Finite-Horizon Stochastic BRL

• Finite-Horizon Stochastic BRL is a framework that establishes necessary and sufficient conditions for disturbance attenuation in stochastic linear systems using Riccati recursions or LMIs.
• It generalizes classical H∞ control by incorporating stochastic inputs, mean-field effects, and distributional uncertainties through anisotropic and induced norm measures.
• Practical implementation involves numerical backward recursions and LMIs to certify performance bounds, aiding robust filtering, controller synthesis, and uncertainty quantification.

The finite-horizon stochastic bounded real lemma provides necessary and sufficient conditions for the disturbance attenuation (in the sense of an induced norm) of discrete- or continuous-time linear systems subject to stochastic inputs, disturbances, and model uncertainties. It extends the classic bounded real lemma (BRL) from deterministic $\mathcal{H}_\infty$ control theory to finite-horizon, stochastic, and possibly mean-field and distributionally uncertain settings. The lemma formalizes the relationship between system parameters, Riccati recursions or LMIs, and norm bounds for worst-case output amplification in the presence of general stochastic uncertainties.

1. System Classes and Stochastic Uncertainty Modeling

Finite-horizon stochastic BRL results apply to a broad class of disturbed systems:

• Linear Discrete-Time Varying (LDTV):

$x_{k+1} = A_k x_k + B_k w_k$, $z_k = C_k x_k + D_k w_k$, $x_0 = 0$ (Maximov et al., 2012).

• Stochastic Difference in Hilbert Spaces:

$$x(k+1) = A_k x(k) + B_k w(k) + [C_k x(k) + D_k w(k)]\, \omega(k)$$, $z(k) = C^z_k x(k) + D^z_k w(k)$, $x(0)=0$ (Li et al., 9 Jan 2026).

• Systems with Mean-Field and Noise:

$x(k+1)=A(k)x(k)+\bar A(k)\mathbb{E}[x(k)] + \cdots$ (mean-field terms and multiplicative noise) (Weihai et al., 2016).

• Continuous-Time Mean-Field Stochastic Systems:

$dX(s) = \cdots$ (Itô equations with mean-field coupling and affine terms) (Fang et al., 26 Jul 2025).

• General Linear-Quadratic Framework:

Covers systems with state, control, and disturbance covariance formulations, including stochastic models as special cases (Bamieh, 2024).

Stochastic uncertainty is modeled either by:

• Color and correlation structure of noise (white, colored, or with imprecisely known distributions).
• Anisotropy of noise—a parameter measuring entropy-theoretic deviation from white Gaussian law (Maximov et al., 2012).
• Distributional ambiguity (admissible input distributions constrained by relative entropy or second-order moments).

2. Norms and Performance Objectives

The framework is built around induced norms quantifying the maximal amplification from disturbance to output in mean-square sense, over the considered time horizon:

• RMS-gain (for operator $F$ and random $W$):

$$\Gamma(F,W) = \sqrt{ \frac{\mathbb{E}|F W|^2}{\mathbb{E}|W|^2} }$$

• $a$-anisotropic norm: Maximal RMS-gain over all inputs $W$ of anisotropy up to $a$:

$\|F\|_a = \sup_{W \in \mathcal{W}_a} \Gamma(F, W)$

where $\mathcal{W}_a$ is the set of input random vectors of anisotropy $\leq a$ (Maximov et al., 2012).

• $\mathcal{H}_\infty$ norm: For general stochastic or mean-field models, the induced norm is:

$$\|z\|_{\ell_2} = \left( \sum_{k=0}^N \mathbb{E}|z(k)|^2 \right)^{1/2}$$

and the performance objective is $\|z\|_{\ell_2} \leq \gamma \|w\|_{\ell_2}$ for all admissible disturbances (Weihai et al., 2016).

As constraints on the noise law are relaxed (e.g., $a \rightarrow \infty$), the anisotropic norm reduces to the classical $\mathcal{H}_\infty$ norm.

