MF Stochastic Bounded Real Lemma

• MF-SBRL is a bounded-real characterization for stochastic systems with mean-field coupling, ensuring the induced ℓ2 gain is below a prescribed level γ.
• It employs coupled backward Riccati-type recursions and positivity constraints to manage the separate energy dynamics of centered and mean state components.
• MF-SBRL underpins mixed H2/H∞ synthesis by shaping disturbance attenuation and controller performance in mean-field stochastic control problems.

Searching arXiv for the cited MF-SBRL papers and closely related formulations. Mean-Field Stochastic Bounded Real Lemma (MF-SBRL) denotes a bounded-real characterization for stochastic systems whose dynamics depend not only on the state and disturbance, but also on their expectations. In the finite-horizon discrete-time setting, the MF-SBRL gives sufficient conditions under which the induced $\ell_2$ gain from an exogenous disturbance to a controlled output is strictly less than a prescribed level $\gamma$, using coupled backward Riccati-type recursions and positivity constraints (Weihai et al., 2016). In a later continuous-time formulation with affine terms, the same bounded-real question is expressed through coupled differential Riccati equations, together with BSDE and ODE components needed to accommodate affine drift and diffusion terms, and the bounded-real condition is stated as an equivalence (Fang et al., 26 Jul 2025).

1. Mean-field structure and system-theoretic setting

In the discrete-time formulation, the state $x_k \in \mathbb{R}^n$, control input $u_k \in \mathbb{R}^q$, exogenous disturbance $v_k \in \mathbb{R}^\ell$, and performance output $z_k \in \mathbb{R}^m$ evolve over a finite horizon $k \in \{0,1,\dots,K\}$ according to

$$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$

with deterministic $x_0$ and scalar stochastic process $w_k$ satisfying

$\gamma$0

The disturbance $\gamma$1 belongs to $\gamma$2 over the horizon, is $\gamma$3-measurable, and is uncorrelated with $\gamma$4 (Weihai et al., 2016).

A standard notational device is the introduction of aggregated matrices

$\gamma$5

These aggregated quantities encode the mean-field coupling and appear systematically in the bounded-real recursions.

The defining structural feature is the decomposition into centered and mean components. Taking expectations yields

$\gamma$6

while the fluctuation dynamics are written in terms of $\gamma$7 and $\gamma$8. This split is not cosmetic: it is the mechanism through which the mean-field coupling enters the bounded-real analysis and produces an additional matrix recursion absent from the classical non-mean-field stochastic BRL (Weihai et al., 2016).

2. Bounded-real performance and induced-gain interpretation

For the discrete-time MF-SBRL, the disturbance-to-output map is defined with zero initial state: $\gamma$9 Its induced $x_k \in \mathbb{R}^n$0 gain is

$x_k \in \mathbb{R}^n$1

The bounded real property with level $x_k \in \mathbb{R}^n$2 is the strict inequality

$x_k \in \mathbb{R}^n$3

equivalently expressed, when $x_k \in \mathbb{R}^n$4, as

$x_k \in \mathbb{R}^n$5

Here the mean-field terms enter through the state trajectory $x_k \in \mathbb{R}^n$6, hence indirectly through the output $x_k \in \mathbb{R}^n$7 (Weihai et al., 2016).

The induced-gain interpretation is the stochastic mean-field analogue of the deterministic and stochastic bounded-real viewpoint: the lemma certifies a dissipative inequality with attenuation level $x_k \in \mathbb{R}^n$8. The nontriviality lies in the fact that the energy balance must be handled separately on the centered and mean subspaces.

In the same discrete-time framework, the $x_k \in \mathbb{R}^n$9 component used later for synthesis is built from the generic output

$u_k \in \mathbb{R}^q$0

with finite-horizon cost

$u_k \in \mathbb{R}^q$1

This places the MF-SBRL inside a mixed $u_k \in \mathbb{R}^q$2 architecture, where the $u_k \in \mathbb{R}^q$3 side controls the disturbance attenuation and the $u_k \in \mathbb{R}^q$4 side determines quadratic optimality (Weihai et al., 2016).

3. Discrete-time MF-SBRL and its Riccati-type recursions

The discrete-time MF-SBRL is formulated through two coupled backward difference equations for symmetric matrices $u_k \in \mathbb{R}^q$5. At each time $u_k \in \mathbb{R}^q$6, the paper defines

$u_k \in \mathbb{R}^q$7

and similarly

$u_k \in \mathbb{R}^q$8

The lemma states that if the constrained backward equations

$u_k \in \mathbb{R}^q$9

with terminal conditions

$v_k \in \mathbb{R}^\ell$0

satisfy

$v_k \in \mathbb{R}^\ell$1

and admit a unique solution $v_k \in \mathbb{R}^\ell$2, then

$v_k \in \mathbb{R}^\ell$3

(Weihai et al., 2016).

