Belavkin Equations in Quantum Filtering

• Belavkin Equations are stochastic evolution equations that model the conditioned dynamics of quantum systems based on continuous measurement records.
• They are derived using approaches like the reference probability method and repeated indirect measurement limits, yielding both diffusive and jump processes.
• The framework leverages quantum Itō calculus and structural representations, facilitating analyses in open systems, coherent state generalizations, and many-body extensions.

Searching arXiv for the cited Belavkin-equation papers to ground the article in current arXiv records. arxiv_search query: "Belavkin equation quantum filtering coherent states squeezed light mean-field repeated measurements" Belavkin equations are stochastic evolution equations for conditioned quantum states under continuous observation. In the standard non-demolition setting, one observes a commutative output process generated by an open quantum system coupled to bosonic fields, and the aim is to compute the conditional expectation of a Heisenberg observable $X$ given the observation history up to time $t$. In this sense, the Belavkin–Kushner–Stratonovich equation is the quantum analogue of nonlinear filtering, while its linear unnormalized counterpart is the quantum analogue of a Zakai equation (Gough et al., 2013). A complementary discrete viewpoint derives the same equations as scaling limits of repeated indirect measurements and iterated Bayesian updates, in either a Brownian diffusive limit or a Poissonian jumpy limit (Bauer et al., 2012).

1. Conditional-state dynamics and the non-demolition setting

The basic object is the conditional estimate

$$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$

where $j_t(X)=V(t)^*[X\otimes I]V(t)$ is the Heisenberg-evolved observable and $\mathfrak{Y}^{\mathrm{out}}_t$ is the commutative von Neumann algebra generated by the observation record up to time $t$ (Gough et al., 2013). The non-demolition property,

$j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$

is the structural condition that makes conditioning meaningful: evolved system observables commute with all past observations (Gough et al., 2013).

In the Hudson–Parthasarathy framework, the system is coupled to bosonic input fields, and the measurement is performed on an output process obtained after interaction. Measured outputs may be quadratures, photon counts, or more general commuting linear combinations of output fields. In the phase-space treatment of multichannel diffusive measurements, the measured process is written as $Z=FY$, with

$FF^T\succ 0,\qquad FJF^T=0,$

so that the observation channels are self-commuting and therefore classical (Vladimirov, 2016).

The same filtering structure emerges from repeated indirect measurements. There, one starts from a discrete Bayesian recursion

$$Q_n(\alpha| i_1,\cdots,i_n)= \frac{p(i_n|\alpha)\, Q_{n-1}(\alpha| i_1,\cdots,i_{n-1})} {\pi_{n-1}(i_n| i_1,\cdots,i_{n-1})},$$

for pointer-state probabilities, and then passes to a continuous-time limit. In the quantum mechanical framework, this continuous time limit leads to Belavkin equations (Bauer et al., 2012). This suggests that the stochastic terms in Belavkin equations are best viewed as continuum innovations extracted from measurement records.

2. Hudson–Parthasarathy structure and derivational methods

For an open Markov quantum system with system Hilbert space $t$0, the unitary cocycle $t$1 is governed by the Hudson–Parthasarathy QSDE

$t$2

with unitary scattering $t$3, coupling operator $t$4, and self-adjoint Hamiltonian $t$5 (Gough et al., 2013). The corresponding Heisenberg dynamics satisfy

$t$6

with Evans–Hudson maps $t$7 and Lindblad generator

$t$8

The input–output relations

$t$9

$$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$0

provide the quantum analogue of classical state and observation equations (Gough et al., 2013).

Two derivational strategies recur throughout the literature. The first is the reference probability approach, the quantum analogue of the classical Girsanov/Kallianpur–Striebel method, based on the operator-valued relation

$$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$1

where $$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$2 is an unnormalized conditional state (Gough et al., 2013). The second is the characteristic-function or generating-map method, in which one introduces an adapted process $$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$3 or $$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$4, expands identities such as

$$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$5

and identifies the drift and stochastic gain by Itō calculus (Gough et al., 2013, Dabrowska et al., 2012).

A different but related algebraic viewpoint appears in Belavkin’s matrix representation of quantum stochastic calculus. There, QSDE coefficients are packaged into Belavkin matrices, the quantum Itō product becomes ordinary matrix multiplication, and physical realizability is encoded by the $$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$6-unitarity condition

$$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$7

In that broader sense, “Belavkin equations” may also refer to structural matrix equations governing quantum stochastic evolutions and feedback networks (Smolyanov et al., 2010).

