mathbf{x};t)\widetilde{\wp}_{\mathrm{F}}(\mathbf{x};t)}{\sum_{\mathbf{x}''\in\mathcal{X}}\mathcal{F}_{\mathbf{y}_t}(\mathbf{x}';t+\mathrm{d}t|\mathbf{x}'';t)\widetilde{\wp}_{\mathrm{F}}(\mathbf{x}'';t)}. \tag*{(5)}

Importantly, one should note that the retrodictive map depends on the filtered state. This is because, in retrodiction, one generally needs a reference prior to help invert the forward dynamical map, in this case Fyt\mathcal{F}_{\mathbf{y}_t}, in the event that the map is not bijective. Generally, the reference prior is chosen to be one’s best guess of the state of the system at time tt, which in this case is the filtered state.

One can write the smoothed state in another form, which is more intuitive from a Bayesian update perspective, by introducing a secondary probability distribution, which I will refer to as the retrofiltered effect for similarity with the quantum terminology, ER(x;t):=℘(O‾t∣x;t)E_{\mathrm{R}}(\mathbf{x};t):=\wp(\overline{\mathcal{O}}_t|\mathbf{x};t) where the future record is defined as O→t={yτ:t∈[t,T)}\overrightarrow{\mathcal{O}}_t=\{\mathbf{y}_\tau:t\in[t,T)\}. This retrofiltered effect evolves as

ER(x;t)=∑x′∈XFyt(x′;t+dt∣x;t)ER(x′;t+dt),(6)E_{\mathrm{R}}(\mathbf{x};t)=\sum_{\mathbf{x}'\in\mathcal{X}}\mathcal{F}_{\mathbf{y}_t}(\mathbf{x}';t+\mathrm{d}t|\mathbf{x};t)E_{\mathrm{R}}(\mathbf{x}';t+\mathrm{d}t), \tag*{(6)}

from a final uninformative state ER(x;T)∝1E_{\mathrm{R}}(\mathbf{x};T)\propto1. With the retrofiltered effect defined, the smoothed state can also be written as

℘S(x;t)=ER(x;t)℘F(x;t)∑x′∈XER(x′;t)℘F(x′;t).(7)\wp_{\mathrm{S}}(\mathbf{x};t)=\frac{E_{\mathrm{R}}(\mathbf{x};t)\wp_{\mathrm{F}}(\mathbf{x};t)}{\sum_{\mathbf{x}'\in\mathcal{X}}E_{\mathrm{R}}(\mathbf{x}';t)\wp_{\mathrm{F}}(\mathbf{x}';t)}. \tag*{(7)}

Quantum Systems

Let us consider a Markovian open quantum system, in the strongest sense [5], where the system is described by a dd-dimensional density operator ρ(t)\rho(t). The evolution of such a system is given by a Lindblad master equation, however, for simplicity, I will just express this as a completely-positive, trace-preserving map E\mathcal{E}. As such, the unconditioned state evolves as

ρ(t+dt)=E[ρ(t)],(8)\rho(t+\mathrm{d}t)=\mathcal{E}[\rho(t)], \tag*{(8)}

with initial condition ρ0\rho_0. As with the classical case, this is the most accurate description of the quantum system in the absence of additional information.

One can gain more information about the system by continuously monitoring the environment that the system is coupled with, providing a vector of classical measurement outcomes yt\mathbf{y}_t. Note, the past, future and past-future records are defined identically to the classical case here. Such a measurement of the environment unravels the unconditioned dynamics E\mathcal{E} into a particular Kraus representation according to the measurement operators on the system M^yt\widehat{M}_{\mathbf{y}_t}, E[∙]=∑yt,iM^ytX^i∙X^i†M^yt†\mathcal{E}[\bullet]=\sum_{\mathbf{y}_t,i}\widehat{M}_{\mathbf{y}_t}\widehat{X}_i\bullet\widehat{X}_i^\dagger\widehat{M}_{\mathbf{y}_t}^\dagger, with the operator X^i\widehat{X}_i describing any remaining dissipation that is affecting the system, either due to measurement inefficiency or unmeasured environments. Repeatedly conditioning the quantum state on the measurement outcomes, one obtained the filtered quantum state, which evolves as

