Introduction
CV photonic systems are a leading platform for quantum information, quantum communication, quantum computing, sensing, and simulation [1 , 2 , 3 , 4 , 5 ] . In these systems, information is encoded in the quadratures of optical modes, which can be prepared as coherent or squeezed states, transformed by Gaussian operations, and measured using homodyne detection [6 , 7 , 8 , 9 ] . In many CV photonic systems, the relevant dynamics is well described in the Gaussian regime, where the state is characterized by its displacement vector and covariance matrix. Noise analysis in this setting is commonly formulated in terms of Gaussian channels, including photon loss, thermal noise and additive Gaussian noise [10 , 11 , 12 ] . These models are powerful because Gaussian noise can be represented compactly via drift and diffusion matrices that govern the evolution of the first and second moments [9 , 13 , 14 , 15 , 16 , 17 ] .
Most standard noise models for CV photonic systems assume a memoryless environment and describe CP-divisible evolution [18 , 19 ] . A physical density operator is positive semidefinite and has unit trace, ensuring non-negative measurement probabilities that sum to one. A trace-preserving quantum dynamical map Λ \Lambda Λ is completely positive (CP) if it maps every physical density operator, including when applied to one part of an entangled composite system while leaving the other part unchanged. For example, attenuating both quadratures of the vacuum without adding the required noise would violate the Heisenberg uncertainty relation. Such a prediction indicates an inconsistent model, not a realizable deterministic process. The evolution is CP-divisible if every intermediate map is also CP and trace-preserving, namely [20 , 21 ]
Λ ( t , 0 ) = Λ ( t , s ) Λ ( s , 0 ) \Lambda(t,0) = \Lambda(t,s)\Lambda(s,0) Λ ( t , 0 ) = Λ ( t , s ) Λ ( s , 0 )
Λ ( t , s ) is CPTP for all t ≥ s ≥ 0 (1) \Lambda(t,s)\ \text{is CPTP for all } t \ge s \ge0 \tag*{(1)} Λ ( t , s ) is CPTP for all t ≥ s ≥ 0 (1)
where CPTP denotes Completely Positive and Trace Preserving dynamics.
Realistic hardware, however, can contain temporally correlated Gaussian noise sources, including resonator memory, delay-line feedback, thermal drift, pump-amplitude and pump-phase fluctuations, hidden spectral modes, correlated dissipation and detector-induced correlations [11 , 12 ] . Non-Markovian open quantum system theory provides a natural language for describing such memory effects, especially through time-local Lindblad-like master equations with time-dependent coefficients. If the instantaneous generator violates the condition for CP-divisibility while the total map remains CP, the dynamics is non-Markovian in the divisibility sense [18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 ] . Physically, the environment can retain information about the system’s earlier evolution through system-environment correlations. Consequently, an intermediate map need not represent a physical operation on an arbitrary system state independently of these correlations. Its failure to be CP does not make the actual evolving state unphysical: the total map from the initial time remains CP. Throughout this work, we focus on non-Markovianity characterized by violations of CP-divisibility, which provides a generator-level diagnostic complementary to broader descriptions based on multi-time correlations [26 ] . Accordingly, CP-divisibility alone does not exclude such correlations.
In CV Gaussian systems, memory effects have been investigated through several complementary diagnostics. Fidelity-based distinguishability measures adapt the information-backflow picture to Gaussian dynamical maps and have been applied to damping and quantum Brownian-motion channels. Divisibility-based approaches provide criteria for non-Markovianity directly at the level of Gaussian channels by testing the CP of intermediate maps, while related Gaussian-process hierarchies distinguish Markovian, weakly non-Markovian, and strongly non-Markovian regimes. Beyond channel-level diagnostics, structured reservoirs and fluctuating environments can lead to non-monotonic behavior of CV resources such as two-mode entanglement, logarithmic negativity, quantum discord, intensity correlations, Gaussian interferometric power and revivals of quantum correlations [27 , 28 , 29 , 30 , 31 , 32 , 33 ] . These results motivate a generator-level tomographic reconstruction strategy in which memory is identified not only through state-dependent witnesses, but also through the reconstructed time-local drift and diffusion matrices.
Tomographic methods provide a well-established framework for reconstructing unknown quantum states and processes from measurement data [34 , 35 ] . In CV systems, homodyne tomography provides direct access to quadrature statistics and enables the reconstruction of Gaussian states, covariance matrices, and Gaussian channels [36 , 37 , 38 , 39 ] . Maximum-likelihood and constrained estimation methods provide statistical tools for quantum tomography [40 , 41 , 42 , 43 ] , including state estimation with additive Gaussian measurement noise [44 ] . Related applications of likelihood-based inference include amplitude estimation on noisy quantum computers [45 ] and noise estimation for quantum error correction [46 ] . For non-Markovian dynamics, process tomography can characterize multi-time correlations [47 ] . Most closely related to our methodology, Varona et al. [48 ] developed Lindblad-like quantum tomography to reconstruct time-local master equations, including temporarily negative decay rates, through constrained maximum-likelihood estimation for qubit dynamics. Aguiar et al. [49 ] introduced regression-based quantum Liouville tomography and applied it to superconducting-qubit data, while Reddy [50 ] investigated structure-preserving reconstruction of non-Markovian time-local generators using basis expansions and neural networks. Related work by Guan et al. [51 ] examined the identifiability of drift and diffusion matrices from temporal distributions of classical linear stochastic systems. Complementarily, Torre et al. [30 ] established CP-divisibility criteria for Gaussian quantum channels, with further characterization of Gaussian non-Markovian dynamics and their classification [33 , 52 ] . Renault et al. [53 ] experimentally characterized environmental memory and non-Markovian effects using squeezing and entanglement correlations in a CV optical platform.
Although existing Gaussian-channel criteria provide a rigorous characterization of CP-divisibility, their integration with data-driven reconstruction of time-local Gaussian generators remains an important methodological problem. Conversely, likelihood-based generator reconstruction has primarily been developed for finite-dimensional quantum systems, without explicitly addressing the drift-diffusion structure and physicality constraints of Gaussian CV dynamics. The central challenge is therefore to reconstruct time-local Gaussian generators from time-resolved homodyne statistics while enforcing CP of the finite-time evolution and independently assessing the CP-divisibility of the reconstructed dynamics.
In this work, we address this challenge by developing a theoretical and simulation-based framework for tomographic reconstruction of non-Markovian Gaussian noise in CV photonic systems. The unknown dynamics is represented by a time-local Gaussian generator with drift matrix, diffusion matrix, and displacement drift. These quantities determine the evolution of the displacement vector and covariance matrix, which completely specify each output Gaussian state and hence its homodyne probability distributions. The dynamics considered here preserves Gaussian states in both the Markovian and non-Markovian regimes. We estimate the generator parameters using a regularized maximum-likelihood procedure constrained by CP of the total Gaussian channel. Non-Markovianity is then diagnosed by testing whether the reconstructed infinitesimal dynamics violates CP-divisibility. Figure 1 summarizes the physical picture considered in this work: a CV photonic circuit is treated as the accessible system, while memory-bearing environmental degrees of freedom can produce time-correlated loss, diffusion, and correlated noise that are reconstructed from homodyne data.
Figure 1. Schematic of a CV photonic circuit subject to possible memory-bearing noise sources. Resonator modes, delay-line feedback, temporally correlated thermal drift, pump fluctuations and hidden spectral modes can generate time-correlated system noise, whereas detector memory can introduce correlations in the measured homodyne record. Auxiliary pseudomodes represent environmental memory and possible information backflow. These mechanisms produce non-Markovian reduced dynamics only when the reconstructed time-local generator violates CP-divisibility while the finite-time Gaussian channel remains CP. Homodyne detection provides the quadrature data used for generator-level tomography.
The paper is organized as follows. Section II introduces the theoretical framework for describing Gaussian Markovian and non-Markovian dynamics in CV photonic systems. Section III connects this framework to physical noise sources and memory mechanisms. Section IV presents the homodyne-based tomographic reconstruction method. Section V develops the diagnostic and model-selection procedures used to identify non-Markovian memory. Section VI provides a numerical validation of the proposed approach. Finally, Section VII summarizes the main results and outlines future directions.
Theoretical Framework
Quadrature representation of CV photonic systems
We consider an m m m -mode bosonic system with annihilation and creation operators 1 ^{1} 1
a = ( a 1 , … , a m ) T , [ a j , a k † ] = δ j k (2) \mathbf{a} = (a_1,\ldots,a_m)^{T}, \qquad[a_j,a_k^\dagger] = \delta_{jk} \tag*{(2)} a = ( a 1 , … , a m ) T , [ a j , a k † ] = δ j k (2)
Define quadrature operators
q j = a j + a j † 2 , p j = a j − a j † i 2 (3) q_j = \frac{a_j+a_j^\dagger}{\sqrt{2}}, \qquad p_j = \frac{a_j-a_j^\dagger}{i\sqrt{2}} \tag*{(3)} q j = 2 a j + a j † , p j = i 2 a j − a j † (3)
and collect them into the quadrature vector
R = ( q 1 , p 1 , … , q m , p m ) T (4) \mathbf{R} = (q_1,p_1,\ldots,q_m,p_m)^{T} \tag*{(4)} R = ( q 1 , p 1 , … , q m , p m ) T (4)
They obey
[ R j , R k ] = i Ω j k (5) [R_j,R_k] = i\Omega_{jk} \tag*{(5)} [ R j , R k ] = i Ω j k (5)
where the symplectic matrix is defined by
Ω = ⨁ j = 1 m ( 0 1 − 1 0 ) (6) \Omega= \bigoplus_{j=1}^{m}
\begin{pmatrix}
0 & 1 \\
-1 & 0
\end{pmatrix} \tag*{(6)} Ω = j = 1 ⨁ m ( 0 − 1 1 0 ) (6)
For Gaussian states, the first moment (displacement) and covariance matrix are defined respectively
d j ( t ) = ⟨ R j ⟩ t (7) d_j(t) = \langle R_j\rangle_t \tag*{(7)} d j ( t ) = ⟨ R j ⟩ t (7)
V j k ( t ) = 1 2 ⟨ { R j − d j ( t ) , R k − d k ( t ) } ⟩ t (8) V_{jk}(t) = \frac{1}{2}\left\langle\{R_j-d_j(t),R_k-d_k(t)\}\right\rangle_t \tag*{(8)} V j k ( t ) = 2 1 ⟨ { R j − d j ( t ) , R k − d k ( t )} ⟩ t (8)
where the covariance matrix is a 2 m × 2 m 2m \times2m 2 m × 2 m real and symmetric matrix which must satisfy the uncertainty principle [54 ]
V ( t ) + i 2 Ω ≥ 0 (9) V(t) + \frac{i}{2}\Omega\ge0 \tag*{(9)} V ( t ) + 2 i Ω ≥ 0 (9)
The coherent Gaussian dynamics of such a photonic system is generated by a quadratic Hamiltonian
H ( t ) = 1 2 R T G ( t ) R (10) H(t) = \frac{1}{2}R^{T}G(t)R \tag*{(10)} H ( t ) = 2 1 R T G ( t ) R (10)
where G ( t ) = G T ( t ) ∈ R 2 m × 2 m G(t) = G^{T}(t) \in\mathbb{R}^{2m\times2m} G ( t ) = G T ( t ) ∈ R 2 m × 2 m is the real symmetric matrix generating beam splitters, phase shifters, squeezers, and mode mixing [6 , 8 ] . In addition to the quadratic Hamiltonian in Eq. (10) , coherent driving is included through the linear term h ( t ) T R h(t)^{T}R h ( t ) T R , where h ( t ) h(t) h ( t ) is a real driving vector, generating the displacement drift u ( t ) = Ω h ( t ) u(t) = \Omega h(t) u ( t ) = Ω h ( t ) introduced below.
Markovian Lindblad master equation
The Markovian evolution of a photonic system coupled to a memoryless environment is described by the Lindblad master equation [55 , 56 , 57 ]
d ρ ( t ) d t = L ρ ( t ) = − i [ H , ρ ( t ) ] + ∑ j γ j D [ L j ] ρ ( t ) = − i [ H , ρ ( t ) ] + ∑ j γ j ( L j ρ ( t ) L j † − 1 2 { L j † L j , ρ ( t ) } ) (11) \begin{aligned}
\frac{d\rho(t)}{dt}=\mathcal{L}\rho(t)=-i[H,\rho(t)]+\sum_j\gamma_j\mathcal{D}[L_j]\rho(t) \\
&=-i[H,\rho(t)]+\sum_j\gamma_j\left(L_j\rho(t)L_j^\dagger-\frac{1}{2}\left\{L_j^\dagger L_j,\rho(t)\right\}\right) \tag*{(11)}
\end{aligned} d t d ρ ( t ) = L ρ ( t ) = − i [ H , ρ ( t )] + j ∑ γ j D [ L j ] ρ ( t ) = − i [ H , ρ ( t )] + j ∑ γ j ( L j ρ ( t ) L j † − 2 1 { L j † L j , ρ ( t ) } ) (11)
Here ρ \rho ρ is the reduced density matrix of the CV photonic system, H H H is the system Hamiltonian, L j L_j L j are Lindblad jump operators, γ j ≥ 0 \gamma_j\ge0 γ j ≥ 0 are relaxation rates or decay rates corresponding to each dissipative channel, D [ L j ] \mathcal{D}[L_j] D [ L j ] is the Lindblad dissipator associated with L j L_j L j , and L \mathcal{L} L denotes the full Lindblad generator acting on ρ ( t ) \rho(t) ρ ( t ) . The Markovian model assumes that the environment has no memory and the instantaneous coefficients are time-independent or externally prescribed non-negative functions.
For standard CV photonic hardware, the most common single-mode linear Lindblad operators describe photon loss and incoherent gain [11 , 12 , 18 ]
L j l o s s = a j , L j g a i n = a j † (12) L_j^{\mathrm{loss}}=a_j,\qquad L_j^{\mathrm{gain}}=a_j^\dagger \tag*{(12)} L j loss = a j , L j gain = a j † (12)
More generally, Gaussian-preserving correlated dissipation and gain can be described by Lindblad operators that are linear combinations of creation and annihilation operators
L μ = ∑ j = 1 m ( c μ j a j + d μ j a j † ) μ = 1 , … , N L (13) L_\mu=\sum_{j=1}^{m}\left(c_{\mu j}a_j+d_{\mu j}a_j^\dagger\right)\qquad\mu=1,\ldots,N_L \tag*{(13)} L μ = j = 1 ∑ m ( c μ j a j + d μ j a j † ) μ = 1 , … , N L (13)
where the coefficients c μ j c_{\mu j} c μ j and d μ j d_{\mu j} d μ j determine which photonic modes j j j are coupled to the dissipative environmental channel μ \mu μ . The special case d μ j = 0 d_{\mu j}=0 d μ j = 0 corresponds to correlated photon loss, while nonzero d μ j d_{\mu j} d μ j also allows gain- or squeezing-type Gaussian noise. Equivalently, correlated photon loss may be written in Kossakowski form as introduced below in Eq. (52) . Other important photonic noise mechanisms are generated by nonlinear Lindblad operators, for example
L j d e p h = a j † a j , L j a b s = a j 2 (14) L_j^{\mathrm{deph}}=a_j^\dagger a_j,\qquad L_j^{\mathrm{abs}}=a_j^2 \tag*{(14)} L j deph = a j † a j , L j abs = a j 2 (14)
which describe phase diffusion and two-photon absorption, respectively. They are useful physical noise models, but they generally go beyond the strictly Gaussian-preserving drift-diffusion formulation used below, unless treated perturbatively or approximated within an effective Gaussian model.
Lindblad-like non-Markovian master equation
Some non-Markovian dynamics can be described by a time-local master equation when a well-defined time-local generator exists. Such equations can be obtained, for example, from the time-convolutionless projection-operator formalism, and may be regarded as time-dependent generalizations of the usual Markovian Lindblad equation [57 ] . In this description, the Hamiltonian, dissipative operators, and decay rates can depend explicitly on time. If the instantaneous dissipative rates remain non-negative at all times, the generator has a Lindblad form at each instant and the corresponding dynamics is CP-divisible [18 , 19 , 22 , 23 ] . Here, CP is stronger than positivity of the reduced system alone: it guarantees that the evolution remains physically valid even when the system is initially correlated with an arbitrary ancilla [18 ] . In contrast, if some instantaneous rates become temporarily negative while the overall dynamical map remains CP, the evolution is no longer Markovian in the CP-divisibility sense. This motivates a Lindblad-like time-local master equation in which non-Markovian dynamics can be described by temporarily negative canonical decay rates, while the overall dynamical map remains CP [22 , 23 , 58 ] . The canonical rates, namely the eigenvalues of the dissipative coefficient matrix, are obtained by diagonalizing this matrix in an orthonormal basis of traceless operators; negative coefficients in an arbitrary Lindblad representation do not by themselves establish CP-indivisibility. For the Gaussian dynamics considered here, we use the matrix criterion M C P ( t ) \mathcal{M}_{\mathrm{CP}}(t) M CP ( t ) introduced below.
Here, we model the reduced dynamics of the photonic modes using an effective time-local master equation [23 ]
d ρ ( t ) d t = K ( t ) ρ ( t ) (15) \frac{d\rho(t)}{dt}=\mathcal{K}(t)\rho(t) \tag*{(15)} d t d ρ ( t ) = K ( t ) ρ ( t ) (15)
whose generator K ( t ) \mathcal{K}(t) K ( t ) contains time-dependent Hamiltonian and dissipative contributions
K ( t ) ρ ( t ) = − i [ H ( t ) , ρ ( t ) ] + ∑ μ , ν C μ ν ( t ) ( F μ ρ ( t ) F ν † − 1 2 { F ν † F μ , ρ ( t ) } ) (16) \mathcal{K}(t)\rho(t)=-i[H(t),\rho(t)]
+\sum_{\mu,\nu}C_{\mu\nu}(t)\left(F_\mu\rho(t)F_\nu^\dagger-\frac{1}{2}\left\{F_\nu^\dagger F_\mu,\rho(t)\right\}\right) \tag*{(16)} K ( t ) ρ ( t ) = − i [ H ( t ) , ρ ( t )] + μ , ν ∑ C μν ( t ) ( F μ ρ ( t ) F ν † − 2 1 { F ν † F μ , ρ ( t ) } ) (16)
where { F μ } \{F_\mu\} { F μ } is a fixed, linearly independent family of noise operators and C μ ν C_{\mu\nu} C μν are the entries of the time-dependent Kossakowski-like matrix C ( t ) = C † ( t ) C(t)=C^\dagger(t) C ( t ) = C † ( t ) that specifies how strongly each pair of noise operators ( F μ , F ν ) (F_\mu,F_\nu) ( F μ , F ν ) contributes to the dissipative dynamics [59 , 60 ] .
C ( t ) ≥ 0 ∀ t (17) C(t)\ge0\qquad\forall t \tag*{(17)} C ( t ) ≥ 0 ∀ t (17)
then the dynamics is CP-divisible and Markovian in the divisibility sense. If C ( t ) C(t) C ( t ) has negative eigenvalues at some times, the generator is not of standard Lindblad form at those times. Nevertheless, the total map
ρ ( t ) = Λ ( t , 0 ) ρ ( 0 ) (18) \rho(t)=\Lambda(t,0)\rho(0) \tag*{(18)} ρ ( t ) = Λ ( t , 0 ) ρ ( 0 ) (18)
may remain CP. Thus, non-Markovianity is indicated by
∃ t , λ min [ C ( t ) ] < 0 (19) \exists t,\qquad\lambda_{\min}[C(t)]<0 \tag*{(19)} ∃ t , λ m i n [ C ( t )] < 0 (19)
where λ min [ C ( t ) ] \lambda_{\min}[C(t)] λ m i n [ C ( t )] denotes the smallest eigenvalue of the matrix C ( t ) C(t) C ( t ) . A negative value of λ min [ C ( t ) ] \lambda_{\min}[C(t)] λ m i n [ C ( t )] indicates that C ( t ) C(t) C ( t ) is not positive semidefinite, which signals a breakdown of CP-divisibility and is used as a witness of non-Markovianity of the quantum evolution. Therefore, we call Eq. (15) a Lindblad-like non-Markovian master equation [22 , 23 ] .
Gaussian-preserving Lindblad-like dynamics
We focus on Gaussian-preserving dynamics, a regime relevant to experiments such as squeezed-state-enhanced interferometry, where squeezed light undergoes linear optical transformations and Gaussian noise before homodyne detection. In this regime, the state is fully characterized by its displacement vector and covariance matrix, allowing us to describe the evolution in phase space rather than through the full density operator.
In the phase-space representation, the quantum state is described by its Wigner function W ( x , t ) W(x,t) W ( x , t ) , where the classical phase-space coordinate is
x = ( q 1 , p 1 , … , q m , p m ) T ∈ R 2 m (20) x = (q_1,p_1,\ldots,q_m,p_m)^{T} \in\mathbb{R}^{2m} \tag*{(20)} x = ( q 1 , p 1 , … , q m , p m ) T ∈ R 2 m (20)
where q j , p j q_j,p_j q j , p j are real phase-space coordinates corresponding to the quadrature operators. For the Gaussian-preserving models considered below, we assume that the Hamiltonian is at most quadratic in the quadrature operators and that the Lindblad noise operators are at most linear in the quadratures. Under these conditions, an initially Gaussian Wigner function remains Gaussian throughout the evolution. Therefore, the dynamics can be described either by a linear Fokker-Planck equation for the Wigner function W ( x , t ) W(x,t) W ( x , t ) , or equivalently by closed evolution equations for the first moment d ( t ) d(t) d ( t ) and covariance matrix V ( t ) V(t) V ( t ) .
In phase space, a time-local Wigner evolution equation of Fokker-Planck form governs how the quasiprobability distribution W ( x , t ) W(x,t) W ( x , t ) evolves under drift and diffusion as [15 , 61 , 62 ]
∂ W ( x , t ) ∂ t = − ∇ x T [ A ( t ) x + u ( t ) ] W ( x , t ) + 1 2 ∇ x T D ( t ) ∇ x W ( x , t ) (21) \frac{\partial W(x,t)}{\partial t}
= -\nabla_x^{T}\left[A(t)x+u(t)\right]W(x,t)
+ \frac{1}{2}\nabla_x^{T}D(t)\nabla_x W(x,t) \tag*{(21)} ∂ t ∂ W ( x , t ) = − ∇ x T [ A ( t ) x + u ( t ) ] W ( x , t ) + 2 1 ∇ x T D ( t ) ∇ x W ( x , t ) (21)
where A ( t ) A(t) A ( t ) is the drift matrix describing rotations, squeezing, mode mixing, and attenuation due to photon loss; u ( t ) u(t) u ( t ) is the displacement drift vector describing coherent driving or other deterministic shifts of the mean field; and D ( t ) D(t) D ( t ) is the diffusion matrix describing the effective noise contribution to covariance evolution, including vacuum noise, thermal noise, and technical fluctuations within the Gaussian approximation. In non-Markovian regimes, D ( t ) D(t) D ( t ) need not be positive semidefinite at every instant. The equation then retains Fokker-Planck form but need not define a classical diffusion process. Physicality is imposed through complete positivity of the accumulated Gaussian channel, while intermediate propagators need not preserve positivity for arbitrary input states.