3. Diffusive and counting Belavkin equations

In the Heisenberg-picture conditional-expectation form used for multichannel diffusive observations, the Belavkin–Kushner–Stratonovich equation takes the form

$$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$8

with innovation

$$\pi_t(X)=\mathbb{E}\big[j_t(X)\mid \mathfrak{Y}^{\mathrm{out}}_t\big],$$9

and

$j_t(X)=V(t)^*[X\otimes I]V(t)$0

(Vladimirov, 2016). This is the standard BKSE for conditional expectations, and phase-space work interprets it as the starting point for posterior quasi-characteristic-function dynamics (Vladimirov, 2016).

For repeated indirect measurements, the continuous diffusive limit yields the normalized stochastic master equation

$j_t(X)=V(t)^*[X\otimes I]V(t)$1

with

$j_t(X)=V(t)^*[X\otimes I]V(t)$2

and

$j_t(X)=V(t)^*[X\otimes I]V(t)$3

while the jump limit yields

$j_t(X)=V(t)^*[X\otimes I]V(t)$4

with

$j_t(X)=V(t)^*[X\otimes I]V(t)$5

(Bauer et al., 2012). This makes explicit that diffusive and counting Belavkin equations are scaling limits of discrete state updates.

Concrete state-vector and density-matrix forms appear in oscillator models. For intense balanced heterodyne detection of a single harmonic oscillator coupled to a coherent input field $j_t(X)=V(t)^*[X\otimes I]V(t)$6, the unnormalized posterior wave function obeys

$j_t(X)=V(t)^*[X\otimes I]V(t)$7

with

$j_t(X)=V(t)^*[X\otimes I]V(t)$8

and the normalized posterior density matrix satisfies the corresponding nonlinear diffusive Belavkin equation (Dabrowska et al., 2012).

4. Coherent, heterodyne, and squeezed-field generalizations

A central extension of the vacuum-input theory replaces the input bosonic field by a coherent state with amplitude $j_t(X)=V(t)^*[X\otimes I]V(t)$9. In that case the effective coupling is shifted to

$\mathfrak{Y}^{\mathrm{out}}_t$0

and the unconditional generator becomes

$\mathfrak{Y}^{\mathrm{out}}_t$1

The resulting filters for both quadrature measurements and photon counting have the same structural form as the vacuum Belavkin equations, but with $\mathfrak{Y}^{\mathrm{out}}_t$2 replaced by $\mathfrak{Y}^{\mathrm{out}}_t$3 and $\mathfrak{Y}^{\mathrm{out}}_t$4 replaced by $\mathfrak{Y}^{\mathrm{out}}_t$5 (Gough et al., 2013).

For quadrature measurement, the normalized coherent-state filter is

$\mathfrak{Y}^{\mathrm{out}}_t$6

with innovations

$\mathfrak{Y}^{\mathrm{out}}_t$7

For photon counting,

$\mathfrak{Y}^{\mathrm{out}}_t$8

with

$\mathfrak{Y}^{\mathrm{out}}_t$9

When $t$0, these reduce exactly to the familiar vacuum Belavkin equations (Gough et al., 2013).

A heterodyne realization for a coherent input field shows additional dynamical consequences. In the cavity-mode model of continuous diffusion observation, the filtering equation is relaxing: any initial square-integrable wave function tends asymptotically to a coherent state whose amplitude depends on the coupling constant $t$1 and the initial coherent state $t$2 of the apparatus through $t$3 (Dabrowska et al., 2012). For squeezed coherent initial states, the posterior state remains within the squeezed coherent family, the squeezing parameter satisfies a Riccati equation, and asymptotically $t$4, so the state becomes coherent (Dabrowska et al., 2012).

Belavkin filtering also extends to squeezed Gaussian fields. When monitored outputs are mixed with squeezed light, or when the system is driven directly by squeezed noise, the innovation channels acquire a nontrivial covariance

$t$5

and the gain is weighted by $t$6. The generalized filter is

$t$7

where $t$8 contains both the usual homodyne covariance terms and additional commutator terms involving squeezed-state correlations $t$9 and $j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$0 (Dabrowska et al., 2014). This shows that squeezing alters not only the observation covariance but also the unconditional generator itself.

