Neural Observer Frameworks

• Neural Observer Frameworks are architectures that reconstruct unmeasured states and unknown parameters from input-output data using a structured, two-layer filter approach.
• They leverage initial excitation over a finite window to relax the persistent excitation condition, enabling exponential convergence in state and parameter estimation.
• The framework employs normalized switched gradient descent and regressor stacking to accelerate learning and ensure stability through rigorous Lyapunov-based analysis.

A neural observer framework defines the architecture, adaptation laws, and theoretical guarantees for reconstructing unmeasured states and unknown parameters of a dynamical system from observed input-output data, typically in discrete time, using recursive computational blocks emulating a “neural” signal processing structure. In the context of adaptive observer theory for discrete-time linear time-invariant (LTI) systems, the neural observer framework is instantiated via hierarchically structured filters with rigorous update schemes, enabling exponential convergence of both state and parameter estimates under finite-time information conditions such as initial excitation. This framework is distinct from standard adaptive observers due to its structured multilayer regressor construction and the ability to relax the stringent persistence of excitation conditions, yielding enhanced practicality in stabilization and regulation tasks (Dey et al., 17 Nov 2025).

1. Formal System Definition and Observer Objective

The underlying plant model in this framework is a discrete-time LTI system of the form: $$x(k+1) = A x(k) + B u(k),\quad y(k) = C x(k) + D u(k),$$ where $x \in \mathbb{R}^n$ (state), $u \in \mathbb{R}^m$ (input), $y \in \mathbb{R}^q$ (output), and $A, B, C, D$ are constant matrices of compatible dimensions. It is assumed that $A, B$ are unknown, $x(k)$ is unmeasured, $C = [I_q \;\; 0]$, $D = 0$, and the output history $\{u(0), y(0), \ldots, u(k), y(k)\}$ is available.

The neural observer’s goal is to estimate simultaneously:

• The unknown plant parameters ($A$, $B$),
• The unmeasured full state ($x(k)$),
• The unknown initial condition ($x(0)$),

using input-output data alone, with guaranteed exponential convergence of both state- and parameter-estimation errors.

2. Excitation Conditions: Persistence versus Initial Excitation

The identification and observer literature traditionally demands persistence of excitation (PE) of the regressor, defined for a regressor $\phi(k) \in \mathbb{R}^h$ as: $$\exists N \ge h,\,\alpha>0:\; \forall k,\;\; \sum_{i=k}^{k+N-1} \phi(i) \phi(i)^\top \succeq \alpha I_h.$$ PE requires continuous, infinite-time nontrivial excitation of the input, which in regulation tasks leads to infinite “probing” effort and can degrade closed-loop performance.

The neural observer framework adopts the notion of initial excitation (IE): $$\exists T\ge h,\;\alpha>0:\; \sum_{k=0}^{T-1} \phi(k)\phi(k)^\top \succeq \alpha I_h,$$ which only requires input richness over a finite time window $[0, T)$. After $T$, no further excitation is needed, so stabilization/regulation can proceed in a standard manner. IE is thus a strictly weaker requirement than PE but suffices for exponential parameter convergence once the excitation window is completed.

3. Two-Layer Neural Observer Architecture

The neural observer is structured via two interconnected filter “layers” which recursively transform the input-output sequence to build rich, informative regressors for adaptation:

Layer 1 (State-Propagation Filter):

$$\Phi_1(k+1) = F\,\Phi_1(k) + G\,u(k),\quad \Phi_1(0)=0,$$

where $F$ is a Schur matrix (user-chosen, structurally similar to $A$), $G$ designates input enter points.

Layer 2 (Regressor Accumulation Filter):

$$S(k+1) = \sigma S(k) + \Psi(k)\Psi(k)^\top, \qquad R(k+1) = \sigma R(k) + \Psi(k) Y(k),$$

with $\sigma\in(-1,1)$ a forgetting factor, and where $\Psi(k)$ is a regressor constructed from filtered $(u, y)$. The recursions in Layer 2 generate a secondary regression problem: $R(k) = S(k) \theta,$ with $\theta$ aggregating the unknowns $\mathrm{vec}(A), \mathrm{vec}(B), x(0)$.

The nested filter structure realizes a multi-layer dynamic “neural” architecture in the sense that the inner filter propagates the measured data, and the outer filter aggregates sufficient excitation statistics to support adaptation.