Since the dynamics is Gaussian-preserving, it is fully characterized by the first moment d ( t ) d(t) d ( t ) and covariance matrix V ( t ) V(t) V ( t ) . Taking the first and second moments of the Fokker-Planck equation gives [15 , 17 , 61 , 62 ]
d ˙ ( t ) = d d t ∫ x W ( x , t ) d x = ∫ x ∂ t W ( x , t ) d x = ∫ [ A ( t ) x + u ( t ) ] W ( x , t ) d x = A ( t ) d ( t ) + u ( t ) (22) \begin{aligned}
\dot{d}(t) &= \frac{d}{dt}\int xW(x,t)\,dx \\
&= \int x\,\partial_tW(x,t)\,dx \\
&= \int[A(t)x+u(t)]W(x,t)\,dx \\
&= A(t)d(t)+u(t) \tag*{(22)}
\end{aligned} d ˙ ( t ) = d t d ∫ x W ( x , t ) d x = ∫ x ∂ t W ( x , t ) d x = ∫ [ A ( t ) x + u ( t )] W ( x , t ) d x = A ( t ) d ( t ) + u ( t ) (22)
and
V ˙ ( t ) = d d t ( ⟨ x x T ⟩ − d d T ) = A ⟨ x x T ⟩ + ⟨ x x T ⟩ A T + u d T + d u T − d d T − d d T = A ( ⟨ x x T ⟩ − d d T ) + ( ⟨ x x T ⟩ − d d T ) A T + D = A ( t ) V ( t ) + V ( t ) A T ( t ) + D ( t ) (23) \begin{aligned}
\dot{V}(t) &= \frac{d}{dt}\left(\langle xx^{T}\rangle-dd^{T}\right) \\
&= A\langle xx^{T}\rangle+\langle xx^{T}\rangle A^{T}+ud^{T}+du^{T} \\
&\quad-dd^{T}-dd^{T} \\
&= A\left(\langle xx^{T}\rangle-dd^{T}\right)+\left(\langle xx^{T}\rangle-dd^{T}\right)A^{T}+D \\
&= A(t)V(t)+V(t)A^{T}(t)+D(t) \tag*{(23)}
\end{aligned} V ˙ ( t ) = d t d ( ⟨ x x T ⟩ − d d T ) = A ⟨ x x T ⟩ + ⟨ x x T ⟩ A T + u d T + d u T − d d T − d d T = A ( ⟨ x x T ⟩ − d d T ) + ( ⟨ x x T ⟩ − d d T ) A T + D = A ( t ) V ( t ) + V ( t ) A T ( t ) + D ( t ) (23)
Therefore, the formal solutions of Eqs. (22) and (23) are
d ( t ) = X ( t ) d ( 0 ) + ∫ 0 t X ( t , s ) u ( s ) d s (24) d(t)=X(t)d(0)+\int_{0}^{t}X(t,s)u(s)\,ds \tag*{(24)} d ( t ) = X ( t ) d ( 0 ) + ∫ 0 t X ( t , s ) u ( s ) d s (24)
V ( t ) = X ( t ) V ( 0 ) X T ( t ) + ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (25) V(t)=X(t)V(0)X^{T}(t)+\int_{0}^{t}X(t,s)D(s)X^{T}(t,s)\,ds \tag*{(25)} V ( t ) = X ( t ) V ( 0 ) X T ( t ) + ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (25)
where X ( t ) : = X ( t , 0 ) X(t):=X(t,0) X ( t ) := X ( t , 0 ) describes the deterministic linear transformation and we have
d d t X ( t , s ) = A ( t ) X ( t , s ) , X ( s , s ) = I (26) \frac{d}{dt}X(t,s)=A(t)X(t,s), \qquad X(s,s)=I \tag*{(26)} d t d X ( t , s ) = A ( t ) X ( t , s ) , X ( s , s ) = I (26)
A Gaussian channel is physical and CP if and only if the following condition is satisfied [14 , 15 ]
Y ( t ) + i 2 Ω − i 2 X ( t ) Ω X T ( t ) ≥ 0 (27) Y(t)+\frac{i}{2}\Omega-\frac{i}{2}X(t)\Omega X^{T}(t)\geq0 \tag*{(27)} Y ( t ) + 2 i Ω − 2 i X ( t ) Ω X T ( t ) ≥ 0 (27)
where Ω \Omega Ω is the symplectic matrix defined in Eq. (6) and
Y ( t ) = ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (28) Y(t)=\int_{0}^{t}X(t,s)D(s)X^{T}(t,s)\,ds \tag*{(28)} Y ( t ) = ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (28)
and so
V ( t ) = X ( t ) V ( 0 ) X T ( t ) + Y ( t ) (29) V(t)=X(t)V(0)X^{T}(t)+Y(t) \tag*{(29)} V ( t ) = X ( t ) V ( 0 ) X T ( t ) + Y ( t ) (29)
Here, Y ( t ) Y(t) Y ( t ) is the accumulated Gaussian noise matrix generated by the diffusion term D ( t ) D(t) D ( t ) . Thus, the pair X ( t ) , Y ( t ) X(t),Y(t) X ( t ) , Y ( t ) defines the Gaussian channel acting on the covariance matrix. While Eq. (27) ensures CP of the total Gaussian channel from the initial time to t t t , Markovianity in the CP-divisibility sense requires a stronger local condition. Namely, every infinitesimal intermediate map from t t t to t + d t t+dt t + d t must itself be CP, which gives (Appendix A)
M C P ( t ) = D ( t ) − i 2 [ A ( t ) Ω + Ω A T ( t ) ] ≥ 0 (30) M_{\mathrm{CP}}(t)=D(t)-\frac{i}{2}\left[A(t)\Omega+\Omega A^{T}(t)\right]\ge0 \tag*{(30)} M CP ( t ) = D ( t ) − 2 i [ A ( t ) Ω + Ω A T ( t ) ] ≥ 0 (30)
Here M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) is the local Gaussian CP matrix for the infinitesimal intermediate map from t t t to t + d t t+dt t + d t . Positive semidefiniteness of M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) is equivalent to CP of this infinitesimal Gaussian map. If M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) has a negative eigenvalue, the intermediate map is not CP, so the dynamics is not CP-divisible. Therefore, violation of Eq. (30) indicates a breakdown of CP-divisibility and is used here as a witness of non-Markovianity [21 , 23 ] . The equivalence between M C P ( t ) ≥ 0 M_{\mathrm{CP}}(t)\ge0 M CP ( t ) ≥ 0 throughout an interval and CP-divisibility on that interval assumes that A ( t ) A(t) A ( t ) is real, D ( t ) D(t) D ( t ) is real symmetric, and both matrices are continuous, so that the intermediate Gaussian propagator is well defined. For such a regular drift matrix
det X ( t , s ) = exp [ ∫ s t Tr A ( τ ) d τ ] ≠ 0. \det X(t,s)=\exp\left[\int_s^t \operatorname{Tr} A(\tau)\,\mathrm{d}\tau\right]\ne0. det X ( t , s ) = exp [ ∫ s t Tr A ( τ ) d τ ] = 0.
The time-local description therefore applies separately on intervals that exclude singular times.
Consequently, the Gaussian-preserving formulation makes the Lindblad-like non-Markovian model experimentally and numerically tractable. Instead of reconstructing the full density operator, the dynamics is reduced to the evolution of the displacement vector d ( t ) d(t) d ( t ) and covariance matrix V ( t ) V(t) V ( t ) , which are directly accessible from homodyne measurements in CV photonic platforms. The reconstructed matrices A ( t ) A(t) A ( t ) and D ( t ) D(t) D ( t ) therefore connect the effective master equation description to measurable phase-space dynamics.
Physical Noise Models
Physical decomposition of the Gaussian noise generator
Here we want to connect the abstract Gaussian generator introduced in the previous section to physical noise mechanisms in CV photonic hardware. This effective generator viewpoint is analogous to microscopic decoherence calculations in other open system settings, where decay rates are derived from the underlying system-environment interaction rather than introduced purely phenomenologically [63 , 64 , 65 ] .
In the Gaussian-preserving framework, the first and second moments evolve according to Eqs. (22) and (23) , with drift matrix A ( t ) A(t) A ( t ) , diffusion matrix D ( t ) D(t) D ( t ) and displacement drift vector u ( t ) u(t) u ( t ) . For the noise mechanisms considered below, we decompose the effective Gaussian generator into physically interpretable contributions phenomenologically as
A ( t ) = A u n i t a r y ( t ) + A l o s s ( t ) + A c o r r ( t ) + A t e c h ( t ) (31) A(t)=A_{\mathrm{unitary}}(t)+A_{\mathrm{loss}}(t)+A_{\mathrm{corr}}(t)+A_{\mathrm{tech}}(t) \tag*{(31)} A ( t ) = A unitary ( t ) + A loss ( t ) + A corr ( t ) + A tech ( t ) (31)
D ( t ) = D v a c ( t ) + D t h ( t ) + D c o r r ( t ) + D c l a s ( t ) (32) D(t)=D_{\mathrm{vac}}(t)+D_{\mathrm{th}}(t)+D_{\mathrm{corr}}(t)+D_{\mathrm{clas}}(t) \tag*{(32)} D ( t ) = D vac ( t ) + D th ( t ) + D corr ( t ) + D clas ( t ) (32)
Here, A u n i t a r y A_{\mathrm{unitary}} A unitary describes the intended unitary circuit dynamics, A l o s s A_{\mathrm{loss}} A loss describes independent photon-loss damping, A c o r r A_{\mathrm{corr}} A corr describes damping through shared loss channels and A t e c h A_{\mathrm{tech}} A tech describes deterministic technical deviations from the intended drift. The diffusion terms include D v a c D_{\mathrm{vac}} D vac , which describes the vacuum fluctuations introduced when each optical mode loses photons to an independent vacuum environment, additional thermal noise D t h D_{\mathrm{th}} D th , vacuum noise from shared loss channels D c o r r D_{\mathrm{corr}} D corr and additive classical Gaussian noise D c l a s D_{\mathrm{clas}} D clas . Coherent driving and deterministic displacement offsets are included separately in u ( t ) u(t) u ( t ) . These expressions organize the selected mechanisms rather than define a unique microscopic decomposition. Independent and shared loss contributions are counted separately, while thermal damping is included in A l o s s A_{\mathrm{loss}} A loss and its additional noise in D t h D_{\mathrm{th}} D th . This separation does not imply that each contribution can be independently identified from homodyne data. Even when the probe set is sufficient to reconstruct the total generator, different sources may produce identical drift and diffusion matrices. For example, independent loss and diagonal contributions from shared loss enter through their sum, while thermal and additive classical diffusion can be indistinguishable without additional constraints. Attribution to individual sources therefore requires independent calibration or explicit model assumptions.
The following subsections describe the selected physical mechanisms and their contributions to the Gaussian generator, providing explicit expressions where applicable and effective models for technical fluctuations.
Unitary photonic dynamics
The coherent dynamics of the CV photonic system is generated by a quadratic Hamiltonian in Eq. (10) . Using the canonical commutation relations [ R j , R k ] = i Ω j k [R_j,R_k]=i\Omega_{jk} [ R j , R k ] = i Ω j k , the Heisenberg equation gives
R ˙ ( t ) = i [ H ( t ) , R ( t ) ] = Ω G ( t ) R ( t ) (33) \dot{R}(t)=i[H(t),R(t)]=\Omega G(t)R(t) \tag*{(33)} R ˙ ( t ) = i [ H ( t ) , R ( t )] = Ω G ( t ) R ( t ) (33)
Considering Eqs. (7) and (22) , the Hamiltonian contribution to the Gaussian drift matrix is
A u n i t a r y ( t ) = Ω G ( t ) A_{\mathrm{unitary}}(t)=\Omega G(t) A unitary ( t ) = Ω G ( t )
u u n i t a r y ( t ) = 0 (34) u_{\mathrm{unitary}}(t)=0 \tag*{(34)} u unitary ( t ) = 0 (34)
Under the quadratic Hamiltonian in Eq. (10) , the covariance transforms as V ( t ) = X ( t ) V ( 0 ) X T ( t ) V(t)=X(t)V(0)X^{T}(t) V ( t ) = X ( t ) V ( 0 ) X T ( t ) . Hence,
V ˙ ( t ) = A u n i t a r y ( t ) V ( t ) + V ( t ) A u n i t a r y T ( t ) (35) \dot{V}(t)=A_{\mathrm{unitary}}(t)V(t)+V(t)A_{\mathrm{unitary}}^{T}(t) \tag*{(35)} V ˙ ( t ) = A unitary ( t ) V ( t ) + V ( t ) A unitary T ( t ) (35)
which compared with Eq. (23) gives
D u n i t a r y ( t ) = 0 (36) D_{\mathrm{unitary}}(t)=0 \tag*{(36)} D unitary ( t ) = 0 (36)
It means that a quadratic Hamiltonian transforms the quadratures without coupling the system to an environment, so it changes the covariance through the drift matrix but introduces no diffusion.
Independent photon loss
For independent photon loss, each mode couples to its own reservoir, with no cross-mode dissipative couplings. The Lindblad superoperator for mode j j j is [12 ]
L l o s s ( j ) ρ = κ j ( t ) D [ a j ] ρ (37) \mathcal{L}_{\mathrm{loss}}^{(j)}\rho=\kappa_j(t)\mathcal{D}[a_j]\rho \tag*{(37)} L loss ( j ) ρ = κ j ( t ) D [ a j ] ρ (37)
where κ j ( t ) \kappa_j(t) κ j ( t ) is the time-dependent loss rate of mode j j j and D [ a j ] \mathcal{D}[a_j] D [ a j ] is the Lindblad dissipator. By independent loss, we mean that the Kossakowski matrix in the basis of mode annihilation operators a j a_j a j is diagonal; each mode j j j has its own loss rate κ j ( t ) \kappa_j(t) κ j ( t ) and cross-mode dissipative terms are absent.
For κ j ( t ) ≥ 0 \kappa_j(t)\ge0 κ j ( t ) ≥ 0 , the Lindblad loss term in Eq. (37) is equivalently represented in the Markov vacuum-bath limit by the Heisenberg-Langevin equation [18 , 66 ]
a ˙ j ( t ) = − κ j ( t ) 2 a j ( t ) + κ j ( t ) a j , i n ( t ) (38) \dot{a}_j(t)=-\frac{\kappa_j(t)}{2}a_j(t)+\sqrt{\kappa_j(t)}a_{j,\mathrm{in}}(t) \tag*{(38)} a ˙ j ( t ) = − 2 κ j ( t ) a j ( t ) + κ j ( t ) a j , in ( t ) (38)
where a j , i n ( t ) a_{j,\mathrm{in}}(t) a j , in ( t ) denotes the vacuum input field of the reservoir coupled to mode j j j . Using a j = 1 2 ( q j + i p j ) a_j=\frac{1}{\sqrt{2}}(q_j+ip_j) a j = 2 1 ( q j + i p j ) and defining the single-mode quadrature vectors R j = ( q j , p j ) T \mathbf{R}_j=(q_j,p_j)^{\mathsf{T}} R j = ( q j , p j ) T , Eq. (38) gives
R ˙ j ( t ) = − κ j ( t ) 2 R j ( t ) + κ j ( t ) R j , i n ( t ) (39) \dot{\mathbf{R}}_j(t)=-\frac{\kappa_j(t)}{2}\mathbf{R}_j(t)+\sqrt{\kappa_j(t)}\mathbf{R}_{j,\mathrm{in}}(t) \tag*{(39)} R ˙ j ( t ) = − 2 κ j ( t ) R j ( t ) + κ j ( t ) R j , in ( t ) (39)
The vacuum input satisfies
⟨ a j , i n ( t ) a j , i n † ( t ′ ) ⟩ = δ ( t − t ′ ) , ⟨ a j , i n † ( t ) a j , i n ( t ′ ) ⟩ = 0 (40) \langle a_{j,\mathrm{in}}(t)a_{j,\mathrm{in}}^\dagger(t')\rangle=\delta(t-t'),\qquad\langle a_{j,\mathrm{in}}^\dagger(t)a_{j,\mathrm{in}}(t')\rangle=0 \tag*{(40)} ⟨ a j , in ( t ) a j , in † ( t ′ )⟩ = δ ( t − t ′ ) , ⟨ a j , in † ( t ) a j , in ( t ′ )⟩ = 0 (40)
and using Eq. (3) , symmetrized correlations can be obtained
1 2 ⟨ { R j , i n , μ ( t ) , R j , i n , ν ( t ′ ) } ⟩ = 1 2 δ μ ν δ ( t − t ′ ) (41) \frac{1}{2}\left\langle\left\{\mathbf{R}_{j,\mathrm{in},\mu}(t),\mathbf{R}_{j,\mathrm{in},\nu}(t')\right\}\right\rangle=\frac{1}{2}\delta_{\mu\nu}\delta(t-t') \tag*{(41)} 2 1 ⟨ { R j , in , μ ( t ) , R j , in , ν ( t ′ ) } ⟩ = 2 1 δ μν δ ( t − t ′ ) (41)
which is consistent with the convention [ q j , p j ] = i [q_j,p_j]=i [ q j , p j ] = i and vacuum covariance I 2 / 2 I_2/2 I 2 /2 . Here μ , ν ∈ { 1 , 2 } \mu,\nu\in\{1,2\} μ , ν ∈ { 1 , 2 } label the components of the reservoir input quadrature vector, corresponding to R j , i n , 1 = q j , i n R_{j,\mathrm{in},1}=q_{j,\mathrm{in}} R j , in , 1 = q j , in and R j , i n , 2 = p j , i n R_{j,\mathrm{in},2}=p_{j,\mathrm{in}} R j , in , 2 = p j , in .
Over a small time interval d t dt d t , a pure-loss channel with rate κ j ( t ) \kappa_j(t) κ j ( t ) has transmissivity [6 ]
η j ( t ) = 1 − κ j ( t ) d t + O ( d t 2 ) (42) \eta_j(t)=1-\kappa_j(t)dt+O(dt^2) \tag*{(42)} η j ( t ) = 1 − κ j ( t ) d t + O ( d t 2 ) (42)
where η j ( t ) \eta_j(t) η j ( t ) quantifies the fraction of the input optical intensity transmitted through the loss channel, with η j = 1 \eta_j=1 η j = 1 corresponding to no loss and η j = 0 \eta_j=0 η j = 0 to complete attenuation. Moreover, a pure-loss Gaussian channel describes attenuation with no added noise beyond the vacuum fluctuations required by quantum mechanics. It can be represented as a beam-splitter interaction between the system mode and an environmental vacuum mode. Accordingly, the quadratures transform as [6 , 11 ]
R j ( t + d t ) = η j ( t ) R j ( t ) + 1 − η j ( t ) R j , v a c (43) \mathbf{R}_j(t+dt)=\sqrt{\eta_j(t)}\mathbf{R}_j(t)+\sqrt{1-\eta_j(t)}\mathbf{R}_{j,\mathrm{vac}} \tag*{(43)} R j ( t + d t ) = η j ( t ) R j ( t ) + 1 − η j ( t ) R j , vac (43)
where R j , v a c \mathbf{R}_{j,\mathrm{vac}} R j , vac is an independent vacuum quadrature vector. Substituting η j ( t ) \eta_j(t) η j ( t ) , taking the expectation value based on Eq. (7) and using ⟨ R j , v a c ⟩ = 0 \langle\mathbf{R}_{j,\mathrm{vac}}\rangle=0 ⟨ R j , vac ⟩ = 0 , the first moment evolves as
d j ( t + d t ) = ( 1 − κ j ( t ) 2 d t ) d j ( t ) + O ( d t 2 ) (44) d_j(t+dt)=\left(1-\frac{\kappa_j(t)}{2}dt\right)d_j(t)+O(dt^2) \tag*{(44)} d j ( t + d t ) = ( 1 − 2 κ j ( t ) d t ) d j ( t ) + O ( d t 2 ) (44)
and the covariance matrix in Eq. (8) evolves as
V j ( t + d t ) = ( 1 − κ j ( t ) d t ) V j ( t ) + κ j ( t ) 2 I 2 d t + O ( d t 2 ) (45) V_j(t+dt)=(1-\kappa_j(t)dt)V_j(t)+\frac{\kappa_j(t)}{2}I_2dt+O(dt^2) \tag*{(45)} V j ( t + d t ) = ( 1 − κ j ( t ) d t ) V j ( t ) + 2 κ j ( t ) I 2 d t + O ( d t 2 ) (45)
Subtracting the moments at time t t t from both sides, dividing by d t dt d t , and taking the limit d t → 0 dt\to0 d t → 0 yields
d ˙ j ( t ) = − κ j ( t ) 2 d j ( t ) , V ˙ j ( t ) = − κ j ( t ) V j ( t ) + κ j ( t ) 2 I 2 (46) \dot{d}_j(t)=-\frac{\kappa_j(t)}{2}d_j(t),\qquad\dot{V}_j(t)=-\kappa_j(t)V_j(t)+\frac{\kappa_j(t)}{2}I_2 \tag*{(46)} d ˙ j ( t ) = − 2 κ j ( t ) d j ( t ) , V ˙ j ( t ) = − κ j ( t ) V j ( t ) + 2 κ j ( t ) I 2 (46)
Therefore, comparison with the Gaussian moment equation in Eqs. (22) and (23) identifies the independent photon-loss contribution as
A l o s s , j ( t ) = − κ j ( t ) 2 I 2 D v a c , j ( t ) = κ j ( t ) 2 I 2 (47) A_{\mathrm{loss},j}(t)=-\frac{\kappa_j(t)}{2}I_2\qquad D_{\mathrm{vac},j}(t)=\frac{\kappa_j(t)}{2}I_2 \tag*{(47)} A loss , j ( t ) = − 2 κ j ( t ) I 2 D vac , j ( t ) = 2 κ j ( t ) I 2 (47)
For independent loss in an m m m -mode system, these single-mode blocks give
K κ ( t ) = ⨁ j = 1 m κ j ( t ) I 2 K_{\kappa}(t)=\bigoplus_{j=1}^{m}\kappa_j(t)I_2 K κ ( t ) = j = 1 ⨁ m κ j ( t ) I 2
A l o s s ( t ) = − 1 2 K κ ( t ) A_{\mathrm{loss}}(t)=-\frac{1}{2}K_{\kappa}(t) A loss ( t ) = − 2 1 K κ ( t )
D v a c ( t ) = 1 2 K κ ( t ) (48) D_{\mathrm{vac}}(t)=\frac{1}{2}K_{\kappa}(t) \tag*{(48)} D vac ( t ) = 2 1 K κ ( t ) (48)
The vacuum Langevin and infinitesimal beam-splitter derivations above assume κ j ( t ) ≥ 0 \kappa_j(t)\ge0 κ j ( t ) ≥ 0 . For negative effective rates, the same drift and diffusion relations follow instead by differentiating the accumulated pure-loss channel, provided its transmissivity remains between zero and one. Such rates describe a temporary increase in the accumulated transmissivity, rather than a loss step with a fresh, independent vacuum input.
Thermal-loss channel
Thermal excitation describes the situation in which the effective environment is not in the vacuum state, but contains a finite average number of photons. In this case, the photonic mode is not only damped by photon loss, but can also absorb excitations from the bath. Therefore, the master equation contains two contributions: a loss term proportional to D [ a j ] ρ \mathcal{D}[a_j]\rho D [ a j ] ρ , and a heating term proportional to D [ a j † ] ρ \mathcal{D}[a_j^\dagger]\rho D [ a j † ] ρ . The strength of the heating contribution is controlled by the effective thermal occupation n ˉ j ( t ) \bar{n}_j(t) n ˉ j ( t ) of the environment coupled to mode j j j . The thermal-loss contribution to the Lindblad master equation can therefore be written as [11 ]
L t h ( j ) ρ = κ j ( t ) ( n ‾ j ( t ) + 1 ) D [ a j ] ρ + κ j ( t ) n ‾ j ( t ) D [ a j † ] ρ (49) \mathcal{L}_{\mathrm{th}}^{(j)}\rho= \kappa_j(t)\left(\overline{n}_j(t)+1\right)\mathcal{D}[a_j]\rho+\kappa_j(t)\overline{n}_j(t)\mathcal{D}[a_j^\dagger]\rho \tag*{(49)} L th ( j ) ρ = κ j ( t ) ( n j ( t ) + 1 ) D [ a j ] ρ + κ j ( t ) n j ( t ) D [ a j † ] ρ (49)
where L t h ( j ) \mathcal{L}_{\mathrm{th}}^{(j)} L th ( j ) means the thermal Lindblad superoperator for mode j j j .
To obtain the additional thermal diffusion, we separate the vacuum-loss contribution in Eq. (37) from the thermal-loss generator in Eq. (49) . The remaining term is κ j ( t ) n ‾ j ( t ) ( D [ a j ] + D [ a j † ] ) ρ \kappa_j(t)\overline{n}_j(t)\left(\mathcal{D}[a_j]+\mathcal{D}[a_j^\dagger]\right)\rho κ j ( t ) n j ( t ) ( D [ a j ] + D [ a j † ] ) ρ . Its loss and gain contributions produce opposite drift terms which cancel, while each contributes κ j ( t ) n ‾ j ( t ) I 2 / 2 \kappa_j(t)\overline{n}_j(t)I_2/2 κ j ( t ) n j ( t ) I 2 /2 to diffusion. Comparison with Eq. (23) therefore gives the additional single-mode diffusion κ j ( t ) n ‾ j ( t ) I 2 \kappa_j(t)\overline{n}_j(t)I_2 κ j ( t ) n j ( t ) I 2 . Collecting these blocks for independent modes yields
D t h ( t ) = ⨁ j = 1 m κ j ( t ) n ‾ j ( t ) I 2 (50) D_{\mathrm{th}}(t)=\bigoplus_{j=1}^{m}\kappa_j(t)\overline{n}_j(t)I_2 \tag*{(50)} D th ( t ) = j = 1 ⨁ m κ j ( t ) n j ( t ) I 2 (50)
Here D t h ( t ) D_{\mathrm{th}}(t) D th ( t ) denotes the additional thermal diffusion beyond the vacuum diffusion D v a c ( t ) D_{\mathrm{vac}}(t) D vac ( t ) already associated with photon loss in Eq. (47) . Therefore, the total diffusion associated with a thermal loss channel is
D l o s s + t h ( t ) = D v a c ( t ) + D t h ( t ) = ⨁ j = 1 m κ j ( t ) ( n ‾ j ( t ) + 1 2 ) I 2 (51) D_{\mathrm{loss+th}}(t)=D_{\mathrm{vac}}(t)+D_{\mathrm{th}}(t)=\bigoplus_{j=1}^{m}\kappa_j(t)\left(\overline{n}_j(t)+\frac{1}{2}\right)I_2 \tag*{(51)} D loss + th ( t ) = D vac ( t ) + D th ( t ) = j = 1 ⨁ m κ j ( t ) ( n j ( t ) + 2 1 ) I 2 (51)
Moreover, the drift matrix A l o s s A_{\mathrm{loss}} A loss remains the same as in Eq. (47) , since the additional thermal loss and excitation terms make equal and opposite contributions to the drift. In CV photonic systems, the thermal occupation n ‾ j ( t ) \overline{n}_j(t) n j ( t ) is often negligible because the environment is well approximated by vacuum. We keep this term in the model for generality, especially for microwave, optomechanical, or hybrid CV platforms where thermal photons may contribute appreciably to the noise.