ρ~F(t+dt)=∑iM^ytX^iρ~F(t)X^i†M^yt†=:F~yt[ρ~F(t)],(9)\widetilde{\rho}_{\mathrm{F}}(t+\mathrm{d}t)=\sum_i\widehat{M}_{\mathbf{y}_t}\widehat{X}_i\widetilde{\rho}_{\mathrm{F}}(t)\widehat{X}_i^\dagger\widehat{M}_{\mathbf{y}_t}^\dagger=:\widetilde{\mathcal{F}}_{\mathbf{y}_t}[\widetilde{\rho}_{\mathrm{F}}(t)], \tag*{(9)}

with initial condition ρ0\rho_0. To describe the future measurement record, one can introduce an effect operator E^R(t)\widehat{E}_{\mathrm{R}}(t), defined such that ℘(O→t∣ρ)=Tr⁡[E^R(t)ρ]\wp(\overrightarrow{\mathcal{O}}_t|\rho)=\operatorname{Tr}[\widehat{E}_{\mathrm{R}}(t)\rho]. This effect operator evolves via

E^R(t)=F^yt†[E^R(t+dt)],(10)\widehat{E}_{\mathrm{R}}(t)=\widehat{\mathcal{F}}_{\mathbf{y}_t}^{\dagger}[\widehat{E}_{\mathrm{R}}(t+\mathrm{d}t)], \tag*{(10)}

from a final uninformative state E^R(T)=1^\hat{E}_{R}(T)=\hat{1}. Note, the retrofiltered effect satisfies the completeness condition

∑O→tE^R(t)=1^.\sum_{\overrightarrow{O}_{t}}\hat{E}_{R}(t)=\hat{1}.

At this point, it should be clear that the filtered state and retrofiltered effect behaves analogously to their classical counterparts. Following this logic, we can apply a quantum retrodictive map analogous to the classical Bayesian map, the Petz recovery map, to obtain a smoothed quantum state,

ρ~SPF(t)=R^γE^R(t)[ρ~FPF(t)],(11)\tilde{\rho}_{S}^{\mathrm{PF}}(t)=\hat{\mathcal{R}}_{\gamma}^{\hat{E}_{R}(t)}[\tilde{\rho}_{F}^{\mathrm{PF}}(t)], \tag*{(11)}

where the final condition is taken to be ρ~SPF(T)=ρ~F(T)\tilde{\rho}_{S}^{\mathrm{PF}}(T)=\tilde{\rho}_{F}(T) and the superscript PF is to denote that this is the Petz-Fuchs smoothed state. Here, the Petz recovery map is defined as

R^γE^t[∙]=γF^t†[F^t†[γ]−1∙F^t†[γ]−1]F^tγ,(12)\hat{\mathcal{R}}_{\gamma}^{\hat{E}_{t}}[\bullet]=\sqrt{\gamma}\hat{F}_{t}^{\dagger}\left[\sqrt{\hat{F}_{t}^{\dagger}[\gamma]^{-1}}\bullet\sqrt{\hat{F}_{t}^{\dagger}[\gamma]^{-1}}\right]\hat{F}_{t}\sqrt{\gamma}, \tag*{(12)}

where the reference prior γ\gamma is to be understood in the same way as the classical reference prior. As it turns out, one can solve this dynamical equation analytically [16] in terms of the filtered state and the retrofiltered effect, where one obtains, after renormalization,

ρSPF(t)=ρF(t)E^R(t)ρF(t)Tr⁡[E^R(t)ρF(t)].(13)\rho_{S}^{\mathrm{PF}}(t)=\frac{\sqrt{\rho_{F}(t)}\hat{E}_{R}(t)\sqrt{\rho_{F}(t)}}{\operatorname{Tr}[\hat{E}_{R}(t)\rho_{F}(t)]}. \tag*{(13)}

One will note that this form is analogous to the classical smoothed state in terms of the filtered state and retrofiltered effect.

