Pith. sign in

REVIEW 4 major objections 4 minor 19 references

Adaptive Consensus with Exponential Decay

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Concurrent learning makes adaptive consensus converge exponentially, and the same guarantee survives uniform quantization of the communication.

desk verdict The concurrent-learning update is written with the unknown parameter, so the exponential consensus claim is for an oracle controller; the fix is straightforward and the rest of the analysis is sound. read the letter →

arxiv 2506.07203 v1 pith:QEO4KOUM submitted 2025-06-08 math.OC

classification math.OC MSC 93D5093C40
keywords adaptiveconsensusconcurrentlearningexponentialconvergenceuniformquantizationmulti-agentsystemsparameterestimationKrasovskiisolutions
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that adding a concurrent-learning term to the standard adaptive consensus controller turns merely asymptotic consensus into exponential consensus, for general stabilizable linear agents with matched nonlinear uncertainty, without requiring persistent excitation. The exponential rate is explicitly $\min\{\gamma,q\}$, where $\gamma$ comes from graph connectivity and $q$ from the richness of stored history. The same control law, when measurements pass through a uniform quantizer, still achieves consensus up to an error of order of the quantization level. The point of the paper is that stored historical data can replace persistent excitation, making exponential convergence achievable in situations where traditional adaptive control would only ensure slow asymptotic agreement.

What carries the argument

The carrying object is the Lyapunov function $V = x^T(L\otimes P)x + \frac12\sum_i \operatorname{tr}(\tilde{\theta}_i^T\tilde{\theta}_i)$. The new piece is the concurrent-learning term $\sum_k \Phi_i(x_{i,k})^T\Phi_i(x_{i,k})\tilde{\theta}_i$ added to the parameter update; Condition 1, full rank of the stored matrix $Z_i=[\Phi_i(x_{i,1}),\dots,\Phi_i(x_{i,p})]$, makes this sum at least $q_i I_m$, so the parameter error produces its own exponential decay. The graph Laplacian supplies the rate $\gamma = \lambda_2(L)/C$ through the algebraic connectivity. For the quantized result, the analysis switches to Krasovskii solutions, and the quantization disturbance, bounded by $\sigma$, is absorbed into an exponentially stable comparison system, yielding the $O(\sigma)$ steady-state bound.

What would settle it

Simulate the proposed closed loop with a single unknown scalar parameter per agent and store only the signals an agent actually measures (state, control, and past inputs). The update (2.8) requires $\tilde{\theta}_i=\hat{\theta}_i-\theta_i$, so the concurrent-learning sum $\sum_k \Phi_i(x_{i,k})^T\Phi_i(x_{i,k})\tilde{\theta}_i$ contains the unknown true $\theta_i$, and no rule in the paper tells the agent how to evaluate it. Alternatively, exhibit a history stack $Z_i=[\Phi_i(x_{i,1}),\dots,\Phi_i(x_{i,p})]$ of rank $m-1$; then $q=0$ in (3.4) and the claimed exponential parameter convergence fails.

Watch

Extended reading notes

Core claim

The central claim is Theorem 2.2: with the parameter update $\dot{\hat{\theta}}_i=\Phi_i^T B^T P \sum_j a_{ij}(x_i-x_j) - \sum_{k=1}^r \Phi_i(x_{i,k})^T \Phi_i(x_{i,k}) \tilde{\theta}_i$, where $\tilde{\theta}_i=\hat{\theta}_i-\theta_i$, the closed loop is exponentially stable. Both the consensus error $x^T(L\otimes P)x$ and the total parameter error $\sum_i \|\hat{\theta}_i-\theta_i\|^2$ decay at rate $\min\{\gamma,q\}$. The proof is a Lyapunov argument in which the concurrent-learning sum contributes a negative definite term proportional to the parameter error, provided the stored regressor matrix has full rank. Theorem 4.1 extends the same guarantee to communication through a uniform quantizer of level $\sigma$, giving exponential convergence to a neighborhood of consensus whose size is $O(\sigma)$.

