Backward Riccati Operator Equation

• Backward Riccati operator equation is a nonlinear, operator-valued differential equation generalizing finite-dimensional Riccati equations, crucial for infinite-dimensional stochastic control.
• It ensures existence, uniqueness, and regularity of solutions using analytic semigroup theory and duality approaches, accommodating unbounded operators and randomness.
• The equation supports optimal feedback control in LQ problems with stochastic and jump dynamics, analyzed through methods like Galerkin approximation, fixed-point techniques, and splitting schemes.

The backward Riccati operator equation is a nonlinear, operator-valued differential or stochastic differential equation fundamental to optimal control, particularly in infinite-dimensional systems governed by partial differential equations or stochastic partial differential equations. Its formulation generalizes the classical Riccati equations found in finite-dimensional linear-quadratic regulator (LQR) theory, accommodating unbounded operators, operator-valued coefficients, randomness, and even jumps. Existence, uniqueness, and regularity properties of solutions, as well as their interplay with optimal feedback synthesis and approximation, are central in the modern theory of infinite-dimensional stochastic control and the analysis of controlled SPDEs.

1. Formal Definition and Structural Variants

The backward Riccati operator equation on a real separable Hilbert space $H$ typically takes the form

$$-\frac{d}{dt}P(t) = \mathcal{A}^*P(t)+P(t)\mathcal{A} + C(t)^*P(t)C(t) + C(t)^*Q(t) + Q(t)C(t) - P(t)B(t)B(t)^*P(t) + S(t),$$

with terminal condition $P(T) = M$. Here,

• $\mathcal{A}: D(\mathcal{A}) \subset H \to H$ is an unbounded self-adjoint operator generating an analytic contraction semigroup,
• $B(t)$ and $C(t)$ are (possibly stochastic) bounded operator-valued processes,
• $S(t)$ and $M$ are bounded, self-adjoint positive semidefinite operator-valued processes.

Stochastic variants, known as backward stochastic Riccati equations (BSREs), incorporate operator-valued stochastic integrals, e.g.,

$$-\mathrm{d}P(t) = [\cdots]\,\mathrm{d}t - Q(t) \, \mathrm{d}W(t),$$

where $Q(t)$ is an additional operator-valued process and $W(t)$ is a Wiener process (Guatteri et al., 2014). Further extensions include jump terms, leading to backward stochastic Riccati equations with jumps (BSREJ) (Zhang et al., 2018).

2. Solution Concepts and Functional Analytic Framework

The infinite-dimensional, operator-valued setting precludes strong (pointwise-defined) solutions in most cases, requiring advanced notions defined in appropriate function spaces:

• Mild solution: Given via the variation-of-constants formula, leveraging the analytic semigroup $e^{t\mathcal{A}}$ on $H$. The mild solution $(P,Q)$ satisfies, almost surely for all $t\in[0,T]$,

$$\begin{aligned} P(t) &= e^{(T-t)\mathcal{A}} M e^{(T-t)\mathcal{A}} + \int_t^T e^{(s-t)\mathcal{A}} S(s) e^{(s-t)\mathcal{A}} ds \ &\quad + \int_t^T e^{(s-t)\mathcal{A}} \left[C(s)^* P(s) C(s) + C(s)^* Q(s) + Q(s)C(s) - P(s) B(s) B(s)^* P(s)\right] e^{(s-t)\mathcal{A}} ds\ &\quad + \int_t^T e^{(s-t)\mathcal{A}} Q(s) e^{(s-t)\mathcal{A}} dW(s). \end{aligned}$$

The solution is in $$L^2(\Omega;C([0,T];E_+(H))) \times L^2(\Omega\times[0,T];K_s)$$, where $E_+(H)$ is the cone of symmetric nonnegative operators and $K_s$ the space of symmetric Hilbert–Schmidt operators (Guatteri et al., 2014).

• Transposition solution: When $L(H)$ lacks desirable properties (nonreflexivity, non-UMD), the BSRE is posed weakly, by duality relations with test processes governed by forward stochastic evolution equations. Existence and uniqueness are established via these duality requirements and positivity/invertibility constraints on associated operators (Lu et al., 2019).

Table 1: Summary of Solution Notions

Solution Notion	Key Mathematical Setting	References
Mild	Analytic semigroup, $L^2$ spaces	(Guatteri et al., 2014, Yoshioka et al., 2022)
Transposition	Duality with forward SEEs	(Lu et al., 2019)

3. Existence, Uniqueness, and Regularity Results

Existence and uniqueness of mild solutions are established under:

• Self-adjointness and spectral gap of $\mathcal{A}$
• Boundedness and adaptedness of $B(\cdot), C(\cdot), S(\cdot), M$
• Square-integrability and measurability conditions on stochastic coefficients

The analytic contraction semigroup generated by $\mathcal{A}$ plays a crucial regularizing role (Guatteri et al., 2014). The main argument employs:

• Approximation by finite-rank or Hilbert–Schmidt operator equations
• Uniform a priori estimates (via Doob’s inequality and energy arguments)
• Contraction mappings on small time intervals, patched globally

Transposition solutions exist and are unique if the associated stochastic linear-quadratic (SLQ) control problem admits an optimal $L^2$ feedback operator. Critical is the positivity/invertibility of

$K(t) = R(t) + D(t)^*P(t) D(t),$

and strong integrability properties of all coefficients. A measurable selection argument ensures $K(t,\omega)$ is almost surely invertible in the infinite-dimensional setting (Lu et al., 2019).

