This paper studies continuous-time distributed adaptive identification of unknown nonlinear systems over directed communication networks without requiring source or neighboring inputs to be shared. Each sensing node reconstructs the source state while adapting local estimates of the unknown linear and nonlinear dynamics using only locally available information and neighboring state estimates. A network-level error representation is derived to characterize the interaction among the communication topology, observer dynamics, parameter uncertainty, and nonlinear mismatch. Norm-based, structured Lyapunov, and robust LMI conditions are established to guarantee that the resulting coupling operator is Hurwitz. Unlike the corresponding discrete-time formulation, the proposed continuous-time adaptive laws do not require normalization. A Lyapunov analysis establishes boundedness of the adaptive parameters and asymptotic convergence of the estimation and source-tracking errors in the unperturbed case. Input-to-state stability is further established under bounded predictor mismatch and source disturbances. Numerical studies over star, cyclic, and path topologies support the theoretical results and demonstrate approximately linear computational scaling with the number of sensing nodes.
Introduction
Distributed sensing networks are increasingly used to monitor dynamic processes whose states cannot be reliably reconstructed by a single sensor. Their development builds on advances in consensus, cooperative control, and networked multi-agent systems, which provide systematic tools for coordinating information over directed and time-varying graphs [1, 2, 3, 4, 5]. Applications include autonomous systems, infrastructure monitoring, industrial processes, environmental sensing, and biological networks [6, 7, 8, 9]. In these settings, multiple sensing nodes combine locally available information with data received from neighboring nodes to estimate the behavior of a common dynamic source. Compared with centralized architectures, distributed estimation can improve spatial coverage, scalability, and resilience while reducing reliance on a single processing unit [10, 11].
Distributed state estimation has consequently received considerable attention. Consensus-based Kalman filters and diffusion strategies distribute the filtering task across sensor networks [12, 13, 14, 15, 16]. For deterministic linear systems, distributed observers have been developed under various detectability, communication, and convergence requirements [17, 18, 19, 20, 21]. These contributions establish important foundations, but most presume that the monitored dynamics are known when the observer is constructed.
A more difficult problem arises when the source dynamics are unknown. Adaptive estimation offers an online alternative to fixed-model observers and offline data-driven training. Classical adaptive-observer designs jointly estimate states and unknown parameters for structured nonlinear systems [22, 23, 24, 25, 26]. At the network level, diffusion least-mean-square, recursive least-squares, and related learning strategies exploit cooperation among spatially distributed data sources [27, 28, 29, 30, 31]. Distributed adaptive protocols have also been used to coordinate agents with uncertain linear or Lipschitz nonlinear dynamics [32]. Nevertheless, jointly reconstructing an unknown nonlinear source and adapting its dynamics remains challenging because the model estimates, state estimates, and graph coupling evolve simultaneously.
Unknown source inputs introduce a further difficulty. Classical unknown-input observers seek to decouple or reconstruct unmeasured exogenous signals [33, 34, 35, 36]. In a sensing network, however, requiring the source input to be available or transmitted to every node may be impractical because of limited bandwidth, privacy constraints, communication failures, or the absence of direct input measurements [37]. Exchanging the adaptive inputs generated by neighboring nodes similarly increases communication and storage requirements. A genuinely distributed method should therefore reconstruct the source state and adapt its unknown dynamics without sharing either the source input or neighboring inputs [38].
The interaction between adaptation and network coupling makes the stability analysis nontrivial [39]. Stability of the individual observer matrices alone does not guarantee stability of the complete estimation network because the graph weights, source dynamics, and local observer dynamics jointly determine the network error operator. Nonlinear terms create an additional mismatch because each node evaluates the nonlinear mapping at a locally aggregated state rather than at the unavailable source state. Moreover, time-varying unknown inputs and external disturbances generally preclude exact convergence and require robustness estimates and small-gain arguments [40]. Related adaptive and consensus-based diagnostic methods illustrate both the value and the difficulty of handling uncertainty in networked nonlinear systems [41, 42, 43].
Continuous-time systems deserve particular attention because they naturally describe many physical, biological, and engineering processes. Continuous-time adaptation also has a useful structural feature: the Lyapunov derivative is linear in the adaptive update rates. The predictor-mismatch terms can therefore be canceled directly, without the quadratic adaptive-increment terms that arise in a discrete-time Lyapunov difference. This permits unnormalized adaptive laws and yields a simpler closed-loop analysis. At the same time, local implementability requires more than the existence of an arbitrary positive-definite Lyapunov matrix: its network-level factor must preserve a block-diagonal structure so that each adaptive direction can be computed from local and neighboring errors.