As mentioned earlier, this is just one of a set of possible smoothed states that one can assign to the system that only depends on the filtered state and retrofiltered effect; another known case is the diagonalized smoothed state (see the Supp. Mat. of Ref. [16]). As such, it is useful to know what criteria a state must satisfy in order for it to be considered a valid smoothed quantum state. There are 4 criteria, introduced by Wiseman [25], that a smoothed quantum state should satisfy:

  1. The theory should give a single smoothed quantum state ρS\rho_{S} analogous to the classical state S\mathcal{S}, and not a pair of states, for example.

  1. The smoothed state ρS\rho_{S} should reduce to its corresponding filtered state after averaging over all possible future measurement records O→t\overrightarrow{O}_{t} given a past measurement record, i.e.,

ρF(t)=EO→t∣O←t{ρS(t)}.\rho_{F}(t)=\mathbb{E}_{\overrightarrow{O}_{t}|\overleftarrow{O}_{t}}\{\rho_{S}(t)\}.
  1. The smoothed quantum state should reduce to its classical counterpart when the initial conditions, final conditions, and dynamics of the system can all be described probabilistically in a fixed basis.

  1. The smoothed quantum state must be an S\mathcal{S}-class quantum state; that is, it must be Hermitian and positive semidefinite with unit trace.

While each of these criteria are fairly natural, there is some nuance in the third criterion. Specifically, this criteria can be interpreted in different ways depending on the context of the smoothing problem. In our case, it should be interpreted as referring to the filtered state and retrofiltered effect being co-diagonal in a fixed basis. Note, in the case of Guevara-Wiseman smoothing, this condition is strictly stronger, also applying to the true state of the system [25]. With these criteria, one can define, as I will in the next section, the set of valid smoothed states. However, as I desire similarity to the classical smoothed state, being that it only depends on the filtered state and retrofiltered effect, I will limit the set to consider quantum states that can be constructed from these operators.

The Set of Possible Smoothed States

Let us first consider the implications of Criteria 1 in the case where we are limiting ourself to the smoothed state only depending on the filtered state and retrofiltered effect. Quite simply, these two conditions together mean any possible smoothed state ρ~S(t)\tilde{\rho}_{S}(t) can be written as ρ~S(t)={f[ρF(t),E^R(t)]}\tilde{\rho}_{S}(t)=\{f[\rho_{F}(t),\hat{E}_{R}(t)]\}, where ff, at this stage, is some arbitrary function. We will now proceed to constrain the possible functions ff based on the remaining criteria to identify the class of all possible smoothed states.

Starting with Criteria 2, ρ~S(t)\tilde{\rho}_{S}(t) must satisfy

ρF(t)=∑O→t℘(O→t∣O←t)ρ~S(t)=∑O→tf~{ρF(t),E^R(t)},(14)\begin{aligned} \rho_{F}(t)=\sum_{\overrightarrow{O}_{t}}\wp(\overrightarrow{O}_{t}|\overleftarrow{O}_{t})\tilde{\rho}_{S}(t) \\ &=\sum_{\overrightarrow{O}_{t}}\tilde{f}\{\rho_{F}(t),\hat{E}_{R}(t)\}, \tag*{(14)} \end{aligned}

where, in the second line, we have used the fact that ℘(O→t∣O←t)=Tr⁡[E^R(t)ρF(t)]\wp(\overrightarrow{O}_{t}|\overleftarrow{O}_{t})=\operatorname{Tr}[\hat{E}_{R}(t)\rho_{F}(t)] is just a function of E^R(t)\hat{E}_{R}(t) and ρF(t)\rho_{F}(t) to define the unnormalized state ρ~S(t)=f~{ρF(t),E^R(t)}\tilde{\rho}_{S}(t)=\tilde{f}\{\rho_{F}(t),\hat{E}_{R}(t)\}, with Tr⁡[f~{ρF(t),E^R(t)}]=Tr⁡[E^R(t)ρF(t)]\operatorname{Tr}[\tilde{f}\{\rho_{F}(t),\hat{E}_{R}(t)\}]=\operatorname{Tr}[\hat{E}_{R}(t)\rho_{F}(t)]. This provides the first constraint on the function f~\tilde{f}.

