Robust Quadratic Equivalent in Distributed Control

Updated 5 July 2026

Robust Quadratic Equivalent is a structural property ensuring that nonconvex robust output-feedback problems can be exactly reformulated into convex disturbance-feedback programs.
It leverages quadratic invariance of the sparsity subspace with respect to the stacked plant map to certify convexity via a finite family of binary inequalities.
The approach eliminates implicit signaling and accommodates various time-varying and intermittent information structures, making distributed controller synthesis tractable.

Robust Quadratic Equivalent, in the sense developed for finite-horizon distributed control, denotes the circumstance under which a nonconvex robust distributed output-feedback problem admits an equivalent convex formulation over disturbance-feedback policies. In "Unified Approach to Convex Robust Distributed Control given Arbitrary Information Structures" (Furieri et al., 2017), this equivalence is governed by quadratic invariance (QI) of a sparsity subspace with respect to the stacked plant map $\mathbf{CB}$ , and QI is shown to be equivalent to a finite family of inequalities over binary matrices. The term therefore refers simultaneously to a structural property of the information pattern, a convexity certificate, and an exact correspondence between the original output-feedback synthesis problem and its robust convex disturbance-feedback reformulation (Furieri et al., 2017).

1. Finite-horizon robust distributed output-feedback setting

The underlying plant is a discrete-time LTI system on a finite horizon $k=0,\dots,N$ ,

$\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$

with state $x_k$ , input $u_k$ , measured output $y_k$ , unknown disturbance $w_k \in \mathcal W$ , and known initial condition $x_0$ . After stacking over the horizon, the dynamics take the form

$\mathbf{x} = \mathbf{A}x_0 + \mathbf{B}\mathbf{u} + \mathbf{E}_D\mathbf{w}, \qquad \mathbf{y} = \mathbf{C}\mathbf{x} + \mathbf{H}\mathbf{w},$

with suitable block-lower-triangular matrices (Furieri et al., 2017).

Robustness is imposed through polytopic safety sets. For all time steps and all disturbances,

$\begin{bmatrix}x_k \ u_k\end{bmatrix} \in \Gamma = \{(x,u): Ux + Vu \le b\}, \qquad x_N \in \mathcal X_f = \{x: Rx \le z\}.$

In stacked form these become

$k=0,\dots,N$ 0

for matrices $k=0,\dots,N$ 1 determined by the plant and constraint data. The key structural point is that this constraint is convex in $k=0,\dots,N$2 because it is a max of affine functions (Furieri et al., 2017).

The distributed controller is an affine output-feedback law,

$k=0,\dots,N$ 3

subject to an information structure encoded by binary matrices $k=0,\dots,N$ 4. The constraint $k=0,\dots,N$ 5 specifies which components of $k=0,\dots,N$ 6 may depend on which components of $k=0,\dots,N$ 7. Because the matrices $k=0,\dots,N$ 8 are arbitrary, the framework covers time-invariant sensing topologies, fixed communication networks with delays, time-varying communication delays, time-varying sensing or communication schedules, intermittent observations, and forgetting mechanisms (Furieri et al., 2017).

2. Disturbance-feedback parametrization and exact convex reformulation

The original search over $k=0,\dots,N$ 9 is nonconvex because output-feedback enters the closed loop nonlinearly. The paper therefore uses a disturbance-feedback parametrization with

$\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 0

where $\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 1 is causal and $\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 2 is a deterministic feedforward term (Furieri et al., 2017).

A bijection links admissible output-feedback and disturbance-feedback policies: $\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 3 with the closed-loop map

$\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 4

Under this change of variables, the cost $\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 5 is convex in $\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 6 and does not depend directly on $\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 7, while the robust polytopic constraints remain convex in $\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 8. The only source of potential nonconvexity becomes the structural condition

$\begin{aligned} x_{k+1} &= A x_k + B u_k + D w_k, \ y_k &= C x_k + H w_k, \end{aligned}$ 9

where $x_k$ 0 is the sparsity subspace induced by the information structure (Furieri et al., 2017).

This is the precise locus of the robust quadratic equivalent. If the structural image $x_k$ 1 coincides with a linear subspace, then the original robust distributed output-feedback problem and the disturbance-feedback problem are equivalent in the strong sense used in the paper: feasible policies correspond bijectively, optimal values coincide, and optimal controllers translate through the $x_k$ 2 mapping (Furieri et al., 2017).

3. Quadratic invariance as the convexity criterion

Let $x_k$ 3. Quadratic invariance with respect to $x_k$ 4 is defined by

$x_k$ 5

The interpretation given in the paper is structural closure under one quadratic interaction through the plant map: feedback interconnections allowed by the information pattern must not generate couplings outside that pattern (Furieri et al., 2017).

Two finite-horizon lemmas establish the decisive equivalence. First,

$x_k$ 6

Second,

$x_k$ 7

Together they yield the proposition that the disturbance-feedback formulation is a convex program equivalent to the original output-feedback problem if and only if $x_k$ 8 is QI with respect to $x_k$ 9 (Furieri et al., 2017).

A common misunderstanding is to view QI as a performance condition. In this formulation it is instead a structural convexity condition: it does not by itself optimize the controller, but it determines whether optimal synthesis under the prescribed information structure is tractable. Another common misunderstanding is that adding hard robust constraints should destroy convexity. The paper shows the opposite for polytopic state and input constraints: once QI holds, those constraints enter naturally and do not affect convexity (Furieri et al., 2017).