Motivated by these considerations, this paper develops a continuous-time distributed adaptive identification framework for unknown nonlinear source systems over directed communication networks. Each sensing node forms an aggregated state from locally available source and neighboring information, updates local estimates of the unknown linear and nonlinear dynamics, and adjusts its own input without accessing the unknown source input or the inputs generated by neighboring nodes. The resulting framework simultaneously addresses state reconstruction, online model adaptation, limited communication, local implementability, and robustness to bounded perturbations.
The main contributions are summarized as follows:
A continuous-time distributed adaptive identification framework is developed for nonlinear source systems with unknown linear dynamics, unknown nonlinear coefficients, and an unavailable source input. Each sensing node uses only locally available information and neighboring state and error signals, without requiring the source input or neighboring adaptive inputs to be shared.
A compact network-level error model is derived to separate the effects of the communication topology, local observer dynamics, adaptive predictor mismatch, nonlinear mapping mismatch, and external disturbances. This representation provides the basis for the subsequent stability and robustness analysis.
Proposition 1 establishes a direct logarithmic-norm condition for the Hurwitz stability of the network coupling operator. Lemma 1 provides a structure-preserving Lyapunov certificate with a diagonal network-level factor, while Corollary 1 extends the certificate to norm-bounded uncertainty in the unknown source matrix.
Locally implementable, unnormalized continuous-time adaptive laws are introduced for the unknown linear matrix, nonlinear coefficient matrix, and local input. The structured Lyapunov matrix makes the network-weighted adaptive direction computable from local and neighboring errors. Unlike the discrete-time formulation, normalization is unnecessary because the adaptive predictor mismatch is canceled exactly in the Lyapunov derivative.
Theorem 1 establishes a general Lyapunov bound and, in the nominal case, boundedness of the adaptive variables and asymptotic convergence of the estimation and source-tracking errors under the proposed Hurwitz and nonlinear small-gain conditions. Corollary 2 provides a finite-time ISS estimate and the corresponding asymptotic bounds under bounded predictor mismatch and source disturbances.
Numerical studies over star, cyclic, and path communication topologies validate the theoretical results. The nominal experiments demonstrate asymptotic state reconstruction, while the perturbed experiments show bounded tracking
under a time-varying unknown input and external disturbance. The proposed Hurwitz certificates are verified numerically, and a fixed-step complexity benchmark demonstrates approximately linear growth in online computational cost with the number of sensing nodes.
The remainder of this paper is organized as follows. Section 2 introduces the directed communication network and its associated matrices. Section 3 formulates the unknown nonlinear source and distributed adaptive observers. Section 4 derives the network error dynamics, establishes the Hurwitz stability conditions, and presents the adaptive stability and robustness results. Section 5 reports the numerical studies and computational benchmarks. Section 6 concludes the paper.
Notation.Rn and Rn×p denote the spaces of real n-vectors and n×p matrices, respectively. The p×p identity matrix is Ip. For matrices Ai∈Rn×q, i∈{1,…,p}, the notation diag{A1,…,Ap}∈Rpn×pq denotes the corresponding block-diagonal matrix. The vectors of ones and zeros in Rp are 1p and 0p, respectively. The symbols A⊗B, tr(A), λ(A), and vec(A) denote the Kronecker product, trace, spectrum, and vectorization of A, respectively.
Communication Network
Consider a network of m+1 agents indexed by V={0,1,…,m}. Agent 0 represents the dynamic source whose state is to be reconstructed and whose dynamics are unknown, whereas agents 1,…,m constitute the sensing network. Each sensing node operates as a sensor–observer: it gathers locally available information, updates its adaptive observer, and exchanges state estimates with its in-neighbors to cooperatively reconstruct the source state while adapting local estimates of its unknown dynamics. The source does not participate in the distributed estimation process and solely generates the state trajectory of interest.
The communication architecture is described by a weighted directed graph G=(V,E,W), where E⊆V×V is the edge set and W={wij} is the collection of nonnegative communication weights. An edge (i,j)∈E indicates that agent i receives information from agent j with weight wij>0; otherwise, wij=0. Accordingly, the in-neighborhood of agent i is defined as Ni:={j∈V∣(i,j)∈E}. The weights {wij} are design parameters of the distributed identification scheme and do not represent physical signal strengths.