Load-bearing premise

The load-bearing premise is that the concurrent-learning term in (2.8), built from $\tilde{\theta}_i=\hat{\theta}_i-\theta_i$, can be computed by an agent that does not know $\theta_i$; the paper never states the recorded measurement that would make this term available, and if no computable form exists the closed loop cannot be run as written. A second load-bearing premise is Condition 1, full rank of the stored regressor stack, because without it the parameter-decay rate $q$ is zero.

Editorial extensions

If this is right

  • The controller from prior asymptotic consensus work, upgraded with concurrent learning, gives exponential consensus and exponential parameter convergence without persistent excitation.
  • The convergence rate is explicit, so a designer can choose the graph and the stored-history stack to meet a required settling time.
  • Uniform quantization does not destroy exponential stability; the consensus error scales linearly with the quantization step $\sigma$.
  • Condition 1 replaces a persistent-excitation requirement with a finite-history rank condition, shifting the burden from online signal richness to offline data richness.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • There is an implementability gap in the stated update (2.8): it uses $\tilde{\theta}_i=\hat{\theta}_i-\theta_i$, and since $\theta_i$ is unknown, the paper would need to specify a recorded quantity, such as a stored model-error signal, that makes the concurrent-learning term computable; no such quantity is defined.
  • If that gap is closed in the standard concurrent-learning way, the exponential rate $\min\{\gamma,q\}$ suggests that graph topology bounds speed independent of how much history is stored, while the history stack alone determines the parameter-decay contribution.
  • The quantized result points toward a direct bit-rate versus steady-state-error trade-off, and the same Lyapunov machinery could likely handle logarithmic quantizers or event-triggered communication.
  • Condition 1 can be checked offline and is weaker than persistent excitation, so it could be monitored online through the smallest singular value of the stored regressor stack.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper studies leaderless consensus for a network of agents with identical linear dynamics and matched, state-dependent nonlinear uncertainties, building on the adaptive consensus controller of [3]. The authors add a concurrent-learning term to the parameter update law and claim, in Theorem 2.2, that the resulting closed-loop system achieves exponential consensus and exponential parameter convergence without persistent excitation, relying only on a rank condition on stored regressor data (Condition 1). Theorem 4.1 extends the claim to uniform quantized inter-agent communication, asserting exponential convergence of the consensus error up to an O(σ) bound. The proofs use a Lyapunov function combining a graph-weighted state term and parameter-error terms, with an algebraic Riccati equation cancellation; the quantized proof is an ISS-type argument with a Krasovskii solution. Numerical experiments for five agents with and without quantization are reported for σ = 5, 10, 15.

Significance. If the proposed update law were realizable, the result would be a useful and clean improvement: it replaces the asymptotic consensus of [3] with exponential rates that are explicit in terms of λ2(L), the ARE solution P, and the rank condition constant q, and it gives a transparent O(σ) quantized-error bound. The Lyapunov algebra in Section 3 is standard and the ARE cancellation in (3.2) checks out; the rate is not fitted, and the σ-scaling experiment is an appropriate falsifiable check. However, the central implementability gap described below means that, as written, the theorems are statements about an oracle system rather than a controller that an agent can actually execute, and the simulations do not exercise the proposed law. The contribution is therefore not yet established, but it is plausibly salvageable with a standard reformulation of the concurrent-learning term.