4. Finite binary test and the elimination of implicit signaling

The paper’s main technical contribution is a finite combinatorial test for QI. For any real matrix $u_k$ 0, $u_k$ 1 records its zero-nonzero pattern. Define

$u_k$ 2

and use structural binary matrix multiplication and entrywise order. Then the finite-horizon Rotkowitz-Lall characterization becomes

$u_k$ 3

The paper decomposes this into local block inequalities: $u_k$ 4 for all

$u_k$ 5

Each inequality is entrywise between $u_k$ 6 binary matrices (Furieri et al., 2017).

Its interpretation is a signaling-chain test. The product

$u_k$ 7

checks whether information available through $u_k$ 8 can propagate through a controller at time $u_k$ 9, then through the plant dynamics $y_k$ 0, then through another controller using $y_k$ 1 at time $y_k$ 2. If such implicit signaling exists, convexity requires that the corresponding information about $y_k$ 3 already be explicitly available at time $y_k$ 4, namely through $y_k$ 5. The inequalities therefore enforce a “signaling-free” information structure in the QI sense (Furieri et al., 2017).

In this form, the robust quadratic equivalent condition becomes purely combinatorial: the robust distributed output-feedback problem is convexly solvable through disturbance-feedback if and only if this finite family of binary inequalities holds. The nonconvexity question is reduced from controller synthesis to verification of a finite test over binary matrices (Furieri et al., 2017).

5. Special cases, arbitrary information structures, and computational scope

The binary criterion unifies several earlier QI characterizations. If the sensing topology is time-invariant, $y_k$ 6 for all $y_k$ 7, the test reduces to

$y_k$ 8

Using Cayley-Hamilton, it suffices to check $y_k$ 9. For fixed sensing plus communication topology, with direct measurement matrix $w_k \in \mathcal W$ 0 and communication matrix $w_k \in \mathcal W$ 1, the general condition reduces to

$w_k \in \mathcal W$ 2

for the range of $w_k \in \mathcal W$ 3 and $w_k \in \mathcal W$ 4 stated in the paper. This makes transparent how communication can offset plant-induced coupling and recover convexity (Furieri et al., 2017).

The same theorem also covers cases not handled by monotone information-growth assumptions. Because the matrices $w_k \in \mathcal W$ 5 are arbitrary, time-varying communication networks, intermittent observations, time-varying delays, and forgetting mechanisms are all admissible. The paper gives an example in which controllers both receive time-varying measurements and forget previously known outputs, yet the binary QI test still holds (Furieri et al., 2017).

Computationally, the binary test scales polynomially with the horizon $w_k \in \mathcal W$ 6, and each inequality is a structural check that can be performed with bitwise operations. In structured cases, the number of inequalities collapses substantially. By contrast, the original output-feedback problem in $w_k \in \mathcal W$ 7 is nonconvex and generally intractable, and testing convexity directly at that level is essentially impossible. Once QI holds, the synthesis problem reduces to a convex disturbance-feedback program with linear sparsity constraints and robust polytopic constraints that can, for polytopic $w_k \in \mathcal W$ 8, be dualized row-wise into linear or SOCP constraints (Furieri et al., 2017).

Several other papers use closely related formulations in which a robust problem with quadratic structure is replaced by an exact conic, matrix-inequality, or deterministic counterpart.

Setting	Equivalent object	Status
Two-stage ARO with ellipsoidal uncertainty and quadratic decision rules	SDP, or SOCP for separable QDRs	exact conic reformulation (Woolnough et al., 2020)
Distributionally robust finite-horizon LQG	standard LQG for a least-favorable Gaussian model	exact in value and optimal strategies (Taşkesen et al., 2023)
Robust adaptive beamforming with general-rank signal model	quadratic matrix inequality problem, then LMI relaxation	exact QMI; relaxed SDP (Huang et al., 2019)
Robust SOS-convex polynomial programs	SDP relaxation; SOCP in restricted quadratic cases	exact under stated uncertainty sets (Jeyakumar et al., 2013)
Matrix-valued uncertain quadratic and conic-quadratic constraints	support-function-based SDP/SOCP reformulations	exact for concave-in-parameter cases; inner/outer for hard quadratics (Marandi et al., 2019)
Guaranteed-cost robust MPC with quadratically bounded uncertainty	single QCQP	deterministic robust counterpart (Massera et al., 2016)

Taken together, these usages suggest a broader pattern rather than a single universal definition. In distributed control, the phrase denotes exact convex equivalence controlled by QI and certified by a finite binary test (Furieri et al., 2017). In adjustable robust optimization, it denotes exact SDP or SOCP reformulations under quadratic decision rules (Woolnough et al., 2020). In distributionally robust LQG, it denotes equivalence to a nominal LQG problem for a least-favorable Gaussian noise model (Taşkesen et al., 2023). In robust beamforming, it denotes an exact QMI representation of the original worst-case problem before LMI relaxation (Huang et al., 2019). Across these settings, the common feature is the conversion of a robust problem with quadratic structure into an equivalent or tightly controlled tractable formulation.

In that sense, the control-theoretic result of (Furieri et al., 2017) is a particularly sharp instance of the idea: the robust quadratic equivalent is not merely a relaxation or approximation, but an exact convex reformulation of a constrained, distributed, robust output-feedback synthesis problem, and the existence of that reformulation is characterized completely by quadratic invariance and its finite binary-matrix test.