Coupled Difference Riccati Equations (CDREs)

• Coupled Difference Riccati Equations (CDREs) are discrete-time recursions that link multiple symmetric matrices through shared feedback gains, harmonic means, or cross-coupled dynamics.
• They underpin key applications in finite-horizon LQ control, distributed filtering, and dynamic game theory, with rigorous existence and convergence results.
• Analytical tools such as semigroup factorizations, invariant subspace decompositions, and pseudoinverse techniques are used to elucidate the structure and dynamics of CDREs.

Searching arXiv for recent and foundational papers directly relevant to coupled difference Riccati equations and close variants. Coupled Difference Riccati Equations (CDREs) are discrete-time Riccati-type matrix recursions or fixed-point systems in which several unknown symmetric matrices are linked through shared feedback gains, inverse-information aggregates, or cross-coupled dynamics. In the literature represented here, that class includes finite-horizon generalized Riccati recursions with singular terms, forward Riccati semigroups, harmonic-coupled covariance recursions arising from distributed filtering, and stochastic cross-coupled backward Riccati recursions governing closed-loop Nash equilibria in nonzero-sum difference games (Ferrante et al., 2013, Qian et al., 2022, Meng et al., 22 Jul 2025). The subject therefore sits at the intersection of discrete-time LQ control, optimal filtering, distributed estimation, and dynamic game theory.

1. Representative formulations

A useful starting point is the single generalized Riccati difference equation (GRDE) for the finite-horizon discrete-time LQ problem

$x_{t+1}=Ax_t+Bu_t,$

with cost

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$

where the Popov matrix

$$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$

The backward recursion is

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$

with terminal condition $X_T=P$. Although this is a single equation, its structural analysis supplies tools that are directly relevant to CDREs in singular and generalized settings (Ferrante et al., 2013).

The genuinely coupled discrete-time formulation in distributed filtering is the harmonic-coupled Riccati equation (HCRE),

$$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$

together with its associated forward iteration

$$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$

Here the coupling is through harmonic means of neighboring information matrices, rather than through direct linear combinations of the $P_j$ (Qian et al., 2022).

In two-person stochastic nonzero-sum difference games with random coefficients, the Riccati objects are a pair of cross-coupled stochastic Riccati recursions, called CCREs, of the form

$$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$

$$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$

and symmetrically for Player 2. These are backward, game-theoretically cross-coupled, and stochastic because the operators $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$0, $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$1, and $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$2 contain conditional expectations of random coefficients (Meng et al., 22 Jul 2025).

A complementary operator viewpoint is given by the autonomous Riccati maps

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$3

whose forward semigroup iterates $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$4 and $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$5 are not themselves coupled systems, but provide a language for analyzing time-varying Riccati flows and their transition operators (Moral et al., 2021).

Formulation	Representative equation	Source
Generalized finite-horizon recursion	$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$6 with pseudoinverse term	(Ferrante et al., 2013)
Harmonic-coupled filtering recursion	$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$7	(Qian et al., 2022)
Stochastic game-theoretic CDRE analogue	$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$8, coupled through opponent gains	(Meng et al., 22 Jul 2025)

2. Mechanisms of coupling

The literature exhibits several distinct coupling mechanisms. In nonzero-sum game formulations, coupling is mediated by feedback gains. In the stochastic setting, $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$9 determines $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$0, but the recursion for $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$1 depends on $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$2; $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$3 determines $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$4, but the recursion for $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$5 depends on $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$6. The coupling is therefore cross-player and closed-loop: each player’s backward Riccati recursion is parameterized by the other player’s feedback law (Meng et al., 22 Jul 2025).

In distributed filtering, the coupling is not through gains but through inverse covariances. The HCRE replaces a local prior information term $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$7 by a network aggregate

$$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$8

so that the covariance variable $$\Pi=\begin{bmatrix}Q&S\S^\top&R\end{bmatrix}=\Pi^\top\ge 0, \qquad P=P^\top>0.$$9 is coupled harmonically to its neighbors. The paper explicitly contrasts this with classical coupled Riccati equations that use arithmetic or algebraic means such as $$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$0. The result is a more nonlinear Riccati-type system induced by information fusion (Qian et al., 2022).