For analysis, we separate G into two subgraphs: (i) the sensing subgraph Gm=(Vm,Em,Wm) with Vm={1,…,m}; (ii) the source–sensing subgraph G0=(V0,E0,W0) with V0={0}∪{i∈Vm∣(i,0)∈E}. The sensing subgraph Gm is assumed bidirected, i.e., (i,j)∈Em if and only if (j,i)∈Em, although the reciprocal weights wij and wji need not be equal. The overall graph G remains directed through the source–sensing links in G0; see Figure 1. This decomposition simplifies the subsequent network-level analysis.
Figure 1. Example of a graph G with m=5, its sensing-to-sensing subgraph Gm, and its source-to-sensing subgraph G0, showing the graph decomposition and assignment of wij. Each line in Gm represents two oppositely directed edges, as explicitly illustrated between nodes 4 and 5. The corresponding reciprocal weights need not be equal.
For the sensing subgraph Gm, define the weighted in-degree of node i as di:=∑j:(i,j)∈Emwij, i∈Vm, and let Dm:=diag{d1,…,dm}. The corresponding adjacency matrix Am has entries [Am]ij=wij whenever (i,j)∈Em, and [Am]ij=0 otherwise. The Laplacian matrix associated with Gm is therefore given by Lm:=Dm−Am.
The source–sensing connections in G0 are represented by the diagonal matrix A0:=diag{w10,…,wm0}, where [A0]ii=wi0 if (i,0)∈E0 and [A0]ii=0 otherwise.
Let W:=diag{w1,…,wm}, where wi:=di+wi0. The network coupling matrix is defined as
L:=Lm+A0=W−Am.
We focus on the balanced settingW=Im, under which
(L−A0)1m=0m⟺(Am+A0)1m=1m.(1)
Hence, the total incoming weight assigned by each sensing node to its neighbors and the source is equal to one.1
Remark 1 (Source reachability).Suppose that at least one sensing node has direct access to the source, that is, wi0>0 for some i∈Vm, and that every sensing node can be reached from the source through a directed path. Under this connectivity condition, the network coupling matrix L=Lm+A0 is positive stable; namely, all its eigenvalues satisfy ℜ{λ(L)}>0.
Problem Formulation
Consider a distributed sensing network consisting of an unknown nonlinear source, indexed by 0, and m sensing nodes indexed by i∈{1,…,m}. Each sensing node processes locally available information and implements an adaptive observer to reconstruct the source state while updating local estimates of its unknown dynamics.
The source dynamics are described by
x˙0=A0x0+f0(x0)+B0u0+δ0,(2)
where x0∈Rn denotes the source state and u0∈Rp is the unknown source input. The unknown matrix A0∈Rn×n is assumed to be Hurwitz, whereas B0∈Rn×p is known. In particular, A0 admits the decomposition A0=A∗+ΔA, where A∗ is known and the uncertainty ΔA satisfies ∥ΔA∥2≤a for some known constant a>0.
References
[1]R. Olfati-Saber and R. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1520–1533, 2004.
[2]R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proceedings of the IEEE, vol. 95, no. 1, pp. 215–233, 2007.
[3]M. K. Wafi, Y. E. Nugraha, B. L. Widjijanto, and K. Indriwati, “Cooperative H∞ fault-tolerant tracking with miss guarantees for networked systems with sensor faults,” in 2026 15th Asian Control Conference (ASCC), 2026, pp. 1468–1473.arxiv.org/abs/2604.03921
[4]W. Ren and R. Beard, “Consensus seeking in multiagent systems under dynamically changing interaction topologies,” IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 655–661, 2005.