For Criteria 3, let us consider a trivial case where the filtered state and retrofiltered effect commute, E^R(t)=1^\hat{E}_{R}(t)=\hat{1}. By Criteria 3, it must be the case that f~{ρF(t),E^R(t)}\tilde{f}\{\rho_{F}(t),\hat{E}_{R}(t)\} reproduces the classical smoothed state

ρScl=∑x℘F(x;t)℘R(x;t)℘(O→∣O←)∣ψx(t)⟩⟨ψx(t)∣,(15)\rho_{S}^{\mathrm{cl}}=\sum_{x}\frac{\wp_{F}(x;t)\wp_{R}(x;t)}{\wp(\overrightarrow{O}|\overleftarrow{O})}|\psi_{x}(t)\rangle\langle\psi_{x}(t)|, \tag*{(15)}

where ∣ψx(t)⟩|\psi_{x}(t)\rangle is the common basis shared by ρF(t)\rho_{F}(t) and E^R(t)\hat{E}_{R}(t) at time tt. Choosing ER(x;t)=1E_{R}(x;t)=1, consistent with E^R(t)=1^\hat{E}_{R}(t)=\hat{1}, the function f~\tilde{f} must also satisfy

f~{ρF(t),1^}=ρF(t),∀ρF(t).(16)\tilde{f}\{\rho_{F}(t),\hat{1}\}=\rho_{F}(t),\quad\forall\rho_{F}(t). \tag*{(16)}

Making use of the completion relation for the retrofiltered effect, ∑O→tE^R(t)=1^\sum_{\overrightarrow{O}_{t}}\hat{E}_{R}(t)=\hat{1}, combining (14) with Eq. (16), we have

∑O→tf~{ρF(t),E^R(t)}=f~{ρF(t),∑O→tE^R(t)},∀ρF(t).(17)\sum_{\overrightarrow{O}_t}\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}=\tilde{f}\left\{\rho_F(t),\sum_{\overrightarrow{O}_t}\hat{E}_R(t)\right\},\quad\forall\rho_F(t). \tag*{(17)}

This property is tremendously useful and will drastically limit the space of potential functions f~\tilde{f}. To see this, let’s consider the most general expansion possible,

f~{ρF(t),E^R(t)}=∑i,i′,i′′,…ci,i′,i′′,…Gi[ρF(t)]×Gi′′[E^R(t)]Gi′′′′[ρF(t)]…,(18)\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}=\sum_{i,i',i'',\ldots}c_{i,i',i'',\ldots}G_i[\rho_F(t)]\times G_{i'}'[\hat{E}_R(t)]G_{i''}''[\rho_F(t)]\ldots, \tag*{(18)}

where the set of functions G={Gi}\mathcal{G}=\{G_i\}, similarly G′\mathcal{G}', etc., comprises all possible functions (discontinuous, basis-dependent, etc.). Applying Eq. (17) to this form heavily limits the possible functions containing E^R\hat{E}_R to

f~{ρF(t),E^R(t)}=∑i,jci,jGi[ρF(t)]E^R(t)Hj[ρF(t)],(19)\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}=\sum_{i,j}c_{i,j}G_i[\rho_F(t)]\hat{E}_R(t)H_j[\rho_F(t)], \tag*{(19)}

where there are now only two sets of possible functions G\mathcal{G} and H={Hj}\mathcal{H}=\{H_j\} remaining.