The structural novelty relative to the classical stochastic bounded real lemma is the second recursion $v_k \in \mathbb{R}^\ell$4. It tracks the expected-state channel and is coupled to $v_k \in \mathbb{R}^\ell$5 through $v_k \in \mathbb{R}^\ell$6, $v_k \in \mathbb{R}^\ell$7, and $v_k \in \mathbb{R}^\ell$8. A common misconception is that MF-SBRL is obtained by a straightforward replacement of state matrices by their mean-field aggregates. The discrete-time result shows otherwise: the mean dynamics require an additional backward equation and additional positivity constraints.

Another important point is the status of the result. In this discrete-time formulation, the MF-SBRL is a sufficient condition for $v_k \in \mathbb{R}^\ell$9, not a necessary one. The reason given is that solvability of the backward recursions depends on positivity constraints that must hold at every step, and these constraints need not be implied solely by the induced-gain inequality (Weihai et al., 2016).

4. Role in mean-field $z_k \in \mathbb{R}^m$0 synthesis

The discrete-time MF-SBRL is used as the disturbance-attenuation component of a finite-horizon mixed $z_k \in \mathbb{R}^m$1 design. On the $z_k \in \mathbb{R}^m$2 side, the mean-field stochastic linear-quadratic problem is posed for

$z_k \in \mathbb{R}^m$3

with cost

$z_k \in \mathbb{R}^m$4

A backward recursion obtained by completion of squares yields solvability conditions and the optimal control $z_k \in \mathbb{R}^m$5, together with $z_k \in \mathbb{R}^m$6 in that LQ setting (Weihai et al., 2016).

For the mixed mean-field $z_k \in \mathbb{R}^m$7 problem, the controller and the worst-case disturbance are taken in state-feedback form: $z_k \in \mathbb{R}^m$8 The existence result is expressed through four coupled matrix-valued backward equations: a controller-side Riccati pair $z_k \in \mathbb{R}^m$9 and a disturbance-side Riccati pair $k \in \{0,1,\dots,K\}$0, all with zero terminal conditions and positivity conditions on

$k \in \{0,1,\dots,K\}$1

The worst-case disturbance gains are

$k \in \{0,1,\dots,K\}$2

whereas the controller gains are

$k \in \{0,1,\dots,K\}$3

Under solvability of these equations, the closed loop satisfies the $k \in \{0,1,\dots,K\}$4 bound $k \in \{0,1,\dots,K\}$5 and the $k \in \{0,1,\dots,K\}$6 optimality requirement against the worst-case disturbance $k \in \{0,1,\dots,K\}$7 (Weihai et al., 2016).

This synthesis mechanism shows how MF-SBRL functions as more than a verification device. It is a constructive ingredient in a two-player finite-horizon design: first the disturbance channel is shaped through a bounded-real argument, and then the controller channel is optimized through an LQ argument, with both steps coupled by the mean-field terms.

5. Proof methodology and numerical realization

The proof strategy in the discrete-time setting combines mean-field decomposition, telescoping identities, and completion of squares. The first step rewrites the system in terms of the centered and mean quantities

$k \in \{0,1,\dots,K\}$8

The second step uses key telescoping identities for quadratic forms in $k \in \{0,1,\dots,K\}$9 and $$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$0, exploiting the independence and zero-mean property of $$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$1 to eliminate cross terms. The resulting summed identity expresses

$$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$2

as a quadratic form in the centered and mean variables, plus boundary terms depending on $$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$3 and $$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$4. Completion of squares then produces the Riccati-type recursions and the positivity conditions ensuring nonnegativity of the residual terms (Weihai et al., 2016).

The sign property $$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$5 in the MF-SBRL is not incidental. It reflects the fact that the expected-state channel enters the dissipation argument with its own storage contribution. This suggests that the mean dynamics are not a perturbative correction to the fluctuation dynamics, but a parallel channel requiring a separate energy accounting.

The synthesis proof repeats the same logic twice: once for the $$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$6 disturbance channel and once for the $$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$7 controller channel. That duplication is what yields four coupled backward equations rather than a single Riccati recursion.

A backward recursive algorithm is given for implementation. Starting from

$$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$8

one computes

$$\begin{aligned} x_{k+1} &= A(k)\,x_k + \bar{A}(k)\,\mathbb{E}[x_k] + B(k)\,v_k + \bar{B}(k)\,\mathbb{E}[v_k] \ &\quad + \big(C(k)\,x_k + \bar{C}(k)\,\mathbb{E}[x_k] + D(k)\,v_k + \bar{D}(k)\,\mathbb{E}[v_k]\big)\,w_k + F_1(k)\,u_k, \ z_k &= \Phi(k)\,x_k, \end{aligned}$$9

checks positive definiteness, inverts the matrices if possible, computes the terminal gains

$x_0$0

updates the Riccati variables at time $x_0$1, and repeats backward for $x_0$2. The numerical remarks recommend symmetric matrix factorizations such as Cholesky for positivity checks and inversion, exploitation of block structure, and describe the overall cost as $x_0$3 Riccati-like steps involving $x_0$4, $x_0$5, and $x_0$6 matrix operations (Weihai et al., 2016).