major comments (4)
  1. [§2.2, Eq. (2.8) and §4, Eq. (4.4)] The update law (2.8) and its quantized counterpart (4.4) are written with the term −Σ_k Φ_i^T(x_{i,k}) Φ_i(x_{i,k}) θ̃_i(t), where θ̃_i = θ̂_i − θ_i. Since θ_i is the unknown true parameter, this term is not computable from measurements available to agent i. The paper never defines a stored recorded output y_{i,k} such that the concurrent-learning term can be rewritten as −Σ_k Φ_i^T(x_{i,k})(Φ_i(x_{i,k})θ̂_i − y_{i,k}), which is the standard form used in concurrent learning [5]. Consequently, Theorems 2.2 and 4.1 currently prove exponential decay for an oracle differential equation whose right-hand side depends on the unknown θ_i, not for a realizable adaptive controller. This is load-bearing and must be fixed by explicitly defining the recorded data and rewriting the update law in a computable form.
  2. [§2.2, Condition 1] Condition 1 is stated as 'For each i ∈ {1,2,...,m}' although the agents are indexed by n, and the stored regressor matrix is written as Z_i = [Φ(x_{i,1}),...,Φ(x_{i,p})] without the agent subscript i on Φ; dimension compatibility of the stack (q p × m) is not stated. In addition, Eq. (2.7) sums over r stored points, while Condition 1 and the proof of Theorem 2.2 use p points, and the relation between r and p is never specified. Since the positivity of q in (3.4), and hence the exponential parameter-estimate decay in V̇, depends on exactly this condition, the theorem statement is not well posed as written. These details need to be corrected and reconciled.
  3. [§4, Eqs. (4.1)-(4.7)] The Krasovskii solution framework is introduced for the state x only, but for Theorem 4.1 the closed loop also includes the parameter update (4.4), which itself contains the discontinuous map x ↦ q_u(x_i). The proof asserts that 'this equation admits a Krasovskii solution' and then works with selections v_i(t) ∈ K[Δ_i](t) while θ̂ evolves simultaneously. A rigorous proof of Theorem 4.1 must define the augmented state (x, θ̂), verify local boundedness and measurability of the combined right-hand side, and establish existence of a Krasovskii solution of the augmented system before the Lyapunov differentiation is performed.
  4. [§5, Simulations] The non-quantized simulation is described as 'We simulate (2.2) involving the control given by (2.3)', i.e., the baseline adaptive controller from [3], not the proposed concurrent-learning update (2.8) that Theorem 2.2 concerns. The simulation therefore cannot validate Theorem 2.2 as stated. In addition, the paper gives no history stack: the number of stored points per agent, their locations, and any verification of Condition 1 are not reported, so the experiment is not reproducible. The quantized experiments similarly do not state which update law was actually implemented.
minor comments (4)
  1. [§4, Eq. (4.17)] Equation (4.17) claims a lower bound of the form x^T{(L⊗I) + [(2αL^2 − L)⊗PBB^T P]}x ≥ γ x^T(L^2⊗PBB^T P)x, but the subsequent estimate (4.20) uses a lower bound on x^T(L⊗I)x. As written, (4.17) is not justified in general and appears to be a typo; it should likely be x^T(L⊗I)x ≥ γ x^T(L⊗P)x, as in (3.6)-(3.7).
  2. [§2.2, Theorem 2.2 and §4, Theorem 4.1] The theorem statements say the 'dynamics (2.2) is exponentially stable,' but the Lyapunov function uses L⊗P and the state x does not converge to zero; the correct claim is exponential consensus, i.e., exponential convergence of the consensus error x − 1⊗x̄, together with exponential parameter error convergence.
  3. [§2 and §5] The symbol q is used both for the control input dimension in (2.2) and for the lower bound in (3.5), and in Section 5 q is initially set to 2 and later to 4; m is also reused for the parameter dimension and for the number of agents in Condition 1. These clashes should be resolved with distinct notation.
  4. [Throughout] There are many typographical and grammatical errors, including 'acheived', 'appled', 'Barbalet’s lemma', and the title 'DECA Y'; the reference list also has inconsistent formatting. These should be corrected in a revision.

Circularity Check

1 steps flagged · score 6.0 of 10

The proposed update law (2.8)/(4.4) contains the unknown true parameter theta_i through theta_tilde_i, so the 'adaptive' controller reduces by construction to an oracle controller and the exponential-consensus claim is not established for an implementable protocol.