Moving onto the final criteria, Criteria 4, our smoothed state must be Hermitian, unit trace and positive semidefinite. Beginning with the Hermiticity, it is easy to see that it must be the case that G=H\mathcal{G}=\mathcal{H}. Moreover, if there exists a Gi∈GG_i\in\mathcal{G} such that Gi≠Gi†G_i\ne G_i^\dagger, then Gi†∈GG_i^\dagger\in\mathcal{G}. Thus, without loss of generality, we can write the function as

f~{ρF(t),E^R(t)}=∑i,jci,jGi[ρF(t)]E^RGj†[ρF(t)].(20)\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}=\sum_{i,j}c_{i,j}G_i[\rho_F(t)]\hat{E}_R G_j^\dagger[\rho_F(t)]. \tag*{(20)}

At this point, we are going to work in the basis that diagonalizes the filtered state, i.e., ρF(t)→∑x℘F(x;t)∣ψx(t)⟩⟨ψx(t)∣\rho_F(t)\to\sum_x\wp_F(x;t)|\psi_x(t)\rangle\langle\psi_x(t)| and E^R(t)→E^R(ψ)(t)\hat{E}_R(t)\to\hat{E}_R^{(\psi)}(t), where the superscript (ψ)(\psi) is a reminder that this operator is in the diagonal basis of the filtered state basis {∣ψx⟩}\{|\psi_x\rangle\}. Substituting this into Eq. (20), we get

f~{ρF(t),E^R(t)}=∑x,zKx,z(t)Ex,z(ψ)(t)∣ψx(t)⟩⟨ψz(t)∣,(21)\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}=\sum_{x,z}K_{x,z}(t)E_{x,z}^{(\psi)}(t)|\psi_x(t)\rangle\langle\psi_z(t)|, \tag*{(21)}

where Kx,z(t)=∑i,jci,jGi[℘F(x;t)]Gj†[℘F(z;t)]K_{x,z}(t)=\sum_{i,j}c_{i,j}G_i[\wp_F(x;t)]G_j^\dagger[\wp_F(z;t)] and Ex,z(ψ)(t)=⟨ψx(t)∣E^R(ψ)(t)∣ψz(t)⟩E_{x,z}^{(\psi)}(t)=\langle\psi_x(t)|\hat{E}_R^{(\psi)}(t)|\psi_z(t)\rangle. Making use of Eq. (16), we see that it must be the case that Kx,x=℘F(x;t)K_{x,x}=\wp_F(x;t). We can simplify this computation by defining the Hermitian operator K^=∑x,zKx,z∣ψx(t)⟩⟨ψz(t)∣\hat{K}=\sum_{x,z}K_{x,z}|\psi_x(t)\rangle\langle\psi_z(t)|, where the smoothed state can be written as

f~{ρF(t),E^R(t)}=K^⊙E^R(ψ)(t),(22)\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}=\hat{K}\odot\hat{E}_R^{(\psi)}(t), \tag*{(22)}

with ⊙\odot being the Hadamard product (or entry-wise matrix product). Here we can see that all of the (unnormalized) functions have the same trace, i.e., Tr⁡[f~{ρF(t),E^R(t)}]=∑x℘F(x;t)Ex,x(ψ)(t)=Tr⁡[ρF(t)ER(t)]\operatorname{Tr}[\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}]=\sum_x\wp_F(x;t)E_{x,x}^{(\psi)}(t)=\operatorname{Tr}[\rho_F(t)E_R(t)]. As the normalized state is defined as f~{ρF(t),E^R(t)}/Tr⁡[ρF(t)E^R(t)]\tilde{f}\{\rho_F(t),\hat{E}_R(t)\}/\operatorname{Tr}[\rho_F(t)\hat{E}_R(t)], this means that ρ~S(t)\tilde{\rho}_S(t) will always be unit trace. For the final part of Criteria 4, that f~{ρF(t),E^R(t)}\tilde{f}\{\rho_F(t),\hat{E}_R(t)\} is positive semidefinite for any physical E^R(t)\hat{E}_R(t), this forces K^≥0\hat{K}\ge0. To see this, one only needs to consider the effect operator E^R(ψ)(t)=∑x,z∣ψx(t)⟩⟨ψz(t)∣\hat{E}_R^{(\psi)}(t)=\sum_{x,z}|\psi_x(t)\rangle\langle\psi_z(t)|.