[5]A. Dorri, S. S. Kanhere, and R. Jurdak, “Multi-agent systems: A survey,” IEEE Access, vol. 6, pp. 28 573–28 593, 2018.DOI
[6]Y. Cao, W. Yu, W. Ren, and G. Chen, “An overview of recent progress in the study of distributed multi-agent coordination,” IEEE Transactions on Industrial Informatics, vol. 9, no. 1, pp. 427–438, 2013.arxiv.org/abs/1207.3231
[7]M. K. Wafi, R. Hajian, B. Shafai, and M. Siami, “Advancing fault-tolerant learning-oriented control for unmanned aerial systems,” in 2023 9th International Conference on Control, Decision and Information Technologies (CoDIT), 2023, pp. 1688–1693.DOI
[8]M. K. Wafi, “Learning-oriented fault-tolerant control for uavs with time-varying fault estimation,” arXiv preprint arXiv:2610.30566, 2026. [Online]. Available: https://arxiv.org/abs/2610.30566
[9]A. Jadbabaie, J. Lin, and A. Morse, “Coordination of groups of mobile autonomous agents using nearest neighbor rules,” IEEE Transactions on Automatic Control, vol. 48, no. 6, pp. 988–1001, 2003.
[10]L. Moreau, “Stability of multiagent systems with time-dependent communication links,” IEEE Transactions on Automatic Control, vol. 50, no. 2, pp. 169–182, 2005.DOI
[11]M. K. Wafi and M. Siami, “Distributed adaptive control of disturbed interconnected systems with high-order tuners,” IEEE Control Systems Letters, vol. 8, pp. 1421–1426, 2024.DOI
[12]R. Olfati-Saber, “Distributed kalman filtering for sensor networks,” in 2007 46th IEEE Conference on Decision and Control, 2007, pp. 5492–5498.
[13]S. Das and J. M. F. Moura, “Distributed kalman filtering with dynamic observations consensus,” IEEE Transactions on Signal Processing, vol. 63, no. 17, pp. 4458–4473, 2015.DOI
[14]——, “Consensus+innovations distributed kalman filter with optimized gains,” IEEE Transactions on Signal Processing, vol. 65, no. 2, pp. 467–481, 2017.DOI
[15]F. S. Cattivelli and A. H. Sayed, “Diffusion strategies for distributed kalman filtering and smoothing,” IEEE Transactions on Automatic Control, vol. 55, no. 9, pp. 2069–2084, 2010.DOI
[16]S. P. Talebi and S. Werner, “Distributed kalman filtering and control through embedded average consensus information fusion,” IEEE Transactions on Automatic Control, vol. 64, no. 10, pp. 4396–4403, 2019.DOI
[17]S. Park and N. C. Martins, “Design of distributed lti observers for state omniscience,” IEEE Transactions on Automatic Control, vol. 62, no. 2, pp. 561–576, 2017.DOI
[18]A. Mitra and S. Sundaram, “Distributed observers for lti systems,” IEEE Transactions on Automatic Control, vol. 63, no. 11, pp. 3689–3704, 2018.arxiv.org/abs/1608.01429
[19]L. Wang and A. S. Morse, “A distributed observer for a time-invariant linear system,” IEEE Transactions on Automatic Control, vol. 63, no. 7, pp. 2123–2130, 2018.
[20]W. Han, H. L. Trentelman, Z. Wang, and Y. Shen, “A simple approach to distributed observer design for linear systems,” IEEE Transactions on Automatic Control, vol. 64, no. 1, pp. 329–336, 2019.arxiv.org/abs/1708.01459
[21]R. Ortega, E. Nuño, and A. Bobtsov, “Distributed observers for lti systems with finite convergence time: A parameter-estimation-based approach,” IEEE Transactions on Automatic Control, vol. 66, no. 10, pp. 4967–4974, 2021.arxiv.org/abs/2005.12940
[22]G. Bastin and M. Gevers, “Stable adaptive observers for nonlinear time-varying systems,” IEEE Transactions on Automatic Control, vol. 33, no. 7, pp. 650–658, 1988.
[23]R. Marino, “Adaptive observers for single output nonlinear systems,” IEEE Transactions on Automatic Control, vol. 35, no. 9, pp. 1054–1058, 1990.DOI
[24]R. Marino and P. Tomei, “Global adaptive observers for nonlinear systems via filtered transformations,” IEEE Transactions on Automatic Control, vol. 37, no. 8, pp. 1239–1245, 1992.DOI
[25]Y. M. Cho and R. Rajamani, “A systematic approach to adaptive observer synthesis for nonlinear systems,” IEEE Transactions on Automatic Control, vol. 42, no. 4, pp. 534–537, 1997.
[26]E. Javanfar, M. Rahmani, and M. K. Wafi, “Robust estimation-based non-fragile control for discrete-time nonlinear systems,” International Journal of Robust and Nonlinear Control, vol. 35, no. 6, pp. 2462–2471, 2025.