  1. self definitional [Section 2.2, Eqs. (2.7) and (2.8); also Eq. (4.4) in Section 4.]
    "Precisely, we propose the following update rule for the uncertain variable: dot(theta_hat_i)(t) = Phi_i^T B^T P sum_j a_ij(x_i(t) - x_j(t)) - sum_k Phi_i^T(x_i,k) Phi_i(x_i,k) theta_tilde_i(t), (2.7) where theta_tilde_i(t) = theta_hat_i(t) - theta_i."

    The update rule for the estimate theta_hat_i is defined in terms of the unknown true parameter theta_i through theta_tilde_i = theta_hat_i - theta_i. Since theta_i is exactly the quantity the adaptive loop is supposed to learn, the right-hand side of (2.7)/(2.8) is not computable from measurements available to agent i. No recorded-data equivalent (e.g., the stored model prediction error used in standard concurrent learning [5]) is defined anywhere in the paper; Condition 1 constrains only the regressor stack, not a computable error signal.

full rationale

The exponential-stability analysis itself is a standard, self-contained Lyapunov argument: the decay rate gamma comes from the algebraic connectivity and the ARE solution P, and q comes from Condition 1's rank assumption; neither is fitted to the simulations. The sigma = {5, 10, 15} runs are an external, falsifiable check of the O(sigma) error bound, not an in-sample fit. There is no load-bearing self-citation: reference [3] supplies the baseline controller and [5] supplies concurrent learning, both external works. The single circular step is that the proposed update law (2.8) is written with theta_tilde_i = theta_hat_i - theta_i, so its right-hand side contains the unknown true parameter; no recorded-data error term is defined to make it computable. Hence the 'adaptive' controller reduces by construction to an oracle controller, and Theorems 2.2 and 4.1 prove convergence for that oracle system only. This is the load-bearing connection from 'proposed controller' to 'exponential consensus', so despite the otherwise independent mathematics, the central claim as stated is partially circular (self-definitional).

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard adaptive-control and concurrent-learning premises: stabilizability of (A,B), a connected undirected graph, bounded continuous matched uncertainty Phi_i, the history-stack rank Condition 1, and, implicitly, that the concurrent-learning term is computable without knowing the true parameter theta_i (the paper never states how). All convergence-rate constants are derived from lambda_2(L), the ARE solution P, and the excitation constant q from Condition 1; they are not fitted to the simulations. No invented entities are introduced.

free parameters (2)
  • control gain alpha = alpha = 0.8019 (simulation)
    Design gain. The proof requires alpha >= 1/(2 lambda_2(L)) so that 2 alpha L^2 - L is PSD (paragraph preceding Eq. (3.6)). The simulation value is chosen in that admissible range, not fitted to achieve the proven rate.
  • history stack (stored points x_i,k; count r vs p) = unspecified (r in (2.7), p in Condition 1 and (3.4))
    The number and selection of stored data points are free choices. The proof needs sum_{k=1}^p Phi_i^T Phi_i >= q_i I_m, i.e., Condition 1. The paper gives no recording procedure and inconsistently writes r versus p, so the proof's stack size may not match the implemented one.
assumptions (5)
  • domain assumption Assumption 1: (A,B) is stabilizable
    Needed for the Algebraic Riccati Equation (2.4) to have a stabilizing positive solution P; invoked in Section 2.2.
  • domain assumption The interaction graph is undirected and connected
    Used so that the Laplacian L has a simple zero eigenvalue, lambda_2 = alpha(L) > 0 appears in the decay rate, and x^T (L tensor I_p) x is a norm of the disagreement; Section 2.1.
  • domain assumption Phi_i is bounded and continuous, and the uncertainty enters matched as B Phi_i theta_i
    The matched-uncertainty structure (2.2) is inherited from [3] and is the premise of the whole problem; boundedness also supports well-posedness of the closed loop.
  • domain assumption Condition 1: rank[Phi(x_i,1), ..., Phi(x_i,p)] = m for the stored data of each agent
    Standard concurrent-learning rank (finite-excitation) condition; it makes sum Phi^T Phi positive definite with constant q_i > 0, giving exponential parameter error decay (Eqs. (3.4) and (3.5)).
  • standard math Existence and continuation of Krasovskii solutions for the quantized closed loop
    Invoked in Section 4 (Eq. (4.2) and following) to handle the discontinuity of the uniform quantizer; measurability and local boundedness of the right-hand side are used as the existence criterion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Consensus with Exponential Decay." pith.science (2026). https://pith.science/paper/QEO4KOUM

