---
title: All-Stabilizing Youla-Kučera Architecture
url: https://www.emergentmind.com/topics/all-stabilizing-youla-kucera-architecture
type: topic
---

# All-Stabilizing Youla-Kučera Architecture

to=arxiv_search  菲律宾申博json
{"query":"All-Stabilizing Youla-Kucera architecture Youla-Kucera parameterization realization-stability lemma", "max_results": 10}
The all-stabilizing Youla–Kučera architecture is the classical controller parameterization in which a plant admitting a coprime or doubly-coprime factorization is paired with a stable free parameter \(Q\), and every internally stabilizing controller is obtained from that parameter. In its standard form, the architecture separates stabilization from performance shaping: internal stability is enforced by the requirement \(Q\in RH_\infty\), while closed-loop objectives are encoded through the choice of \(Q\). Recent work places this architecture inside a broader family of realization-theoretic, kernel, system-level, input-output, operator-theoretic, data-driven, and neural constructions, but the core claim remains the same: the parameterization covers all stabilizing controllers for the specified plant class [2009.14328].

## 1. Classical transfer-matrix form

For a right coprime factorization
\[
P(s)=N(s)M(s)^{-1},
\]
with \(N,M\in RH_\infty\) proper, with no common unstable zeros, and with Bézout factors \(X(s),Y(s)\in RH_\infty\) satisfying
\[
X(s)N(s)+Y(s)M(s)=I,
\]
the set of all internally stabilizing controllers can be written uniquely as
\[
K(s)=\bigl(Y(s)-M(s)Q(s)\bigr)\bigl(X(s)+N(s)Q(s)\bigr)^{-1},
\]
where \(Q(s)\in RH_\infty\) is an arbitrary stable proper transfer matrix [2009.14328]. The “all-stabilizing” designation refers precisely to this one-to-one correspondence: every such \(Q\) yields internal stability, and every internally stabilizing controller arises in this way.

Equivalent formulas appear under different coprime-factor conventions. With a doubly-coprime factorization
\[
P=N_rM_r^{-1}=M_l^{-1}N_l,\qquad
\begin{bmatrix}U_l&-V_l\\-N_l&M_l\end{bmatrix}
\begin{bmatrix}M_r&V_r\\N_r&U_r\end{bmatrix}=I,
\]
one obtains
\[
K(s)=\bigl(V_r(s)-M_r(s)Q(s)\bigr)\bigl(U_r(s)-N_r(s)Q(s)\bigr)^{-1},
\qquad Q\in \mathcal{RH}_\infty,
\]
and conversely every \(Q\in\mathcal{RH}_\infty\) so plugged in yields a stabilizing \(K\) [1909.12346]. The sign pattern and factor naming vary across sources, but the underlying content is the same: a stable free parameter spans the full internally stabilizing set.

The architecture is also a shaping mechanism. In the precision-motion formulation, with a nominal stabilizing controller \(K_0(s)\), loop gain \(L_0(s)=P(s)K_0(s)\), and sensitivity \(S_0(s)=1/(1+L_0(s))\), introducing \(Q\) modifies the sensitivity as
\[
S(s)=S_0(s)\,[\,Y(s)-Q(s)D(s)\,],
\]
so notches or peaks in \(S(s)\) can be placed by shaping \(Y-QD\) [2508.14309]. This makes explicit that the free parameter is not merely a proof device; it is a direct design coordinate for closed-loop frequency response.

## 2. Realization–stability viewpoint and equivalence with other parameterizations

A unifying interpretation is provided by the realization–stability lemma. If all internal signals are collected into a vector \(\mathbf z\), a realization matrix \(\mathbf R\) and an internal-stability matrix \(\mathbf S\) satisfy
\[
(I-\mathbf R)\mathbf S=I,\qquad \mathbf S(I-\mathbf R)=I.
\]
Internal stability is then enforced by requiring every transfer from the external signal to each internal signal to lie in \(RH_\infty\), together with causality of the off-diagonal entries of \(\mathbf R\) [2009.14328]. In the Youla setting, substituting \(P=N M^{-1}\) reduces the corresponding conditions to a doubly-coprime Bézout identity, from which the classical parameterization follows directly [2112.02005].