[27]C. G. Lopes and A. H. Sayed, “Diffusion least-mean squares over adaptive networks: Formulation and performance analysis,” IEEE Transactions on Signal Processing, vol. 56, no. 7, pp. 3122–3136, 2008.DOI
[28]F. S. Cattivelli, C. G. Lopes, and A. H. Sayed, “Diffusion recursive least-squares for distributed estimation over adaptive networks,” IEEE Transactions on Signal Processing, vol. 56, no. 5, pp. 1865–1877, 2008.DOI
[29]G. Mateos and G. B. Giannakis, “Distributed recursive least-squares: Stability and performance analysis,” IEEE Transactions on Signal Processing, vol. 60, no. 7, pp. 3740–3754, 2012.DOI
[30]A. H. Sayed, S.-Y. Tu, J. Chen, X. Zhao, and Z. J. Towfic, “Diffusion strategies for adaptation and learning over networks: an examination of distributed strategies and network behavior,” IEEE Signal Processing Magazine, vol. 30, no. 3, pp. 155–171, 2013.DOI
[31]A. H. Sayed, “Adaptive networks,” Proceedings of the IEEE, vol. 102, no. 4, pp. 460–497, 2014.DOI
[32]Z. Li, W. Ren, X. Liu, and M. Fu, “Consensus of multi-agent systems with general linear and lipschitz nonlinear dynamics using distributed adaptive protocols,” IEEE Transactions on Automatic Control, vol. 58, no. 7, pp. 1786–1791, 2013.DOI
[33]M. Hou and P. Muller, “Design of observers for linear systems with unknown inputs,” IEEE Transactions on Automatic Control, vol. 37, no. 6, pp. 871–875, 1992.DOI
[34]M. Darouach, M. Zasadzinski, and S. Xu, “Full-order observers for linear systems with unknown inputs,” IEEE Transactions on Automatic Control, vol. 39, no. 3, pp. 606–609, 1994.DOI
[35]A. Chakrabarty, M. J. Corless, G. T. Buzzard, S. H. Žak, and A. E. Rundell, “State and unknown input observers for nonlinear systems with bounded exogenous inputs,” IEEE Transactions on Automatic Control, vol. 62, no. 11, pp. 5497–5510, 2017.DOI
[36]K. C. Veluvolu and Y. C. Soh, “Discrete-time sliding-mode state and unknown input estimations for nonlinear systems,” IEEE Transactions on Industrial Electronics, vol. 56, no. 9, pp. 3443–3452, 2009.DOI
[37]M. K. Wafi, A. Ataka, Y. Y. Nazaruddin, and B. Jayawardhana, “Distributed nonlinear control of networked two-wheeled robots under adversarial interactions,” in 2026 15th Asian Control Conference (ASCC), 2026, pp. 1524–1529.arxiv.org/abs/2604.03917
[38]M. K. Wafi and M. Siami, “Distributed adaptive estimation of unknown nonlinear systems without input sharing,” arXiv preprint arXiv:2607.16510, 2026. [Online]. Available: https://arxiv.org/abs/2607.16510
[39]M. K. Wafi, H. M. Hedesh, and M. Siami, “A passivity-agnostic framework for distributed adaptive synchronization under unknown leader dynamics,” in 2026 American Control Conference (ACC), 2026, pp. 3385–3391.arxiv.org/abs/2603.02344
[40]A. Teel, “A nonlinear small gain theorem for the analysis of control systems with saturation,” IEEE Transactions on Automatic Control, vol. 41, no. 9, pp. 1256–1270, 1996.DOI
[41]M. K. Wafi and M. Siami, “A comparative analysis of reinforcement learning and adaptive control techniques for linear uncertain systems,” in 2023 Proceedings of the Conference on Control and its Applications (CT), 2023, pp. 25–32.DOI
[42]E. Franco, R. Olfati-Saber, T. Parisini, and M. M. Polycarpou, “Distributed fault diagnosis using sensor networks and consensus-based filters,” in Proceedings of the 45th IEEE Conference on Decision and Control, 2006, pp. 386–391.DOI
[43]R. M. G. Ferrari, T. Parisini, and M. M. Polycarpou, “Distributed fault detection and isolation of large-scale discrete-time nonlinear systems: An adaptive approximation approach,” IEEE Transactions on Automatic Control, vol. 57, no. 2, pp. 275–290, 2012.DOI