In CV photonic systems, loss may not be completely independent between modes. For example, neighboring waveguides, common scattering regions or shared spectral modes can create correlated dissipation. In this case, the standard correlated photon-loss Lindblad form is
L c o r r ρ = ∑ j , k = 1 m Γ j k ( t ) ( a j ρ a k † − 1 2 { a k † a j , ρ } ) (52) \mathcal{L}_{\mathrm{corr}}\rho=\sum_{j,k=1}^{m}\Gamma_{jk}(t)\left(a_j\rho a_k^\dagger-\frac{1}{2}\{a_k^\dagger a_j,\rho\}\right) \tag*{(52)} L corr ρ = j , k = 1 ∑ m Γ j k ( t ) ( a j ρ a k † − 2 1 { a k † a j , ρ } ) (52)
where Γ ( t ) = Γ † ( t ) \Gamma(t)=\Gamma^\dagger(t) Γ ( t ) = Γ † ( t ) is the correlated-loss matrix. The diagonal terms Γ j j ( t ) \Gamma_{jj}(t) Γ j j ( t ) describe ordinary single-mode loss, while the off-diagonal terms Γ j k ( t ) \Gamma_{jk}(t) Γ j k ( t ) describe correlated loss between modes j j j and k k k . For a Markovian correlated loss channel, the matrix must satisfy [18 , 19 , 22 , 23 ] .
Γ ( t ) ≥ 0 (53) \Gamma(t)\ge0 \tag*{(53)} Γ ( t ) ≥ 0 (53)
In a time-local description, Γ ( t ) \Gamma(t) Γ ( t ) may temporarily have negative eigenvalues. For the isolated correlated-loss model, this implies CP-indivisibility, although the accumulated channel may remain CP. When other loss or diffusion mechanisms are present, their contributions may compensate these negative directions. CP-divisibility must then be assessed using M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) constructed from the total drift and diffusion matrices, rather than from Γ ( t ) \Gamma(t) Γ ( t ) alone.
For simplicity, we take the correlated-loss matrix Γ ( t ) \Gamma(t) Γ ( t ) to be real and symmetric. It then produces identical dissipative coupling among the q q q quadratures and among the p p p quadratures, without mixing the two. With the interleaved quadrature ordering defined in Eq. (4) , the corresponding matrix in quadrature space is
B Γ ( t ) = Γ ( t ) ⊗ I 2 (54) B_{\Gamma}(t)=\Gamma(t)\otimes I_2 \tag*{(54)} B Γ ( t ) = Γ ( t ) ⊗ I 2 (54)
To determine how the correlated-loss matrix enters the Gaussian drift, we evaluate the evolution of the mode amplitudes. From Eq. (52)
d d t ⟨ a n ⟩ = Tr [ a n L c o r r ρ ] = − 1 2 ∑ j = 1 m Γ j n ( t ) ⟨ a j ⟩ (55) \frac{d}{dt}\langle a_n\rangle=\operatorname{Tr}[a_n\mathcal{L}_{\mathrm{corr}}\rho]=-\frac{1}{2}\sum_{j=1}^{m}\Gamma_{jn}(t)\langle a_j\rangle \tag*{(55)} d t d ⟨ a n ⟩ = Tr [ a n L corr ρ ] = − 2 1 j = 1 ∑ m Γ j n ( t ) ⟨ a j ⟩ (55)
For a general Hermitian correlated-loss matrix, this corresponds to the amplitude drift − Γ T ( t ) / 2 -\Gamma^{\mathsf{T}}(t)/2 − Γ T ( t ) /2 . Since we restrict Γ ( t ) \Gamma(t) Γ ( t ) to be real and symmetric, Γ T ( t ) = Γ ( t ) \Gamma^{\mathsf{T}}(t)=\Gamma(t) Γ T ( t ) = Γ ( t ) , and the quadrature-space representation remains unchanged. Thus, the diagonal elements of Γ ( t ) \Gamma(t) Γ ( t ) describe ordinary modal attenuation, while the off-diagonal elements generate dissipative coupling between different mode amplitudes. This relation provides the connection between the correlated-loss master equation and the phase-space drift matrix. Using the quadrature definitions in Eq. (3) and the quadrature-space loss matrix B Γ ( t ) B_{\Gamma}(t) B Γ ( t ) , Eq. (55) becomes
d ˙ ( t ) = − 1 2 B Γ ( t ) d ( t ) (56) \dot{d}(t)=-\frac{1}{2}B_{\Gamma}(t)d(t) \tag*{(56)} d ˙ ( t ) = − 2 1 B Γ ( t ) d ( t ) (56)
Comparison with Eq. (22) then shows that the correlated-loss contribution to the drift matrix is
A c o r r ( t ) = − 1 2 B Γ ( t ) (57) A_{\mathrm{corr}}(t)=-\frac{1}{2}B_{\Gamma}(t) \tag*{(57)} A corr ( t ) = − 2 1 B Γ ( t ) (57)
Since the loss generator preserves the vacuum covariance, Eq. (23) requires
0 = ( A c o r r + A c o r r T ) / 2 + D c o r r (58) 0=(A_{\mathrm{corr}}+A_{\mathrm{corr}}^{\mathsf{T}})/2+D_{\mathrm{corr}} \tag*{(58)} 0 = ( A corr + A corr T ) /2 + D corr (58)
Substituting Eq. (57) and using B Γ T = B Γ B_{\Gamma}^{\mathsf{T}}=B_{\Gamma} B Γ T = B Γ gives
D c o r r ( t ) = 1 2 B Γ ( t ) (59) D_{\mathrm{corr}}(t)=\frac{1}{2}B_{\Gamma}(t) \tag*{(59)} D corr ( t ) = 2 1 B Γ ( t ) (59)
Eq. (57) is the multimode analogue of the single-mode photon-loss relations in Eq. (47) . In this case, the scalar loss rate κ j ( t ) \kappa_j(t) κ j ( t ) is replaced by the quadrature-space loss matrix B Γ ( t ) B_{\Gamma}(t) B Γ ( t ) . The drift term A c o r r ( t ) A_{\mathrm{corr}}(t) A corr ( t ) describes the contraction of the multimode phase-space distribution due to to correlated loss, while the diffusion term D c o r r ( t ) D_{\mathrm{corr}}(t) D corr ( t ) represents the accompanying vacuum noise required by the loss channel. The factor 1 / 2 1/2 1/2 in A c o r r A_{\mathrm{corr}} A corr reflects amplitude decay at half the photon-number decay rate. The corresponding D c o r r D_{\mathrm{corr}} D corr follows from the same loss generator and ensures that the vacuum covariance remains unchanged.
Classical technical fluctuations
In addition to quantum loss and thermal excitation, CV photonic systems are also affected by classical technical noise. Important examples include pump-amplitude fluctuations, pump-phase noise, slow thermal drift, imperfect phase stabilization, and fluctuations of external control fields. These noise sources are classical in the sense that they can often be modeled as stochastic variations of system parameters rather than as independent quantum bath modes. For example, pump-amplitude fluctuations can change the effective squeezing strength, while phase fluctuations can perturb phase shifters, interferometers or homodyne reference phases.
To describe technical fluctuations within the Gaussian framework, we consider additive noise represented by a real, zero-mean Gaussian stochastic vector ξ ( t ) \xi(t) ξ ( t ) , coupled to the quadratures through a real matrix B ξ ( t ) B_{\xi}(t) B ξ ( t ) . This assumption applies directly to Gaussian random displacements, but does not generally describe fluctuations of phase or squeezing strength, which act multiplicatively and require an additional linearization or effective Gaussian approximation. A nonzero mean can be included in the displacement drift u ( t ) u(t) u ( t ) . In the white-noise limit, valid when the noise correlation time is much shorter than the relevant system evolution timescales, we have [11 , 18 ]
E [ ξ ( t ) ξ T ( s ) ] = Q ξ ( t ) δ ( t − s ) , Q ξ ( t ) ≥ 0 (60) \mathbb{E}[\xi(t)\xi^{T}(s)] = Q_{\xi}(t)\delta(t-s), \qquad Q_{\xi}(t) \ge0 \tag*{(60)} E [ ξ ( t ) ξ T ( s )] = Q ξ ( t ) δ ( t − s ) , Q ξ ( t ) ≥ 0 (60)
specifies the noise strengths and correlations. Here E \mathbb{E} E denotes an ensemble average over noise realizations. Since the additive noise has zero mean, it contributes neither to A ( t ) A(t) A ( t ) nor to u ( t ) u(t) u ( t ) in the first-moment equation. It nevertheless changes the covariance through the additional diffusion matrix
D c l a s ( t ) = B ξ ( t ) Q ξ ( t ) B ξ T ( t ) ≥ 0 (61) D_{\mathrm{clas}}(t) = B_{\xi}(t)Q_{\xi}(t)B_{\xi}^{T}(t) \ge0 \tag*{(61)} D clas ( t ) = B ξ ( t ) Q ξ ( t ) B ξ T ( t ) ≥ 0 (61)
More generally, for noise with finite temporal correlations, the white-noise relation in Eq. (60) is replaced by [11 ]
E [ ξ μ ( t ) ξ ν ( s ) ] = C μ ν ( t , s ) (62) \mathbb{E}[\xi_{\mu}(t)\xi_{\nu}(s)] = C_{\mu\nu}(t,s) \tag*{(62)} E [ ξ μ ( t ) ξ ν ( s )] = C μν ( t , s ) (62)
where μ \mu μ and ν \nu ν label the classical noise components. Eq. (60) is recovered when C μ ν ( t , s ) = [ Q ξ ( t ) ] μ ν δ ( t − s ) C_{\mu\nu}(t,s) = [Q_{\xi}(t)]_{\mu\nu}\delta(t-s) C μν ( t , s ) = [ Q ξ ( t ) ] μν δ ( t − s ) . When the correlation time is much shorter than the system evolution timescale, the white-noise approximation above is appropriate. Otherwise, the effective diffusion depends on the noise correlations and the intervening system evolution, and cannot generally be expressed by the same instantaneous formula. This description includes finite-correlation-time processes whose reduced dynamics preserves Gaussian states and admits a well-defined time-local generator; it does not cover arbitrary temporally correlated noise. Within this class, temporal correlations may lead to violations of CP-divisibility, but a finite correlation time alone does not guarantee such a violation.
In the time-local Gaussian model, we do not explicitly reconstruct the underlying classical stochastic process. Instead, zero-mean additive fluctuations are represented through the effective diffusion term D c l a s ( t ) D_{\mathrm{clas}}(t) D clas ( t ) in the covariance dynamics of Eq. (23) . For temporally correlated fluctuations, D c l a s ( t ) D_{\mathrm{clas}}(t) D clas ( t ) should be interpreted as an effective time-local coefficient rather than as the microscopic two-time correlation function. Such a coefficient may be time dependent while the dynamics remains CP-divisible. Non-Markovianity is identified only when the reconstructed time-local generator violates the infinitesimal Gaussian CP condition while the accumulated finite-time channel remains CP.
The noise mechanisms discussed above were introduced through time-local coefficients in the Gaussian generator. However, in realistic CV photonic hardware these coefficients can originate from underlying stochastic processes with finite temporal correlations. The goal of this section is to distinguish the microscopic two-time noise kernels from the effective local rates that enter the reconstructed matrices A ( t ) A(t) A ( t ) and D ( t ) D(t) D ( t ) .
Temporal correlations can enter any of the hardware noise coefficients introduced above, including photon loss, thermal occupation, correlated dissipation, and classical control fluctuations. We denote these noise parameters collectively by
ζ ( t ) = { κ j ( t ) , n ˉ j ( t ) , Γ j k ( t ) , ξ μ ( t ) } (63) \zeta(t) = \{\kappa_j(t), \bar{n}_j(t), \Gamma_{jk}(t), \xi_{\mu}(t)\} \tag*{(63)} ζ ( t ) = { κ j ( t ) , n ˉ j ( t ) , Γ j k ( t ) , ξ μ ( t )} (63)
where κ j ( t ) \kappa_j(t) κ j ( t ) is the effective photon-loss rate of mode j j j and n ˉ j ( t ) \bar{n}_j(t) n ˉ j ( t ) the effective occupation of its thermal environment. The entries Γ j k ( t ) \Gamma_{jk}(t) Γ j k ( t ) describe correlated loss between modes j j j and k k k , while ξ μ ( t ) \xi_{\mu}(t) ξ μ ( t ) is the μ \mu μ th component of the classical stochastic noise vector introduced in Eq. (60) . When matrix-valued quantities such as Γ ( t ) \Gamma(t) Γ ( t ) are included in ζ ( t ) \zeta(t) ζ ( t ) , their independent entries are collected into a single parameter vector. The microscopic parameter fluctuations are characterized by the two-time covariance matrix [18 ]
C ζ ( t , s ) = E [ δ ζ ( t ) δ ζ T ( s ) ] , δ ζ ( t ) = ζ ( t ) − E [ ζ ( t ) ] (64) C_{\zeta}(t,s) = \mathbb{E}[\delta\zeta(t)\delta\zeta^{T}(s)], \qquad\delta\zeta(t) = \zeta(t) - \mathbb{E}[\zeta(t)] \tag*{(64)} C ζ ( t , s ) = E [ δ ζ ( t ) δ ζ T ( s )] , δ ζ ( t ) = ζ ( t ) − E [ ζ ( t )] (64)
Here ζ ( t ) \zeta(t) ζ ( t ) is understood as a column vector obtained by collecting the microscopic stochastic noise parameters, with matrix-valued quantities such as Γ j k ( t ) \Gamma_{jk}(t) Γ j k ( t ) vectorized when necessary. For a stationary noise source, the correlation depends only on the time delay τ = t − s \tau= t-s τ = t − s , so that C ζ ( t , s ) = C ζ ( τ ) C_{\zeta}(t,s) = C_{\zeta}(\tau) C ζ ( t , s ) = C ζ ( τ ) . The same information can then be represented by the noise spectral-density matrix [18 , 23 ]
C ζ ( τ ) = ∫ d ω 2 π S ζ ( ω ) e − i ω τ , τ = t − s (65) C_{\zeta}(\tau) = \int\frac{d\omega}{2\pi}\,S_{\zeta}(\omega)e^{-i\omega\tau}, \qquad\tau= t-s \tag*{(65)} C ζ ( τ ) = ∫ 2 π d ω S ζ ( ω ) e − iω τ , τ = t − s (65)
where ω \omega ω is the angular-frequency variable, and S ζ ( ω ) S_{\zeta}(\omega) S ζ ( ω ) describes the frequency content of the correlated noise. Structured features of S ζ ( ω ) S_{\zeta}(\omega) S ζ ( ω ) correspond to finite bandwidth reservoirs, resonator memory, delay-line feedback, or slowly relaxing technical noise.
The effect of temporal correlations depends on the physical sector. Photon loss and correlated dissipation modify both the drift and the associated vacuum diffusion, whereas thermal occupation and classical fluctuations primarily contribute additional diffusion. For example, using the same coupling matrix B ξ ( t ) B_{\xi}(t) B ξ ( t ) introduced above, classical fluctuations with finite temporal correlations give the phase-space noise kernel [11 , 18 ]
N c l ( t , s ) = B ξ ( t ) C ξ ( t , s ) B ξ T ( s ) (66) \mathcal{N}_{\mathrm{cl}}(t,s) = B_{\xi}(t)C_{\xi}(t,s)B_{\xi}^{T}(s) \tag*{(66)} N cl ( t , s ) = B ξ ( t ) C ξ ( t , s ) B ξ T ( s ) (66)
with C ξ ( t , s ) = E [ ξ ( t ) ξ T ( s ) ] C_{\xi}(t,s) = \mathbb{E}[\xi(t)\xi^{T}(s)] C ξ ( t , s ) = E [ ξ ( t ) ξ T ( s )] . This relation follows by defining the effective phase-space noise force as f ( t ) = B ξ ( t ) ξ ( t ) \mathbf{f}(t) = B_{\xi}(t)\xi(t) f ( t ) = B ξ ( t ) ξ ( t ) and evaluating its two-time covariance E [ f ( t ) f T ( s ) ] \mathbb{E}[\mathbf{f}(t)\mathbf{f}^{T}(s)] E [ f ( t ) f T ( s )] .
In the white-noise limit, this kernel becomes local in time and reduces to an ordinary diffusion term. For finite-bandwidth fluctuations, however, the noise is more naturally described by a memory kernel; in the time-local reconstruction this appears as a time-dependent effective diffusion. Thus, the quantities κ j ( t ) \kappa_j(t) κ j ( t ) , Γ j k ( t ) \Gamma_{jk}(t) Γ j k ( t ) , and D c l a s ( t ) D_{\mathrm{clas}}(t) D clas ( t ) should be interpreted as effective local coefficients associated with underlying two-time noise correlations, not as the microscopic memory kernels themselves. This distinction is important because time dependence alone does not imply non-Markovianity. A time-dependent Gaussian generator can still be CP-divisible while non-Markovian behavior requires violation of the infinitesimal CP-divisibility condition.
Markovian and non-Markovian regimes
The physical noise models above can generate either Markovian or non-Markovian Gaussian dynamics, depending on the local divisibility properties of the effective generator. Time dependence of the coefficients is not sufficient to identify memory: a process with time-dependent loss, thermal occupation, dephasing, correlated dissipation, or classical diffusion may still be CP-divisible.
Using the infinitesimal Gaussian CP matrix M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) defined in Eq. (30) , we classify the dynamics in the divisibility sense as
{ λ min [ M C P ( t ) ] ≥ 0 ∀ t , CP-divisible regime , λ min [ M C P ( t ) ] < 0 for some t , non-CP-divisible regime (67) \begin{cases}
\lambda_{\min}[M_{\mathrm{CP}}(t)] \ge0 & \forall t,\qquad\text{CP-divisible regime},\\
\lambda_{\min}[M_{\mathrm{CP}}(t)] < 0 & \text{for some }t,\qquad\text{non-CP-divisible regime} \tag*{(67)}
\end{cases} { λ m i n [ M CP ( t )] ≥ 0 λ m i n [ M CP ( t )] < 0 ∀ t , CP-divisible regime , for some t , non-CP-divisible regime (67)
In the second case, the negative eigenvalue indicates that the infinitesimal intermediate map from t t t to t + d t t+dt t + d t is not CP. This is the Gaussian CP-divisibility witness used here to identify non-Markovianity in the divisibility sense [22 , 23 ] . This local violation must be distinguished from loss of physicality of the full evolution. The finite-time Gaussian channel must still satisfy the CP condition in Eq. (27) . Within the present framework, the dynamics is therefore identified as non-Markovian only when M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) has a negative eigenvalue over a finite time interval while the finite-time Gaussian channel remains CP. Time-dependent coefficients or finite temporal correlations alone are not sufficient to establish non-Markovianity in the CP-divisibility sense.
Markovian embedding of non-Markovian Gaussian noise
The time-local Lindblad-like generator gives an effective description of the physical photonic modes alone. When some effective coefficients become temporarily negative, the reduced generator is not a standard Lindblad generator at that instant. This does not imply that the full dynamics is unphysical. A standard interpretation is that the observed modes are coupled to unobserved memory-bearing degrees of freedom, such as resonator modes, delay-line modes, hidden spectral modes, or structured reservoirs [67 , 68 , 69 , 70 ] .
Let a j a_j a j denote the physical photonic modes and b ℓ b_{\ell} b ℓ auxiliary bosonic memory modes. A Gaussian-preserving embedded model may be written as [69 ]
H e m b = H p s + ∑ ℓ ω b , ℓ b ℓ † b ℓ + ∑ j , ℓ ( g j ℓ a j † b ℓ + g j ℓ ∗ a j b ℓ † ) (68) H_{\mathrm{emb}} = H_{\mathrm{ps}} + \sum_{\ell}\omega_{b,\ell}b_{\ell}^{\dagger}b_{\ell} + \sum_{j,\ell}\left(g_{j\ell}a_j^{\dagger}b_{\ell} + g_{j\ell}^{*}a_jb_{\ell}^{\dagger}\right) \tag*{(68)} H emb = H ps + ℓ ∑ ω b , ℓ b ℓ † b ℓ + j , ℓ ∑ ( g j ℓ a j † b ℓ + g j ℓ ∗ a j b ℓ † ) (68)
where H p s H_{\mathrm{ps}} H ps is the quadratic Hamiltonian of the photonic system and g j ℓ g_{j\ell} g j ℓ controls the coupling to the memory modes and ω b , ℓ \omega_{b,\ell} ω b , ℓ is the resonance frequency of the ℓ \ell ℓ -th auxiliary memory mode. The auxiliary modes b ℓ b_{\ell} b ℓ act as memory-bearing degrees of freedom, while their coupling to external Markovian reservoirs is represented through the damping terms introduced below. For each input probe, the initial joint state is ρ ( 0 ) = ρ S ( 0 ) ⊗ ρ M \rho(0) = \rho_{\mathrm{S}}(0) \otimes\rho_{\mathrm{M}} ρ ( 0 ) = ρ S ( 0 ) ⊗ ρ M , where ρ S ( 0 ) \rho_{\mathrm{S}}(0) ρ S ( 0 ) is the photonic input state and ρ M \rho_{\mathrm{M}} ρ M is a fixed Gaussian memory state, identical for all probes. The pseudo-modes couple to independent Markovian reservoirs, initially uncorrelated with the system and memory. Under the Gaussian-preserving enlarged dynamics, these assumptions define a Gaussian CPTP reduced channel for arbitrary system inputs.
The joint density operator ρ ( t ) \rho(t) ρ ( t ) of the physical photonic modes and the auxiliary memory modes evolves according to the Markovian Lindblad equation
d ϱ d t = − i [ H e m b , ϱ ] + ∑ ℓ γ ℓ D [ b ℓ ] ϱ (69) \frac{d\varrho}{dt}=-i[H_{\mathrm{emb}},\varrho]+\sum_{\ell}\gamma_{\ell}\mathcal{D}[b_{\ell}]\varrho \tag*{(69)} d t d ϱ = − i [ H emb , ϱ ] + ℓ ∑ γ ℓ D [ b ℓ ] ϱ (69)
with positive pseudomode decay rates γ ℓ \gamma_{\ell} γ ℓ . Thus, the external environment does not appear as an explicit dynamical mode in the enlarged Hamiltonian; instead, its influence is encoded in the positive pseudomode decay rates γ ℓ \gamma_{\ell} γ ℓ and the associated Lindblad dissipators D [ b ℓ ] \mathcal{D}[b_{\ell}] D [ b ℓ ] . The physical photonic state is obtained after tracing out the memory modes
ρ ( t ) = Tr b { ϱ ( t ) } (70) \rho(t)=\operatorname{Tr}_{b}\{\varrho(t)\} \tag*{(70)} ρ ( t ) = Tr b { ϱ ( t )} (70)
Formally solving the damped equation of motion for the auxiliary modes b ℓ b_{\ell} b ℓ and substituting the result into the equations for the photonic modes produces a time-nonlocal term proportional to ∑ k ∫ 0 t K j k ( t − s ) a k ( s ) d s \sum_k\int_0^t K_{jk}(t-s)a_k(s)\,ds ∑ k ∫ 0 t K j k ( t − s ) a k ( s ) d s , together with a reservoir-noise term. For the passive, number-preserving coupling in Eq. (68) , eliminating the damped pseudomodes gives the memory kernel [67 , 70 ]
K j k ( t − s ) = ∑ ℓ g j ℓ g k ℓ ∗ e − ( i ω b , ℓ + γ ℓ / 2 ) ( t − s ) Θ H ( t − s ) (71) K_{jk}(t-s)=\sum_{\ell}g_{j\ell}g_{k\ell}^{*}e^{-(i\omega_{b,\ell}+\gamma_{\ell}/2)(t-s)}\Theta_{H}(t-s) \tag*{(71)} K j k ( t − s ) = ℓ ∑ g j ℓ g k ℓ ∗ e − ( i ω b , ℓ + γ ℓ /2 ) ( t − s ) Θ H ( t − s ) (71)
Here, γ ℓ \gamma_{\ell} γ ℓ is the damping rate of pseudomode ℓ \ell ℓ and the Heaviside function Θ H ( t − s ) \Theta_{H}(t-s) Θ H ( t − s ) enforces causality, while the exponential describes oscillation at frequency ω b , ℓ \omega_{b,\ell} ω b , ℓ and amplitude decay as e − γ ℓ ( t − s ) / 2 e^{-\gamma_{\ell}(t-s)/2} e − γ ℓ ( t − s ) /2 , giving an amplitude-memory decay time 2 / γ ℓ 2/\gamma_{\ell} 2/ γ ℓ for γ ℓ > 0 \gamma_{\ell}>0 γ ℓ > 0 . The diagonal elements K j j ( t − s ) K_{jj}(t-s) K j j ( t − s ) describe memory in individual loss channels, while the off-diagonal elements K j k ( t − s ) K_{jk}(t-s) K j k ( t − s ) describe correlated dissipative memory shared between modes. Thus, a reduced photonic generator may become temporarily non-CP-divisible even though the joint system evolves under a CP Markovian Lindblad equation.