This viewpoint matters because the Youla parameterization is not isolated. The realization–stability lemma shows that existing controller synthesis methods and realization proposals are all special cases of a simple lemma, and it enables easier equivalence proofs among existing methods [2009.14328]. In particular, explicit affine mappings link the Youla parameter \(Q\), the System-Level Parameterization (SLP), and the Input-Output Parameterization (IOP), so any convex problem in one coordinate system can be reformulated equivalently in the others [1907.06256]. A common misconception is that these are competing notions of stabilizing design; the cited work instead presents them as affine reparameterizations of the same stabilizing controller set.

The non-convexity of the controller set itself is preserved throughout. What becomes convex is the description in terms of closed-loop maps. One line of work identifies four groups of stable closed-loop transfer matrices equivalent to internal stability: one used in SLP, one used in IOP, and two new mixed forms, leading to four convex parameterizations of \(\mathcal C_{\text{stab}}\) [1909.12346]. Under a fixed FIR horizon \(T\), these parameterizations yield different inner approximations, and the IOP has the best ability of approximating \(\mathcal C_{\text{stab}}\) given FIR constraints [1909.12346]. This suggests that the practical distinction among parameterizations is often computational and structural rather than fundamental.

## 3. Kernel and convex reformulations

A major modern development is the kernel version of the Youla parameterization. Focusing on left factors \(\mathbf G=\mathbf M_\ell^{-1}\mathbf N_\ell\), one obtains
\[
\mathcal C_{\rm stab}
=
\{\,K=Y\,X^{-1}\mid X,Y\in RH_\infty,\; M_\ell X-N_\ell Y=I\,\}.
\]
Thus, instead of one free parameter \(Q\), all stabilizing controllers are described by two stable transfer matrices \(X,Y\) constrained by a single affine equation [2203.17145]. This is still an all-stabilizing architecture, but it relocates the freedom from a fractional formula into an affine kernel relation.

The same paper shows that allowing the residual
\[
\Delta=M_\ell X-N_\ell Y-I
\]
and imposing
\[
\|\Delta\|_\infty<1
\]
is equivalent to internal stability, by the small-gain theorem [2203.17145]. In that form, stabilizing-controller synthesis becomes a right \(\mathcal H_\infty\)-filtering problem,
\[
\|P_1F-P_2\|_\infty<1,\qquad
P_1=[\,M_\ell\;\;-\!N_\ell\,],\quad
P_2=I,\quad
F=[\,X\;\;Y\,].
\]
This reinterpretation is technically significant because it imports finite-dimensional \(\mathcal H_\infty\)-LMI machinery into an otherwise infinite-dimensional parameterization.

Using a minimal state-space realization of \([P_1\;\;P_2]\) and a standard finite-dimensional \(\mathcal H_\infty\)-LMI lemma from Geromel–Bernussou or the extended version by de Scherer, the paper derives a single LMI of size \((2n+n_f)\times(2n+n_f)\), with \(n_f=n\), that is necessary and sufficient for \(\|M_\ell X-N_\ell Y-I\|_\infty<1\) [2203.17145]. Once \(F(z)=\begin{bmatrix}X(z)&Y(z)\end{bmatrix}\) is obtained in state-space form with \(\hat D_X\) invertible, the controller \(K=Y X^{-1}\) is reconstructed explicitly and has internal order \(\dim(\hat A)=n\), matching the plant order [2203.17145]. The paper states that this yields the first efficient Linear Matrix Inequality implicit parametrization of stabilizing controllers [2203.17145].