The reported illustrative case is a two-step example with $x_0$7 and $x_0$8, where the recursive algorithm computes $x_0$9, $w_k$0, $w_k$1, $w_k$2 and the matrices $w_k$3, $w_k$4, $w_k$5, $w_k$6, verifies the positivity conditions, and yields the mixed $w_k$7 controller. Detailed numerical values are tabulated in the source (Weihai et al., 2016).

6. Relation to classical stochastic BRL and intrinsic limitations

When the mean-field terms vanish,

$w_k$8

the discrete-time MF-SBRL reduces to the standard stochastic bounded real lemma for discrete-time systems with state- and disturbance-dependent noise. Under appropriate assumptions, the non-mean-field problem admits equivalence-type results, whereas the mean-field version in the discrete-time formulation remains sufficient only (Weihai et al., 2016).

The main structural difference is therefore not merely the presence of expectation terms in the state equation, but the appearance of the additional recursion $w_k$9 and the associated matrices $\gamma$00. These encode the coupling between the expected dynamics and the fluctuation dynamics. A plausible implication is that mean-field bounded-real analysis should be viewed as a genuinely two-layer Riccati problem, even when the original state equation appears close to a classical stochastic model.

Several limitations are explicit in the discrete-time result. The horizon is finite, the system is discrete-time, the multiplicative noise is scalar, the disturbance is exogenous and uncorrelated with the noise, and no explicit detectability or stabilizability assumptions are formulated beyond solvability and positivity of the backward recursions. Moreover, the Riccati-like equations may fail the intermediate positivity tests even when $\gamma$01 holds, which is precisely why the theorem is framed as sufficient rather than necessary (Weihai et al., 2016).

A further misconception is that open-loop and closed-loop solvability should coincide automatically in mean-field stochastic $\gamma$02 problems. The later continuous-time affine theory explicitly distinguishes these notions and states that, in that setting, open-loop solvability does not imply closed-loop solvability because mean dynamics and affine terms require additional structural equations beyond the bounded-real Riccati pair (Fang et al., 26 Jul 2025).

7. Continuous-time affine extension and later MF-SBRL formulations

A continuous-time mean-field stochastic system with affine terms is given on $\gamma$03 by

$\gamma$04

with deterministic bounded coefficients, affine terms $\gamma$05, and $\gamma$06 (Fang et al., 26 Jul 2025).

For the $\gamma$07 analysis, one sets $\gamma$08, $\gamma$09, $\gamma$10, $\gamma$11, and defines

$\gamma$12

where

$\gamma$13

The bounded-real inequality is

$\gamma$14

(Fang et al., 26 Jul 2025).

The continuous-time MF-SBRL introduces the aggregated matrices

$\gamma$15

and the functions

$\gamma$16

$\gamma$17

It then states that, for a given $\gamma$18, the following are equivalent:

1. $\gamma$19.
2. There exist continuous symmetric matrix solutions $\gamma$20 to the coupled differential Riccati equations with $\gamma$21, $\gamma$22,

$\gamma$23

This is a substantive contrast with the discrete-time 2016 formulation: the continuous-time affine result is stated as an equivalence rather than only a sufficiency condition (Fang et al., 26 Jul 2025).

Affine terms require an additional BSDE/ODE layer. If $\gamma$24 and $\gamma$25 solve the specified backward stochastic and deterministic equations, the minimizing disturbance is

$\gamma$26

with

$\gamma$27

The associated optimal cost is given explicitly in terms of $\gamma$28 (Fang et al., 26 Jul 2025).

For the full joint $\gamma$29 problem, open-loop equilibrium is characterized by an MF-FBSDE together with stationarity conditions

$\gamma$30

while closed-loop synthesis is derived from four coupled CDREs $\gamma$31, two BSDEs, and two ODEs. The resulting feedback laws are

$\gamma$32

with

$\gamma$33

$\gamma$34

Under solvability of the four CDREs, the BSDEs, and the ODEs, the closed loop achieves the bounded-real property and $\gamma$35 optimality (Fang et al., 26 Jul 2025).

Taken together, the discrete-time and continuous-time formulations show a consistent pattern: mean-field coupling forces a separation between centered and mean dynamics; bounded-real certification is expressed through coupled Riccati objects rather than a single equation; and once affine or mixed $\gamma$36 features are introduced, additional backward equations become necessary. The precise status of the lemma, however, depends on the model class: the discrete-time finite-horizon version is sufficient, whereas the continuous-time affine version is formulated as necessary and sufficient under its stated conditions.