The pseudomode embedding plays three roles in the present framework. First, it provides a physical mechanism for the time-dependent and possibly negative coefficients that appear in the reduced Lindblad-like generator. Second, it shows how reduced non-CP-divisible dynamics can arise from a CP Markovian evolution on an enlarged system. Third, it connects phenomenological parameters such as γ ℓ \gamma_{\ell} γ ℓ , ω b , ℓ \omega_{b,\ell} ω b , ℓ , and g j ℓ g_{j\ell} g j ℓ to hardware-level quantities such as pseudomode lifetime, detuning, and coupling strength. Therefore, pseudomode embedding provides a bridge between the phenomenological time-local tomography model and a microscopic hardware model of non-Markovian photonic noise.
The theoretical framework developed in the preceding sections motivates the reconstruction strategy introduced below. Rather than directly identifying all microscopic reservoirs, hidden memory modes, or two-time noise kernels, we reconstruct their effect through the time-local Gaussian generator A ( t ) A(t) A ( t ) , D ( t ) D(t) D ( t ) , and u ( t ) u(t) u ( t ) .
Tomographic Reconstruction Method
Quantum tomography provides a framework for reconstructing unknown quantum states, processes, or effective dynamical models from measurement outcomes [34 , 35 ] . Here, we develop a theoretical and simulation-based tomography framework for CV photonic systems whose central objective is to move beyond fixed-time Gaussian state reconstruction and instead learn an effective time-local description of the noise dynamics. The reconstruction procedure combines Gaussian moment propagation, homodyne probability distributions, maximum-likelihood estimation, physicality constraints, and statistical model comparison. This makes it possible to determine whether the observed or simulated homodyne data are consistent with CP-divisible Markovian Gaussian noise, or whether they require a non-Markovian time-local Gaussian description.
Tomographic reconstruction of non-Markovian Gaussian noise dynamics
The unknown noise dynamics is modeled by an effective time-local Gaussian generator. In the Gaussian-preserving regime, this generator is fully characterized at the level of first and second moments in Eq. (25) by the drift matrix A ( t ) A(t) A ( t ) , the diffusion matrix D ( t ) D(t) D ( t ) , and the displacement drift u ( t ) u(t) u ( t ) . When the coherent part of the system is also unknown, the Hamiltonian matrix G ( t ) G(t) G ( t ) may be included through the relation A u n i t a r y ( t ) = Ω G ( t ) A_{\mathrm{unitary}}(t)=\Omega G(t) A unitary ( t ) = Ω G ( t ) . These quantities determine the evolution of the displacement vector and covariance matrix, which in turn determine the predicted homodyne statistics at each measurement time. The reconstruction problem is therefore formulated as a statistical inverse problem: choose the generator parameters so that the predicted quadrature distributions best match the measured or synthetic time-resolved homodyne data while preserving the physicality of the total Gaussian channel.
Suppose one prepares a set of input Gaussian states
S i n = { ρ α ( 0 ) } α = 1 N i n (72) \mathcal{S}_{\mathrm{in}}=\{\rho_{\alpha}(0)\}_{\alpha=1}^{N_{\mathrm{in}}} \tag*{(72)} S in = { ρ α ( 0 ) } α = 1 N in (72)
including coherent states, squeezed states, displaced squeezed states, and two-mode squeezed states. At measurement times
T = { t 1 , t 2 , … , t K } (73) \mathcal{T}=\{t_{1},t_{2},\ldots,t_{K}\} \tag*{(73)} T = { t 1 , t 2 , … , t K } (73)
one performs homodyne measurements of quadratures
X j , θ = q j cos θ + p j sin θ = e j , θ T R (74) X_{j,\theta}=q_j\cos\theta+p_j\sin\theta=\mathbf{e}_{j,\theta}^{T}\mathbf{R} \tag*{(74)} X j , θ = q j cos θ + p j sin θ = e j , θ T R (74)
where j j j labels the measured mode, θ \theta θ is the homodyne measurement phase, and e j , θ \mathbf{e}_{j,\theta} e j , θ is the corresponding unit vector in phase space. A homodyne measurement of X j , θ X_{j,\theta} X j , θ yields a real-valued outcome x x x . For a Gaussian state, this outcome x x x is normally distributed [28 ]
p ( x ∣ j , θ , t , α ; Θ ) = 1 2 π σ j , θ , α 2 ( t ) exp [ − ( x − μ j , θ , α ( t ) ) 2 2 σ j , θ , α 2 ( t ) ] (75) p(x|j,\theta,t,\alpha;\Theta)=\frac{1}{\sqrt{2\pi\sigma_{j,\theta,\alpha}^{2}(t)}}\exp\left[-\frac{(x-\mu_{j,\theta,\alpha}(t))^{2}}{2\sigma_{j,\theta,\alpha}^{2}(t)}\right] \tag*{(75)} p ( x ∣ j , θ , t , α ; Θ ) = 2 π σ j , θ , α 2 ( t ) 1 exp [ − 2 σ j , θ , α 2 ( t ) ( x − μ j , θ , α ( t ) ) 2 ] (75)
where
μ j , θ , α ( t ) = e j , θ T d α ( t ) (76) \mu_{j,\theta,\alpha}(t)=\mathbf{e}_{j,\theta}^{T}\mathbf{d}_{\alpha}(t) \tag*{(76)} μ j , θ , α ( t ) = e j , θ T d α ( t ) (76)
σ j , θ , α 2 ( t ) = e j , θ T V α ( t ) e j , θ (77) \sigma_{j,\theta,\alpha}^{2}(t)=\mathbf{e}_{j,\theta}^{T}V_{\alpha}(t)\mathbf{e}_{j,\theta} \tag*{(77)} σ j , θ , α 2 ( t ) = e j , θ T V α ( t ) e j , θ (77)
Here, Θ \Theta Θ denotes the set of unknown model parameters to be estimated, including the parameters that define A ( t ) A(t) A ( t ) , D ( t ) D(t) D ( t ) , u ( t ) u(t) u ( t ) and possibly G ( t ) G(t) G ( t ) . The vector e θ \mathbf{e}_{\theta} e θ is the corresponding unit vector in phase space; for a single mode e θ = ( cos θ , sin θ ) T \mathbf{e}_{\theta}=(\cos\theta,\sin\theta)^{T} e θ = ( cos θ , sin θ ) T . For an m m m -mode system, e θ ∈ R 2 m \mathbf{e}_{\theta}\in\mathbb{R}^{2m} e θ ∈ R 2 m selects the measured mode and quadrature direction. The vector d α ( t ) \mathbf{d}_{\alpha}(t) d α ( t ) is the first-moment displacement vector of the evolved state that originated from the input probe state ρ α ( 0 ) \rho_{\alpha}(0) ρ α ( 0 ) and V α ( t ) V_{\alpha}(t) V α ( t ) is its covariance matrix. The label α \alpha α identifies which input probe state was prepared. The quantity μ j , θ , α ( t ) \mu_{j,\theta,\alpha}(t) μ j , θ , α ( t ) is the predicted mean value of the measured quadrature and σ j , θ , α 2 ( t ) \sigma_{j,\theta,\alpha}^{2}(t) σ j , θ , α 2 ( t ) is the corresponding quadrature variance.
For multimode reconstruction, we also record joint homodyne outcomes of commuting quadratures on distinct modes. For a setting represented by a matrix H H H whose rows select these quadratures, the joint outcome is Gaussian with mean H d α ( t ) H\mathbf{d}_{\alpha}(t) H d α ( t ) and covariance H V α ( t ) H T HV_{\alpha}(t)H^{T} H V α ( t ) H T . Measuring each mode pair at phase combinations ( 0 , 0 ) (0,0) ( 0 , 0 ) , ( 0 , π / 2 ) (0,\pi/2) ( 0 , π /2 ) , ( π / 2 , 0 ) (\pi/2,0) ( π /2 , 0 ) , and ( π / 2 , π / 2 ) (\pi/2,\pi/2) ( π /2 , π /2 ) determines its intermode covariance block. Accordingly, the likelihood below uses joint Gaussian densities for multimode measurements, retaining correlations within each shot and assuming independence only between shots. The scalar formulas apply to the single-mode benchmarks; identification of the full generator additionally requires sufficiently informative probes and a specified temporal parameterization.
Let the homodyne data set, either experimentally collected or synthetically generated in numerical simulations, be
D = { x n , j n , θ n , t n , α n } n = 1 N (78) \mathcal{D}=\left\{x_n,j_n,\theta_n,t_n,\alpha_n\right\}_{n=1}^{N} \tag*{(78)} D = { x n , j n , θ n , t n , α n } n = 1 N (78)
Here, n n n labels the individual homodyne samples in the data set. For a fixed choice of unknown model parameters Θ \Theta Θ , the Gaussian moment relations in Eqs. (22) and (23) are solved for every input state ρ α ( 0 ) \rho_{\alpha}(0) ρ α ( 0 ) . This gives the predicted mean vector d α ( t ) \mathbf{d}_{\alpha}(t) d α ( t ) and covariance matrix V α ( t ) V_{\alpha}(t) V α ( t ) at all measurement times.
We adopt a maximum-likelihood approach to estimate the unknown generator parameters Θ \Theta Θ from homodyne data [39 , 40 , 41 , 42 , 43 ] . For each trial value of Θ \Theta Θ , the Gaussian moment equations determine the predicted mean and variance of every measured quadrature, and hence the probability density for each homodyne outcome. The likelihood is defined as the joint probability density of obtaining the observed data set under these model predictions.
Assuming statistically independent homodyne shots conditioned on Θ \Theta Θ , the total likelihood is therefore given by the product of the single-shot probability densities
L ( Θ ∣ D ) = ∏ n = 1 N p ( x n ∣ j n , θ n , t n , α n ; Θ ) (79) \mathcal{L}(\Theta|\mathcal{D})=\prod_{n=1}^{N}p(x_n|j_n,\theta_n,t_n,\alpha_n;\Theta) \tag*{(79)} L ( Θ∣ D ) = n = 1 ∏ N p ( x n ∣ j n , θ n , t n , α n ; Θ ) (79)
For numerical optimization, we work with the negative log-likelihood. Equivalently, one minimizes the negative log-likelihood
J ( Θ ) = − ∑ n = 1 N log p ( x n ∣ j n , θ n , t n , α n ; Θ ) (80) \mathcal{J}(\Theta)=-\sum_{n=1}^{N}\log p(x_n|j_n,\theta_n,t_n,\alpha_n;\Theta) \tag*{(80)} J ( Θ ) = − n = 1 ∑ N log p ( x n ∣ j n , θ n , t n , α n ; Θ ) (80)
The maximum-likelihood estimator is therefore
Θ M L = arg min Θ J ( Θ ) (81) \Theta_{\mathrm{ML}}=\underset{\Theta}{\operatorname{arg\,min}}\ \mathcal{J}(\Theta) \tag*{(81)} Θ ML = Θ arg min J ( Θ ) (81)
where Θ M L \Theta_{\mathrm{ML}} Θ ML is the set of model parameters that gives the best fit to the homodyne data. However, reconstructing a time-dependent generator from finite data is an ill-conditioned inverse problem. So regularization can be used to obtain stable estimates [71 ] . Without additional constraints, many rapidly oscillating functions A ( t ) A(t) A ( t ) and D ( t ) D(t) D ( t ) may fit the same noisy data. Therefore, we introduce a regularized estimator for general generator reconstruction [41 , 71 ]
Θ r e g = arg min Θ [ J ( Θ ) + λ s R s m o o t h ( Θ ) + λ p R p h y s ( Θ ) ] (82) \Theta_{\mathrm{reg}}=\underset{\Theta}{\operatorname{arg\,min}}\left[\mathcal{J}(\Theta)+\lambda_sR_{\mathrm{smooth}}(\Theta)+\lambda_pR_{\mathrm{phys}}(\Theta)\right] \tag*{(82)} Θ reg = Θ arg min [ J ( Θ ) + λ s R smooth ( Θ ) + λ p R phys ( Θ ) ] (82)
where the first term, J ( Θ ) \mathcal{J}(\Theta) J ( Θ ) is the negative log-likelihood and measures how well the model explains the homodyne data; the term R s m o o t h ( Θ ) R_{\mathrm{smooth}}(\Theta) R smooth ( Θ ) penalizes unphysical rapid temporal variations of the reconstructed generator, while R p h y s ( Θ ) R_{\mathrm{phys}}(\Theta) R phys ( Θ ) penalizes violations of CP of the total Gaussian channel. The non-negative hyperparameters λ s \lambda_s λ s and λ p \lambda_p λ p control the relative importance of smoothness and physicality compared with data fitting. Thus, Eq. (82) estimates the most likely generator that is simultaneously consistent with the measured data, temporally regular, and physically valid.
We use a first-order Tikhonov regularizer [71 ] which penalizes the squared temporal derivatives of the reconstructed generator coefficients
R s m o o t h ( Θ ) = ∫ 0 T [ ∥ A ˙ ( t ) ∥ F 2 + ∥ D ˙ ( t ) ∥ F 2 + ∥ u ˙ ( t ) ∥ 2 2 ] d t (83) R_{\mathrm{smooth}}(\Theta)=\int_{0}^{T}\left[\left\|\dot{A}(t)\right\|_{F}^{2}+\left\|\dot{D}(t)\right\|_{F}^{2}+\left\|\dot{u}(t)\right\|_{2}^{2}\right]\,dt \tag*{(83)} R smooth ( Θ ) = ∫ 0 T [ A ˙ ( t ) F 2 + D ˙ ( t ) F 2 + ∥ u ˙ ( t ) ∥ 2 2 ] d t (83)
Here ∥ ⋅ ∥ F \|\cdot\|_{F} ∥ ⋅ ∥ F denotes the Frobenius norm, while ∥ ⋅ ∥ 2 \|\cdot\|_{2} ∥ ⋅ ∥ 2 denotes the Euclidean norm for vectors. This term encodes the physical assumption that hardware noise parameters do not vary arbitrarily fast compared with the experimental sampling time.
The second regularizer penalizes departures from physicality of the reconstructed Gaussian channel. Let X ( t ) X(t) X ( t ) and Y ( t ) Y(t) Y ( t ) be defined through Eqs. (26) and (28) . Moreover, a Gaussian channel is CP if Eq. (27) is satisfied. We therefore introduce the penalty
R p h y s ( Θ ) = ∑ k = 1 K ∥ [ − ( Y ( t k ) + i 2 Ω − i 2 X ( t k ) Ω X T ( t k ) ) ] + ∥ F 2 (84) R_{\mathrm{phys}}(\Theta)=\sum_{k=1}^{K}\left\|\left[-\left(Y(t_k)+\frac{i}{2}\Omega-\frac{i}{2}X(t_k)\Omega X^{T}(t_k)\right)\right]_{+}\right\|_{F}^{2} \tag*{(84)} R phys ( Θ ) = k = 1 ∑ K [ − ( Y ( t k ) + 2 i Ω − 2 i X ( t k ) Ω X T ( t k ) ) ] + F 2 (84)
where [ M ] + [M]_{+} [ M ] + denotes the positive part of the Hermitian matrix M M M . This penalty vanishes when the reconstructed map is CP at all sampled times.
For numerical implementation, the time-dependent quantities may be expanded in a finite set of basis functions
A ( t ) = ∑ ℓ = 1 N A A ℓ f ℓ ( t ) (85) A(t)=\sum_{\ell=1}^{N_A}A_{\ell}f_{\ell}(t) \tag*{(85)} A ( t ) = ℓ = 1 ∑ N A A ℓ f ℓ ( t ) (85)
D ( t ) = ∑ ℓ = 1 N D D ℓ g ℓ ( t ) (86) D(t)=\sum_{\ell=1}^{N_D}D_{\ell}g_{\ell}(t) \tag*{(86)} D ( t ) = ℓ = 1 ∑ N D D ℓ g ℓ ( t ) (86)
u ( t ) = ∑ ℓ = 1 N u u ℓ h ℓ ( t ) (87) \boldsymbol{u}(t)=\sum_{\ell=1}^{N_u}\boldsymbol{u}_{\ell}h_{\ell}(t) \tag*{(87)} u ( t ) = ℓ = 1 ∑ N u u ℓ h ℓ ( t ) (87)
where N A N_A N A , N D N_D N D , and N u N_u N u denote the numbers of basis functions used for the drift, diffusion, and displacement-drift terms, respectively and f ℓ ( t ) f_{\ell}(t) f ℓ ( t ) , g ℓ ( t ) g_{\ell}(t) g ℓ ( t ) , and h ℓ ( t ) h_{\ell}(t) h ℓ ( t ) are scalar basis functions and may be splines, Fourier modes, exponentials, or hardware-motivated memory functions. The unknown optimization vector is obtained by collecting the independent entries of the coefficient matrices and vectors
Θ = { A ℓ } ℓ = 1 N A ∪ { D ℓ } ℓ = 1 N D ∪ { u ℓ } ℓ = 1 N u (88) \Theta=\{A_{\ell}\}_{\ell=1}^{N_A}\cup\{D_{\ell}\}_{\ell=1}^{N_D}\cup\{\boldsymbol{u}_{\ell}\}_{\ell=1}^{N_u} \tag*{(88)} Θ = { A ℓ } ℓ = 1 N A ∪ { D ℓ } ℓ = 1 N D ∪ { u ℓ } ℓ = 1 N u (88)
together with any additional parameters appearing in the Hamiltonian or in the noise model. The coefficients A ℓ A_{\ell} A ℓ and u ℓ \boldsymbol{u}_{\ell} u ℓ are real, and each D ℓ D_{\ell} D ℓ is real symmetric. When fitting an unrestricted drift matrix A ( t ) A(t) A ( t ) , its coherent contribution is already included; hence G ( t ) G(t) G ( t ) is not fitted separately. A separate Hamiltonian parameterization requires a fixed decomposition of the drift to avoid redundancy.
The reconstruction is identifiable only to the extent allowed by the chosen probe states, measurement times, and homodyne directions. In practice, the input Gaussian states and measured quadratures must be sufficiently rich to distinguish the unknown parameters of the selected generator model. For this reason, we use a finite-dimensional, physically motivated parameterization of the time-dependent noise rates instead of attempting to reconstruct arbitrary matrix-valued functions A ( t ) A(t) A ( t ) , D ( t ) D(t) D ( t ) , and u ( t ) \boldsymbol{u}(t) u ( t ) . The reliability of the reconstruction is assessed through the conditioning of the likelihood landscape and through predictive model comparison on test synthetic data. The reconstructed time-local generator then provides the input for the CP-divisibility analysis carried out in the next section.
From Reconstructed Dynamics to Memory Detection
Diagnosing non-Markovianity from the reconstructed generator
After the tomographic reconstruction, the inferred time-local Gaussian generator is used to decide whether the observed dynamics is compatible with a Markovian Gaussian Lindblad description or requires a non-Markovian interpretation. In this work, non-Markovianity is diagnosed at the generator level through CP-divisibility, rather than through state-dependent distinguishability measures.
The first step is to verify that the reconstructed finite-time Gaussian channel is physical. This is done by checking the CP condition for the total Gaussian map, given in Eq. (27) at all sampled times. This global physicality check is necessary because a temporary violation of divisibility is meaningful only if the overall reconstructed map remains CP. The second step is to test the infinitesimal CP-divisibility condition using Eq. (30) . If this condition is satisfied for all times, the reconstructed dynamics is consistent with a CP-divisible Gaussian Lindblad process. If, however, this condition is violated over a finite time interval while the total Gaussian channel still satisfies CP condition, then the dynamics is identified as non-Markovian in the CP-divisibility sense [23 , 52 ] . Operationally, we compute the eigenvalues of the local CP matrix M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) defined in Eq. (30) . Negative eigenvalues indicate time intervals in which the intermediate map from t t t to t + d t t+dt t + d t is not CP.
To quantify the total strength of the violation, we introduce a CP-divisibility non-Markovianity score, inspired by the Gaussian-channel non-Markovianity analysis of Torre et al. [30 ]
N C P = ∫ 0 T ∑ λ ℓ ( t ) < 0 ∣ λ ℓ ( t ) ∣ d t (89) \mathcal{N}_{\mathrm{CP}}=\int_{0}^{T}\sum_{\lambda_{\ell}(t)<0}\left|\lambda_{\ell}(t)\right|\,dt \tag*{(89)} N CP = ∫ 0 T λ ℓ ( t ) < 0 ∑ ∣ λ ℓ ( t ) ∣ d t (89)
Here λ ℓ ( t ) \lambda_{\ell}(t) λ ℓ ( t ) are the real eigenvalues of the Hermitian matrix M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) , ℓ = 1 , … , 2 m \ell=1,\ldots,2m ℓ = 1 , … , 2 m labels the eigenvalues at each time and T T T is the final reconstruction time. At each time, the sum includes only the negative eigenvalues. Therefore, N C P = 0 \mathcal{N}_{\mathrm{CP}}=0 N CP = 0 for CP-divisible Gaussian dynamics, whereas N C P > 0 \mathcal{N}_{\mathrm{CP}}>0 N CP > 0 quantifies an accumulated violation of CP-divisibility over a time set of nonzero duration.
For discretely sampled data, Eq. (89) is evaluated numerically as a sum over the reconstruction times
N C P ≃ ∑ k = 1 K − 1 Δ t k ∑ λ ℓ ( t k ) < 0 ∣ λ ℓ ( t k ) ∣ , Δ t k = t k + 1 − t k (90) \mathcal{N}_{\mathrm{CP}}\simeq\sum_{k=1}^{K-1}\Delta t_k\sum_{\lambda_{\ell}(t_k)<0}\left|\lambda_{\ell}(t_k)\right|,\qquad\Delta t_k=t_{k+1}-t_k \tag*{(90)} N CP ≃ k = 1 ∑ K − 1 Δ t k λ ℓ ( t k ) < 0 ∑ ∣ λ ℓ ( t k ) ∣ , Δ t k = t k + 1 − t k (90)
where Δ t k \Delta t_k Δ t k is the time step associated with the sampled point t k t_k t k and k k k is the index labeling the discrete reconstruction times. This quantity measures the accumulated negative part of the spectrum of M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) .
In numerical reconstructions, small negative eigenvalues may arise from finite sampling noise or optimization error. Therefore, a non-Markovian interpretation should be assigned only when the violation of Eq. (30) , or equivalently the positive value of N C P \mathcal{N}_{\mathrm{CP}} N CP is larger than the estimated reconstruction uncertainty. In practice, this can be assessed using bootstrap resampling, confidence intervals, or repeated synthetic data sets. Here, the reconstruction uncertainty refers to the statistical and numerical uncertainty in the reconstructed eigenvalues of M C P ( t ) \mathcal{M}_{\mathrm{CP}}(t) M CP ( t ) , arising from finite homodyne sampling, regularization and numerical optimization.
Candidate Markovian and non-Markovian Gaussian noise models
The tomography procedure described above reconstructs an effective time-local Gaussian noise generator from time-resolved homodyne data. However, observing time-dependent noise parameters does not automatically imply that the dynamics is genuinely non-Markovian. A time-dependent but CP-divisible generator may still describe a memoryless process with externally varying control parameters. Therefore, it is important to compare a collection of candidate models with increasing physical complexity.