A different convex state-space variant appears in robust \(\mathcal H_\infty\) guaranteed-cost control under parametric uncertainty. There, a variant of the Youla-Kučera parameterization is expressed by a symmetric matrix variable \(W\) and scalar \(\mu=1/\gamma^2\), with the convex set
\[
\mathscr C=\{(W,\mu):W\succeq 0,\;\Theta_1(W,\mu)\preceq 0,\;\mu>0\},
\]
and each feasible pair generates the stabilizing gain
\[
K=W_2^T W_1^{-1}
\]
such that the closed loop is stable and \(\|H(s)\|_\infty\le \gamma=1/\sqrt{\mu}\) [2001.00708]. In the uncertain case, the construction extends to all extreme vertices simultaneously, preserving convexity.

## 4. Numerical computation and robustness

The computational appeal of all-stabilizing parameterizations lies in the possibility of finite-dimensional convex synthesis, but numerical behavior differs sharply across formulations. For the kernel-LMI approach, the finite-dimensional problem is a single semidefinite program of size \(O(n)\) returning an \(n\)th-order controller, whereas FIR or Ritz approximations of Youla/SLP/IOP lead to SDPs of size \(O(nT)\) and controllers of order \(nT\) [2203.17145]. In chain-graph examples up to 14 subsystems, the LMI approach was an order of magnitude faster, in seconds rather than minutes, and produced 2nd-order local controllers, whereas a length-20 FIR SLP produced 1092nd-order controllers [2203.17145].

Numerical robustness is not uniform across parameterizations. For IOP, small mismatches in the affine constraints do not compromise closed-loop stability when the plant is open-loop stable; in that sense the IOP is numerically robust for open-loop stable plants [1909.12346]. The same work proves that a direct IOP implementation will fail to stabilize open-loop unstable systems in practice [1909.12346]. For SLP, numerical robustness of the two-block state-feedback controller is established, but numerical robustness of the four-block SLP controller requires case-by-case analysis even when the plant is open-loop stable [1909.12346]. A direct implication is that exact affine equivalence at the symbolic level does not imply equal floating-point behavior.

For large-scale robust programs, the state-space convex variant is paired with a symmetric Gauss–Seidel ADMM. The reformulated problem introduces consensus variables over the cone
\[
\mathcal K=S_+^p\times (S_+^r)^N\times \mathbb R_+,
\]
uses closed-form projection steps for the \(Y\)-update, a backward–forward symmetric Gauss–Seidel sweep for the \((W,\mu)\)-update, and a dual update with \(0<\tau<(1+\sqrt 5)/2\), with \(\tau=1.618\) used in the paper [2001.00708]. The paper states that direct multi-block ADMM may fail to converge, whereas the symmetric Gauss–Seidel sweep recovers a two-block-like structure and yields linear convergence under the linear-quadratic non-smooth setting [2001.00708].

## 5. Data-driven and learning-based architectures

The all-stabilizing idea has also been recast in purely data-driven form. Using one long input-output trajectory \(\{u_t,y_t\}_{t=0}^{N-1}\) with \(u\) persistently exciting of order \(L+n\), one builds block-Hankel matrices \(H_L(u)\) and \(H_L(y)\). By Willems’ Fundamental Lemma, any length-\(L\) trajectory of the unknown LTI plant arises if and only if there exists \(\alpha\) solving
\[
\begin{bmatrix}H_L(u)\\H_L(y)\end{bmatrix}\alpha
=
\begin{bmatrix}\hat u\\\hat y\end{bmatrix},
\]
and, in the strictly proper case, the next output is predicted by \(H'_L(y)\alpha\) [2304.03422]. This yields a data-driven internal model that can be inserted directly into a Youla realization: at each time step, the controller forms the internal Youla input \(\hat r_t=e_t+\bar y_L\), feeds it into a stable operator \(Q\), and reproduces exactly the classical Youla–Kučera controller \(u_t=Q(e+P_0u)\) for the unknown plant [2304.03422].