At this point, we have identified the entire set of valid smoothed states that that only depend on ρF(t)\rho_F(t) and E^R(t)\hat{E}_R(t) to be

ρ~S(ψ)(t)=K^⊙E^R(ψ)(t)Tr⁡[ρF(t)E^R(t)],(23)\tilde{\rho}_S^{(\psi)}(t)=\frac{\hat{K}\odot\hat{E}_R^{(\psi)}(t)}{\operatorname{Tr}[\rho_F(t)\hat{E}_R(t)]}, \tag*{(23)}

where K^≥0\hat{K}\ge0 and Kx,x(t)=℘F(x;t)K_{x,x}(t)=\wp_F(x;t), where I remind the reader that the parenthesis denotes that an operator is represented in the diagonal basis of the filtered state at time tt.

Optimality of the Petz-Fuchs Smoothed State

Now that we have identified the set of valid smoothed quantum state that satisfy the smoothing criteria, we can show that the Petz-Fuchs smoothed state majorizes this set. Before I prove this statement, I will remind the reader what it means for a state ρ\rho to majorize another state σ\sigma, denoted as ρ≻σ\rho\succ\sigma. Let us assume that ρ\rho and σ\sigma are dd-dimensional quantum states. ρ\rho is said to majorize [26] σ\sigma if ∑i=0mλi(ρ)>∑i=0mλi(σ)\sum_{i=0}^{m}\lambda_i(\rho)>\sum_{i=0}^{m}\lambda_i(\sigma) for all m≤dm\le d, where λi(A)\lambda_i(A) denotes the iith eigenvalue of AA in decreasing order, i.e., λi(A)≥λi+1(A)\lambda_i(A)\ge\lambda_{i+1}(A). Importantly, it should be noted that the following two statements are equivalent [26]:

(i) ρ≻σ\rho\succ\sigma.

(ii) There exists a unital channel Φ\Phi the turns ρ\rho into σ\sigma, i.e., σ=Φ[ρ]\sigma=\Phi[\rho].

With the preliminaries out of the way, I will now show that the set of valid smoothed states can be expressed as ρ~S(t)=Φ[ρSPF(t)]\tilde{\rho}_S(t)=\Phi[\rho_S^{\mathrm{PF}}(t)] where Φ\Phi is a unital channel. First note that one obtains the Petz-Fuchs smoothed state in the diagonal basis of the filtered state when Kx,z=℘F(x;t)℘F(z;t)K_{x,z}=\sqrt{\wp_F(x;t)\wp_F(z;t)}. Let us define the correlation matrix Cx,z=Kx,z/℘F(x;t)℘F(z;t)C_{x,z}=K_{x,z}/\sqrt{\wp_F(x;t)\wp_F(z;t)}, where a correlation matrix is defined as a symmetric positive semidefinite with unit elements on the diagonal, i.e., C≥0C\ge0 and Cx,x=1C_{x,x}=1. It is easy to show that the correlation matrix Cx,z=Kx,z/℘F(x;t)℘F(z;t)C_{x,z}=K_{x,z}/\sqrt{\wp_F(x;t)\wp_F(z;t)} satisfies both of these conditions. With this correlation matrix defined we can now express any valid smoothed state as ρ~S(ψ)(t)=C⊙ρSPF(ψ)(t)\tilde{\rho}_S^{(\psi)}(t)=C\odot\rho_S^{\mathrm{PF}(\psi)}(t). Since the Hadamard product between a quantum state and a correlation matrix is equivalent to a unital channel Φ[∙]\Phi[\bullet] applied to ρSPF(ψ)(t)\rho_S^{\mathrm{PF}(\psi)}(t) [26], we have that ρ~S(ψ)(t)=Φ[ρSPF(ψ)(t)]\tilde{\rho}_S^{(\psi)}(t)=\Phi[\rho_S^{\mathrm{PF}(\psi)}(t)]. Since unitary transformations of a unital channel are result in another unital channel, i.e., UΦ[V∙V†]U†=Φ′[∙]U\Phi[V \bullet V^\dagger]U^\dagger= \Phi'[\bullet] for any unitaries UU and VV, we can move out of the diagonal basis of the filtered state to any basis. This gives that any valid smoothed state that is a function of ρF(t)\rho_F(t) and E^R(t)\hat{E}_R(t) is given by