The purpose of this subsection is to compare candidate Gaussian noise models with different levels of physical complexity and to determine which model gives the best predictive description of the observed or simulated homodyne quadrature data without overfitting. We consider four models. The first is a stationary Markovian baseline. The second allows time-dependent rates but still enforces CP-divisibility. The third allows Lindblad-like non-Markovian dynamics through temporary violation of the infinitesimal CP condition. The fourth uses an enlarged Markovian system with auxiliary memory modes, thereby providing a microscopic interpretation of the observed memory effects.
Model M0: Stationary CP-divisible Gaussian noise
The simplest model assumes that the noise is stationary and memoryless. In this case, the Gaussian generator is time independent:
A ( t ) = A 0 , D ( t ) = D 0 , u ( t ) = u 0 (91) A(t) = A_0,\qquad D(t) = D_0,\qquad u(t) = u_0 \tag*{(91)} A ( t ) = A 0 , D ( t ) = D 0 , u ( t ) = u 0 (91)
The infinitesimal CP condition must hold (M C P ≥ 0 \mathcal{M}_{\mathrm{CP}} \ge0 M CP ≥ 0 ). This model corresponds to ordinary Markovian Lindblad noise with constant photon-loss, heating, and diffusion rates. It is the appropriate description when the environment has no memory and the hardware parameters are stable over the experimental time window. This model should be accepted if it already explains the data within statistical uncertainty. If it fails, then either the noise is time-dependent, correlated or non-Markovian.
Model M1: Time-dependent CP-divisible Gaussian noise
The second model allows the generator to vary in time
A ( t ) , D ( t ) , u ( t ) time dependent M C P ( t ) ≥ 0 for all t ∈ [ 0 , T ] (92) \begin{aligned}
&A(t),\ D(t),\ u(t)\quad\text{time dependent}\\
&\mathcal{M}_{\mathrm{CP}}(t) \ge0\quad\text{for all }t \in[0,T] \tag*{(92)}
\end{aligned} A ( t ) , D ( t ) , u ( t ) time dependent M CP ( t ) ≥ 0 for all t ∈ [ 0 , T ] (92)
but still requires the infinitesimal CP condition at all times. This model describes noise whose strength changes in time but remains CP-divisible. Physically, it can represent slow calibration drift, controlled changes in optical loss, or externally modulated Gaussian noise. Although the coefficients depend on time, the dynamics can still be interpreted as Markovian in the CP-divisibility sense because every intermediate map is CP. This model is useful for separating ordinary time-dependent hardware drift from genuine memory effects. If M 1 M_1 M 1 fits the data well, there is no need to invoke non-Markovian information backflow.
Model M2: Time-local non-CP-divisible Gaussian noise
The third model allows the reconstructed generator to violate CP-divisibility:
A ( t ) , D ( t ) , u ( t ) time dependent , M C P ( t ) ≱ 0 for some t ∈ [ 0 , T ] (93) \begin{aligned}
&A(t),\ D(t),\ u(t)\quad\text{time dependent},\\
&\mathcal{M}_{\mathrm{CP}}(t) \not\ge0\quad\text{for some }t \in[0,T] \tag*{(93)}
\end{aligned} A ( t ) , D ( t ) , u ( t ) time dependent , M CP ( t ) ≥ 0 for some t ∈ [ 0 , T ] (93)
Equivalently, in the operator form of the generator, the Kossakowski-like matrix may have temporarily negative eigenvalues, as in Eq. (19) . This model is called Lindblad-like because it retains the operator structure of a Lindblad generator, but relaxes the requirement that all instantaneous rates be positive. Physically, it can describe optical modes coupled to a structured reservoir or a delayed feedback path, where temporary storage and return of excitations can produce non-CP-divisible reduced dynamics. The total map from the initial time to the final time must still remain physical, satisfying Eq. (27) . Therefore, M 2 M_2 M 2 describes a situation in which the full evolution is CP, while the intermediate evolution is not CP-divisible. In the present framework, this combination of a physical finite-time Gaussian channel with a non-CP-divisible time-local generator is the main signature of non-Markovian memory. This model should be preferred over M 1 M_1 M 1 only if the data require negative instantaneous rates or violation of Eq. (30) to obtain a significantly better fit.
Note that Eq. (93) defines the non-CP-divisible regime. In the benchmarks, the label M 2 M_2 M 2 denotes the broader finite-time CP fitting family in which local CP violations are allowed but not imposed; a fitted member belongs to the regime of Eq. (93) only when such a violation occurs.
Model M3: Pseudomode-embedded Markovian model
The fourth model gives a microscopic explanation for non-Markovianity by embedding the photonic system into a larger Markovian system. The physical photonic modes a j a_j a j are coupled to auxiliary memory modes b ℓ b_\ell b ℓ through Eq. (68) . The enlarged density operator ϱ ( t ) \varrho(t) ϱ ( t ) evolves according to the Markovian Lindblad equation in Eq. (69) , while the observed photonic state is obtained by tracing out the memory modes, ρ ( t ) = Tr b [ ϱ ( t ) ] \rho(t)=\operatorname{Tr}_{b}[\varrho(t)] ρ ( t ) = Tr b [ ϱ ( t )] . Although the enlarged dynamics is Markovian, the reduced photonic dynamics may be non-Markovian. This model is physically motivated when the photonic circuit couples to auxiliary degrees of freedom, such as ring resonators, delay modes, or spectral side modes, that retain information about its earlier evolution. Their subsequent influence on the circuit can generate memory effects, although their presence alone does not guarantee non-CP-divisible dynamics. Compared with the phenomenological time-local model M 2 M_2 M 2 , the pseudomode model M 3 M_3 M 3 is more structured because it attributes the observed memory effects to explicit auxiliary modes.
Statistical model selection
After fitting the candidate models, we compare them using predictive performance and information criteria. This step is necessary because a more flexible time-dependent generator can always improve the training likelihood, even when the improvement is not statistically meaningful. Model selection therefore determines whether the data justify moving from a stationary Markovian description to a time-dependent CP-divisible model, or further to a non-CP-divisible time-local model.
Let Θ M \Theta_M Θ M be the fitted parameters of model M M M , ν M \nu_M ν M denotes the number of fitted parameters in each model, N N N is the number of homodyne samples, and J M ( Θ M ) \mathcal{J}_M(\Theta_M) J M ( Θ M ) denotes the negative log-likelihood of the data for model M M M in Eq. (80) , evaluated at the fitted parameters and excluding the regularization penalties. We use the Akaike information criterion (AIC) and Bayesian information criterion (BIC) as standard model-selection criteria that penalize goodness of fit by the number of fitted parameters [72 , 73 , 74 , 75 ]
A I C M = 2 ν M + 2 J M ( Θ M ) (94) \mathrm{AIC}_M = 2\nu_M + 2\mathcal{J}_M(\Theta_M) \tag*{(94)} AIC M = 2 ν M + 2 J M ( Θ M ) (94)
B I C M = ν M log N + 2 J M ( Θ M ) (95) \mathrm{BIC}_M = \nu_M \log N + 2\mathcal{J}_M(\Theta_M) \tag*{(95)} BIC M = ν M log N + 2 J M ( Θ M ) (95)
Lower values of A I C M \mathrm{AIC}_M AIC M and B I C M \mathrm{BIC}_M BIC M indicate a better balance between how well the model fits the data and the size of its parameterization. The regularizers influence the fitted parameters Θ M \Theta_M Θ M and therefore indirectly affect the value of J M ( Θ M ) \mathcal{J}_M(\Theta_M) J M ( Θ M ) . However, their numerical values are not added directly to the information criteria: A I C M \mathrm{AIC}_M AIC M and B I C M \mathrm{BIC}_M BIC M use the data negative log-likelihood evaluated at the fitted parameters, together with their respective ν M \nu_M ν M -dependent terms.
We also evaluate the predictive negative log-likelihood on held-out test data
ϵ M t e s t = 1 N t e s t J M t e s t ( Θ M ) = 1 N t e s t ∑ n ∈ t e s t − log p ( x n ∣ j n , θ n , t n , α n ; Θ M ) (96) \begin{aligned}
\epsilon_M^{\mathrm{test}} = \frac{1}{N_{\mathrm{test}}}\mathcal{J}_M^{\mathrm{test}}(\Theta_M) \\
&= \frac{1}{N_{\mathrm{test}}}\sum_{n\in\mathrm{test}}-\log p(x_n\mid j_n,\theta_n,t_n,\alpha_n;\Theta_M) \tag*{(96)}
\end{aligned} ϵ M test = N test 1 J M test ( Θ M ) = N test 1 n ∈ test ∑ − log p ( x n ∣ j n , θ n , t n , α n ; Θ M ) (96)
Here p p p is the model-predicted probability density for one homodyne outcome x n x_n x n and ϵ M t e s t \epsilon_M^{\mathrm{test}} ϵ M test is the average negative log-likelihood per test sample. A smaller value means that the fitted model assigns, on average, higher probability to homodyne outcomes that were not used during parameter estimation, and therefore has better predictive performance. This comparison helps determine whether an improvement on the training data reflects genuine predictive structure or merely overfitting.
Considering these criteria, we compare the candidate Gaussian noise models according to both their statistical support and their physical interpretation.
Algorithm for non-Markovian Gaussian noise reconstruction
The reconstruction framework introduced above defines the inverse problem at the level of homodyne probability distributions and Gaussian moment dynamics. To apply this framework to simulated or experimental data, we summarize the procedure as an explicit algorithm. The purpose of the algorithm is to convert time-resolved quadrature measurements into an estimate of the time-local Gaussian generator, and then to test whether the reconstructed generator is compatible with CP-divisible Markovian dynamics or shows signatures of non-CP-divisible non-Markovian behavior.
Algorithm 1 summarizes the operational implementation of the proposed tomographic reconstruction workflow. The algorithm has three conceptual stages. First, one specifies the probe states, measurement times, homodyne angles, and a parameterized model for the unknown Gaussian generator. Second, for each trial parameter set, the Gaussian moment equations are solved and the likelihood of the observed data is evaluated. Third, the optimized generator is checked for physical consistency of the finite-time Gaussian channel and then used to test CP-divisibility through M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) and the corresponding non-Markovianity score N C P \mathcal{N}_{\mathrm{CP}} N CP .
Numerical validation
We validate the framework using synthetic homodyne data generated from known Gaussian dynamics. The benchmarks represent distinct noise mechanisms rather than a specific device, allowing direct comparison with the reference dynamics and testing the distinction between time-dependent CP-divisible noise and non-CP-divisible memory. In these benchmarks, source attribuAlgorithm 1: Tomographic reconstruction and CP-divisibility diagnosis of Gaussian noise dynamics
Choose a set of Gaussian input probe states { ρ α ( 0 ) } \{\rho_\alpha(0)\} { ρ α ( 0 )} .
Choose measurement times { t k } \{t_k\} { t k } and homodyne settings { X j , θ } \{X_j,\theta\} { X j , θ } .
Collect the time-resolved homodyne data D = { x n , j n , θ n , t n , α n } n = 1 N D=\{x_n,j_n,\theta_n,t_n,\alpha_n\}_{n=1}^{N} D = { x n , j n , θ n , t n , α n } n = 1 N .
Split D D D into training and test sets, D t r D_{\mathrm{tr}} D tr and D t e s t D_{\mathrm{test}} D test .
for each candidate model M ∈ { M 0 , M 1 , M 2 , M 3 } M \in\{M_0,M_1,M_2,M_3\} M ∈ { M 0 , M 1 , M 2 , M 3 } do
Parameterize the candidate dynamics through the model parameters Θ M \Theta_M Θ M , either directly through A M ( t ) A_M(t) A M ( t ) , D M ( t ) D_M(t) D M ( t ) and u M ( t ) u_M(t) u M ( t ) , or through an equivalent reduced Gaussian channel or pseudomode embedding.
For each numerical evaluation of Θ M \Theta_M Θ M , propagate d α ( t ) d_\alpha(t) d α ( t ) and V α ( t ) V_\alpha(t) V α ( t ) using Eqs. (22) and (23) , or an analytically equivalent Gaussian-channel propagation.
Compute the predicted homodyne means and variances using Eqs. (76) and (77) .
Evaluate the training negative log-likelihood in Eq. (80) .
Minimize the training objective, using Eq. (82) when regularization is required, and impose the physical constraints of the selected model.
Verify CP of the reconstructed finite-time channel using Eq. (27) .
Compute M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) using Eq. (30) and evaluate the CP-divisibility witness λ min [ M C P ( t ) ] \lambda_{\min}[M_{\mathrm{CP}}(t)] λ m i n [ M CP ( t )] .
Compute A I C M \mathrm{AIC}_M AIC M , B I C M \mathrm{BIC}_M BIC M and ϵ M t e s t \epsilon^{\mathrm{test}}_M ϵ M test using Eqs. (94) , (95) and (96) , respectively.
end for
Compare the candidate models using A I C M \mathrm{AIC}_M AIC M , B I C M \mathrm{BIC}_M BIC M and ϵ M t e s t \epsilon^{\mathrm{test}}_M ϵ M test , and select the best-supported physically valid model.
Estimate reconstruction uncertainty by generating parametric bootstrap samples from the fitted selected model, refitting its parameters and constructing confidence intervals for λ min [ M C P ( t ) ] \lambda_{\min}[M_{\mathrm{CP}}(t)] λ m i n [ M CP ( t )] .
Report the selected model.
tion is conditional on the specified model families and environmental assumptions; it does not establish a unique microscopic identification of the noise mechanism. Candidate models are fitted to training data and compared using information criteria and predictive performance on independent test data. Numerical details are provided in Appendix B. All simulations, data generation, parameter fitting, and bootstrap analyses were implemented in Python using NumPy and SciPy.
Benchmark 1: single-mode squeezed state under memory-dependent loss
The first benchmark considers a single-mode squeezed vacuum state with zero displacement and covariance matrix [6 ]
V ( 0 ) = 1 2 ( e − 2 r 0 0 e 2 r ) (97) V(0)=\frac{1}{2}\begin{pmatrix}e^{-2r} & 0 \\ 0 & e^{2r}\end{pmatrix} \tag*{(97)} V ( 0 ) = 2 1 ( e − 2 r 0 0 e 2 r ) (97)
The squeezing parameter r = 0.80 r=0.80 r = 0.80 is chosen as a benchmark input parameter rather than fitted from the data and corresponds to approximately 10 log 10 ( e 2 r ) ≃ 6.95 d B 10\log_{10}(e^{2r})\simeq6.95\,\mathrm{dB} 10 log 10 ( e 2 r ) ≃ 6.95 dB of quadrature squeezing, giving a moderately squeezed input state whose variance is sensitive to memory-dependent loss.
Candidate models
The reference loss rate is chosen as a positive background loss rate κ 0 \kappa_0 κ 0 plus a damped oscillatory correction with amplitude α κ \alpha_\kappa α κ , memory decay time τ c \tau_c τ c , and oscillation frequency ω c \omega_c ω c . A physical motivation is an optical mode subject to background loss and coupled to a damped auxiliary resonator: coherent excitation exchange can produce oscillations in the effective loss rate, while resonator damping suppresses these oscillations over time. The coupling between the modes need not vary in time; a time-dependent reduced loss rate can arise even for constant coupling. We represent this behavior through the following phenomenological effective loss rate, without assuming an exact correspondence to a specific mode-resonator system [18 ]
κ r e f ( t ) = κ 0 + α κ e − t / τ c cos ( ω c t ) (98) \kappa_{\mathrm{ref}}(t)=\kappa_0+\alpha_\kappa e^{-t/\tau_c}\cos(\omega_c t) \tag*{(98)} κ ref ( t ) = κ 0 + α κ e − t / τ c cos ( ω c t ) (98)
with
κ 0 = 0.08 , α κ = 0.18 , τ c = 3.0 , ω c = 2.2 (99) \kappa_0=0.08,\qquad\alpha_\kappa=0.18,\qquad\tau_c=3.0,\qquad\omega_c=2.2 \tag*{(99)} κ 0 = 0.08 , α κ = 0.18 , τ c = 3.0 , ω c = 2.2 (99)
The parameters are chosen such that κ r e f ( t ) \kappa_{\mathrm{ref}}(t) κ ref ( t ) becomes temporarily negative between its two zero crossings, producing a violation of CP-divisibility, while the accumulated loss (defined below) remains non-negative so that the finite-time Gaussian channel remains CP and physical. This non-negativity is verified over the complete observation interval by checking its endpoints and all stationary points of the accumulated loss.
The time-dependent loss rate produces the accumulated loss
Γ r e f ( t ) = ∫ 0 t κ r e f ( s ) d s (100) \Gamma_{\mathrm{ref}}(t)=\int_0^t \kappa_{\mathrm{ref}}(s)\,\mathrm{d}s \tag*{(100)} Γ ref ( t ) = ∫ 0 t κ ref ( s ) d s (100)
and the transmissivity of the finite-time pure loss channel is therefore
η r e f ( t ) = e − Γ r e f ( t ) , 0 ≤ η r e f ( t ) ≤ 1 (101) \eta_{\mathrm{ref}}(t)=e^{-\Gamma_{\mathrm{ref}}(t)},\qquad0\leq\eta_{\mathrm{ref}}(t)\leq1 \tag*{(101)} η ref ( t ) = e − Γ ref ( t ) , 0 ≤ η ref ( t ) ≤ 1 (101)
A pure-loss Gaussian channel is equivalent to mixing the input mode with a vacuum environmental mode at a beam splitter of transmissivity η ( t ) \eta(t) η ( t ) . Using V v a c = I 2 / 2 V_{\mathrm{vac}} = I_2/2 V vac = I 2 /2 , the covariance matrix therefore evolves as [6 ]
V ( t ) = η ( t ) V ( 0 ) + [ 1 − η ( t ) ] I 2 2 (102) V(t) = \eta(t)V(0) + [1-\eta(t)]\frac{I_2}{2} \tag*{(102)} V ( t ) = η ( t ) V ( 0 ) + [ 1 − η ( t )] 2 I 2 (102)
To make the comparison between the time-dependent CP-divisible and non-CP-divisible descriptions as controlled as possible, models M 1 M_1 M 1 and M 2 M_2 M 2 are assigned the same parameter temporal family
κ M ( t ) = κ 0 , M + α M e − t / τ M cos ( ω M t ) (103) \kappa_M(t) = \kappa_{0,M} + \alpha_M e^{-t/\tau_M}\cos(\omega_M t) \tag*{(103)} κ M ( t ) = κ 0 , M + α M e − t / τ M cos ( ω M t ) (103)
Thus, M 1 M_1 M 1 and M 2 M_2 M 2 differ only through their physical constraints rather than through different fitting functions. The time-dependent CP-divisible model M 1 M_1 M 1 uses Eq. (103) subject to
κ M 1 ( t ) ≥ 0 for all t ∈ [ 0 , T ] (104) \kappa_{M_1}(t) \ge0 \quad\text{for all } t \in[0,T] \tag*{(104)} κ M 1 ( t ) ≥ 0 for all t ∈ [ 0 , T ] (104)
The time-local model M 2 M_2 M 2 uses the same Eq. (103) , but allows the instantaneous loss rate to take either positive or negative values while enforcing finite-time physicality.
Γ M 2 ( t ) = ∫ 0 t κ M 2 ( s ) d s ≥ 0 for all t ∈ [ 0 , T ] (105) \Gamma_{M_2}(t) = \int_0^t \kappa_{M_2}(s)\,ds \ge0 \quad\text{for all } t \in[0,T] \tag*{(105)} Γ M 2 ( t ) = ∫ 0 t κ M 2 ( s ) d s ≥ 0 for all t ∈ [ 0 , T ] (105)
Models M 1 M_1 M 1 and M 2 M_2 M 2 share the same temporal ansatz and parameter count, but allowing negative rates enlarges the admissible parameter region of M 2 M_2 M 2 . Consequently, an improved likelihood indicates better relative model support but does not independently establish the statistical necessity of negative rates. Since M 1 M_1 M 1 is nested within M 2 M_2 M 2 with a boundary constraint, the information criteria are interpreted as descriptive model comparisons rather than formal significance tests.
The stationary model M 0 M_0 M 0 assumes a constant Markovian non-negative loss rate
κ M 0 ( t ) = κ M 0 (106) \kappa_{M_0}(t) = \kappa_{M_0} \tag*{(106)} κ M 0 ( t ) = κ M 0 (106)
Model M 3 M_3 M 3 is conceptually different from M 0 M_0 M 0 –M 2 M_2 M 2 . It does not postulate a phenomenological time-dependent loss rate for the observed mode. Instead, the observed photonic mode a a a is coupled to an auxiliary bosonic memory mode b b b and the joint system is assumed to evolve according to a Markovian Lindblad equation. This is the standard pseudomode mechanism by which non-Markovian reduced dynamics can arise from Markovian dynamics on an enlarged Hilbert space [67 , 68 , 69 ] .
For Benchmark 1, we use one resonant auxiliary mode, initially in the vacuum state and based on Eq. (69) we write
d ϱ ( t ) d t = − i [ g ( a † b + a b † ) , ϱ ( t ) ] + γ m D [ b ] ϱ ( t ) (107) \frac{d\varrho(t)}{dt}=-i\left[g\left(a^\dagger b+ab^\dagger\right),\varrho(t)\right]+\gamma_m\mathcal{D}[b]\varrho(t) \tag*{(107)} d t d ϱ ( t ) = − i [ g ( a † b + a b † ) , ϱ ( t ) ] + γ m D [ b ] ϱ ( t ) (107)
where ϱ ( t ) \varrho(t) ϱ ( t ) is the joint density operator, γ m ≥ 0 \gamma_m \ge0 γ m ≥ 0 is the damping rate of the auxiliary memory mode, and g ≥ 0 g \ge0 g ≥ 0 is the real coherent system-memory coupling. The decay rates in Eq. (107) are non-negative, so the joint two-mode dynamics is Markovian and CP.