@misc{pith2026250607203,
  author       = {Pith},
  title        = {Pith review of: Adaptive Consensus with Exponential Decay},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QEO4KOUM}},
  note         = {Machine review of arXiv:2506.07203}
}
read the original abstract

This paper addresses the adaptive consensus problem in uncertain multi-agent systems, particularly under challenges posed by quantized communication. We consider agents with general linear dynamics subject to nonlinear uncertainties and propose an adaptive consensus control framework that integrates concurrent learning. Unlike traditional methods relying solely on instantaneous data, concurrent learning leverages stored historical data to enhance parameter estimation without requiring persistent excitation. We establish that the proposed controller ensures exponential convergence of both consensus and parameter estimation. Furthermore, we extend the analysis to scenarios where inter-agent communication is quantized using a uniform quantizer. We prove that the system still achieves consensus up to an error proportional to the quantization level, with exponential convergence rate.

Figures

Figures reproduced from arXiv: 2506.07203 by the authors.

Figure 1
Figure 1. Undirected communication graph We use the same Laplacian matrix and the nonlinear function Φ as in [19]. Specifically, the Laplacian matrix L ∈ R5×5 is given by L =   2.168 −1.037 0 −0.865 −0.266 −1.037 1.037 0 0 0 0 0 1.651 −1.651 0 −0.865 0 −1.651 2.863 −0.347 −0.266 0 0 −0.347 0.613   , and the solution of Algebraic Riccati Equation P ∈ R4×4 is P =   2.8917 −0.3741 −1.8010 1.2765 −0.3741 0.52… view at source ↗
Figure 2
Figure 2. Consensus error of states and parameter estimation [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Consensus of xi,1 and xi,3. Next we test the control (2.12) involving the quantization operator. We set σ = 5 for the quantization level in (4.3). The graph of consensus error and the convergence of parameters to references with quantization are presented in [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Consensus error and parameter estimation with quantization σ = 5 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Consensus of xi,1 and xi,3 with quantization σ = 5 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Consensus error and parameter estimation with quantization σ = 10 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Consensus error and parameter estimation with quantization σ = 15. 6. Conclusion In this study, we demonstrated that by introducing the concurrent learning technique, it is possible to achieve exponential convergence in adaptive consensus algorithms. By leveraging hist…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

  1. [3]

    &Song, Y

    Mei, J., Ren, W. &Song, Y. (2021). A unified framework for adaptive leaderless consensus of uncertain multiagent systems under directed graphs. IEEE Transactions on Automatic Control, 66(12), 6179- 6186

  2. [5]

    & Johnson, E

    Chowdhary, G. & Johnson, E. (2010, December). Concurrent learning for convergence in adaptive control without persistency of excitation. In 49th IEEE Conference on Decision and Control (CDC) (pp. 3674-3679). IEEE

  3. [1]

    J., & Slotine, J

    Chung, S. J., & Slotine, J. J. E. (2009). Cooperative robot control and concurrent synchronization of Lagrangian systems. IEEE transactions on Robotics, 25(3), 686-700

  4. [2]

    & Cao, J

    Yue, D., Baldi, S. & Cao, J. (2022, August). Robust Model Reference Adaptive Consensus with Neural Networks. In 2022 34th Chinese Control and Decision Conference (CCDC) (pp. 2503-2508). IEEE