A third mechanism appears in generalized Riccati theory. The GRDE and the associated constrained generalized discrete algebraic Riccati equation,

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$1

with constraint

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$2

show that singularity enters the recursion through Moore–Penrose pseudoinverses and kernel constraints. While this is not yet a coupled system, it identifies the generalized algebraic structure that coupled systems inherit whenever singular control or estimation penalties are present (Ferrante et al., 2013).

A further coupling architecture is visible in the two-player continuous-time Nash CARE system

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$3

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$4

together with the stationarity condition

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$5

This is not a CDRE result, but it makes the Nash coupling pattern explicit and suggests a direct discrete-time analogue based on alternating single-player Riccati solves (Li et al., 2020).

3. Algebraic backbone, singularity, and invariant subspaces

The algebraic backbone of generalized discrete-time Riccati theory is the constrained generalized discrete algebraic Riccati equation, denoted CGDARE$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$6, together with the extended symplectic pencil

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$7

where

$$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$8

If $$X_t=A^\top X_{t+1}A-(A^\top X_{t+1}B+S)(R+B^\top X_{t+1}B)^\dagger(B^\top X_{t+1}A+S^\top)+Q,$$9 is a symmetric solution of CGDARE$X_T=P$0, there exist invertible matrices $X_T=P$1 such that

$X_T=P$2

with

$X_T=P$3

This factorization ties the generalized eigenstructure of the pencil to the closed-loop matrix $X_T=P$4 (Ferrante et al., 2013).

A central structural result is that, although CGDARE may admit many symmetric solutions, they all coincide on a canonical subspace associated with the zero eigenvalue of the closed loop. If

$X_T=P$5

then all solutions $X_T=P$6 and $X_T=P$7 of CGDARE$X_T=P$8 satisfy

$X_T=P$9

and the subspace $$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$0 is independent of the chosen solution. In a basis $$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$1 adapted to $$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$2, any two solutions have the form

$$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$3

The blocks corresponding to $$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$4 and the cross terms are therefore common to all solutions; only the complementary lower-right block may vary (Ferrante et al., 2013).

The same paper also gives a generalized singularity criterion. The matrix $$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$5 is singular if and only if at least one of

$$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$6

is singular. For a symmetric CGDARE solution $$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$7, the closed-loop matrix $$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$8 is singular if and only if

$$P_i=A\Big(\sum_{j=1}^{N}l_{ij}P_j^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q, \qquad i=1,\dots,N,$$9

The paper remarks that the algebraic multiplicity at zero, and hence the nilpotent structure of $$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$0, is invariant over all symmetric solutions of CGDARE$$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$1 (Ferrante et al., 2013).

For CDREs, the immediate significance is structural rather than direct. The literature does not state a coupled analogue of this theorem, but the result identifies a canonical singular or nilpotent component that is independent of which algebraic Riccati solution is chosen. This suggests that coupled recursions with generalized or singular terms may admit a comparable decomposition into a common algebraic singular part and a variable reduced-order part.

4. Dynamic representations and Riccati flow viewpoints

The forward semigroup treatment of discrete-time Riccati maps provides exact identities for the propagation of solution differences. For the autonomous map $$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$2,

$$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$3

where $$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$4 is the directed product generated by

$$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$5

The same framework gives the Fréchet derivative formula

$$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$6

These identities turn the nonlinear Riccati orbit into a transition-operator problem and make sensitivity to initial data explicit (Moral et al., 2021).

Under controllability of $$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$7 and observability of $$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$8, the maps $$P_{i,k+1}=A\Big(\sum_{j=1}^{N}l_{ij}P_{j,k}^{-1}+l_{ij}C_j^\top R_j^{-1}C_j\Big)^{-1}A^\top+Q.$$9 and $P_j$0 each have a unique positive definite fixed point,

$P_j$1

and the associated closed-loop matrices

$P_j$2

satisfy

$P_j$3

The paper’s semigroup duality formula,

$P_j$4

with

$P_j$5

relates the primal and dual Riccati semigroups through a nonlinear harmonic-mean identity. The authors state that this is the first result of this type for discrete-time Riccati difference equations (Moral et al., 2021).

