In a theoretical study published in Frontiers in Signal Processing on November 16, 2026, Augusto Sarti of the Dipartimento di Elettronica, Informazione e Bioingegneria at Politecnico di Milano formulated a representation-aware framework for inverting nonlinear dynamical systems with memory. Constructing accurate, computationally tractable inverse models is fundamental across signal processing, physical modeling, and control engineering, finding application in digital predistortion of radio-frequency power amplifiers, loudspeaker linearization, and sensor compensation. Historically, research tackling nonlinear inversion evolved along isolated disciplinary tracks, divided by divergent mathematical formalisms and computational tools. Sarti’s central thesis establishes that intrinsic local input-output invertibility is an invariant property of the underlying system map once signal spaces and operating points are fixed, whereas constructibility, numerical conditioning, and computational cost depend strictly upon the chosen representation and its discrete realization. By comparing functional, structural, and geometric representations, the study demonstrates that scalable inversion relies on synthesizing inverse systems through progressive, modular operations rather than attempting intractable global analytical solutions.
Classical functional approaches to nonlinear system inversion predominantly rely on Volterra series expansions, which model causal, time-invariant systems as polynomial functionals constructed from multidimensional convolution operators of increasing order. While Volterra series provide a flexible functional foundation, standard inversion methodologies encounter severe computational obstacles arising from combinatorial kernel growth. In a conventional dense Volterra representation, a symmetric kernel of order p possessing memory depth M requires independent coefficients governed by combinatorial selection with replacement. Consequently, an inverse truncated at order five for a system with memory depth of 64 discrete taps demands exactly 8,531,168 independent coefficients in a dense, fully expanded representation. Attempting to determine or manipulate these high-dimensional multilinear operators simultaneously creates an insurmountable computational burden. Historically, recursive operatorial inversion mitigated this bottleneck by avoiding global system equations, instead progressively synthesizing inverse operators order by order to cancel residual nonlinear distortion terms, allowing researchers to relax restrictive minimal-order constraints and achieve controlled approximations.
In contrast to functional expansions, physical modeling traditions address nonlinear inversion from a structural perspective by mapping dynamical systems into equivalent electrical networks. In these equivalent circuits, system state variables correspond to voltages and currents governed by topological interconnections among linear reactive components and localized nonlinear elements. Within this structural paradigm, system inversion can be achieved at the circuit level by inserting idealized two-port nullors—comprising paired nullators and norators—which reassign the functional roles of input and output variables while preserving internal network topology. Although nullor-based transformations yield exact structural inverses, they initially produce implicit nonlinear Kirchhoff equations that traditionally required costly iterative numerical solvers at each discrete time step. By converting these circuit models into Wave Digital Filter architectures using wave scattering variables and port resistance adaptation, implicit relations transform into explicit, sample-by-sample update equations. This structural transformation eliminates iterative algebraic loops, ensuring numerical stability and predictable latency for demanding real-time digital signal processing applications.
The framework incorporates a third complementary perspective rooted in differential geometry and Lie theory, which formulates nonlinear dynamical systems in state-space on differentiable manifolds governed by smooth vector fields. In this geometric setting, system inversion corresponds to reconstructing the inverse dynamical flow along system trajectories, a task that remains globally intractable but admits systematic local approximations within the tangent space. Local invertibility and controllability are characterized through the algebraic structure of associated Lie brackets and iterated Lie derivatives, establishing a direct connection to classical feedback linearization. Feedback linearization achieves exact closed-loop linearization through strictly local operations, where well-posedness requires a well-defined relative degree—the geometric analogue of an invertible first-order linear transfer operator. Furthermore, this geometric viewpoint provides a theoretical foundation for dynamic adaptation in Wave Digital Filters, where nonlinear elements are recursively replaced by their local tangent models as the state trajectory advances, effectively updating the computational representation itself in real time.
To formalize the connections among these three computational languages, the study defines a representation not as a simple coordinate change, but as an operational medium possessing unique primitives, composition rules, and notions of locality. In functional representations, the primitives are multilinear convolution kernels combined through operator composition, with locality defined order by order. In geometric representations, the primitives are vector fields composed through flow concatenation, with locality existing in the trajectory’s tangent space. In structural representations, the primitives are physical ports and interconnections operating under wave scattering, where locality means a nonlinearity occupies an isolated topological node. Building on these definitions, Hypothesis H1 posits structure-preserving constructibility: whenever an invertible nonlinear system exhibits exploitable compositional structure, such as localized nonlinearities, separability, or low-rank factorizations, there exists an admissible representation preserving this structure, ensuring inversion complexity scales with retained components rather than dense global expansions. Sarti defines a clear falsification criterion: identifying an invertible structured system whose faithful realization unavoidably incurs combinatorial expansion.