  5. [4]

    Yang, Y., Mei, J. & Ma, G. (2025). Adaptive leaderless consensus of MIMO multi-agent systems with unknown dynamics and nonlinear dynamic uncertainties. Systems and Control Letters, 196, 105983

  6. [6]

    & De Schutter, B

    Yue, D., Baldi, S., Cao, J. & De Schutter, B. (2023). Model reference adaptive stabilizing control with application to leaderless consensus. IEEE Transactions on Automatic Control, 69(3), 2052-2059

  7. [7]

    & Ren, W

    Yuan, C., Stegagno, P., He, H. & Ren, W. (2021). Cooperative adaptive containment control with pa- rameter convergence via cooperative finite-time excitation. IEEE Transactions on Automatic Control, 66(11), 5612-5618

  8. [8]

    & Wen, C

    Chen, W., Li, X., Ren, W. & Wen, C. (2013). Adaptive consensus of multi-agent systems with unknown identical control directions based on a novel Nussbaum-type function. IEEE Transactions on Automatic Control, 59(7), 1887-1892

Show all 19 references
  1. [9]

    G., Cheng, L

    Hou, Z. G., Cheng, L. & Tan, M. (2009). Decentralized robust adaptive control for the multiagent sys- tem consensus problem using neural networks. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(3), 636-647

  2. [10]

    P., Wen, G

    Chen, C. P., Wen, G. X., Liu, Y. J. & Wang, F. Y. (2014). Adaptive consensus control for a class of nonlinear multiagent time-delay systems using neural networks. IEEE Transactions on Neural Net- works and Learning Systems, 25(6), 1217-1226

  3. [11]

    & Wei, Q

    Ma, H., Wang, Z., Wang, D., Liu, D., Yan, P. & Wei, Q. (2015). Neural-network-based distributed adaptive robust control for a class of nonlinear multiagent systems with time delays and external noises. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 46(6), 750-758

  4. [12]

    Psillakis, H. E. (2019). PI consensus error transformation for adaptive cooperative control of nonlinear multi-agent systems. Journal of the Franklin Institute, 356(18), 11581-11604

  5. [13]

    & Huang, J

    Cai, H. & Huang, J. (2015). The leader-following consensus for multiple uncertain Euler-Lagrange systems with an adaptive distributed observer. IEEE Transactions on Automatic Control, 61(10), 3152-3157

  6. [14]

    & Ahn, C

    Zhu, Y., Wang, Z., Liang, H. & Ahn, C. K. (2023). Neural-network-based predefined-time adaptive consensus in nonlinear multi-agent systems with switching topologies. IEEE Transactions on Neural Networks and Learning Systems

  7. [15]

    & Zhou, J

    Zou, W. & Zhou, J. (2023). A novel neural-network-based consensus protocol of nonlinear multiagent systems. IEEE Transactions on Automatic Control, 69(3), 1713-1720

  8. [16]

    & Tempo, R

    Ishii, H. & Tempo, R. (2014). The PageRank problem, multiagent consensus, and web aggregation: A systems and control viewpoint. IEEE Control Systems Magazine, 34(3), 34-53

  9. [17]

    & Murray, R

    Lavaei, J. & Murray, R. M. (2011). Quantized consensus by means of gossip algorithm. IEEE Trans- actions on Automatic Control, 57(1), 19-32

  10. [18]

    & Wang, L

    Zhang, Z., Zhang, L., Hao, F. & Wang, L. (2016). Leader-following consensus for linear and Lipschitz nonlinear multiagent systems with quantized communication. IEEE transactions on cybernetics, 47(8), 1970-1982. 15

  11. [19]

    W., Li, L., & Liu, M

    Chen, S., Ho, D. W., Li, L., & Liu, M. (2014). Fault-tolerant consensus of multi-agent system with distributed adaptive protocol. IEEE transactions on cybernetics, 45(10), 2142-2155. Email address: choiwc@skku.edu

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.