The same paper derives a Floquet-type representation even though the sequence $P_j$6 is generally aperiodic: $P_j$7 This factors an aperiodic Riccati transition product into a stable fixed matrix power and a uniformly bounded invertible correction (Moral et al., 2021).

For CDREs, these identities are not direct theorems, because the paper treats a single autonomous recursion rather than a coupled family. Their importance lies in methodology. They suggest that coupled systems may be studied through semigroup factorizations, directed products, and duality identities that separate asymptotic closed-loop structure from finite-time coupling effects.

5. Solvability, uniqueness, and convergence

The strongest explicit existence-and-convergence result in the coupled discrete-time literature summarized here concerns HCREs. Under the assumptions that $P_j$8 is invertible, $P_j$9 is observable with

$$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$0

that $$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$1 is primitive and row stochastic, and that $$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$2 and $$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$3, the HCRE

$$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$4

has exactly one solution. Moreover, for any initial $$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$5, the iteration

$$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$6

converges to that unique solution. The proof combines eventual boundedness, monotone iteration from below for sufficiently small initial conditions, and decay of induced matrix products $$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$7 (Qian et al., 2022).

The same paper establishes an exact steady-state covariance result for the CIDF setting. If the local filtering covariances converge to the HCRE solution, then the stacked estimation error covariance converges to $$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$8 satisfying the discrete-time Lyapunov equation

$$T^{1}_{k}=\Delta\left(T^{1}_{k+1},\Pi^{2}_{k}\right)-\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right)^{\top}\Upsilon^{-1}(T^{1}_{k+1})\Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$9

where $$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$0 is shown to be Schur stable by constructing

$$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$1

from the Perron–Frobenius left eigenvector $$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$2 of $$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$3 and proving

$$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$4

This shifts the performance characterization from upper bounds to an exact DLE description (Qian et al., 2022).

In stochastic nonzero-sum difference games, solvability is expressed through regularity of the cross-coupled Riccati system. A solution pair $$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$5 is regular if

$$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$6

the corresponding range conditions hold, and the induced feedback operators belong to the admissible classes. The equations are strongly regularly solvable if there exists $$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$7 such that

$$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$8

Under strong regularity of the CCREs and solvability of the associated cross-coupled BS$$\Pi^{1}_{k}=-\Upsilon^{-1}(T^{1}_{k+1}) \Lambda\left(T^{1}_{k+1},\Pi^{2}_{k}\right),$$9Es, the game is closed-loop solvable and the resulting feedback quadruple is a closed-loop Nash equilibrium. Theorem 5.2 further gives necessary and sufficient conditions through cross-coupled Lyapunov-type equations, semidefinite conditions $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$00, and stationarity identities

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$01

for each player (Meng et al., 22 Jul 2025).

By contrast, the alternating solver proposed for the two-player continuous-time coupled CARE system is algorithmically clear but does not come with a full convergence theorem. The paper explicitly states that, due to nonlinearity in $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$02, numerical theory is difficult and is left for future research. This matters for CDREs because it distinguishes computational templates from proved convergence results (Li et al., 2020).

6. Reduction methods, computation, and applications

A major computational idea in generalized Riccati theory is decomposition by a reference algebraic solution. If $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$03 solves CGDARE$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$04, and $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$05 has nilpotency index $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$06 on

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$07

then for the GRDE solution $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$08 the difference

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$09

satisfies

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$10

and therefore, after $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$11 backward steps,

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$12

In coordinates adapted to $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$13,

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$14

with $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$15 nilpotent and $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$16 nonsingular, the recursion reduces for $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$17 to

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$18

where $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$19 obeys the reduced homogeneous GRDE

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$20

The paper emphasizes two computational consequences: dimension reduction from $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$21 to $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$22, and elimination of the singular or nilpotent closed-loop part from the reduced equation (Ferrante et al., 2013).