The corresponding Heisenberg-Langevin equations are [66 ]
a ˙ ( t ) = − i g b ( t ) (108) \dot{a}(t)=-igb(t) \tag*{(108)} a ˙ ( t ) = − i g b ( t ) (108)
b ˙ ( t ) = − γ m 2 b ( t ) − i g a ( t ) + γ m b i n ( t ) (109) \dot{b}(t)=-\frac{\gamma_m}{2}b(t)-iga(t)+\sqrt{\gamma_m}\,b_{\mathrm{in}}(t) \tag*{(109)} b ˙ ( t ) = − 2 γ m b ( t ) − i g a ( t ) + γ m b in ( t ) (109)
where b i n ( t ) b_{\mathrm{in}}(t) b in ( t ) is the input-field operator of the independent Markovian vacuum reservoirs coupled to the pseudomodes. Because the input reservoirs are in vacuum, their mean values vanish. The first moments therefore obey
d d t ( ⟨ a ( t ) ⟩ ⟨ b ( t ) ⟩ ) = P ( ⟨ a ( t ) ⟩ ⟨ b ( t ) ⟩ ) (110) \frac{d}{dt}
\begin{pmatrix}
\langle a(t)\rangle\\
\langle b(t)\rangle
\end{pmatrix}
=
\mathbf{P}
\begin{pmatrix}
\langle a(t)\rangle\\
\langle b(t)\rangle
\end{pmatrix} \tag*{(110)} d t d ( ⟨ a ( t )⟩ ⟨ b ( t )⟩ ) = P ( ⟨ a ( t )⟩ ⟨ b ( t )⟩ ) (110)
where
P = ( 0 − i g − i g − γ m / 2 ) (111) \mathbf{P}=
\begin{pmatrix}
0 & -ig\\
-ig & -\gamma_m/2
\end{pmatrix} \tag*{(111)} P = ( 0 − i g − i g − γ m /2 ) (111)
Since the auxiliary mode is initially in vacuum, ⟨ b ( 0 ) ⟩ = 0 \langle b(0)\rangle=0 ⟨ b ( 0 )⟩ = 0 , the solution of Eq. (110) gives
⟨ a ( t ) ⟩ = χ 3 ( t ) ⟨ a ( 0 ) ⟩ , χ 3 ( t ) = [ e P t ] 11 (112) \langle a(t)\rangle=\chi_3(t)\langle a(0)\rangle,\qquad\chi_3(t)=\left[e^{\mathbf{P}t}\right]_{11} \tag*{(112)} ⟨ a ( t )⟩ = χ 3 ( t ) ⟨ a ( 0 )⟩ , χ 3 ( t ) = [ e P t ] 11 (112)
Here, the subscript 11 11 11 denotes the first-row, first-column element of the matrix exponential, corresponding to the transfer amplitude from the initial system mode a ( 0 ) a(0) a ( 0 ) back to the observed mode a ( t ) a(t) a ( t ) . For the resonant two-mode model in Eq. (111) , the matrix exponential can be evaluated analytically. Using Eqs. (108) and (109) , the auxiliary-mode amplitude ⟨ b ( t ) ⟩ \langle b(t)\rangle ⟨ b ( t )⟩ can be eliminated, giving
⟨ a ¨ ( t ) ⟩ + γ m 2 ⟨ a ˙ ( t ) ⟩ + g 2 ⟨ a ( t ) ⟩ = 0 (113) \langle\ddot{a}(t)\rangle+\frac{\gamma_m}{2}\langle\dot{a}(t)\rangle+g^2\langle a(t)\rangle=0 \tag*{(113)} ⟨ a ¨ ( t )⟩ + 2 γ m ⟨ a ˙ ( t )⟩ + g 2 ⟨ a ( t )⟩ = 0 (113)
The two characteristic exponents are − γ m / 4 ± Δ -\gamma_m/4\pm\Delta − γ m /4 ± Δ , where we define the pseudomode spectral discriminant
Δ = γ m 2 16 − g 2 (114) \Delta=\sqrt{\frac{\gamma_m^2}{16}-g^2} \tag*{(114)} Δ = 16 γ m 2 − g 2 (114)
which determines whether the embedded dynamics is overdamped or oscillatory. Using χ 3 ( 0 ) = 1 \chi_3(0)=1 χ 3 ( 0 ) = 1 and χ ˙ 3 ( 0 ) = 0 \dot{\chi}_3(0)=0 χ ˙ 3 ( 0 ) = 0 , which follow from ⟨ b ( 0 ) ⟩ = 0 \langle b(0)\rangle=0 ⟨ b ( 0 )⟩ = 0 , the general solution is
χ 3 ( t ) = e − γ m t / 4 [ cosh ( Δ t ) + γ m 4 Δ sinh ( Δ t ) ] (115) \chi_3(t)=e^{-\gamma_m t/4}\left[\cosh(\Delta t)+\frac{\gamma_m}{4\Delta}\sinh(\Delta t)\right] \tag*{(115)} χ 3 ( t ) = e − γ m t /4 [ cosh ( Δ t ) + 4Δ γ m sinh ( Δ t ) ] (115)
At critical damping (Δ = 0 \Delta=0 Δ = 0 ), this expression is understood through its continuous limit, χ 3 ( t ) = e − γ m t / 4 ( 1 + γ m t / 4 ) \chi_3(t)=e^{-\gamma_m t/4}(1+\gamma_m t/4) χ 3 ( t ) = e − γ m t /4 ( 1 + γ m t /4 ) . For g > ∣ γ m ∣ 4 g>\frac{\lvert\gamma_m\rvert}{4} g > 4 ∣ γ m ∣ , Δ \Delta Δ is imaginary. Writing
ω o s c = g 2 − γ m 2 16 (116) \omega_{\mathrm{osc}}=\sqrt{g^2-\frac{\gamma_m^2}{16}} \tag*{(116)} ω osc = g 2 − 16 γ m 2 (116)
as the oscillation frequency generated by the coupling between the system mode a a a and the pseudomode. So, Eq. (115) becomes
χ 3 ( t ) = e − γ m t / 4 [ cos ( ω o s c t ) + γ m 4 ω o s c sin ( ω o s c t ) ] (117) \chi_3(t) = e^{-\gamma_m t/4}\left[\cos(\omega_{\mathrm{osc}}t) + \frac{\gamma_m}{4\omega_{\mathrm{osc}}}\sin(\omega_{\mathrm{osc}}t)\right] \tag*{(117)} χ 3 ( t ) = e − γ m t /4 [ cos ( ω osc t ) + 4 ω osc γ m sin ( ω osc t ) ] (117)
Thus, the damped oscillations appearing in the reduced dynamics of M 3 M_3 M 3 arise from coherent excitation exchange with the auxiliary memory mode. This mode and its Markovian reservoir are initially in vacuum. The coupling between the system and memory modes is passive; since χ 3 ( t ) \chi_3(t) χ 3 ( t ) is the field-amplitude transfer coefficient in Eq. (112) , while transmissivity refers to the corresponding intensity attenuation, the reduced quantum-limited attenuation channel has transmissivity
η 3 ( t ) = ∣ χ 3 ( t ) ∣ 2 (118) \eta_3(t) = |\chi_3(t)|^2 \tag*{(118)} η 3 ( t ) = ∣ χ 3 ( t ) ∣ 2 (118)
Consequently, the covariance matrix of the observed mode evolves as Eq. (102) with η 3 ( t ) \eta_3(t) η 3 ( t ) . This is the quantity used to predict the homodyne variances and hence to evaluate the likelihood of model M 3 M_3 M 3 . To compare the embedded model with the reduced loss-rate descriptions M 0 M_0 M 0 –M 2 M_2 M 2 , we define the effective reduced loss rate based on Eq. (101)
κ 3 e f f ( t ) = − d d t ln η 3 ( t ) (119) \kappa_3^{\mathrm{eff}}(t) = -\frac{d}{dt}\ln\eta_3(t) \tag*{(119)} κ 3 eff ( t ) = − d t d ln η 3 ( t ) (119)
which, using Eq. (118) gives
κ 3 e f f ( t ) = − 2 Re [ χ ˙ 3 ( t ) χ 3 ( t ) ] (120) \kappa_3^{\mathrm{eff}}(t) = -2\operatorname{Re}\left[\frac{\dot{\chi}_3(t)}{\chi_3(t)}\right] \tag*{(120)} κ 3 eff ( t ) = − 2 Re [ χ 3 ( t ) χ ˙ 3 ( t ) ] (120)
Eq. (120) is defined on intervals for which χ 3 ( t ) ≠ 0 \chi_3(t) \ne0 χ 3 ( t ) = 0 . A temporary increase of the reduced transmissivity η ˙ 3 ( t ) > 0 \dot{\eta}_3(t) > 0 η ˙ 3 ( t ) > 0 is equivalent to κ 3 e f f ( t ) < 0 \kappa_3^{\mathrm{eff}}(t) < 0 κ 3 eff ( t ) < 0 , and therefore signals temporary information or excitation return from the memory mode to the observed mode. The reduced evolution can thus become non-CP-divisible even though the joint evolution in Eq. (107) remains a CP Markovian Lindblad dynamics. Importantly, κ 3 e f f ( t ) \kappa_3^{\mathrm{eff}}(t) κ 3 eff ( t ) is a derived quantity and is not assigned the damped-cosine form of Eq. (103) . Model M 3 M_3 M 3 is therefore statistically compared with M 0 M_0 M 0 –M 2 M_2 M 2 at the level of the same measured homodyne data, while representing a genuinely different physical hypothesis for the origin of the observed memory.
Fit results and uncertainty assessment
The resulting fits are compared using the information criteria and held-out predictive score introduced above, together with a bootstrap assessment of the reconstructed CP-divisibility violation. Each candidate model M 0 M_0 M 0 –M 3 M_3 M 3 was fitted to the training data by minimizing the training negative log-likelihood in Eq. (80) , using numerical optimization routines implemented in Python with NumPy/SciPy. For these low-dimensional benchmark models, no regularization is required and the parameters are obtained by minimizing the training negative log-likelihood subject to the model-specific physicality constraints. So, after optimization the unregularized training negative log-likelihood was used to compute the AIC and BIC values in Eqs. (94) and (95) , the corresponding relative scores Δ A I C \Delta\mathrm{AIC} Δ AIC and Δ B I C \Delta\mathrm{BIC} Δ BIC were obtained from Eq. (162) , and the predictive negative log-likelihood was evaluated independently on the held-out test data using Eq. (96) .
Table 1. Model comparison for benchmark 1. Relative AIC and BIC are reported with respect to the best model; ϵ M ( t e s t ) \epsilon_M^{(\mathrm{test})} ϵ M ( test ) is reported per test sample and ν M \nu_M ν M denotes the number of independent real parameters in model M M M .
The results in Table I were obtained from the numerical reconstruction pipeline described above in which ν M \nu_M ν M counts the fitted real parameters: κ 0 , M \kappa_{0,M} κ 0 , M for M 0 M_0 M 0 ; ( κ 0 , M , α M , τ M , ω M ) (\kappa_{0,M},\alpha_M,\tau_M,\omega_M) ( κ 0 , M , α M , τ M , ω M ) for M 1 M_1 M 1 and M 2 M_2 M 2 ; and ( g , γ m ) (g,\gamma_m) ( g , γ m ) for the resonant pseudomode model M 3 M_3 M 3 . As shown, the time-local model M 2 M_2 M 2 is selected by both information criteria, with Δ A I C = Δ B I C = 0 \Delta\mathrm{AIC}=\Delta\mathrm{BIC}=0 Δ AIC = Δ BIC = 0 , and also gives the lowest test score. Importantly, M 1 M_1 M 1 and M 2 M_2 M 2 use the same four-parameter damped-oscillatory family. The preference for M 2 M_2 M 2 is consistent with its ability to capture the temporary negative-rate interval excluded by M 1 M_1 M 1 , although this preference alone does not establish the statistical necessity of non-CP-divisible dynamics. The stationary model M 0 M_0 M 0 is strongly disfavored because it cannot reproduce the temporal memory structure. The pseudomode model M 3 M_3 M 3 is also strongly disfavored. Its fitted parameters place the pseudomode dynamics in the overdamped regime, so the effective loss builds up monotonically through coupling to the damped memory mode and does not reproduce the oscillatory negative-rate structure of the reference process. This is consistent with the fact that the synthetic data in this benchmark are generated from the signed time-local family rather than from a pseudomode-embedded dynamics. To assess the robustness of the reconstructed negative-rate interval against finite-sampling uncertainty, we perform a parametric bootstrap with N b o o t = 300 N_{\mathrm{boot}}=300 N boot = 300 . Each synthetic training data set is generated from the fitted M 2 M_2 M 2 channel, using the same probe states, measurement times, homodyne settings, and number of shots as in the original data. Model M 2 M_2 M 2 is independently refitted to each bootstrap realization, and the resulting distribution is used to construct confidence intervals for the reconstructed negative-rate interval. The 95% bootstrap interval for the minimum reconstructed rate is
min t κ M 2 ( t ) ∈ [ − 0.0494 , − 0.0165 ] (121) \min_t \kappa_{M_2}(t) \in[-0.0494,-0.0165] \tag*{(121)} t min κ M 2 ( t ) ∈ [ − 0.0494 , − 0.0165 ] (121)
which lies entirely below zero. Moreover, the pointwise 97.5% upper bootstrap bound remains negative approximately over
t ∈ [ 1.12 , 1.63 ] (122) t \in[1.12,1.63] \tag*{(122)} t ∈ [ 1.12 , 1.63 ] (122)
These bootstrap results support a robust negative-rate interval within the fitted M 2 M_2 M 2 model, although they do not constitute a calibrated hypothesis test against the CP-divisible model M 1 M_1 M 1 .
Figure 2 illustrates the memory-dependent loss dynamics at three different physical levels. Panel (a) compares the reference loss rate with the fitted models. The stationary Markovian model M 0 M_0 M 0 captures only the average attenuation, while the CP-divisible model M 1 M_1 M 1 follows part of the temporal variation but cannot enter the negative-rate region because of its positivity constraint. In contrast, the signed time-local model M 2 M_2 M 2 closely reconstructs the damped oscillatory reference profile, including the interval of negative loss rate. The pseudomode model M 3 M_3 M 3 is obtained from an explicit enlarged Markovian dynamics and, for this benchmark, does not reproduce the full oscillatory structure as accurately as M 2 M_2 M 2 , which is expected because the synthetic data are generated directly from the time-local reference model. The monotonic M 3 M_3 M 3 profile is physically consistent with the fitted overdamped pseudomode dynamics: the observed mode has no direct loss channel, and its effective attenuation develops only through coherent transfer to the damped pseudomode.
Figure 2. Benchmark 1: single-mode squeezed state under memory-dependent loss. (a) Reference loss rate in Eq. (98) , fitted M 0 M_0 M 0 –M 2 M_2 M 2 rates, and the effective reduced loss rate derived from the fitted pseudomode model M 3 M_3 M 3 . Models M 1 M_1 M 1 and M 2 M_2 M 2 use the same damped-cosine family, with M 1 M_1 M 1 constrained to remain CP-divisible. (b) Smallest eigenvalue of the infinitesimal Gaussian CP matrix. Negative values identify the interval of non-CP-divisible reduced dynamics; the light band shows the 95% pointwise bootstrap uncertainty for M 2 M_2 M 2 . (c) Squeezed-quadrature variance V q ( t ) V_q(t) V q ( t ) obtained from the corresponding finite-time transmissivities. The memory-dependent dynamics produces a small transient recovery of the variance during the interval of negative effective loss.
Panel (b) shows the corresponding local CP-divisibility witness λ min [ M C P ( t ) ] \lambda_{\min}[\mathrm{M}_{\mathrm{CP}}(t)] λ m i n [ M CP ( t )] . The witness becomes negative for the reference dynamics and for M 2 M_2 M 2 , identifying the interval in which the infinitesimal reduced evolution is not CP-divisible, while M 0 M_0 M 0 , M 1 M_1 M 1 , and M 3 M_3 M 3 remain non-negative. The vertical shaded region marks the time interval in which the reference loss rate is negative, whereas the shaded band around the M 2 M_2 M 2 witness represents its bootstrap uncertainty. The persistence of the negative region beyond this uncertainty supports a statistically resolved violation of CP-divisibility.
Finally, panel (c) shows the variance V q ( t ) V_q(t) V q ( t ) of the initially squeezed q q q quadrature, with V q ( 0 ) = e − 2 r / 2 V_q(0)=e^{-2r}/2 V q ( 0 ) = e − 2 r /2 ; for notational simplicity, this quantity is denoted by V ( t ) V(t) V ( t ) in the figure. For a pure-loss channel, V q ( t ) = η ( t ) V q ( 0 ) + [ 1 − η ( t ) ] / 2 V_q(t)=\eta(t)V_q(0)+[1-\eta(t)]/2 V q ( t ) = η ( t ) V q ( 0 ) + [ 1 − η ( t )] /2 . Since V q ( 0 ) < 1 / 2 V_q(0)<1/2 V q ( 0 ) < 1/2 , ordinary attenuation decreases η ( t ) \eta(t) η ( t ) and therefore increases V q ( t ) V_q(t) V q ( t ) toward the vacuum value 1 / 2 1/2 1/2 , corresponding to a gradual degradation of the initial squeezing. The reference dynamics and M 2 M_2 M 2 exhibit a small temporary decrease of V q ( t ) V_q(t) V q ( t ) during the negative-rate interval. This corresponds to an increase of the transmissivity and hence to a partial recovery of the initial squeezing. In contrast, M 0 M_0 M 0 gives a smooth monotonic degradation determined by its constant effective loss rate, while M 1 M_1 M 1 remains monotonic because its instantaneous loss rate is constrained to be non-negative. Model M 3 M_3 M 3 also shows no corresponding recovery in this benchmark. Its fitted overdamped pseudomode dynamics produces a monotonic effective attenuation, so the quadrature variance increases smoothly rather than reproducing the transient recovery present in the reference dynamics and M 2 M_2 M 2 .
Benchmark 2: nonstationary thermal excitation
Benchmark 1 tests whether the reconstruction can resolve a genuinely non-CP-divisible loss process. Here we consider the complementary specificity test: the noise is clearly nonstationary, but the local Gaussian generator remains CP-divisible throughout the observation window. This benchmark therefore tests whether the reconstruction distinguishes ordinary time-dependent thermal excitation from a genuine local breakdown of CP-divisibility.
Candidate models
The reference channel is the thermal-loss process introduced in Eq. (49) , with fixed damping rate κ t h = 0.20 \kappa_{\mathrm{th}} = 0.20 κ th = 0.20 and time-dependent bath occupation. A physical example is a microwave resonator coupled to a broadband thermal reservoir whose noise level is externally varied while the coupling strength remains constant. This changes the reservoir’s effective thermal occupation and hence the noise injected into the resonator, without changing its damping rate. Assuming that the reservoir correlation time remains short compared with the relevant evolution timescales, we model its occupation as [18 ]
n ˉ r e f ( t ) = n ˉ 0 , r e f + α r e f e − t / τ r e f cos ( ω r e f t ) \bar{n}_{\mathrm{ref}}(t) = \bar{n}_{0,\mathrm{ref}} + \alpha_{\mathrm{ref}} e^{-t/\tau_{\mathrm{ref}}} \cos(\omega_{\mathrm{ref}}t) n ˉ ref ( t ) = n ˉ 0 , ref + α ref e − t / τ ref cos ( ω ref t )
n ˉ r e f ( t ) = 0.30 + 0.25 e − t / 2.8 cos ( 1.8 t ) (123) \bar{n}_{\mathrm{ref}}(t) = 0.30 + 0.25e^{-t/2.8}\cos(1.8t) \tag*{(123)} n ˉ ref ( t ) = 0.30 + 0.25 e − t /2.8 cos ( 1.8 t ) (123)
No additional dynamical relation is required here: the covariance is propagated directly with Eq. (23) using the drift and diffusion terms already given in Eqs. (47) and (51) . The profile in Eq. (123) remains positive over the entire interval; numerically
min t ∈ [ 0 , T ] n ˉ r e f ( t ) ≃ 0.1633 (124) \min_{t\in[0,T]} \bar{n}_{\mathrm{ref}}(t) \simeq0.1633 \tag*{(124)} t ∈ [ 0 , T ] min n ˉ ref ( t ) ≃ 0.1633 (124)
For this thermal-loss channel, the two nonzero infinitesimal Kossakowski rates are κ t h [ n ˉ ( t ) + 1 ] \kappa_{\mathrm{th}}[\bar{n}(t)+1] κ th [ n ˉ ( t ) + 1 ] and κ t h n ˉ ( t ) \kappa_{\mathrm{th}}\bar{n}(t) κ th n ˉ ( t ) . Hence the smallest eigenvalue of the local Gaussian CP matrix is
λ min [ M C P ( t ) ] = κ t h n ˉ ( t ) (125) \lambda_{\min}[M_{\mathrm{CP}}(t)] = \kappa_{\mathrm{th}}\bar{n}(t) \tag*{(125)} λ m i n [ M CP ( t )] = κ th n ˉ ( t ) (125)
which is strictly positive for the reference process. Time dependence of the bath occupation therefore does not by itself imply non-Markovianity in the CP-divisibility sense [18 , 23 ] .
The candidate models follow the discussed hierarchy. Model M 0 M_0 M 0 uses a constant thermal occupation and therefore contains one fitted parameter. Model M 1 M_1 M 1 preserves the thermal-loss structure of Eq. (49) and uses
n ˉ M 1 ( t ) = n ˉ 0 , M 1 + α M 1 e − t / τ M 1 cos ( ω M 1 t ) , n ˉ M 1 ( t ) ≥ 0 (126) \bar{n}_{M_1}(t) = \bar{n}_{0,M_1} + \alpha_{M_1}e^{-t/\tau_{M_1}}\cos(\omega_{M_1}t), \qquad\bar{n}_{M_1}(t) \ge0 \tag*{(126)} n ˉ M 1 ( t ) = n ˉ 0 , M 1 + α M 1 e − t / τ M 1 cos ( ω M 1 t ) , n ˉ M 1 ( t ) ≥ 0 (126)
with four fitted parameters ( n ˉ 0 , M 1 , α M 1 , τ M 1 , ω M 1 ) (\bar{n}_{0,M_1},\alpha_{M_1},\tau_{M_1},\omega_{M_1}) ( n ˉ 0 , M 1 , α M 1 , τ M 1 , ω M 1 ) . Thus M 1 M_1 M 1 allows nonstationary thermal noise while enforcing local CP-divisibility. In model M 1 M_1 M 1 , the decay and excitation rates are not fitted independently. Because M 1 M_1 M 1 assumes a thermal-loss channel, both are determined by the same reconstructed bath occupation n ˉ M 1 ( t ) \bar{n}_{M_1}(t) n ˉ M 1 ( t ) with the damping rate κ t h \kappa_{\mathrm{th}} κ th fixed.
Model M 2 M_2 M 2 is the more general time-local, phase-insensitive Gaussian description. In contrast to M 1 M_1 M 1 , it does not impose the thermal relation between the loss and excitation processes. We denote by κ ↓ ( t ) \kappa_{\downarrow}(t) κ ↓ ( t ) the decay rate multiplying D [ a ] \mathcal{D}[a] D [ a ] , which describes photon loss from the system, and by κ ↑ ( t ) \kappa_{\uparrow}(t) κ ↑ ( t ) the excitation rate multiplying D [ a † ] \mathcal{D}[a^\dagger] D [ a † ] , which describes thermal excitation of the system by the environment. For an ordinary thermal bath, these rates are not independent but satisfy
κ ↓ ( t ) = κ t h [ n ˉ ( t ) + 1 ] , κ ↑ ( t ) = κ t h n ˉ ( t ) (127) \kappa_{\downarrow}(t) = \kappa_{\mathrm{th}}\left[\bar{n}(t)+1\right], \qquad\kappa_{\uparrow}(t) = \kappa_{\mathrm{th}}\bar{n}(t) \tag*{(127)} κ ↓ ( t ) = κ th [ n ˉ ( t ) + 1 ] , κ ↑ ( t ) = κ th n ˉ ( t ) (127)
as in the thermal-loss model introduced above. In M 2 M_2 M 2 , this constraint is relaxed and the two rates are reconstructed independently using
κ ↓ , M 2 ( t ) = κ ↓ , 0 + α ↓ e − t / τ cos ( ω t ) \kappa_{\downarrow,M_2}(t) = \kappa_{\downarrow,0} + \alpha_{\downarrow}e^{-t/\tau}\cos(\omega t) κ ↓ , M 2 ( t ) = κ ↓ , 0 + α ↓ e − t / τ cos ( ω t )
κ ↑ , M 2 ( t ) = κ ↑ , 0 + α ↑ e − t / τ cos ( ω t ) (128) \kappa_{\uparrow,M_2}(t) = \kappa_{\uparrow,0} + \alpha_{\uparrow}e^{-t/\tau}\cos(\omega t) \tag*{(128)} κ ↑ , M 2 ( t ) = κ ↑ , 0 + α ↑ e − t / τ cos ( ω t ) (128)
with the six fitted parameters ( κ ↓ , 0 , κ ↑ , 0 , α ↓ , α ↑ , τ , ω ) (\kappa_{\downarrow,0},\kappa_{\uparrow,0},\alpha_{\downarrow},\alpha_{\uparrow},\tau,\omega) ( κ ↓ , 0 , κ ↑ , 0 , α ↓ , α ↑ , τ , ω ) .
For this phase-insensitive generator, local CP-divisibility requires both instantaneous rates to be non-negative. Equivalently, the smallest eigenvalue of the local Gaussian CP matrix is
λ min [ M C P M 2 ( t ) ] = min { κ ↓ , M 2 ( t ) , κ ↑ , M 2 ( t ) } (129) \lambda_{\min}\left[M_{\mathrm{CP}}^{M_2}(t)\right] = \min\left\{\kappa_{\downarrow,M_2}(t),\kappa_{\uparrow,M_2}(t)\right\} \tag*{(129)} λ m i n [ M CP M 2 ( t ) ] = min { κ ↓ , M 2 ( t ) , κ ↑ , M 2 ( t ) } (129)
Model M 2 M_2 M 2 does not impose these pointwise positivity conditions; instead, the accumulated Gaussian channel is constrained to satisfy the finite-time CP condition in Eq. (27) . It can therefore represent a locally non-CP-divisible effective dynamics while remaining physical at every observation time. For the thermal subfamily corresponding to M 1 M_1 M 1 , one has κ ↓ = κ ↑ = κ t h \kappa_{\downarrow}=\kappa_{\uparrow}=\kappa_{\mathrm{th}} κ ↓ = κ ↑ = κ th and the above witness reduces to λ min [ M C P ( t ) ] = κ t h n ˉ ( t ) \lambda_{\min}[M_{\mathrm{CP}}(t)]=\kappa_{\mathrm{th}}\bar{n}(t) λ m i n [ M CP ( t )] = κ th n ˉ ( t ) , consistently with Eq. (125) . Crucially, model M 3 M_3 M 3 is not modified for this benchmark. We use exactly the resonant one-pseudomode Markovian embedding introduced in Benchmark 1 with fitted parameters ( g , γ m ) (g,\gamma_m) ( g , γ m ) . The auxiliary mode and its Markovian reservoir are in vacuum, exactly as in Benchmark 1; no thermal pseudomode or additional thermal parameter is introduced. Moreover, the reference data in the present benchmark contain no auxiliary mode: M 3 M_3 M 3 is included only as a competing physical hypothesis for the same homodyne data. Since M 3 M_3 M 3 is a quantum-limited attenuation model, it cannot in general reproduce the excess diffusion generated by a nonzero thermal occupation.