3. Finite-Horizon Stochastic Bounded Real Lemma: Main Results

The lemma gives a test for the norm bound in terms of the feasibility of recursion relations (Ricatti equations) or LMIs:

3.1. Anisotropic-Norm BRL (LDTV, (Maximov et al., 2012))

For an LDTV system with output operator $F_{0:N}$, the ANBRL states:

• For prescribed $\gamma>0$, anisotropy bound $a\geq 0$, $\|F\|_a < \gamma$ iff there exists $q \in (0, \gamma^{-2})$ such that:
  • Forward Riccati recursion

  $$\begin{cases} S_k = I_r - C_k R_k C_k^T - q D_k D_k^T \ M_k = - (A_k R_k C_k^T + q B_k D_k^T) S_k^{-1} \ R_{k+1} = A_k R_k A_k^T + q B_k B_k^T + M_k S_k M_k^T \ R_0 = 0 \end{cases}$$

  with $S_k \succ 0$ for all $k$. - Determinant inequality:

  $$\sum_{k=0}^N \ln \det S_k \geq m (N+1) \ln (1 - q \gamma^2) + 2a.$$

• As $a \to \infty$, the determinant constraint becomes redundant and the result reduces to standard $\mathcal{H}_\infty$-BRL.

3.2. Hilbert Space Stochastic BRL (Li et al., 9 Jan 2026)

For systems in separable Hilbert spaces with additive and multiplicative noise, $\|\mathcal{L}\| < \gamma$ if and only if there exists a sequence of self-adjoint operators $\{Y_k\}_{k=0}^{N+1}$ with $Y_{N+1}=0$ such that:

• Backward Riccati recursion:

$$\begin{aligned} Y_k &= A_k^* Y_{k+1} A_k + C_k^* Y_{k+1} C_k + (C_k^z)^* C_k^z \ &\quad - (A_k^* Y_{k+1} B_k + C_k^* Y_{k+1} D_k + (C_k^z)^* D_k^z) \ &\quad \times [\gamma^2 I - (D_k^z)^* D_k^z + B_k^* Y_{k+1} B_k + D_k^* Y_{k+1} D_k]^{-1} \ &\quad \times (B_k^* Y_{k+1} A_k + D_k^* Y_{k+1} C_k + (D_k^z)^* C_k^z) \end{aligned}$$

• For all $k$, the “denominator” $$R_k = \gamma^2 I - (D_k^z)^* D_k^z + B_k^* Y_{k+1} B_k + D_k^* Y_{k+1} D_k \succ 0$$.

3.3. Mean-Field Stochastic and Continuous-Time Extensions (Weihai et al., 2016, Fang et al., 26 Jul 2025)

In mean-field discrete-time settings, two coupled backward Riccati recursions for $P(k)$ and $Q(k)$ (with $P(N+1)=Q(N+1)=0$) and LMI positivity conditions are required for BRL feasibility. In continuous-time mean-field Itô systems, the lemma relies on the existence, uniqueness, and positive definiteness of solutions to coupled Riccati differential equations (CDREs) over the finite time horizon, along with solvability of auxiliary backward SDEs and deterministic ODEs capturing affine input effects.

3.4. Covariance Representation and LMI Duality (Bamieh, 2024)

Within the linear-conic duality framework, the stochastic finite-horizon $\mathcal{H}_\infty$ analysis is cast as a semidefinite program (SDP) over joint state-input-noise covariances, with dual LMIs and equivalence to Riccati recursions. Specifically, the dual problem yields for each $k$: $$\begin{bmatrix} -A_{k+1}^T P_{k+1} A_{k+1} + C_k^T C_k & A_k^T P_{k+1} B_k + C_k^T D_k \ * & -P_k + B_k^T P_{k+1} B_k + D_k^T D_k + \gamma^2 I \end{bmatrix} \succeq 0$$ with backward Riccati recursion for $P_k$.

