Operator-Valued Backward Stochastic Integral Equations

• Operator-valued backward stochastic integral equations are defined on Hilbert spaces, where unknowns and coefficients act as operators, introducing noncommutative dynamics in infinite dimensions.
• They employ advanced stochastic integration techniques and solution concepts such as mild semigroup formulations and transposition methods to address challenges in nonreflexive, nonseparable operator spaces.
• These equations play a pivotal role in stochastic control theory by providing second-order adjoints and feedback laws for infinite-dimensional systems.

Operator-valued backward stochastic integral equations are backward stochastic equations in which the unknowns, coefficients, or martingale integrands act as operators on a Hilbert space rather than as scalar or vector processes. In infinite dimensions, they typically appear as operator-valued backward stochastic evolution equations, backward stochastic Riccati equations, or related mild backward integral systems with terminal condition at time $T$. The subject combines three distinct but tightly connected ingredients: a stochastic integration theory for operator-valued integrands, solution concepts that remain meaningful when the natural operator space is not Hilbertian, and semigroup-based well-posedness mechanisms for concrete infinite-dimensional equations (Lü et al., 2012, Guatteri et al., 2014, Tesko, 2016).

1. Formal structure and mathematical meaning

A canonical operator-valued backward stochastic evolution equation on a Hilbert space $H$ has the formal form

$$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$

where $P(t)\in \mathcal L(H)$ is the operator-valued unknown and $Q$ is the martingale coefficient (Lü et al., 2012). In control-theoretic applications, this equation is the infinite-dimensional analogue of the finite-dimensional matrix-valued second-order adjoint BSDE.

A more specialized but structurally central class is the operator-valued backward stochastic Riccati equation

$$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$

with terminal condition

$P(T)=M,$

posed on an infinite-dimensional Hilbert space and interpreted through a mild backward integral formula (Guatteri et al., 2014).

The phrase “backward stochastic integral equation” emphasizes that these objects are not merely differential expressions written backward in time. Their rigorous content is usually given either by a mild variation-of-constants representation,

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$

or by a duality identity against forward test equations (Lü et al., 2012, Guatteri et al., 2014). In this sense, the integral formulation is primary and the differential notation is often shorthand.

Two recurring features distinguish the operator-valued setting from standard Hilbert-valued BSDEs. First, the unknown $P(t)$ acts on the state space $H$, so the drift contains noncommutative compositions such as $H$0, $H$1, $H$2, and $H$3. Second, in infinite dimensions the natural space $H$4 of bounded operators is a Banach space that is not, in general, an adequate Hilbertian environment for classical stochastic calculus (Lü et al., 2012, Guatteri et al., 2014).

2. Operator-valued stochastic integration as a foundational layer

A foundational integration framework is provided by the Hilbert-space-valued stochastic integral of operator-valued functions introduced in "A stochastic integral of operator-valued functions" (Tesko, 2016). The setting starts with a complex Hilbert space $H$5, a nonzero vector $H$6, and a right-continuous increasing family of orthogonal projections

$H$7

which determines a projector-valued measure $H$8. The associated abstract martingale is

$H$9

and the corresponding $$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$0-valued set function is

$$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$1

The control measure is

$$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$2

The integrands are operator-valued. For $$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$3, the increment subspace is

$$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$4

and the relevant operator space is

$$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$5

An operator is $$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$6-measurable if it is bounded on $$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$7, its operator norm is consistent across later times, and it partially commutes with the resolution of identity: $$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$8 This is the paper’s substitute for adaptedness.

For a simple $$\begin{cases} dP = - (A^*+J^*)P\,dt - P(A+J)\,dt - K^*PK\,dt -(K^*Q+QK)\,dt +F\,dt +Q\,dw(t), & \text{in } [0,T),\ P(T)=P_T, \end{cases}$$9-adapted operator-valued function

$P(t)\in \mathcal L(H)$0

the stochastic integral is defined by

$P(t)\in \mathcal L(H)$1

The $P(t)\in \mathcal L(H)$2-type quasinorm on simple integrands is

$P(t)\in \mathcal L(H)$3

and completion yields the admissible class $P(t)\in \mathcal L(H)$4. The key estimate is

$P(t)\in \mathcal L(H)$5

together with linearity of the integral. Orthogonality of projector increments and the partial commutation condition are the structural devices behind this bound.