4. Normalized Switched Gradient-Based Parameter Update

The neural observer updates the parameter estimate $\hat\theta(k)$ by a normalized, switched gradient-descent law based on the outputs of both layers: $$\hat\theta(k+1) = \hat\theta(k) + \frac{ \kappa_1\,\Psi(k+1)^\top\left[Y(k+1)-\Psi(k+1)\hat\theta(k)\right] + \kappa_2\,S(k+1)\left[R(k+1)-S(k+1)\hat\theta(k)\right] + \kappa_3\,\eta\,S(T)\left[R(T)-S(T)\hat\theta(k)\right] } {1+\|\Gamma(k)\|_F^2},$$ where $\kappa_1,\kappa_2,\kappa_3>0$ are adaptation gains, and $\Gamma(k)$ collects all regressor blocks. Crucially, the "IE switch" term (proportional to $\kappa_3$ and activated via $\eta=1$ at $k\geq T$) gives an eigenvalue boost that drives exponential convergence after the IE window. The normalization denominator controls adaptation in the presence of high regressor magnitudes.

5. Regressor Modification and Convergence Acceleration

For further acceleration, the neural observer modifies regressors by stacking multiple (exponentially weighted) past values: $$W(k) = \begin{bmatrix} \Psi(k) \ \alpha\Psi(k-1) + (1-\alpha)\Psi(k-2) \ \vdots \end{bmatrix},\quad 0<\alpha<1,$$ such that

$Y_W(k) = W(k)\,\theta,$

creating multiple independent error directions at each update. This “matrix-regressor” approach (strong-IE) enriches the aggregate excitation along distinct components of $\theta$, ensuring $W^\top W \succ 0$ over the finite IE window, which further expedites learning.

6. Theoretical Guarantees: Lyapunov Analysis and Exponential Convergence

Under boundedness of all signals (achieved by design of $F$, $\sigma$, and adaptation gains), and assuming the strong-IE condition on $W$, the Lyapunov function

$V(k) = \|\tilde\theta(k)\|^2$

admits the decay estimate

$V(k+1)\le (1-\kappa)V(k),$

for some $\kappa>0$ after the IE switch at $T$. The filtered estimator structure ensures that the state estimation error $\tilde x(k)$ converges exponentially due to bounded input and exponentially decaying parameter error (Dey et al., 17 Nov 2025). The exponential rate is tunable via adaptation gains and the regressor stacking parameter $\alpha$.

7. Practical Implications, Limitations, and Applications

The neural observer framework—realized as described—offers several practical advantages:

• Regulation/Control Synergy: Only finite-time probing (via initial excitation) is needed. After the excitation phase, inputs can focus on regulation tasks without continuous probing signals, which would otherwise conflict with stabilizer requirements.
• Computational Parity: The neural observer’s filter complexity remains $O(n\times p)$, matching classical PE-based designs while yielding greater practical flexibility.
• Convergence Speed: Explicit eigenvalue “switch-on” after IE also results in faster convergence than standard PE-based approaches, as observed in simulation.
• Limitations: IE verification must be conducted online. Insufficient initial input richness leads to absence of the switch and loss of exponential convergence. Noise robustness and adaptation to unmodeled time variation require further extension beyond the core framework.

In exemplars, such as a 3-state, 2-input high-dimensional plant with unstable $A,B$, the IE-driven neural observer achieves exponential decay of both parameter and state errors within the first 12 steps of initial excitation; PE or non-PE-based observers prove inferior in convergence or require continuous excitation (Dey et al., 17 Nov 2025).

8. Summary Table: Key Elements of Neural Observer Frameworks

Component	Role	Mathematical Feature
Two-layer filtering	Regressor/statistics design	Recursive, nested filters
IE-based excitation	Guarantees convergence	Finite-time (windowed) PE
Normalized switched law	Robust, fast adaptation	Gradient + IE-activated boost
Matrix-regressor stacking	Accelerate learning	Strong-IE, multi-sample fusion
Exponential Lyapunov	Proof of guarantee	Boundedness & rate tuning

The neural observer framework thus provides a systematic, scalable, and analytically rigorous approach for discrete-time, output-only adaptive state and parameter estimation in LTI systems. It is practically deployable in high-dimensional, regulation-centric control architectures where infinite-time excitation is infeasible or undesirable, and forms a foundation for advanced, high-performance observer synthesis (Dey et al., 17 Nov 2025).