Regularity in time, operator-norm continuity, and self-adjointness are inherited from the smoothing properties of the semigroup and the structure of the data (Guatteri et al., 2014, Hansen et al., 25 Apr 2025).

4. Applications to Infinite-Dimensional Stochastic LQ Control

The backward Riccati operator equation provides necessary and sufficient conditions for optimal feedback synthesis in infinite-dimensional LQ problems: $$\min_{u \in L^2} \mathbb{E} \left[ \int_0^T \left( \|y(t)\|^2 + \|u(t)\|^2 \right) dt + (M y(T), y(T)) \right], \quad dy(t) = [\mathcal{A} y(t) + B(t) u(t)]dt + C(t) y(t) dW(t).$$

Given the unique solution $(P,Q)$, the value function is quadratic,

$V(x) = (P(0)x,x)_H,$

and the optimal control is in state feedback form,

$u^*(t) = -B(t)^* P(t) y^*(t),$

where $y^*$ solves the closed-loop stochastic evolution (Guatteri et al., 2014). In settings with jumps or random coefficients, the associated operator-valued BSRE or BSREJ governs the value function and feedback law (Zhang et al., 2018, Yoshioka et al., 2022). The connection extends to Markovian performance evaluation via the Kolmogorov backward equation and Hamilton–Jacobi–Bellman (HJB) theory (Yoshioka et al., 2022).

5. Analytical and Numerical Solution Techniques

Analytical methods rest on the regularity and contractivity of the semigroup, monotonicity properties of the Lyapunov operator, and the accretivity of the Gelfand triple embeddings involved:

• Galerkin/Faedo-approximation: Projection onto finite-dimensional subspaces, yielding matrix-valued Riccati equations solvable by classical BSDE theory, then passage to the infinite-dimensional limit (Lu et al., 2019).
• Fixed-point techniques: With Banach fixed-point theorem on cones of self-adjoint operators for the algebraic Riccati step in the finite-difference time-discretization (Eisenmann et al., 2018).
• Monotonicity and energy arguments: Essential for rigorous control of quadratic nonlinearities.

For temporal discretization, Lie and Strang splitting schemes decompose the Riccati operator flow into linear (Lyapunov) and nonlinear (quadratic) components. Under sufficient regularity (analytic semigroup, smooth initial data or smoothing nonlinearity), operator-norm convergence rates of $O(\tau)$ (Lie) and $O(\tau^2)$ (Strang) are obtained (Hansen et al., 25 Apr 2025). The backward Euler scheme guarantees convergence in $C([0,T];H_{op})$ and $L^2(0,T;V_{op})$ for a wide class of right-hand sides and initial data (Eisenmann et al., 2018).

Finite-dimensional kernel discretization and explicit backward Euler time-stepping enable practical computation of integral operator Riccati equations arising in control of spatially continuous stochastic systems (Yoshioka et al., 2022).

Table 2: Numerical Schemes for Operator-valued Backward Riccati Equations

Method	Key Features	Order / Convergence	References
Lie Splitting	Linear & quadratic subflows	$O(\tau)$/$O(\tau \log \tau)$	(Hansen et al., 25 Apr 2025)
Strang Splitting	Symmetric operator splitting	$O(\tau^2)$	(Hansen et al., 25 Apr 2025)
Backward Euler	Implicit time-stepping	Strong/weak convergence	(Eisenmann et al., 2018)
Finite Difference	State kernel discretization	$O(n^{-1}), O(\Delta t)$	(Yoshioka et al., 2022)

6. Extensions: Stochasticity, Jumps, and Indefiniteness

BSREs accommodate Brownian and Poisson (jump) noise, random and possibly unbounded coefficients, and even indefiniteness in the cost weights:

• BSREJ: Incorporate both Brownian and jump noise. The solution is a càdlàg (right-continuous with left limits), nonnegative, progressively measurable process triple $(K,L,R)$ with BMO-martingale properties. Existence and uniqueness rely on positivity of block-matrices built from $K$, $L$, $R$ (Zhang et al., 2018).
• Indefinite problems: Allow weighting operators in the cost (and cross-terms) to be indefinite. Well-posedness hinges on strong uniform convexity/positivity conditions on certain block-matrix combinations, ensuring invertibility in the feedback synthesis and quadratic value representation (Sun et al., 2022).
• Transposition and verification results: Dual formulations guarantee that the optimal feedback and value function construction is equivalent to solvability of the associated operator-valued BSRE or Riccati ODE (Lu et al., 2019, Guatteri et al., 2014).

7. Relation to Forward and Forward-Backward Riccati Equations

In indefinite and backward stochastic LQ optimization, the operator-valued Riccati object can also arise as a forward-in-time equation after completion of squares or suitable transformation of variables. Specifically, for problems with indefinite cross-terms or nonhomogeneous data, the construction typically produces a coupled Riccati equation and an ancillary affine equation, linking backward SDEs and classic forward LQ theory. This structural flexibility is central in multidimensional backward stochastic control and FBSDE systems (Sun et al., 2022).

The backward Riccati operator equation thus serves as the foundational analytic object linking infinite-dimensional control, stochastic analysis, and the synthesis of optimal feedback in the presence of unbounded, random, and discontinuous system dynamics. Its study unifies techniques from operator semigroup theory, stochastic calculus, functional analysis, and numerical analysis.