This construction is not itself a theory of backward stochastic equations. The paper explicitly does not formulate or solve BSDEs or backward stochastic integral equations. Nevertheless, it shows that the abstract integral recovers both the classical Itô integral with respect to a normal martingale and the Itô integral in symmetric Fock space. In the normal-martingale realization, $P(t)\in \mathcal L(H)$6-measurability coincides exactly with ordinary adaptedness of multiplication operators, and the admissible class $P(t)\in \mathcal L(H)$7 becomes the classical $P(t)\in \mathcal L(H)$8 space (Tesko, 2016). This suggests that operator-valued backward equations can be supplied with a rigorous stochastic integral term once an appropriate notion of operator adaptedness and $P(t)\in \mathcal L(H)$9-control is available.

3. Transposition and relaxed transposition formulations

A general theory for operator-valued backward stochastic evolution equations is developed by Lü and Zhang through transposition methods (Lü et al., 2012). The central motivation is that direct $Q$0-valued stochastic integration is unavailable in the needed level of generality. The paper works on a complete filtered probability space with a one-dimensional standard Brownian motion and explicitly allows a general filtration, without assuming natural filtration or quasi-left continuity.

The key obstacle is functional-analytic. In the infinite-dimensional setting, $Q$1 is nonreflexive and nonseparable, so one cannot directly invoke the standard Hilbertian martingale representation or the usual BSDE machinery. To bypass this, the equation is defined through bilinear identities against forward stochastic evolution equations. For forward test systems

$Q$2

the operator-valued transposition solution $Q$3 is characterized by the identity

$Q$4

The backward martingale term is therefore encoded by duality rather than by constructing a strong $Q$5-valued stochastic integral.

The theory has two levels. In the Hilbert–Schmidt setting, where

$Q$6

the equation admits a unique transposition solution

$Q$7

and the problem can be treated by lifting it to the Hilbert space $Q$8 (Lü et al., 2012).

For general bounded-operator data, the paper proves uniqueness of transposition solutions but establishes existence only in a relaxed sense. The relaxed transposition solution replaces a bona fide process $Q$9 by operator families

$$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$0

which act on forward test triples and enter a modified duality identity. Theorem 6.1 yields a unique relaxed transposition solution

$$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$1

This is one of the defining features of the subject: in full operator generality, the martingale integrand may be representable only indirectly.

The methodological core consists of finite-dimensional approximations, Yosida regularization, and new weakly sequential Banach-Alaoglu-type compactness results for uniformly bounded operator families. Those compactness theorems make it possible to pass from matrix-valued approximating BSDEs to an infinite-dimensional operator-valued limit (Lü et al., 2012).

4. Mild semigroup formulations and operator-valued BSREs

A different resolution of the operator-valued backward problem is given by the mild theory of infinite-dimensional backward stochastic Riccati equations (Guatteri et al., 2014). The underlying space is a real separable Hilbert space $$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$2, and the unbounded drift operator $$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$3 is assumed self-adjoint, diagonalizable, and such that

$$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$4

with

$$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$5

Under these assumptions, $$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$6 generates an analytic contraction semigroup $$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$7.

The decisive idea is not to seek the martingale coefficient $$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$8 in $$-dP(t) = \Big( A^*P(t)+P(t)A +C^*(t)P(t)C(t) +C^*(t)Q(t)+Q(t)C(t) -P(t)B(t)B^*(t)P(t) +S(t) \Big)\,dt - Q(t)\,dW(t),$$9 itself. Instead, the paper introduces

$P(T)=M,$0

forming a Hilbert triple

$P(T)=M,$1

and defines the Hilbert space of operators

$P(T)=M,$2

A central structural fact is

$P(T)=M,$3

Its symmetric subspace $P(T)=M,$4 is used for $P(T)=M,$5. The semigroup regularization bounds

$P(T)=M,$6

are then used to control sandwiched operator expressions of the form

$P(T)=M,$7

even though $P(T)=M,$8 need not belong to $P(T)=M,$9.