4. Theoretical Derivation and Underlying Methods

• Maximum Entropy and Duality: The worst-case disturbance is often attained at a Gaussian law with covariance derived via Kullback-Leibler duality and entropy maximization over input distributions (Maximov et al., 2012).
• Riccati Equations: The completion of squares and optimal state feedback structure in the performance index require solvability of Riccati difference or differential equations. In mean-field cases, these are coupled systems tracking both mean and fluctuation contributions (Weihai et al., 2016, Fang et al., 26 Jul 2025).
• LMI Formulation: In the covariance representation approach, the stochastic BRL emerges as the dual feasibility of a family of finite-horizon LMIs, connecting state-input covariance constraints to performance bounds (Bamieh, 2024).
• Auxiliary Equations: The presence of affine or mean-field terms introduces auxiliary backward SDEs or ODEs whose solvability is required for the disturbance attenuation certificate (Fang et al., 26 Jul 2025).

5. Comparison to the Classical Deterministic Bounded Real Lemma

The finite-horizon stochastic BRL generalizes the traditional deterministic BRL along several axes:

• Distributional Robustness: Stochastic versions quantify robustness to unknown noise distributions via entropy or anisotropy constraints (Maximov et al., 2012).
• Statistical Uncertainty: The Riccati recursions and determinant/global constraints (e.g., $\sum_k \ln \det S_k$) encode the effect of statistical uncertainty, not present in deterministic BRL.
• Inheritance: As the constraint on statistical uncertainty is relaxed or disappears (e.g., $a\to\infty$ in the anisotropic norm, noise law becomes completely unknown), the stochastic BRL reduces to the classical $\mathcal{H}_\infty$ BRL as a special case (Maximov et al., 2012).

Lemma Flavor	Key Riccati Type / Constraint	Noise Model	Extra Condition
Deterministic $\mathcal{H}_\infty$	Riccati with $q=\gamma^{-2}$	Deterministic/worst-case	None
Anisotropic/Robust BRL	Riccati with $q \in (0, \gamma^{-2})$	Uncertain law, bound $a$	Determinant-inequality
Mean-field/BSDE variants	Coupled Riccati or CDRE, Aux. SDE/ODE	Mean-field/multiplicative	Invertibility (gain)

6. Practical Implementation and Applications

In practical terms, verification and controller synthesis via the finite-horizon stochastic BRL proceeds as follows:

• Specify system dynamics (possibly with mean-field or multiplicative noise).
• Formulate RMS or $\ell_2$-induced norm-based performance requirements.
• Solve the relevant Riccati difference/differential equations or check LMI feasibility (with determinant/invertibility constraints if anisotropy or mean-field structure is present).
• Verify positive-definiteness of key block or denominator terms at each recursion step.
• Compute worst-case disturbance and, if required, feedback gains for optimal attenuation.
• Use numerical backward recursions and matrix computations to determine critical norm thresholds and establish explicit performance bounds (Maximov et al., 2012, Li et al., 9 Jan 2026, Fang et al., 26 Jul 2025).

Applications span robust filtering, $\mathcal{H}_2/\mathcal{H}_\infty$ mixed control, mean-field games, and stochastic optimization for uncertain environments in engineering systems.

7. Extensions and Related Results

• The framework naturally extends to continuous-time, infinite-dimensional, and operator-theoretic systems (e.g., Hilbert-space setups) (Li et al., 9 Jan 2026).
• The mean-field stochastic BRL is the analytical engine underlying finite-horizon $H_2/H_\infty$ control with mean-field and affine effects (Fang et al., 26 Jul 2025), and mean-field mixed stochastic systems (Weihai et al., 2016).
• The anisotropic norm approach provides a refined interpolation between purely stochastic norm bounds and the deterministic robust setting, quantifying probabilistic uncertainty via entropy (Maximov et al., 2012).

The finite-horizon stochastic bounded real lemma thus represents a comprehensive unification of performance analysis, uncertainty modeling, LQ theory, and duality-based optimization in stochastic control and robust system theory.