Hypothesis H2 extends this conceptual foundation by addressing representation transfer, asserting that when an underlying input-output map possesses a unique local inverse and admits faithful, regular descriptions across multiple representation classes, each class can furnish a corresponding well-posed realization of that inverse. To provide a rigorous anchor, Proposition 1 proves the uniqueness and representation consistency of the formal local input-output inverse under the assumption that the first-order linear operator is a bounded linear isomorphism. When expanded order by order using homogeneous degrees, the formal inverse is uniquely determined through recursive operator composition. Consequently, any geometric or structural realization that projects onto a formal input-output inverse must strictly agree with the functional expansion order by order. However, the study emphasizes that Proposition 1 proves algebraic consistency rather than existence transfer, leaving structural realization transfer conjectural because internal scattering Jacobians can encounter avoidable singularities despite regular external input-output behavior.
Bridging theoretical existence and practical engineering implementation, the author introduces Operational Principle P1, which states that analytical invertibility alone is insufficient for real-time usefulness. A practically viable inverse realization must be strictly causal or employ a clearly defined preview horizon, remain stable over the operating signal space, exhibit robust numerical conditioning against parameter perturbations, and compute within deterministic latency budgets. Within practical signal processing pipelines, structured inverses operate primarily within Direct-Inverse-Direct architectures, where a digital predistortion module is placed in cascade with a physical forward plant, preceded by an optional reference target model. In this configuration, the inverse functions operationally as a right inverse across the reachable target range, enforcing identity mapping to cancel physical distortion. This operational right-inverse requirement differs from the left-inverse identity used in formal recursive proofs, though both coincide over common regular domains, providing the architectural foundation for advanced applications such as electroacoustic loudspeaker linearization.
The practical consequences of the framework are demonstrated through a canonical physical model of an electroacoustic transducer featuring a dominant cubic stiffness characteristic. Proposition 2 establishes structure-preserving constructibility for this cubic element, contrasting a local series expansion against an exact localized structural realization. When inverted through a formal Taylor series about the origin, the resulting series exhibits a finite convergence radius of two divided by three times the square root of three times the stiffness coefficient. This strict analytic boundary is dictated by complex branch points where the forward derivative vanishes. If the driving input exceeds this critical radius, truncated Taylor series fail to converge, and adding higher-order terms degrades approximation accuracy. Conversely, because the real cubic function is strictly monotonic and constitutes a global diffeomorphism on the real line, retaining the cubic nonlinearity as a single localized structural element permits an exact, closed-form algebraic evaluation that requires a fixed, order-independent computational cost per sample.
To evaluate the limits of invertibility, the study examines a non-monotone softening stiffness characteristic where restoring force diminishes with increasing displacement. In this softening regime, the forward derivative vanishes at a physical fold bifurcation, occurring at a critical displacement of one over the square root of three times the stiffness coefficient, corresponding to a maximum reachable output threshold. Beyond this fold, demanded displacements become physically unreachable on the principal solution branch, causing the system to exhibit discontinuous jump phenomena. Sarti highlights that this fold represents an intrinsic loss of mathematical invertibility belonging to the underlying map itself, rather than a computational failure of any specific representation. In structural Wave Digital realizations, proximity to this fold is exposed directly through the localized tangent slope, which is continuously monitored during dynamic adaptation. This allows engineers to implement explicit guard thresholds that limit demanded outputs to the reachable set, maintaining bounded scalar inverse gains and ensuring numerical stability.
Addressing dynamic systems with resonant memory, the study analyzes the classic Duffing oscillator driven by a cubic restoring force embedded within second-order resonant differential equations. When configured in a Direct-Inverse-Direct architecture targeting the response of an idealized linear oscillator, the exact third-order functional predistorter possesses a separable Volterra kernel that is strictly rank-one rather than a dense multidimensional tensor. In the structural circuit domain, the identical inverse computation is achieved by simulating the linear target block and directly adding the localized cubic restoring force. This finding reveals that the structural and factorized functional implementations are algebraically identical, disproving the earlier assumption that structural methods hold an inherent complexity advantage over all functional models. The author extends these insights to digital predistortion in communications, noting how machine learning architectures act as modern functional realizations, while establishing clear boundaries: structural inversion cannot operate on unmodeled black-box data or distributed continuous media where nonlinearities cannot be spatially localized.
Subject of Research: Representation-aware constructive principles for structured nonlinear system inversion
Article Title: Structured nonlinear inversion of systems with memory: a representation-aware constructive principle across operator, structural, and geometric views
Article References: Sarti, A. (2026). Structured nonlinear inversion of systems with memory: a representation-aware constructive principle across operator, structural, and geometric views. Frontiers in Signal Processing, 6, 1951118. https://doi.org/10.3389/frsip.2026.1951118
Image Credits: AI Generated
DOI: 10.3389/frsip.2026.1951118
Keywords: nonlinear system inversion, systems with memory, Volterra series, wave digital filters, Lie groups, structural inversion, recursive operatorial inversion, feedback linearization, Direct-Inverse-Direct architecture, Duffing oscillator, nullor-based circuit inversion, digital predistortion, dynamic adaptation, cubic stiffness, representation-aware inversion, computational complexity, real-time signal processing, state-space manifolds
News Source: Denise Maddox. (November 16, 2026). Modular Structural Representations Enable Scalable Real-Time Inversion of Nonlinear Systems. Scienmag.