For linear \(Q\), one paper gives an explicit unconstrained matrix factorization of all stable state matrices
\[
A_q=L^{-1}\bigl(U\tanh(D)V^{\mathsf T}\bigr)L,
\]
with \(L>0\) lower-triangular and \(U,V\) orthogonal, so that \(\rho(A_q)<1\) is certified by a quadratic Lyapunov function [2310.14098]. For nonlinear \(Q\), a neural Lyapunov correction is used:
\[
z_{t+1}=\gamma(z_t)\,\hat f_\theta(z_t),
\]
with \(\gamma\) chosen in closed form to ensure \(V_\theta(z_{t+1})\le \beta V_\theta(z_t)\) for \(0<\beta<1\), guaranteeing global exponential stability of the nonlinear dynamics [2310.14098]. The same paper shows that if the measurement data are corrupted by i.i.d. Gaussian noise, then the induced operator \(H^+H'\) has spectral radius less than \(1+\varepsilon_N\) with high probability, with \(\varepsilon_N\to 0\) as \(N\to\infty\), so the open-loop data-driven predictor remains BIBO-stable despite noise [2310.14098].

These constructions have been integrated with reinforcement learning. In the two-tank study, the stable \(Q\)-dynamics were trained with TD3, using a two-layer MLP for \(\hat f_\theta\), a two-layer input-convex network for \(V_\theta\), and a linear readout \(h_\theta\); 20 independent runs of 100 episodes each, at 0.5 s sampling and Gaussian measurement noise variance \(0.015\), showed median cumulative reward converging after approximately 40 episodes, with no training runs exhibiting divergence [2304.03422]. A related line based on recurrent equilibrium networks (RENs) proves that RENs are universal approximators of contracting and Lipschitz operators, and therefore that the Youla–REN architecture is a universal approximator of stabilizing nonlinear controllers for a partially-observed linear system [2112.04219]. The central design principle is unchanged: optimization proceeds over a parameterization of stable \(Q\), so stability is guaranteed during learning transients [2112.04219].

## 6. Nonlinear, distributed, quantum, and application-specific extensions

In the operator-theoretic nonlinear setting, for a strictly-causal plant operator \(G\in \mathcal L_p\), the all-stabilizing architecture takes the form
\[
K=(I+GQ)^{-1}Q,
\qquad Q\in \mathcal L_p,
\]
with closed-loop maps
\[
\Phi_y=I+GQ,\qquad \Phi_u=Q.
\]
Under these conditions, every \(\ell_p\)-stabilizing controller arises in this form for a unique \(Q\in\mathcal L_p\), and small-gain robustness to model mismatch is expressed by \(\gamma_\Delta\|Q\|_p<1\) [2412.19280]. This is an exact nonlinear analogue of the stable-parameter principle, though it assumes the nominal plant itself is \(\ell_p\)-stable.

For partially-observed nonlinear systems, the Youla architecture is built around a base stabilizing controller, a contracting and Lipschitz observer, the innovation \(\tilde y_t=y_t-\hat y_t\), and a free contracting and Lipschitz parameter \(\mathcal Q\), often implemented as a REN [2506.01226]. The resulting augmented controller is
\[
u_t=k(s_t,\eta_t,y_t)+h_q\bigl(q_t,\eta_t,y_t-h(\hat x_t)\bigr).
\]
With any two among nonlinear dynamics, partial observation, and incremental closed-loop stability requirements, a contracting and Lipschitz Youla parameter leads to contracting and Lipschitz closed loops; if all three hold and exogenous disturbances are present, full incremental stability can be lost, and the preserved property is d-tube contraction and Lipschitzness [2506.01226]. The same work proves a disturbance-free converse and a partial converse under certainty-equivalence structure, which qualifies the meaning of “all-stabilizing” in the nonlinear disturbed case [2506.01226]. A local continuous-time state-space counterpart shows that every dynamic state-feedback controller that locally exponentially stabilizes a nonlinear input-affine system can be rewritten as a linear baseline \(u=Kx\) plus the output of internal locally exponentially stable controller dynamics, yielding a local nonlinear Youla-type parametrization specialized to local exponential stability [2601.02244].