Fit results and uncertainty assessment
We use the same single-mode squeezed-vacuum probe as in Benchmark 1, with r = 0.80 r = 0.80 r = 0.80 , and the numerical protocol detailed in Appendix B: T = 10 T = 10 T = 10 , K = 80 K = 80 K = 80 , four calibrated homodyne phases, N s h o t = 2000 N_{\mathrm{shot}} = 2000 N shot = 2000 samples per time and phase, and an 80/20 training/test split.
The result in Table II shows that the time-dependent CP-divisible model M 1 M_1 M 1 is the best model by both information criteria. The stationary model M 0 M_0 M 0 is disfavored because it cannot resolve the temporal heating profile. The more general model M 2 M_2 M 2 obtains a slightly lower training negative log-likelihood, as expected from its additional freedom, but the improvement is too small to compensate for its two additional parameters. Importantly, its fitted local witness remains positive:
Table II. Model comparison for Benchmark 2. The data are generated from a nonstationary thermal excitation process. The table shows that the temporal structure in the data is explained without requiring a non-CP-divisible reduced generator.
min t λ min [ M C P M 2 ( t ) ] ≃ 0.0302 > 0 (130) \min_t \lambda_{\min}[M_{\mathrm{CP}}^{M_2}(t)] \simeq0.0302 > 0 \tag*{(130)} t min λ m i n [ M CP M 2 ( t )] ≃ 0.0302 > 0 (130)
Thus, the data do not support a non-CP-divisible interpretation. The pseudomode model M 3 M_3 M 3 is also strongly disfavored. Its fitted dynamics remains a quantum-limited vacuum attenuation process and therefore cannot reproduce the excess diffusion generated by the time-dependent thermal occupation in the reference model. This is physically consistent with the fact that the synthetic reference contains nonstationary thermal excitation rather than vacuum pseudomode memory.
To quantify finite-sampling uncertainty, we perform a parametric bootstrap with N b o o t = 300 N_{\mathrm{boot}} = 300 N boot = 300 . Synthetic training data sets are generated from the fitted M 1 M_1 M 1 channel using the original measurement settings and sample size. Model M 1 M_1 M 1 is independently refitted to each realization, and the resulting distribution is used to construct confidence intervals for the minimum local CP-divisibility witness. The resulting 95% bootstrap interval for the
minimum local CP witness is
min t λ min [ M C P M 1 ( t ) ] ∈ [ 0.0242 , 0.0413 ] (131) \min_t \lambda_{\min}[M_{\mathrm{CP}}^{M_1}(t)] \in[0.0242, 0.0413] \tag*{(131)} t min λ m i n [ M CP M 1 ( t )] ∈ [ 0.0242 , 0.0413 ] (131)
which remains entirely above zero. The benchmark results are consistent with the selected CP-divisible model.
Figure 3 uses the same three-panel structure as Benchmark 1, but here the panels emphasize the distinction between nonstationary thermal excitation and genuinely non-CP-divisible dynamics. Panel (a) compares the reference thermal occupation with the profiles reconstructed by the fitted models. The stationary model M 0 M_0 M 0 can reproduce only an effective average heating level, whereas M 1 M_1 M 1 follows the temporal modulation of the bath occupation while remaining within the CP-divisible thermal-loss sector. For M 2 M_2 M 2 , we plot the effective occupation n ˉ e f f ( t ) = κ ↑ ( t ) / [ κ ↓ ( t ) − κ ↑ ( t ) ] \bar{n}_{\mathrm{eff}}(t) = \kappa_{\uparrow}(t)/[\kappa_{\downarrow}(t) - \kappa_{\uparrow}(t)] n ˉ eff ( t ) = κ ↑ ( t ) / [ κ ↓ ( t ) − κ ↑ ( t )] wherever the denominator is nonzero. This quantity provides a direct way of comparing the more general time-local model with the thermal description, even though in M 2 M_2 M 2 the upward and downward rates are fitted independently. Model M 3 M_3 M 3 is not shown in panel (a), because it is the same vacuum-pseudomode model used in Benchmark 1 and therefore does not possess an independent thermal-occupation parameter.
Figure 3. Benchmark 2: reconstruction of nonstationary thermal excitation. (a) Reference thermal occupation and fitted model classes. The reference profile is time dependent but remains positive. For M 2 M_2 M 2 , the plotted thermal occupation is the effective quantity n ˉ e f f ( t ) = κ ↑ ( t ) / [ κ ↓ ( t ) − κ ↑ ( t ) ] \bar{n}_{\mathrm{eff}}(t) = \kappa_{\uparrow}(t)/[\kappa_{\downarrow}(t) - \kappa_{\uparrow}(t)] n ˉ eff ( t ) = κ ↑ ( t ) / [ κ ↓ ( t ) − κ ↑ ( t )] . Model M 3 M_3 M 3 is the same vacuum-pseudomode model as in Benchmark 1 and therefore has no thermal-occupation parameter. (b) Smallest eigenvalue of the local Gaussian CP matrix. The witness remains non-negative for the selected model M 1 M_1 M 1 ; the light green band shows its 95% pointwise bootstrap uncertainty. (c) Squeezed-quadrature variance V q ( t ) V_q(t) V q ( t ) obtained from the corresponding fitted dynamics. The M 1 M_1 M 1 and M 2 M_2 M 2 variance curves nearly overlap the reference.
Panel (b) shows the corresponding local CP-divisibility witness. The reference process and the selected model M 1 M_1 M 1 remain strictly non-negative throughout the observation window, demonstrating that the pronounced time dependence in panel (a) does not by itself imply non-Markovianity in the CP-divisibility sense. The light green region denotes the 95% pointwise bootstrap uncertainty for M 1 M_1 M 1 and remains above zero, showing that this conclusion is robust against finite-sampling fluctuations. The fitted M 2 M_2 M 2 trajectory also remains in the CP-divisible region, so its additional freedom to describe negative instantaneous rates is not required by the data. For M 3 M_3 M 3 , the smallest local CP eigenvalue vanishes provided that κ e f f M 3 ( t ) ≥ 0 \kappa_{\mathrm{eff}}^{M_3}(t) \ge0 κ eff M 3 ( t ) ≥ 0 ; this zero value reflects saturation of the local CP bound rather than the absence of dissipative evolution.
Panel (c) translates the reconstructed dynamics into the variance of the initially squeezed q q q -quadrature. More precisely, the plotted quantity is V q ( t ) V_q(t) V q ( t ) , denoted simply by V ( t ) V(t) V ( t ) in the figure, with V q ( 0 ) = e − 2 r / 2 V_q(0) = e^{-2r}/2 V q ( 0 ) = e − 2 r /2 . For the reference dynamics and M 1 M_1 M 1 , the covariance is propagated using Eq. (23) with the thermal-loss drift and diffusion terms in Eqs. (47) and (51) , where the time dependence enters through the reconstructed bath occupation n ˉ ( t ) \bar{n}(t) n ˉ ( t ) . Model M 0 M_0 M 0 uses the same thermal-loss evolution but with a constant fitted occupation. For M 2 M_2 M 2 , the decay and excitation rates κ ↓ ( t ) \kappa_{\downarrow}(t) κ ↓ ( t ) and κ ↑ ( t ) \kappa_{\uparrow}(t) κ ↑ ( t ) are reconstructed independently according to Eq. (128) , and the corresponding covariance evolution is obtained from these two rates without imposing the thermal-bath relation in Eq. (127) . In contrast, M 3 M_3 M 3 is the vacuum-pseudomode model of Benchmark 1 and therefore describes a quantum-limited attenuation channel; its variance is obtained from the pseudomode transmissivity in Eq. (118) using the covariance transformation in Eq. (102) . The reference dynamics and M 1 M_1 M 1 show the expected gradual increase of V ( t ) V(t) V ( t ) , corresponding to degradation of the initial squeezing by thermal excitation, together with a weak modulation inherited from the time-dependent bath occupation. Unlike Benchmark 1, the local rates remain non-negative and no transient recovery of the squeezed-quadrature variance occurs. Thus, the behavior in panel (c) is consistent with the CP-divisibility witness in panel (b): the observed time-dependent heating is explained by a nonstationary but CP-divisible thermal process, without requiring a non-CP-divisible reduced dynamics.
Benchmark 3: resonant two-pseudomode memory
Benchmarks 1 and 2 test, respectively, a phenomenological non-CP-divisible time-local loss process and a nonstationary but CP-divisible thermal process. We now consider a complementary microscopic benchmark in which the synthetic homodyne data are generated directly from the pseudomode embedding introduced in Eq. (68) . The purpose of this benchmark is to test whether the reconstruction can identify an explicit physical memory mechanism when the reduced photonic dynamics is non-CP-divisible and the embedded model M 3 M_3 M 3 is the correct description.
Physically, benchmark 3 represents a structured vacuum-loss environment with two hidden resonant modes. A possible realization is an observed photonic mode coupled coherently to two weakly damped auxiliary resonances, representing, for example, parasitic cavity modes, spectral side modes, or delay-line modes of the photonic circuit. Each auxiliary mode is itself coupled to an independent Markovian vacuum reservoir. Photons can therefore leave the observed mode, remain temporarily stored in the auxiliary modes, and subsequently return before being irreversibly lost to the external reservoirs. After eliminating the auxiliary modes, this coherent storage-and-return mechanism appears as a time-correlated vacuum-loss channel with a two-exponential memory kernel. Thus, benchmark 3 provides a microscopic realization of non-Markovian photon loss rather than thermal excitation or classical technical noise.
Candidate models
We consider one observed photonic mode a a a coupled to two auxiliary bosonic memory modes b 1 b_1 b 1 and b 2 b_2 b 2 . The three modes are taken to be resonant. The interaction in Eq. (68) is written in the rotating-wave approximation (RWA), retaining only the excitation-conserving terms a † b ℓ a^\dagger b_\ell a † b ℓ and a b ℓ † a b_\ell^\dagger a b ℓ † while neglecting the counter-rotating terms a b ℓ a b_\ell a b ℓ and a † b ℓ † a^\dagger b_\ell^\dagger a † b ℓ † [18 ] . In a frame rotating at the common resonance frequency,3 the free-frequency terms vanish and the detunings are zero, so the embedded Hamiltonian becomes [11 , 76 ]
H e m b ( 2 ) = g 1 ( a † b 1 + a b 1 † ) + g 2 ( a † b 2 + a b 2 † ) (132) H_{\mathrm{emb}}^{(2)} = g_1\left(a^\dagger b_1 + a b_1^\dagger\right) + g_2\left(a^\dagger b_2 + a b_2^\dagger\right) \tag*{(132)} H emb ( 2 ) = g 1 ( a † b 1 + a b 1 † ) + g 2 ( a † b 2 + a b 2 † ) (132)
where the coupling constants g 1 g_1 g 1 and g 2 g_2 g 2 are taken to be real. Their phases can be absorbed into independent phase definitions of b 1 b_1 b 1 and b 2 b_2 b 2 .
To match the embedding of Eqs. (68) –(71) exactly, the observed mode a a a is not coupled directly to an additional Markovian reservoir. Instead, environmental damping acts only on the two pseudomodes. In this benchmark we denote the pseudomode decay rates by γ 1 \gamma_1 γ 1 and γ 2 \gamma_2 γ 2 . After tracing out the two memoryless vacuum reservoirs coupled to b 1 b_1 b 1 and b 2 b_2 b 2 , the joint density operator of the explicit a + b 1 + b 2 a+b_1+b_2 a + b 1 + b 2 subsystem obeys
d ϱ d t = − i [ H e m b ( 2 ) , ϱ ] + γ 1 D [ b 1 ] ϱ + γ 2 D [ b 2 ] ϱ (133) \frac{d\varrho}{dt}=-i\left[H_{\mathrm{emb}}^{(2)},\varrho\right]+\gamma_1\mathcal{D}[b_1]\varrho+\gamma_2\mathcal{D}[b_2]\varrho \tag*{(133)} d t d ϱ = − i [ H emb ( 2 ) , ϱ ] + γ 1 D [ b 1 ] ϱ + γ 2 D [ b 2 ] ϱ (133)
Thus, the enlarged a + b 1 + b 2 a+b_1+b_2 a + b 1 + b 2 subsystem evolves according to a standard CP Markovian Lindblad equation even when the reduced dynamics of the observed mode a a a is not CP divisible.
The corresponding Heisenberg-Langevin equations are [18 ]
a ˙ ( t ) = − i g 1 b 1 ( t ) − i g 2 b 2 ( t ) (134) \dot{a}(t)=-ig_1b_1(t)-ig_2b_2(t) \tag*{(134)} a ˙ ( t ) = − i g 1 b 1 ( t ) − i g 2 b 2 ( t ) (134)
b ˙ 1 ( t ) = − γ 1 2 b 1 ( t ) − i g 1 a ( t ) + γ 1 b 1 , i n ( t ) (135) \dot{b}_1(t)=-\frac{\gamma_1}{2}b_1(t)-ig_1a(t)+\sqrt{\gamma_1}\,b_{1,\mathrm{in}}(t) \tag*{(135)} b ˙ 1 ( t ) = − 2 γ 1 b 1 ( t ) − i g 1 a ( t ) + γ 1 b 1 , in ( t ) (135)
b ˙ 2 ( t ) = − γ 2 2 b 2 ( t ) − i g 2 a ( t ) + γ 2 b 2 , i n ( t ) (136) \dot{b}_2(t)=-\frac{\gamma_2}{2}b_2(t)-ig_2a(t)+\sqrt{\gamma_2}\,b_{2,\mathrm{in}}(t) \tag*{(136)} b ˙ 2 ( t ) = − 2 γ 2 b 2 ( t ) − i g 2 a ( t ) + γ 2 b 2 , in ( t ) (136)
Here a ( t ) a(t) a ( t ) is the annihilation operator of the observed photonic mode and b 1 ( t ) , b 2 ( t ) b_1(t),b_2(t) b 1 ( t ) , b 2 ( t ) are the two explicit pseudomode operators retained in the enlarged system. In contrast, b 1 , i n ( t ) b_{1,\mathrm{in}}(t) b 1 , in ( t ) and b 2 , i n ( t ) b_{2,\mathrm{in}}(t) b 2 , in ( t ) are the input-field operators of the independent Markovian vacuum reservoirs coupled to the pseudomodes. The terms − γ ℓ b ℓ / 2 -\gamma_\ell b_\ell/2 − γ ℓ b ℓ /2 describe pseudomode damping and the terms γ ℓ b ℓ , i n \sqrt{\gamma_\ell}\,b_{\ell,\mathrm{in}} γ ℓ b ℓ , in provide the corresponding vacuum Langevin noise required by the damping channel.
For c μ , i n ∈ { b 1 , i n , b 2 , i n } c_{\mu,\mathrm{in}}\in\{b_{1,\mathrm{in}},b_{2,\mathrm{in}}\} c μ , in ∈ { b 1 , in , b 2 , in } , the independent vacuum inputs satisfy
⟨ c μ , i n ( t ) c ν , i n † ( s ) ⟩ = δ μ ν δ ( t − s ) (137) \left\langle c_{\mu,\mathrm{in}}(t)c_{\nu,\mathrm{in}}^\dagger(s)\right\rangle=\delta_{\mu\nu}\delta(t-s) \tag*{(137)} ⟨ c μ , in ( t ) c ν , in † ( s ) ⟩ = δ μν δ ( t − s ) (137)
⟨ c μ , i n † ( t ) c ν , i n ( s ) ⟩ = 0 (138) \left\langle c_{\mu,\mathrm{in}}^\dagger(t)c_{\nu,\mathrm{in}}(s)\right\rangle=0 \tag*{(138)} ⟨ c μ , in † ( t ) c ν , in ( s ) ⟩ = 0 (138)
The homogeneous part of Eqs. (134) –(136) is generated by
P 3 = ( 0 − i g 1 − i g 2 − i g 1 − γ 1 / 2 0 − i g 2 0 − γ 2 / 2 ) (139) P_3=
\begin{pmatrix}
0 & -ig_1 & -ig_2\\
-ig_1 & -\gamma_1/2 & 0\\
-ig_2 & 0 & -\gamma_2/2
\end{pmatrix} \tag*{(139)} P 3 = 0 − i g 1 − i g 2 − i g 1 − γ 1 /2 0 − i g 2 0 − γ 2 /2 (139)
The absence of diagonal frequency offsets is the zero-detuning resonance condition, while the zero in the ( 1 , 1 ) (1,1) ( 1 , 1 ) entry also reflects the absence of a direct Markovian loss channel on a a a .
The two-memory kernel follows by formally integrating the pseudomode equations. For ℓ = 1 , 2 \ell=1,2 ℓ = 1 , 2 , we have
b ℓ ( t ) = e − γ ℓ t / 2 b ℓ ( 0 ) − i g ℓ ∫ 0 t e − γ ℓ ( t − s ) / 2 a ( s ) d s + γ ℓ ∫ 0 t e − γ ℓ ( t − s ) / 2 b ℓ , i n ( s ) d s (140) b_\ell(t)=e^{-\gamma_\ell t/2}b_\ell(0)-ig_\ell\int_0^t e^{-\gamma_\ell(t-s)/2}a(s)\,ds+\sqrt{\gamma_\ell}\int_0^t e^{-\gamma_\ell(t-s)/2}b_{\ell,\mathrm{in}}(s)\,ds \tag*{(140)} b ℓ ( t ) = e − γ ℓ t /2 b ℓ ( 0 ) − i g ℓ ∫ 0 t e − γ ℓ ( t − s ) /2 a ( s ) d s + γ ℓ ∫ 0 t e − γ ℓ ( t − s ) /2 b ℓ , in ( s ) d s (140)
Substituting Eq. (140) into Eq. (134) gives
a ˙ ( t ) = − ∫ 0 t K ( t − s ) a ( s ) d s − i ∑ ℓ = 1 2 g ℓ e − γ ℓ t / 2 b ℓ ( 0 ) − i ∑ ℓ = 1 2 g ℓ γ ℓ ∫ 0 t e − γ ℓ ( t − s ) / 2 b ℓ , i n ( s ) d s (141) \dot{a}(t)=-\int_0^tK(t-s)a(s)\,ds-i\sum_{\ell=1}^{2}g_\ell e^{-\gamma_\ell t/2}b_\ell(0)-i\sum_{\ell=1}^{2}g_\ell\sqrt{\gamma_\ell}\int_0^t e^{-\gamma_\ell(t-s)/2}b_{\ell,\mathrm{in}}(s)\,ds \tag*{(141)} a ˙ ( t ) = − ∫ 0 t K ( t − s ) a ( s ) d s − i ℓ = 1 ∑ 2 g ℓ e − γ ℓ t /2 b ℓ ( 0 ) − i ℓ = 1 ∑ 2 g ℓ γ ℓ ∫ 0 t e − γ ℓ ( t − s ) /2 b ℓ , in ( s ) d s (141)
where the resonant memory kernel is
K ( t ) = g 1 2 e − γ 1 t / 2 + g 2 2 e − γ 2 t / 2 , t ≥ 0 (142) K(t)=g_1^2e^{-\gamma_1t/2}+g_2^2e^{-\gamma_2t/2},\qquad t\geq0 \tag*{(142)} K ( t ) = g 1 2 e − γ 1 t /2 + g 2 2 e − γ 2 t /2 , t ≥ 0 (142)
Eq. (142) is the two-pseudomode specialization of Eq. (71) , with the same decay-rate notation γ ℓ \gamma_\ell γ ℓ . The two exponentials introduce the memory timescales 2 / γ 1 2/\gamma_1 2/ γ 1 and 2 / γ 2 2/\gamma_2 2/ γ 2 .
Let χ 3 ( t ) \chi_3(t) χ 3 ( t ) denote the coefficient multiplying a ( 0 ) a(0) a ( 0 ) in the solution for a ( t ) a(t) a ( t ) . Since the pseudomode initial operators and vacuum input fields in Eq. (141) do not contain a ( 0 ) a(0) a ( 0 ) , this coefficient obeys
χ ˙ 3 ( t ) = − ∫ 0 t K ( t − s ) χ 3 ( s ) d s , χ 3 ( 0 ) = 1 (143) \dot{\chi}_3(t)=-\int_0^tK(t-s)\chi_3(s)\,ds,\qquad\chi_3(0)=1 \tag*{(143)} χ ˙ 3 ( t ) = − ∫ 0 t K ( t − s ) χ 3 ( s ) d s , χ 3 ( 0 ) = 1 (143)
Equivalently
χ 3 ( t ) = [ e P 3 t ] 11 (144) \chi_3(t)=\left[e^{P_3t}\right]_{11} \tag*{(144)} χ 3 ( t ) = [ e P 3 t ] 11 (144)
Therefore, χ 3 ( t ) \chi_3(t) χ 3 ( t ) is determined by the three eigenmodes of P 3 P_3 P 3 , giving a three-pole response that is more structured than the single damped-oscillatory parameterization used for M 1 M_1 M 1 and M 2 M_2 M 2 . The transfer amplitude χ 3 ( t ) \chi_3(t) χ 3 ( t ) has isolated zeros, at which the reduced time-local generator is singular. Fitting therefore uses the embedded dynamics and the finite-time channel, which remain well defined through these points; generator reconstruction and its diagnostics are restricted to nonsingular intervals. By contrast, a bounded damped-cosine rate yields strictly positive transmissivity at every finite time and cannot reproduce these exact zeros.
The two pseudomodes and their two Markovian reservoirs are initially in vacuum. Because the embedding is passive and number preserving, the reduced channel of the observed mode is therefore a quantum-limited attenuation channel with transmissivity η \eta η that obeys Eq. (118) . For the squeezed-vacuum probe used in the previous benchmarks, the initial variances of the q q q -... and p p p -quadratures of the squeezed-vacuum input state V q ( 0 ) = 1 2 e − 2 r V_q(0) = \frac{1}{2}e^{-2r} V q ( 0 ) = 2 1 e − 2 r and V p ( 0 ) = 1 2 e 2 r V_p(0) = \frac{1}{2}e^{2r} V p ( 0 ) = 2 1 e 2 r and they evolve as Eq. (102) . Away from isolated zeros of χ 3 ( t ) \chi_3(t) χ 3 ( t ) , the corresponding time-local reduced loss rate is defined in Eq. (120) . Generator-level diagnostics are evaluated only away from transfer-amplitude zeros; evolution through these zeros is described using the finite-time Gaussian channel.
Table III. Model comparison for Benchmark 3. The reference homodyne data are generated by the resonant two-pseudomode embedding without a direct Markovian loss channel on the observed mode. Relative AIC and BIC are reported with respect to the best model, while ϵ M t e s t \epsilon_M^{\mathrm{test}} ϵ M test is the held-out negative log-likelihood per test sample.
For a vacuum attenuation channel, the two eigenvalues of the local Gaussian CP matrix are 0 0 0 and κ 3 e f f ( t ) \kappa_3^{\mathrm{eff}}(t) κ 3 eff ( t ) . Hence
λ min [ M C P ( t ) ] = min { 0 , κ 3 e f f ( t ) } (145) \lambda_{\min}[M_{\mathrm{CP}}(t)] = \min\left\{0,\kappa_3^{\mathrm{eff}}(t)\right\} \tag*{(145)} λ m i n [ M CP ( t )] = min { 0 , κ 3 eff ( t ) } (145)
Whenever η ( t ) \eta(t) η ( t ) increases, η ˙ ( t ) > 0 \dot{\eta}(t)>0 η ˙ ( t ) > 0 and κ 3 e f f ( t ) < 0 \kappa_3^{\mathrm{eff}}(t)<0 κ 3 eff ( t ) < 0 . A transmissivity revival therefore identifies non-CP-divisible reduced dynamics, while the enlarged dynamics in Eq. (133) remains CP and Markovian.
Fit results and uncertainty assessment
We use the same numerical protocol as in the previous benchmarks. The reference parameters are
g 1 = 0.60 , g 2 = 0.20 , γ 1 = 0.12 , γ 2 = 0.45 (146) g_1 = 0.60,\qquad g_2 = 0.20,\qquad\gamma_1 = 0.12,\qquad\gamma_2 = 0.45 \tag*{(146)} g 1 = 0.60 , g 2 = 0.20 , γ 1 = 0.12 , γ 2 = 0.45 (146)
The first pseudomode is relatively long-lived, with a memory lifetime 1 / γ 1 ≃ 8.3 1/\gamma_1 \simeq8.3 1/ γ 1 ≃ 8.3 , and provides the dominant coherent storage-and-return pathway, whereas the second pseudomode introduces a shorter memory component. Their combined effect produces two attenuation revivals over the observation window.
