Lur'e Systems with Feedthrough Analysis

• Lur'e systems with feedthrough are feedback interconnections defined by an implicit output equation involving a direct nonlinear term, complicating standard analysis.
• They require rigorous conditions like global invertibility and monotonicity to ensure existence, uniqueness, and forward completeness of solutions.
• These systems are key in applications such as power systems, neural networks, and hybrid controls where instantaneous coupling and algebraic feedback are prevalent.

A Lur'e system with feedthrough is a feedback interconnection in which a finite-dimensional linear time-invariant (LTI) system is coupled in feedback with a static nonlinearity, and crucially, the nonlinearity may act on the output via an algebraic relation that includes a direct (feedthrough) term. Unlike classical Lur'e systems—where the output equation is explicit (feedthrough-free)—the presence of a direct nonlinear feedthrough creates an implicit algebraic coupling between the system's output, state, and input, significantly complicating well-posedness, analysis, and control synthesis. Such systems frequently arise in control applications where instantaneous coupling, measurement loops, or strict causality constraints are not assumed, and are also encountered in modern control of physical networks, neural, biological, and quantum systems. Recent advances have focused on establishing existence, uniqueness, continuation, and forward completeness of solutions to Lur'e systems with feedthrough under minimal and verifiable analytical conditions (Guiver et al., 22 Aug 2025).

1. Mathematical Structure and Characterization

A general Lur'e system with feedthrough is described by the coupled equations: $$\begin{aligned} \dot{x}(t) &= A x(t) + B f(t, y(t)) + B_e v(t), \ y(t) &= C x(t) + D f(t, y(t)) + D_e v(t), \end{aligned}$$ where $x(t) \in \mathbb{R}^n$ is the state, $y(t) \in \mathbb{R}^p$ is the output, $v(t)$ is an exogenous input, and $f$ is a (possibly time-varying) nonlinearity. The presence of $D \neq 0$ (the feedthrough term) results in output equations which are implicit in $y(t)$: $y(t) = C x(t) + D f(t, y(t)) + D_e v(t).$ For a fixed state and input, $y$ solves $F_t(y) = Cx + D_e v$ where $x(t) \in \mathbb{R}^n$0. If $x(t) \in \mathbb{R}^n$1, this reduces to the classic feedthrough-free case $x(t) \in \mathbb{R}^n$2.

Feedthrough arises naturally in models where measurement, interconnection, or energy flows are not strictly delayed, as in power systems, biomedical circuits, analog and mixed-signal controllers, or networks with high-frequency or algebraic coupling.

2. Well-posedness: Existence, Uniqueness, and Continuation

Well-posedness of Lur'e systems with feedthrough is nontrivial due to the implicit definition of $x(t) \in \mathbb{R}^n$3. The principal issues are:

• Existence: For initial condition $x(t) \in \mathbb{R}^n$4 and input $x(t) \in \mathbb{R}^n$5, does a solution pair $x(t) \in \mathbb{R}^n$6 exist on an interval?
• Uniqueness: Is the solution unique, or can multiple solutions arise from the same initial state and input?
• Continuation / Forward Completeness: Can maximal solutions be continued for all future times (i.e., global existence), or do solutions blow up in finite time?

While, for feedthrough-free systems, global existence and uniqueness can often be guaranteed by standard Lipschitz conditions on $x(t) \in \mathbb{R}^n$7, in the presence of feedthrough this is generally insufficient (Guiver et al., 22 Aug 2025). Simple examples demonstrate absence of solutions (non-surjectivity of $x(t) \in \mathbb{R}^n$8), non-uniqueness (multiple preimages under $x(t) \in \mathbb{R}^n$9), or finite-time blowup even when $y(t) \in \mathbb{R}^p$0 is globally Lipschitz.

Sufficient conditions ensuring well-posedness include:

• Global Invertibility: For each $y(t) \in \mathbb{R}^p$1, the map $y(t) \in \mathbb{R}^p$2 is a homeomorphism (e.g., via uniform monotonicity or lower Lipschitz conditions on $y(t) \in \mathbb{R}^p$3), together with local Lipschitz and boundedness.
• Radial Unboundedness: $y(t) \in \mathbb{R}^p$4 as $y(t) \in \mathbb{R}^p$5 (uniformly on compacts of $y(t) \in \mathbb{R}^p$6), preventing solution escape in finite time.
• Set-valued Formulations: If $y(t) \in \mathbb{R}^p$7 is not invertible, the system can be recast as a differential inclusion using measurable selections and upper semicontinuity techniques from nonsmooth analysis. This ensures existence, and—under further monotonicity of the set-valued feedback—uniqueness.

Table 1 summarizes the key scenarios:

Structural Property of $y(t) \in \mathbb{R}^p$8	Well-posedness Result
Globally injective, radially unbounded, Carathéodory $y(t) \in \mathbb{R}^p$9	Existence and uniqueness. Blow-up iff state or $v(t)$0 diverges.
Non-invertible but set-valued, Carathéodory $v(t)$1, convex $v(t)$2 fibers	Existence of maximal solution. Uniqueness under monotonicity.