5. Alternative formulations, repeated-measurement limits, and structural generalizations

One major reformulation replaces operator-valued filtering dynamics by phase-space evolution. For nonlinear quantum stochastic systems with Weyl-quantized Hamiltonian and coupling operators, the posterior quasi-characteristic function

$j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$1

satisfies the stochastic integro-differential equation

$j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$2

This is a spatial Fourier-domain representation of the Belavkin–Kushner–Stratonovich equation, and its inverse Fourier transform yields a corresponding equation for the posterior quasi-probability density (Vladimirov, 2016). In the linear-Gaussian case, the same framework reproduces the quantum Kalman filter (Vladimirov, 2016).

A second formulation starts from repeated indirect measurements. In that picture, a complete measurement consists of infinitely many partial measurements, each partial outcome updates the state by Bayes’ rule, and the continuous-time Brownian or Poisson limits produce the diffusive and jump Belavkin equations (Bauer et al., 2012). This discrete-to-continuous derivation makes the innovation process directly interpretable as the centered measurement record. A more recent repeated-measurement analysis shows that the same construction can also converge to non-Markovian Volterra stochastic equations when memory is introduced at the microscopic level; in that case the limit is no longer a standard Markovian Belavkin SDE but a Volterra-type stochastic equation with memory kernel (Jacquier et al., 12 Dec 2025).

A third structural extension uses Belavkin matrices. In the Gough–James theory of quantum feedback networks, the coefficient matrix

$j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$3

encodes the Hudson–Parthasarathy coefficients $j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$4, and physical realizability is equivalent to

$j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$5

Feedback reduction then becomes a non-commutative Möbius transformation

$j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$6

which again preserves $j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$7-unitarity (Smolyanov et al., 2010). In this broader algebraic usage, Belavkin equations refer to matrix identities governing quantum stochastic evolutions and feedback interconnections.

6. Many-body, mean-field, infinite-dimensional, and terminological developments

Belavkin equations have also been extended to many-particle systems under continuous observation. For $j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$8 identical particles with mean-field interaction, the normalized $j_t(X)\in (\mathfrak{Y}^{\mathrm{out}}_t)',$9-particle state obeys a many-body stochastic Schrödinger equation or, equivalently, a many-body stochastic master equation, and as $Z=FY$0 the one-particle marginal converges to a nonlinear stochastic equation of McKean–Vlasov type (Kolokoltsov, 2020). In density-matrix form, the limiting equation is

$Z=FY$1

$Z=FY$2

with $Z=FY$3 (Kolokoltsov, 2020). This gives a quantum mean-field analogue of a McKean–Vlasov diffusion.

A finite-dimensional controlled version writes the mean-field Belavkin equation directly for density matrices,

$Z=FY$4

$Z=FY$5

with $Z=FY$6 (Chalal et al., 2023). Under boundedness and Lipschitz assumptions, the equation is well posed and, for $Z=FY$7, propagation of chaos is proved under purification assumption (Chalal et al., 2023).

The infinite-dimensional wave-function version on $Z=FY$8 has now been derived rigorously. For $Z=FY$9 interacting particles under continuous diffusive measurement, the mean-field limit is the stochastic Hartree-type Belavkin equation

$FF^T\succ 0,\qquad FJF^T=0,$0

with global well-posedness proved directly by fixed-point methods and convergence from the $FF^T\succ 0,\qquad FJF^T=0,$1-body stochastic Schrödinger dynamics established in trace norm for reduced marginals (Bouard et al., 25 Jul 2025). A related mixed-state construction with unbounded Hamiltonians and unbounded interaction operators uses a system of stochastic interacting wave functions and proves that the reconstructed mixed state satisfies the diffusive stochastic quantum master equation, which is also known as Belavkin equation (Mora, 19 Mar 2025).

A recurrent source of confusion is terminological. Several papers concern the Belavkin–Staszewski relative entropy, BS-conditional mutual information, or BS-quantum Markov chains; these works are about quantum information divergences and recoverability, not about Belavkin stochastic filtering equations (Bluhm et al., 2019, Bluhm et al., 16 Jan 2025). This suggests a useful distinction between Belavkin equations in the filtering sense—stochastic evolution equations for conditioned quantum states—and Belavkin–Staszewski constructions in the information-theoretic sense.