Distributed and domain-specific versions further broaden the architecture. The Youla Operator State-Space framework characterizes stably realizable structured controllers over a network through stable operators \(Q\) and \(Z\), with
\[
K=Z(I+Q)^{-1},
\]
and under subspace-like assumptions the corresponding performance problem becomes a convex model-matching problem [1910.01045]. A later distributed RL formulation embeds Graph Neural Networks into a Youla-like magnitude-direction parameterization, where a stable disturbance-feedback operator supplies the magnitude and a GNN on local observations supplies a unit-ball direction, guaranteeing network-level closed-loop stability by design [2512.18540]. In sparse add-on design, all stabilizing system-level add-on controllers around a fixed decentralized baseline \(K_0\) are parameterized by a static hollow Youla matrix \(X\), giving
\[
\bar K(X)=K_0-(S\hat G)^{-1}ZX(I+ZX)^{-1},
\]
and an affine closed-loop map \(T_{zw}(X)=T_{zw}^0+T_{12}XT_{21}\) that yields a convex \(\mathcal H_2+\ell_1\) synthesis problem [2606.01226].

Application-specific loop shaping uses the same architecture for multi-band disturbance rejection. Starting from a baseline loop \(L(z)=P(z)C(z)\) and a stable approximate inverse \(\hat L^{-1}(z)\), the augmenting controller becomes
\[
C_Q(z)=\frac{1+z^{-m}\hat L^{-1}(z)Q(z)}{1-z^{-m}Q(z)},
\]
and the sensitivity is approximately \(\tilde S(z)\approx S_0(z)[1-z^{-m}Q(z)]\) [2508.14309]. Direct single-shot design becomes severely ill-conditioned beyond 4–5 notches, so the paper proposes an iterative multi-stage algorithm adding 1–2 notches at a time; in the hard-disk-drive case study, grouping frequencies into 6 groups of 2 each, with bandwidth around 30 Hz and stagewise order reduction to 2nd or 4th order, yields final 24th- or 48th-order controllers rather than a direct 240th-order design, with mean notch depth around 50 dB, peak high-frequency amplification of +6 dB max, and approximately 45% position-error-signal RMS reduction over baseline [2508.14309].

The architecture also extends to coherent quantum control. For a modified plant \(\mathbf P(s)\) with loop block \(P_{22}(s)\), every internally stabilizing controller has the form
\[
K(s)=(U+MQ)(V+NQ)^{-1},\qquad Q\in RH_\infty,
\]
but physical realizability requires the controller to satisfy \(J\)-unitarity, which translates into the quadratic constraint
\[
\Phi+Q^\sim\Lambda+\Lambda^\sim Q+Q^\sim \Pi Q=0
\]
together with \(\det(V+NQ)(\infty)\neq 0\) [1503.02118]. In that setting, the all-stabilizing parameterization remains intact, but admissible \(Q\) is restricted by quantum physical realizability.

Taken across these variants, the all-stabilizing Youla–Kučera architecture is best understood as a structural principle rather than a single formula. In classical LTI synthesis it is a stable transfer-matrix parameterization of all internally stabilizing controllers; in convex and kernel formulations it becomes an affine feasibility problem; in operator and neural settings it becomes a stable-operator search space; and in specialized domains it is adapted by additional constraints such as sparsity, network realizability, disturbance-feedback structure, or physical realizability. The common thread is the same one established by the classical theory: stabilization is encoded in the architecture, and performance optimization is carried out entirely within the stabilizing set.

Source: https://www.emergentmind.com/topics/all-stabilizing-youla-kucera-architecture