3. Mathematical and Analytical Framework

The analysis employs a blend of classical ODE theory, global inversion theorems, and tools from differential inclusions.

• When $v(t)$3 is invertible, one may define $v(t)$4 and rewrite the system as a standard ODE in $v(t)$5, replacing $v(t)$6 by $v(t)$7, and defining an effective nonlinearity $v(t)$8.
• For set-valued inverses $v(t)$9, the system is formulated as a differential inclusion

$f$0

with $f$1, and measurable selection theorems (e.g., Filippov's theorem) are invoked to establish solution properties.

• Lower Lipschitz constant conditions:

$f$2

provide uniform invertibility of $f$3 via Clarke's generalized derivative (for nonsmooth $f$4), a critical condition for well-posedness.

4. Illustrative Counterexamples and Structural Pitfalls

The pathological behaviors are illustrated by explicit examples (Guiver et al., 22 Aug 2025):

• Non-existence: There are time-varying, piecewise-affine nonlinearities with feedthrough such that, for some initial states, the output equation $f$5 has no solution (the range of $f$6 does not cover $f$7).
• Non-uniqueness: Piecewise-linear $f$8 with feedthrough may lead to multiple values of $f$9 solving the same output equation for fixed $D \neq 0$0 and $D \neq 0$1, resulting in non-unique system trajectories.
• Finite-time Blow-up: Even with globally Lipschitz $D \neq 0$2, the algebraic part can experience finite-time singularity, where solution norms diverge as the trajectory approaches a non-invertible point of $D \neq 0$3.
• Feedthrough-free contrast: For comparison, in the absence of feedthrough, classical ODE theory applies (uniqueness, continuation by Grönwall or Cauchy–Lipschitz) under weaker regularity conditions.

These examples underscore that standard "feedforward" well-posedness and robustness guarantees do not extend automatically to the feedthrough case.

5. Sufficient Conditions via Invertibility and Monotonicity

The main sufficient conditions for well-posedness are formulated in terms of properties of the mapping $D \neq 0$4:

• Uniform local injectivity of $D \neq 0$5 (for all $D \neq 0$6 and $D \neq 0$7 in compact intervals), or more generally, that the lower Clarke derivative (or the minimal singular value of $D \neq 0$8) remains bounded away from zero.
• Radial unboundedness (the norm of $D \neq 0$9 grows at infinity), to prevent finite escape.
• Convexity of fibers of $y(t)$0, facilitating selection of measurable branches for existence.
• Monotonicity of $y(t)$1 in $y(t)$2 (or sector boundedness), i.e., inner product $y(t)$3, ensures uniqueness in the set-valued context.

Notably, invertibility of $y(t)$4 for all subgradients $y(t)$5 is operationally analogous to nonsingularity of the linear algebraic constraint in the classic case, and arises in Lyapunov and LMI-based analysis.

6. Implications for Analysis, Controller Synthesis, and Applications

The rigorous theory for Lur’e systems with feedthrough illuminates several aspects:

• Extension of Absolute Stability Criteria: Classical circle and Popov criteria, which depend critically on standard output equations, cannot be directly transferred; passivity-type conditions and Lyapunov-based arguments require adaptation to implicit output mappings and algebraic constraints.
• Robustness and Design: Controller and filter design procedures for nonlinear circuits, biological systems, or networks with algebraic feedback must ensure that their feedback interconnections satisfy invertibility or monotonicity conditions for the resulting coupled algebraic equations.
• Nonlinear Networks and Neural Models: Non-unique, non-existent, or explosive behaviors in neural architectures or hybrid physical systems often trace back to failures of global invertibility or monotonicity in the feedthrough mapping.

The analytical framework is crucial for constraining model classes in robust design and networked control applications, ensuring that adverse behaviors associated with feedthrough are avoided, and providing verifiable criteria for closed-loop correctness.

7. Extensions and Connections to Differential Inclusions and Nonsmooth Analysis

The methodology naturally generalizes to systems with nonsmooth dynamics, time-varying or set-valued nonlinearities, and even hybrid inclusions.

• Set-valued feedbacks: The differential inclusion framework accommodates feedbacks defined via maximally monotone operators, subdifferentials, or normal cones (as in projected dynamics or network equilibria).
• Lyapunov Stability and ISS: The adaptation of Lyapunov and passivity arguments to systems with feedthrough leverages convexity, monotonicity, and sector conditions on the nonlinearity, often encoded in algebraic and incremental dissipativity properties.
• Applications in Uncertain and Data-driven Models: The well-posedness results provide a mathematical basis for data-driven and neural network feedback system design, where identification or learning may produce models with implicit or non-invertible feedthrough structure.

---

The study of Lur'e systems with feedthrough thus requires a departure from standard ODE and feedback theory. Global invertibility and monotonicity conditions on the output equation's implicit map $y(t)$6, as detailed in (Guiver et al., 22 Aug 2025), form the basis for ensuring existence, uniqueness, and robustness properties in such systems. This has direct ramifications for mathematical modeling and synthesis in energy systems, cyber-physical networks, and nonlinear circuit and control architectures where direct feedthrough cannot be neglected.