The candidate models retain the same hierarchy used above. Model M 0 M_0 M 0 is stationary pure loss with one fitted rate. Models M 1 M_1 M 1 and M 2 M_2 M 2 use the same four-parameter damped-cosine rate in Eq. (103) . Model M 1 M_1 M 1 imposes κ M 1 ( t ) ≥ 0 \kappa_{M_1}(t) \geq0 κ M 1 ( t ) ≥ 0 throughout the observation window. Model M 2 M_2 M 2 allows a signed instantaneous rate but requires the accumulated attenuation channel to remain completely positive. Model M 3 M_3 M 3 is the resonant two-pseudomode embedding in Eq. (133) and has four fitted real parameters, ( g 1 , g 2 , γ 1 , γ 2 ) (g_1,g_2,\gamma_1,\gamma_2) ( g 1 , g 2 , γ 1 , γ 2 ) .
tbl:III exhibits the intended hierarchy of model performance. The stationary model M 0 M_0 M 0 cannot reproduce the coherent memory revivals. Model M 1 M_1 M 1 follows part of the temporal variation, but its pointwise positivity constraint forbids information return. Model M 2 M_2 M 2 performs substantially better than M 1 M_1 M 1 despite having the same number of fitted parameters. The best-fit solution for M 2 M_2 M 2 , obtained by minimizing the training negative log-likelihood, has an instantaneous loss rate that becomes negative
min t ∈ [ 0 , T ] κ M 2 ( t ) ≃ − 0.470 (147) \min_{t\in[0,T]} \kappa_{M_2}(t) \simeq-0.470 \tag*{(147)} t ∈ [ 0 , T ] min κ M 2 ( t ) ≃ − 0.470 (147)
This negative reconstructed rate is consistent with non-CP-divisible reduced dynamics, although the model comparison alone does not establish their statistical necessity. Nevertheless, the single damped-cosine rate of M 2 M_2 M 2 cannot reproduce the complete three-pole response produced by the two pseudomodes. Model M 3 M_3 M 3 is selected decisively by both information criteria and gives the best held-out predictive score. Since M 1 M_1 M 1 , M 2 M_2 M 2 , and M 3 M_3 M 3 all contain four fitted parameters, this preference reflects the two-pseudomode memory structure rather than a larger parameter count.
To quantify the statistical uncertainty of the selected memory model, we perform a parametric bootstrap with N b o o t = 300 N_{\mathrm{boot}}=300 N boot = 300 . Each bootstrap training set is generated from the fitted M 3 M_3 M 3 channel using the same measurement times, homodyne phases, and number of training shots as the original benchmark, and M 3 M_3 M 3 is refitted independently to every bootstrap realization. At each resonance, χ 3 ( t ) \chi_3(t) χ 3 ( t ) has isolated zeros near the coherent-exchange minima, where the effective loss rate κ 3 e f f ( t ) \kappa_3^{\mathrm{eff}}(t) κ 3 eff ( t ) becomes singular, although the finite-time transmissivity remains regular. We therefore evaluate the local CP-divisibility witness on the regular interval I = [ 3.3 , 4.8 ] I=[3.3,4.8] I = [ 3.3 , 4.8 ] , away from these singularities. Since the minimum witness value depends on the chosen interval boundary, we instead assess whether its bootstrap upper confidence bound remains negative throughout this interval. For N b o o t = 300 N_{\mathrm{boot}}=300 N boot = 300 parametric bootstrap refits of the selected M 3 M_3 M 3 model, the maximum pointwise 97.5% upper bound is − 0.139 -0.139 − 0.139 . The reconstructed witness therefore remains negative across the considered interval within the pointwise bootstrap uncertainty, supporting non-CP-divisible reduced dynamics conditional on the fitted model.
Figure 4 illustrates how the two pseudomodes generate and store memory in the reduced photonic dynamics. In panel (a), the effective transmissivity of the observed mode, defined in Eq. (118) , exhibits pronounced revivals due to coherent excitation exchange with the two damped pseudomodes. The fitted M 3 M_3 M 3 model closely reproduces the reference dynamics. We plot the transmissivity rather than the effective loss rate κ 3 e f f ( t ) \kappa_3^{\mathrm{eff}}(t) κ 3 eff ( t ) because the latter is singular at isolated zeros of χ 3 ( t ) \chi_3(t) χ 3 ( t ) , whereas the transmissivity remains well defined. Panel (b) shows that these revivals are accompanied by negative values of the local CP-divisibility witness, demonstrating in formation backflow in the reduced mode while the enlarged pseudomode dynamics remains Markovian. Panel (c) shows the corresponding evolution of the squeezed-quadrature variance. As the transmissivity decreases, the initially squeezed variance increases toward the vacuum value, reflecting degradation of the squeezing. When the transmissivity subsequently revives because excitation returns from the pseudomodes to the observed mode, the quadrature variance decreases again and part of the squeezing is recovered. The close agreement between the reference and fitted M 3 M_3 M 3 curves therefore shows that the reconstructed pseudomode model reproduces not only the underlying memory dynamics but also its directly observable signature in the homodyne variance.
Figure 4. Benchmark 3: resonant two-pseudomode memory without direct Markovian loss of the observed mode. (a) Reference and fitted finite-time transmissivities. The blue Reference curve is the ground-truth two-pseudomode model, while the purple M 3 M_3 M 3 curve is the fitted reconstruction within the same model class. The transmissivity is used instead of the full-time effective loss rate because the latter has isolated representation poles when χ 3 ( t ) \chi_3(t) χ 3 ( t ) crosses zero. (b) Smallest eigenvalue of the local Gaussian CP matrix on the regular interval I = [ 3.3 , 4.8 ] I=[3.3,4.8] I = [ 3.3 , 4.8 ] , chosen away from singularities of the effective time-local generator. The shaded region represents the 95% pointwise bootstrap interval for M 3 M_3 M 3 . The upper bound remains negative throughout the displayed interval, supporting the reconstructed violation of CP-divisibility. (c) Squeezed-quadrature variance V ( t ) ≡ V q ( t ) V(t)\equiv V_q(t) V ( t ) ≡ V q ( t ) . The quadrature recovery follows the transmissivity revival and provides an observable signature of pseudomode-mediated information return.
Consequently, the three benchmarks demonstrate discrimination between CP-divisible and non-CP-divisible dynamics within the tested model families. The preference for a pseudomode description does not uniquely identify the underlying environment, since different environmental realizations can produce identical single-time homodyne statistics.
Conclusion
The main contribution of this work is a physically interpretable framework for reconstructing non-Markovian Gaussian noise in CV photonic quantum systems. The theoretical framework connects reconstructed Gaussian generators to physical noise mechanisms, including photon loss, thermal excitation, correlated dissipation, and coupling to auxiliary memory modes, while also discussing effective descriptions of technical fluctuations. The numerical benchmarks demonstrate the distinction between CP-divisible and non-CP-divisible dynamics for single-mode attenuation and thermal-noise processes. A key advantage of the framework is its model hierarchy, which ranges from static Markovian noise to time-dependent Markovian dynamics, Lindblad-like non-Markovian generators, and pseudomode-embedded Markovian models. This hierarchy provides a systematic framework for comparing CP-divisible and non-CP-divisible descriptions of the observed dynamics. Non-Markovianity in the CP-divisibility sense is diagnosed by negative eigenvalues of M C P ( t ) M_{\mathrm{CP}}(t) M CP ( t ) . Covariance revivals and improved predictive performance provide supporting evidence in the present benchmarks, but are not general stand-alone criteria for non-Markovianity.
The proposed reconstruction procedure combines Gaussian moment propagation, likelihood-based parameter estimation, physicality constraints, and model-selection criteria. In the numerical benchmarks, this combination distinguishes genuinely non-CP-divisible reduced dynamics from ordinary time-dependent Gaussian noise across three attenuation-related physical scenarios: time-dependent photon loss, nonstationary thermal loss with excitation, and a structured non-Markovian vacuum-loss environment with two hidden resonant modes. The thermal benchmark shows that time-dependent diffusion and heating need not imply non-Markovianity, whereas the two-pseudomode benchmark provides a microscopic interpretation of genuine memory in terms of coherent storage and return of excitation from hidden environmental modes. As a result, the method provides both a quantitative reconstruction of the unknown noise dynamics and a physical interpretation of its possible microscopic origin. This is especially useful for CV photonic platforms, where memory effects may arise from structured reservoirs, spectral side modes, imperfect filtering, feedback paths, or slowly fluctuating control fields.
The present study is restricted to Gaussian-preserving dynamics and simulation-based validation. Although the theoretical framework includes correlated dissipation and effective descriptions of technical fluctuations, the numerical benchmarks focus on single-mode attenuation, thermal excitation, and pseudomode-mediated memory. Multiplicative fluctuations, such as pump-phase and squeezing-strength noise, require additional approximations and are not explicitly validated here. Future work may extend the numerical validation to correlated noise, multimode photonic systems, and experimental homodyne data, as well as investigate adaptive tomography and non-Gaussian noise.
Acknowledgments. The author would like to thank Sevag Gharibian and Michael Stefszky for their valuable discussions, insightful comments, and constructive suggestions that contributed to the development of this work. The author acknowledges support from the Pho-Quant project, funded by the Bundesministerium für Forschung, Technologie und Raumfahrt (BMFTR).
Data availability. Further details of the numerical implementation, simulation parameters, and validation procedures are available from the corresponding author upon reasonable request.
Derivation of the Gaussian propagator and local CP condition
Starting from the Gaussian moment relations in Eqs. (22) and (23)
d ˙ ( t ) = A ( t ) d ( t ) + u ( t ) , (148) \dot{d}(t) = A(t)d(t) + u(t), \tag*{(148)} d ˙ ( t ) = A ( t ) d ( t ) + u ( t ) , (148)
V ˙ ( t ) = A ( t ) V ( t ) + V ( t ) A T ( t ) + D ( t ) (149) \dot{V}(t) = A(t)V(t) + V(t)A^{T}(t) + D(t) \tag*{(149)} V ˙ ( t ) = A ( t ) V ( t ) + V ( t ) A T ( t ) + D ( t ) (149)
we introduce the fundamental matrix X ( t , s ) X(t,s) X ( t , s ) , defined as the solution of
d d t X ( t , s ) = A ( t ) X ( t , s ) , X ( s , s ) = I (150) \frac{d}{dt}X(t,s) = A(t)X(t,s), \qquad X(s,s) = I \tag*{(150)} d t d X ( t , s ) = A ( t ) X ( t , s ) , X ( s , s ) = I (150)
Multiplying the first-moment equation by X − 1 ( t , 0 ) X^{-1}(t,0) X − 1 ( t , 0 ) and integrating from 0 0 0 to t t t gives
d ( t ) = X ( t , 0 ) d ( 0 ) + ∫ 0 t X ( t , s ) u ( s ) d s (151) d(t) = X(t,0)d(0) + \int_{0}^{t} X(t,s)u(s) \, ds \tag*{(151)} d ( t ) = X ( t , 0 ) d ( 0 ) + ∫ 0 t X ( t , s ) u ( s ) d s (151)
For the covariance matrix, we define
V ~ ( t ) = X − 1 ( t , 0 ) V ( t ) X − T ( t , 0 ) (152) \widetilde{V}(t) = X^{-1}(t,0)V(t)X^{-T}(t,0) \tag*{(152)} V ( t ) = X − 1 ( t , 0 ) V ( t ) X − T ( t , 0 ) (152)
Differentiating this expression and using the covariance equation cancels the drift terms, leaving
V ~ ˙ ( t ) = X − 1 ( t , 0 ) D ( t ) X − T ( t , 0 ) (153) \dot{\widetilde{V}}(t) = X^{-1}(t,0)D(t)X^{-T}(t,0) \tag*{(153)} V ˙ ( t ) = X − 1 ( t , 0 ) D ( t ) X − T ( t , 0 ) (153)
Integrating and multiplying back by X ( t , 0 ) X(t,0) X ( t , 0 ) and X T ( t , 0 ) X^{T}(t,0) X T ( t , 0 ) gives
V ( t ) = X ( t , 0 ) V ( 0 ) X T ( t , 0 ) + ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (154) V(t) = X(t,0)V(0)X^{T}(t,0) + \int_{0}^{t} X(t,s)D(s)X^{T}(t,s) \, ds \tag*{(154)} V ( t ) = X ( t , 0 ) V ( 0 ) X T ( t , 0 ) + ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (154)
Thus, the Gaussian channel can be written in the standard form
V ( t ) = X ( t ) V ( 0 ) X T ( t ) + Y ( t ) (155) V(t) = X(t)V(0)X^{T}(t) + Y(t) \tag*{(155)} V ( t ) = X ( t ) V ( 0 ) X T ( t ) + Y ( t ) (155)
where
Y ( t ) = ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (156) Y(t) = \int_{0}^{t} X(t,s)D(s)X^{T}(t,s) \, ds \tag*{(156)} Y ( t ) = ∫ 0 t X ( t , s ) D ( s ) X T ( t , s ) d s (156)
The CP condition for this Gaussian channel can be defined [14 , 15 ]
Y ( t ) + i 2 Ω − i 2 X ( t ) Ω X T ( t ) ≥ 0 (157) Y(t) + \frac{i}{2}\Omega- \frac{i}{2}X(t)\Omega X^{T}(t) \ge0 \tag*{(157)} Y ( t ) + 2 i Ω − 2 i X ( t ) Ω X T ( t ) ≥ 0 (157)
For an infinitesimal map from t t t to t + d t t+dt t + d t , assuming sufficiently smooth drift and diffusion matrices, one has
X ( t + d t , t ) = I + A ( t ) d t + O ( d t 2 ) , X(t+dt,t) = I + A(t)dt + O(dt^{2}), X ( t + d t , t ) = I + A ( t ) d t + O ( d t 2 ) ,
Y ( t + d t , t ) = D ( t ) d t + O ( d t 2 ) (158) Y(t+dt,t) = D(t)dt + O(dt^{2}) \tag*{(158)} Y ( t + d t , t ) = D ( t ) d t + O ( d t 2 ) (158)
Substituting these expressions into the Gaussian CP condition and keeping only first-order terms in d t dt d t gives
D ( t ) − i 2 [ A ( t ) Ω + Ω A T ( t ) ] ≥ 0 (159) D(t) - \frac{i}{2}\left[A(t)\Omega+ \Omega A^{T}(t)\right] \ge0 \tag*{(159)} D ( t ) − 2 i [ A ( t ) Ω + Ω A T ( t ) ] ≥ 0 (159)
This is the infinitesimal CP condition used to test CP-divisibility of the Gaussian dynamics.
Numerical implementation and synthetic-data generation
All numerical simulations are performed in dimensionless units, with the vacuum covariance normalized as I 2 / 2 I_{2}/2 I 2 /2 . Time is measured in units of an arbitrary reference scale t 0 t_{0} t 0 , so that t = t p h y s / t 0 t=t_{\mathrm{phys}}/t_{0} t = t phys / t 0 . Accordingly, the dimensionless rates and frequencies reported below correspond to physical rates and frequencies measured in units of t 0 − 1 t_{0}^{-1} t 0 − 1 . The continuous interval t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] is sampled at discrete measurement times t k t_{k} t k , k = 1 , … , K k=1,\ldots,K k = 1 , … , K , at which the homodyne data are generated and the reconstructed dynamics are evaluated. The total simulation time is T = 10 T=10 T = 10 and the interval [ 0 , T ] [0,T] [ 0 , T ] is discretized into $K=80 equally spaced measurement times. These choices provide sufficient temporal resolution to resolve the benchmark memory times and oscillation frequencies. Their conversion to laboratory units depends on the platform-dependent reference time t 0 t_0 t 0 . For each time, homodyne data are generated for the phases
θ ∈ { 0 , π / 4 , π / 2 , 3 π / 4 } (160) \theta\in\{0,\pi/4,\pi/2,3\pi/4\} \tag*{(160)} θ ∈ { 0 , π /4 , π /2 , 3 π /4 } (160)
For every time, phase and input probe state, we use N s h o t = 2000 N_{\mathrm{shot}}=2000 N shot = 2000 synthetic homodyne samples. Each homodyne shot samples a separately prepared evolution, with the system reinitialized and the memory modes and environment restored to the same reference state. Different measurement times therefore correspond to independent preparations, not repeated destructive measurements of one trajectory. The product likelihood and random-shot train/test split assume that resetting and sufficient waiting between shots suppress residual environmental and detector correlations, with negligible drift during acquisition. If these conditions fail, a likelihood accounting for correlations between shots and block-based splitting and resampling are required.
The data are split into training and test sets using an 80/20 split. The training data are used to estimate the parameters of each candidate model, whereas the test data are kept separate from the fitting procedure and are used only to evaluate how well the fitted model predicts independent homodyne samples [77 , 78 ] . This separation helps determine whether the improved training fit of a more flexible model reflects genuine dynamical structure or merely overfitting to statistical noise in the training data.
For an input probe state labelled by α \alpha α , the real-valued synthetic homodyne outcomes are generated as
x k ℓ s ( α ) ∼ N ( e ℓ T d α t r u e ( t k ) , e ℓ T V α t r u e ( t k ) e ℓ ) (161) x_{k\ell s}^{(\alpha)} \sim\mathcal{N}\left(\mathbf{e}_{\ell}^{\mathsf{T}}\mathbf{d}_{\alpha}^{\mathrm{true}}(t_k),\mathbf{e}_{\ell}^{\mathsf{T}}\mathbf{V}_{\alpha}^{\mathrm{true}}(t_k)\mathbf{e}_{\ell}\right) \tag*{(161)} x k ℓ s ( α ) ∼ N ( e ℓ T d α true ( t k ) , e ℓ T V α true ( t k ) e ℓ ) (161)
where N \mathcal{N} N denotes a normal distribution, t k t_k t k is the measurement time, ℓ = 1 , … , L \ell=1,\ldots,L ℓ = 1 , … , L labels the homodyne phase θ ℓ \theta_\ell θ ℓ in Eq. (160) , e ℓ \mathbf{e}_\ell e ℓ is the homodyne direction and s s s labels the shot. The reference moments d α t r u e ( t ) \mathbf{d}_{\alpha}^{\mathrm{true}}(t) d α true ( t ) and V α t r u e ( t ) \mathbf{V}_{\alpha}^{\mathrm{true}}(t) V α true ( t ) are obtained by solving the Gaussian moment equations of the preceding sections. The reference dynamics is used consistently for data generation, model fitting and plotting.
The following steps clarify how the reference dynamics, synthetic homodyne data, fitted models, and plotted curves are connected. First, the exact benchmark model defines d α t r u e ( t ) \mathbf{d}_{\alpha}^{\mathrm{true}}(t) d α true ( t ) and V α t r u e ( t ) \mathbf{V}_{\alpha}^{\mathrm{true}}(t) V α true ( t ) through the Gaussian moment equations. Synthetic homodyne samples are then drawn from the corresponding Gaussian distributions in Eq. (161) . The candidate models M 0 M_0 M 0 –M 3 M_3 M 3 are fitted only to the training data, while the test data are used only for predictive comparison. The curves shown in the figures of each benchmark are therefore not independent assumptions: they are the reference or fitted time-local coefficients and the corresponding propagated observables.
The numerical values are constrained benchmark parameters chosen to represent moderate optical loss or diffusion over the full observation window, together with finite memory times comparable to a few circuit time units. The memory lifetimes τ c = 2.5 \tau_c=2.5 τ c = 2.5 –3.5 3.5 3.5 and oscillation frequencies ω c = 1.8 \omega_c=1.8 ω c = 1.8 –2.2 2.2 2.2 describe damped memory kernels of the type produced by resonator modes, delay-line feedback, finite-bandwidth technical noise, or pseudomode embeddings [18 , 67 , 69 ] . The oscillatory components represent a moderate memory regime in which the instantaneous time-local generator can become temporarily non-CP-divisible, while the accumulated finite-time Gaussian channel remains CP.
In all benchmarks, ν M \nu_M ν M denotes the number of real scalar parameters that are varied during the fit. Parameters fixed by the benchmark design, such as the sampled time grid, homodyne angles, and input-state squeezing, are not counted in ν M \nu_M ν M . For the numerical benchmarks, each candidate model is described by a low-dimensional physically motivated parameter set Θ M \Theta_M Θ M . Accordingly, the parameters are obtained by minimizing the training negative log-likelihood in Eq. (80) , subject to the model-specific physicality constraints described below. The regularized estimator in Eq. (82) is intended for more general reconstructions in which A ( t ) A(t) A ( t ) , D ( t ) D(t) D ( t ) , and u ( t ) u(t) u ( t ) are represented by flexible basis expansions in Eqs. (85) –(87) .
The AIC and BIC criteria defined in Eqs. (94) and (95) are evaluated using the unregularized training negative log-likelihood, J M t r ( Θ M ) \mathcal{J}^{\mathrm{tr}}_M(\Theta_M) J M tr ( Θ M ) and N N N is replaced by the number of independent training homodyne samples N t r N_{\mathrm{tr}} N tr . The test predictive score ϵ M t e s t \epsilon_M^{\mathrm{test}} ϵ M test is evaluated separately using Eq. (96) . To compare the candidate models independently of the absolute scale of the information criteria, we subtract the smallest AIC and BIC values obtained among the four candidate models as a reference.4
Δ A I C M = A I C M − min M ′ A I C M ′ Δ B I C M = B I C M − min M ′ B I C M ′ (162) \begin{aligned}
\Delta\mathrm{AIC}_M &= \mathrm{AIC}_M-\min_{M'}\mathrm{AIC}_{M'}\\
\Delta\mathrm{BIC}_M &= \mathrm{BIC}_M-\min_{M'}\mathrm{BIC}_{M'} \tag*{(162)}
\end{aligned} Δ AIC M Δ BIC M = AIC M − M ′ min AIC M ′ = BIC M − M ′ min BIC M ′ (162)
where M ′ M' M ′ is a dummy model index over the full model set { M 0 , M 1 , M 2 , M 3 } \{M_0,M_1,M_2,M_3\} { M 0 , M 1 , M 2 , M 3 } used to identify the best AIC or BIC value. Lower values of Δ A I C M \Delta\mathrm{AIC}_M Δ AIC M and Δ B I C M \Delta\mathrm{BIC}_M Δ BIC M indicate stronger information-criterion support relative to the best candidate model, while a lower ϵ M t e s t \epsilon_M^{\mathrm{test}} ϵ M test indicates better test predictive performance.
In the numerical benchmarks, the matrices A M ( t ) A_M(t) A M ( t ) and D M ( t ) D_M(t) D M ( t ) are not treated as arbitrary matrix-valued functions. Instead, each candidate model is described by a low-dimensional, physically motivated parameter set Θ M \Theta_M Θ M . These parameters determine the corresponding drift and diffusion matrices or, for pseudomode models, the corresponding reduced Gaussian channel from which the moments are propagated. In all three benchmarks, the displacement drift vanishes, u M ( t ) = 0 u_M(t)=0 u M ( t ) = 0 . For the time-dependent phenomenological models M 1 M_1 M 1 and M 2 M_2 M 2 , we employ damped oscillatory parameterizations as controlled low-dimensional ansätze. This choice facilitates comparison between models and should not be interpreted as implying either a common microscopic origin or non-Markovianity from oscillatory time dependence alone.
The numerical benchmarks use SciPy routines to minimize the Gaussian training negative log-likelihood within prescribed parameter bounds. The stationary model M 0 M_0 M 0 uses the limited-memory Broyden-Fletcher-Goldfarb-Shanno algorithm with bound constraints (L-BFGS-B). In Benchmarks 1 and 2, M 1 M_1 M 1 and M 2 M_2 M 2 use sequential least squares programming (SLSQP) with multiple initial guesses, whereas M 3 M_3 M 3 uses differential evolution followed by L-BFGS-B refinement. In Benchmark 3, M 1 M_1 M 1 , M 2 M_2 M 2 , and M 3 M_3 M 3 use differential evolution with multiple random seeds and local polishing.
The numerical optimization employs model-specific parameter bounds, multiple initial guesses, and prescribed stopping tolerances, with different random initializations where applicable. Physicality constraints are evaluated on separate grids containing 600, 300, and 1001 points for Benchmarks 1, 2, and 3, respectively, using explicit inequalities or large objective penalties. These checks provide discrete approximations to the continuous-time constraints. Channel propagation uses analytical expressions where available; the diffusion contribution for M 2 M_2 M 2 in Benchmark 2 is integrated using an adaptive Runge-Kutta method of order 5(4), with relative and absolute tolerances 2 × 10 − 8 2\times10^{-8} 2 × 1 0 − 8 and 10 − 10 10^{-10} 1 0 − 10 . Pseudomode models propagate the finite-time channel directly, avoiding division by the transfer amplitude during fitting. In Benchmark 3, the time-local witness is evaluated on a restricted interval away from transfer-amplitude zeros. Independent test data are used only for predictive evaluation.
References
(2012).