ρ˘S(t)=Φ[ρSPF(t)],(24)\breve{\rho}_S(t) = \Phi[\rho_S^{\mathrm{PF}}(t)], \tag*{(24)}

where Φ\Phi is a unital channel, which is equivalent to ρSPF(t)≻ρ˘S(t)\rho_S^{\mathrm{PF}}(t) \succ\breve{\rho}_S(t).

An immediate consequence of the fact that the Petz-Fuchs smoothed state majorizes all valid smoothed states means that any Schur concave function of ρSPF\rho_S^{\mathrm{PF}} will be minimized, like the von Neumann, Rényi entropy. Similarly, any Schur convex will be maximized, like the purity. This means that, the Petz-Fuchs state will be the purest smoothed assignment that one can attribute to an open Markovian quantum system given a past and future measurement record. Another way to understand this is that the Petz-Fuchs smoothed state uses maximally utilises the information contained in the past-future measurement record to reconstruct the purest description of the system. It should be noted, however, that on a individual trajectory level, this does not mean that the Petz-Fuchs smoothed state (or any valid smoothed state for that matter) will be purer than the filtered state (See Ref. [16] for an example). Nevertheless, on average, the smoothed state always performs better than the filtered state [16].

Conclusion

In this paper, I have characterized the complete set of possible valid smoothed states that only depend on the filtered state and retrofiltered effect, similar to the classical case. Subsequently, I have shown that the Petz-Fuchs smoothed state majorizes all other smoothed states in this set, showing that it is the purest possible smoothed reconstruction of the system state. As this is the purest assignment possible to an open quantum system given past and future measurement information, it would be interesting to see whether this knowledge of the system can help in understanding particular physical phenomenon better. As an example, could one learn better the state of a light field scattered of a atom coupled to a bosonic bath.

Acknowledgments

I would like to thank Mingxuan Liu, Samyak P. Prasad, Kian Hwee Lim, Nathan Shettell and Howard M. Wiseman for discussions on the initial idea of this paper. This project is supported by the National Research Foundation, Singapore through the National Quantum Office, hosted in A*STAR, under its Centre for Quantum Technologies Funding Initiative (S24Q2d0009), and the Plan France 2030 through the projects NISQ2LSQ (Grant ANR-22-PETQ-0006), OQuLuS (Grant ANR-22-PETQ-0013), and OECQ through BPI France.