For HCREs, the practical solver is simply the CIDF covariance recursion itself,

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$23

which converges globally to the unique HCRE solution for any positive definite initialization. In this setting, the Riccati iteration is simultaneously a model of distributed filtering dynamics and a constructive algorithm for computing the steady-state covariance matrices (Qian et al., 2022).

In the two-player Nash setting, the continuous-time alternating method proceeds by freezing one player’s feedback, reducing the other player’s equation to a standard Riccati equation with effective data such as

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$24

and then updating the feedback via the stationarity relation. The same structure applies symmetrically to Player 2. This is not a discrete-time theorem, but it exposes a solver architecture that is directly suggestive for CDREs in dynamic games: an outer fixed-point loop over players wrapped around inner standard Riccati solves (Li et al., 2020).

The stochastic game paper combines dynamic programming with Hamiltonian-system decoupling. The equilibrium controls are given explicitly by

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$25

with gains and affine terms determined by the CCREs and the cross-coupled BS$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$26Es. The equilibrium state then evolves according to a closed-loop stochastic difference equation with both drift and multiplicative-noise terms modified by $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$27. This places CDREs at the center of closed-loop equilibrium synthesis under random coefficients (Meng et al., 22 Jul 2025).

Across these settings, a recurrent theme is that Riccati coupling is rarely a purely formal complication. It is the mechanism through which network information fusion, opponent feedback, singular control penalties, or random adapted coefficients enter the recursion itself. The principal analytical responses in the literature are likewise structural: invariant-subspace reduction, semigroup factorization, monotone or product-decay convergence arguments, and Lyapunov-type reformulations.

7. Conceptual boundaries and recurring misconceptions

One recurring misconception is that every CDRE is a finite-horizon backward recursion of standard LQ type. The literature here shows a broader landscape. The generalized Riccati difference equation is indeed a backward finite-horizon recursion, but HCREs are forward coupled Riccati-type iterations whose steady-state limit defines the algebraic equation, and the semigroup treatment studies forward iterates of an autonomous Riccati map rather than backward dynamic programming (Ferrante et al., 2013, Qian et al., 2022, Moral et al., 2021).

A second misconception is that coupling must occur through arithmetic means or Markov-mode averaging. The HCRE framework demonstrates coupling through the matrix harmonic mean

$J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$28

which is natural in information-form distributed filtering and leads to a distinct nonlinear fixed-point structure (Qian et al., 2022).

A third misconception is that singularity or nonuniqueness of algebraic Riccati solutions necessarily destroys structure. In the generalized setting, all CGDARE solutions coincide on the canonical subspace $J(x_0,u)=\sum_{t=0}^{T-1} \begin{bmatrix}x_t\u_t\end{bmatrix}^{\!\top} \begin{bmatrix}Q&S\S^\top&R\end{bmatrix} \begin{bmatrix}x_t\u_t\end{bmatrix} +x_T^\top P x_T,$29, and the nilpotent structure of the closed-loop matrix is solution-independent. This shows that singularity may coexist with strong structural invariants (Ferrante et al., 2013).

Finally, not every computational scheme for coupled Riccati equations comes with a theorem guaranteeing convergence or uniqueness. The HCRE paper proves existence, uniqueness, and global convergence of the coupled iteration under its assumptions, while the two-player CARE paper supplies a practical alternating method but explicitly does not establish a full convergence theorem for the outer coupled iteration. The stochastic nonzero-sum game paper, in turn, gives equilibrium characterizations through regular solvability and Lyapunov-type conditions rather than through a generic iterative convergence theory (Qian et al., 2022, Li et al., 2020, Meng et al., 22 Jul 2025).

In aggregate, these works present CDREs not as a single equation template but as a family of structurally related discrete-time Riccati systems. Their unifying features are backward or forward matrix recursions, nonlinear coupling through gains or information terms, and a persistent algebraic backbone supplied by generalized Riccati equations, closed-loop matrices, and Lyapunov or symplectic-penciled representations.