For the Lyapunov equation, the mild solution is a pair

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$0

satisfying

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$1

Theorem 3.4 proves existence and uniqueness of this mild solution.

For the Riccati equation, the mild form becomes

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$2

Theorem 4.4 establishes existence and uniqueness of a mild solution $$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$3 on $$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$4, with

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$5

The proof is local-to-global. Regularized approximants based on

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$6

produce Hilbert-Schmidt-valued backward equations, for which standard Hilbert-space BSDE arguments are available. A priori estimates are obtained first on short backward intervals, including the local bound

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$7

and continuation then yields global existence (Guatteri et al., 2014).

5. Control-theoretic roles: second-order adjoints and feedback operators

Operator-valued backward equations enter stochastic control in two distinct but related ways.

In the Pontryagin framework for general stochastic evolution equations,

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$8

with cost

$$P(t)=\text{terminal term}+\int_t^T (\cdots)\,ds+\int_t^T (\cdots)\,dW(s),$$9

the first adjoint is vector-valued, but control-dependent diffusion and nonconvex control domains require a second-order adjoint equation that is operator-valued (Lü et al., 2012). The Hamiltonian is

$P(t)$0

and the second-order adjoint data are

$P(t)$1

$P(t)$2

The resulting operator-valued BSEE supplies the quadratic correction term in the maximum condition,

$P(t)$3

where

$P(t)$4

This term records the second-order effect of diffusion perturbations and is indispensable in the nonconvex, control-dependent diffusion regime.

In stochastic linear-quadratic control, the operator-valued BSRE plays the synthesis role usually occupied by the Riccati equation in deterministic control. The controlled state equation is

$P(t)$5

with cost

$P(t)$6

The mild BSRE solution $P(t)$7 determines the value function through

$P(t)$8

and the optimal feedback law is

$P(t)$9

The closed-loop state equation is then

$H$0

Thus, in the maximum-principle setting the backward operator equation functions as a second-order adjoint, whereas in the LQ setting it functions as a feedback-generating Riccati equation (Lü et al., 2012, Guatteri et al., 2014).

6. Functional-analytic difficulties, scope, and limitations

The theory is shaped by several obstructions that do not arise in finite-dimensional matrix-valued BSDEs.

The first is that the natural operator space $H$1 is only a Banach space and, in infinite dimensions, is nonseparable and nonreflexive. Lü and Zhang explicitly identify this as the reason classical operator-valued martingale representation and standard BSDE methods fail in the general bounded-operator setting (Lü et al., 2012). Their response is the transposition framework and, in full generality, the relaxed transposition solution. A common misconception is that every operator-valued backward stochastic equation should admit a strong $H$2-valued martingale integrand; the general theory in fact does not provide this. It provides a surrogate object $H$3 defined through dual operator families.

The second is that even when one enlarges the operator space to a Hilbert space, algebraic drift terms need not preserve that space. In the BSRE theory, $H$4 is placed in

$H$5

but terms such as $H$6 need not belong to $H$7. The solution is not to close the drift in $H$8 abstractly, but to exploit analytic semigroup smoothing after sandwiching by $H$9 and $H$00 (Guatteri et al., 2014). This is a highly specific infinite-dimensional mechanism rather than a generic feature of operator-valued BSDEs.

The third is that stochastic integration of operator-valued integrands requires an adaptedness notion that remains meaningful beyond scalar or vector processes. Tesko’s construction supplies exactly such a notion through $H$01-measurability and an $H$02-type control measure, but the paper does not develop backward equations or existence and uniqueness for operator-valued BSDEs (Tesko, 2016). It is therefore best interpreted as a preparatory integration theory.

These three strands delineate the current scope reflected in the cited works. One strand gives an abstract operator-valued stochastic integral with classical and Fock-space specializations. A second gives a general operator-valued backward evolution theory in duality form, with relaxed existence in the bounded-operator regime. A third gives direct mild well-posedness for Lyapunov and Riccati equations under analytic semigroup regularization. This suggests that a fully unified theory of operator-valued backward stochastic integral equations would need to combine operator-valued integration, weak or mild notions of solution, and problem-specific regularization mechanisms; however, that synthesis is not proved in these papers (Lü et al., 2012, Guatteri et al., 2014, Tesko, 2016).