References

References

  1. [1]G. A. Einicke, Smoothing, filtering and prediction: Estimating the past, present and future (InTech, Rijeka Croatia, 2012).DOI
  2. [2]S. Särkkä, Bayesian filtering and smoothing, Vol. 3 (Cambridge University Press, Cambridge, England, 2013).
  3. [3]A. H. Jazwinski, Stochastic processes and filtering theory (Dover Publications, New York, 2007).
  4. [4]R. G. Brown and P. Y. C. Hwang, Introduction to random signals and applied Kalman filtering, 4th ed. (Wiley, New York, 2012).
  5. [5]L. Li, M. J. W. Hall, and H. M. Wiseman, “Concepts of quantum non-Markovianity: A hierarchy,” Phys. Rep. 759, 1–51 (2018).DOI
  6. [6]H. J. Carmichael, An open systems approach to quantum optics (Springer, Berlin, 1993).DOI
  7. [7]V. P. Belavkin, “Quantum continual measurements and a posteriori collapse on CCR,” Commun. Math. Phys. 146, 611–635 (1992).DOI
  8. [8]V. P. Belavkin, “Measurement, filtering and control in quantum open dynamical systems,” Rep. Math. Phys. 43, A405 (1999).DOI
  9. [9]H. M. Wiseman and G. J. Milburn, Quantum measurement and control (Cambridge University Press, Cambridge, England, 2010).DOI
  10. [10]S. Watanabe, “Symmetry of physical laws. Part III. Prediction and retrodiction,” Rev. Mod. Phys. 27, 179 (1955).DOI
  11. [11]Y. Aharonov, P. G. Bergmann, and J. L. Lebowitz, “Time symmetry in the Quantum process of measurement,” Phys. Rev. 134, B1410–B1416 (1964).
  12. [12]M. Tsang, “Optimal waveform estimation for classical and quantum systems via time-symmetric smoothing,” Phys. Rev. A 80, 033840 (2009).DOI
  13. [13]A. Chantrasri, J. Dressel, and A. N. Jordan, “Action principle for continuous quantum measurement,” Phys. Rev. A 88, 042110 (2013).
  14. [14]S. Gammelmark, B. Julsgaard, and K. Mølmer, “Past quantum states as a monitored system,” Phys. Rev. Lett. 111, 160401 (2013).pubmed.ncbi.nlm.nih.gov/24182235
  15. [15]A. Chantrasri, I. Guevara, K. T. Laverick, and H. M. Wiseman, “Unifying theory of quantum state estimation using past and future information,” Phys. Rep. 930, 1–40 (2021).
  16. [16]M. Liu, V. Scarani, A. Auffèves, and K. T. Laverick, “Retrodictive approach to quantum state smoothing,” Phys. Rev. A 112, L030203 (2025).DOI
  17. [17]D. Petz, “Sufficient subalgebras and the relative entropy of states of a von neumann algebra,” Commun. Math. Phys. 105, 123–131 (1986).DOI
  18. [18]D. Petz, “Sufficiency of channels over von Neumann algebra,” Q. J. Math. 39, 97–108 (1988).DOI
  19. [19]S. K. Manikandan and A. N. Jordan, “Time reversal symmetry of generalized quantum measurements with past and future boundary conditions,” Quantum Stud.: Math. Found. 6, 241–268 (2019).DOI
  20. [20]J. Sau,ce and M. Scandi, “State retrieval beyond Bayes’ retrodiction,” Quantum 7, 990 (2023).
  21. [21]A. J. Parzygnat and J. Fullwood, “From time-reversal symmetry to quantum bayes’ rules,” PRX Quantum 4, 020334 (2023).DOI
  22. [22]A. J. Parzygnat and F. Buscemi, “Axioms for retrodiction: achieving time-reversal symmetry with a prior,” Quantum 7, 1013 (2023).DOI
  23. [23]I. Guevara and H. M. Wiseman, “Quantum state smoothing,” Phys. Rev. Lett. 115, 180407 (2015).
  24. [24]F. Buscemi and V. Scarani, “Fluctuation theorems from bayesian retrodiction,” Phys. Rev. E 103, 052111 (2021).DOI
  25. [25]K. T. Laverick, A. Chantasri, and H. M. Wiseman, “General criteria for quantum state smoothing with necessary and sufficient criteria for linear Gaussian quantum systems,” Quantum Stud.: Math. Found. 8, 37–50 (2021).DOI
  26. [26]J. Watrous et al., The theory of quantum information, Vol. 1 (Cambridge university press Cambridge, 2018).DOI

Paper details

vlnc:W_mupj42nd
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents