--- title: Consensus Error-Dependent Laplace Noise url: https://www.emergentmind.com/topics/consensus-error-dependent-laplace-noise type: topic --- # Consensus Error-Dependent Laplace Noise Searching arXiv for the exact paper and closely related consensus/privacy/noise papers to ground the article in current arXiv records. Consensus error-dependent Laplace noise is a privacy mechanism for networked consensus control in which the Laplace scale is tied to a local disagreement signal rather than fixed a priori. In the arXiv literature summarized here, its most direct formulation appears in the distributed finite-horizon LQR framework of "Distributed Finite-Horizon Optimal Control for Consensus with Differential Privacy Guarantees" [2509.11917]. There, the mechanism perturbs communicated state information with Laplace noise whose magnitude depends on the current local consensus mismatch, while a time-dependent cost scaling is introduced so that the closed-loop system preserves differential privacy of the local LQR weighting matrices and still achieves bounded consensus. Closely related literatures study decaying Laplace noise, fixed-design Laplace initialization noise, or state-dependent Brownian diffusion, but those mechanisms are distinct in both noise law and adaptation rule [1512.09039], [2309.08464], [1602.06467]. ## 1. Conceptual scope and distinguishing features In the formulation of [2509.11917], the multi-agent system has discrete-time single-integrator dynamics \[ x_i(t+1)=x_i(t)+u_i(t), \qquad i=1,2,\ldots,N,\; t\in\mathbb{N}, \] with \(x_i(t)\in \mathbb{R}^n\) and \(u_i(t)\in \mathbb{R}^n\). The agents communicate over a directed graph \[ \mathcal{G}(\mathcal{V},\mathcal{E},\mathcal{A}), \] with adjacency matrix \(\mathcal{A}=[a_{ij}]\), in-neighbor set \[ \mathcal{N}_i=\{j\in\mathcal{V}\mid (j,i)\in\mathcal{E}\}, \] and Laplacian \(\mathcal{L}=[l_{ij}]\). The standing graph condition is that \(\mathcal{G}\) contains a directed spanning tree [2509.11917]. The expression “consensus error-dependent Laplace noise” refers here to a mechanism in which the Laplace scale is proportional to a local consensus mismatch term, \[ \left\|\sum_{j\in\mathcal{N}_i} a_{ij}\bigl[\hat{x}_j(t)-x_i(t)\bigr]\right\|_\infty, \] rather than to a fixed initialization variance or to an exogenous time-decay schedule [2509.11917]. The noisy communicated state is \[ \hat{x}_j(t)\triangleq x_j(t)+\eta_j(t), \] and the local mismatch is the quantity through which the current state of disagreement enters the privacy mechanism. This construction differs sharply from several adjacent mechanisms. In "Differentially Private Average Consensus: Obstructions, Trade-Offs, and Optimal Algorithm Design" [1512.09039], the Laplace scale is purely time-dependent, \[ b_i(k)=c_i q_i^k, \] with exponentially decaying Laplace noise and no dependence on the realized consensus mismatch. In "Differentially Private Consensus for Time-Delay Multi-agent Systems" [2606.15135], the Laplace scale is also time-varying only and may even increase with time, but it is designed from delayed-history sensitivity rather than from the current disagreement. In "Observer-based Differentially Private Consensus for Linear Multi-agent Systems" [2512.16736], the Laplace schedule is exogenous, \[ b_i(k)=c_i p_i(k),\qquad \sum_{k=0}^\infty p_i(k)<\infty, \] or \[ b_i(k)=c_i g_i^k,\qquad 00\): \[ \begin{aligned} J_i(x(t),c(t)) &\triangleq \sum_{k=0}^{T-1}\Biggl( u_i^\top(t+k)R_i u_i(t+k) \\ &\qquad + c(t)\sum_{j\in \mathcal{N}_i} a_{ij} \bigl(x_i(t+k)-\hat{x}_j(t)\bigr)^\top Q_i \bigl(x_i(t+k)-\hat{x}_j(t)\bigr) \Biggr). \end{aligned} \tag{1} \] Here \[ \hat{x}_j(t)\triangleq x_j(t)+\eta_j(t). \tag{2} \] The resulting receding-horizon distributed control law is \[ u_i(t)= c(t)K_{Q_i,R_i,t}\sum_{j\in\mathcal{N}_i} a_{ij}\bigl(\hat{x}_j(t)-x_i(t)\bigr), \tag{3} \] where \(K_{Q_i,R_i,t}\) is the finite-horizon LQR gain determined by \(Q_i,R_i\), the horizon \(T\), and the finite-horizon recursion [2509.11917]. The closed-loop dynamics are then \[ x_i(t+1)=x_i(t)+ c(t)\sum_{j\in\mathcal{N}_i} a_{ij}K_{Q_i,R_i,t}\bigl(\hat{x}_j(t)-x_i(t)\bigr). \tag{4} \] This architecture is essential to the meaning of consensus error-dependent Laplace noise in [2509.11917]. The gain sensitivity with respect to \((Q_i,R_i)\) determines how a private change enters the dynamics, while the local mismatch term determines the magnitude of the noisy perturbation. ## 3. Error-dependent Laplace mechanism The network-level disagreement operator is \[ P\triangleq I_N-\frac{1}{N}\mathbf{1}\mathbf{1}^\top, \] and with \[ X(t)= \begin{bmatrix} x_1^\top(t) & x_2^\top(t) & \cdots & x_N^\top(t) \end{bmatrix}^\top, \] the consensus error is \[ \delta(t)\triangleq (P\otimes I_n)X(t). \tag{5} \] Its squared norm \[ V(t)=\delta^\top(t)\delta(t) \] is the disagreement energy, \[ V(t)=\sum_{i=1}^N \left\|x_i(t)-\frac{1}{N}\sum_{k=1}^N x_k(t)\right\|^2. \] The paper’s noise law is local rather than global, but this consensus-error formalism clarifies what “error-dependent” means [2509.11917]. The sensitivity of the finite-horizon LQR gain is defined by \[ \begin{aligned} \triangle_{i,t} &\triangleq \max_{W,W'} \Bigl\{ \|K_{Q_i,R_i,t}-K_{Q'_i,R'_i,t}\|_\infty : (Q_i,R_i)\in W, \\ &\hspace{27mm} (Q'_i,R'_i)\in W', \; W \text{ and } W' \text{ satisfy Definition 1} \Bigr\}. \end{aligned} \tag{6} \] This quantity is bounded when the feasible sets \(\mathcal{Q}_i,\mathcal{R}_i\) are bounded [2509.11917]. The Laplace mechanism is then defined by \[ \eta_i(t+1)\sim \operatorname{Lap}\!\left( \frac{n\,c(t)\,p(t)\,\triangle_{i,t}\, \left\|\sum_{j\in\mathcal{N}_i} a_{ij}\bigl[\hat{x}_j(t)-x_i(t)\bigr]\right\|_\infty}{\epsilon} \right), \qquad t\in\mathbb{N}, \tag{7} \] with \[ \eta_i(0)=0_n. \] Equivalently, each component of \(\eta_i(t+1)\) is sampled i.i.d. from a zero-mean Laplace law with scale \[ b_{i,t+1} = \frac{n\,c(t)\,p(t)\,\triangle_{i,t}\, \left\|\sum_{j\in\mathcal{N}_i} a_{ij}\bigl[\hat{x}_j(t)-x_i(t)\bigr]\right\|_\infty}{\epsilon}. \tag{8} \] This formula exhibits four coupled dependencies already stated in [2509.11917]: the time-varying state/cost scaling \(c(t)\), the auxiliary design sequence \(p(t)\), the gain sensitivity \(\triangle_{i,t}\), and the current local consensus mismatch. The mechanism is therefore simultaneously time-varying, agent-specific, privacy-budget-dependent, and disagreement-dependent. A common misconception is to equate any decaying Laplace mechanism with an error-dependent one. The literature summarized here does not support that identification. In [1512.09039], the scale is fixed by \(c_i q_i^k\), not by realized disagreement. In [2512.16736], the scale is \(b_i(k)=c_i p_i(k)\) or \(c_i g_i^k\), again not a function of \(\delta(k)\). In [2606.15135], the scale may increase with time, \[ b_j(k)=\underline b (k+a_2)^\gamma, \] but still remains time-index-dependent rather than disagreement-dependent. The precise hallmark of the mechanism in [2509.11917] is the appearance of the local consensus mismatch directly inside the Laplace scale. ## 4. Differential privacy of local weight matrices The adjacency relation in [2509.11917] is defined on the database \[ W=\{(Q_i,R_i)\}_{i=1}^N \] after canonical normalization of each pair. The text states that privacy is defined for the pair \((Q_i,R_i)\) up to a normalization because the LQR gain is scale-invariant with respect to simultaneous scaling of \(Q_i\) and \(R_i\) [2509.11917]. The mechanism satisfies the standard definition of \(\epsilon\)-differential privacy: \[ \mathbb{P}\bigl(\hat{\mathcal{M}(W)\in \mathcal{S}\bigr) \le e^\epsilon \mathbb{P}\bigl(\hat{\mathcal{M}(W')\in \mathcal{S}\bigr), \tag{9} \] for adjacent \(W,W'\) and measurable \(\mathcal{S}\). At time \(t\), the relevant mechanism output is \[ \mathcal{M}_t(W)= \begin{bmatrix} x_1^\top(t+1) & x_2^\top(t+1) & \cdots & x_N^\top(t+1) \end{bmatrix}^\top. \] If \(W\) and \(W'\) differ only in agent \(i\)’s weight pair, the sensitivity of \(\mathcal{M}_t\) is bounded by \[ \max_{W,W'} \|\mathcal{M}_t(W)-\mathcal{M}_t(W')\|_\infty \le c(t)\left\|\sum_{j\in\mathcal{N}_i} a_{ij}\bigl[\hat{x}_j(t)-x_i(t)\bigr]\right\|_\infty \triangle_{i,t}. \tag{10} \] The scale in the Laplace law is calibrated exactly to this quantity, with the factor \(n p(t)/\epsilon\), and the paper invokes the standard Laplace mechanism lemma to obtain per-step privacy [2509.11917]. Under the summability condition \[ \sum_{t=0}^\infty \frac{1}{p(t)}<\infty, \tag{11} \] Theorem 1 of [2509.11917] states that \(\mathcal{M}_t\) preserves \[ \frac{\epsilon}{p(t)} \] -differential privacy at time \(t\), and the full infinite-horizon mechanism \[ \mathcal{M}=\{\mathcal{M}_0,\mathcal{M}_1,\mathcal{M}_2,\ldots\} \] preserves \[ \sum_{t=0}^\infty \frac{\epsilon}{p(t)} \] -differential privacy. The proof uses the Adaptive Sequential Composition Theorem because later states depend on earlier noisy states [2509.11917]. This privacy formulation is structurally different from earlier DP consensus papers. In [1512.09039], privacy concerns agents’ initial states and the global eavesdropper’s observation of the message trajectory. In [2309.08464], privacy also targets initial data and uses a correlated-zero-sum initialization mechanism, with a dedicated Laplace-noise design at the initialization stage rather than continual message perturbation. By contrast, [2509.11917] uses continual state perturbation to hide local LQR preference matrices rather than initial conditions. ## 5. Consensus dynamics, second moments, and privacy–performance trade-off With stacked variables \[ \eta(t)= \begin{bmatrix} \eta_1^\top(t) & \eta_2^\top(t) & \cdots & \eta_N^\top(t) \end{bmatrix}^\top, \] block-diagonal gain matrix \[ K_t= \operatorname{diag}\{K_{Q_1,R_1,t},\ldots,K_{Q_N,R_N,t}\}, \] and shorthand \[ \bar{\mathcal{L}}=\mathcal{L}\otimes I_n,\qquad \bar{\mathcal{A}}=\mathcal{A}\otimes I_n,\qquad \bar{P}=P\otimes I_n, \] the closed-loop network dynamics are \[ X(t+1)=\left(I_{nN}-c(t)K_t\bar{\mathcal{L}}\right)X(t)+c(t)K_t\bar{\mathcal{A}}\eta(t), \tag{12} \] and the disagreement dynamics are \[ \delta(t+1)=\left(I_{nN}-c(t)\bar{P}K_t\bar{\mathcal{L}}\right)\delta(t) +c(t)\bar{P}K_t\bar{\mathcal{A}}\eta(t). \tag{13} \] The Lyapunov function is \[ V(t)\triangleq \delta^\top(t)\delta(t). \tag{14} \] A key technical constant is \[ \hat{\triangle}\triangleq \max_{i=1,\ldots,N}\hat{\triangle}_i, \qquad \hat{\triangle}_i\triangleq \sup_{t\ge 0}\triangle_{i,t}, \] together with \[ \bar{\triangle}\triangleq \max_i \bar{\triangle}_i,\qquad \bar{\triangle}_i\triangleq \sup_{t\ge 0}\|K_{Q_i,R_i,t}\|. \] The privacy-dependent coefficient is \[ \kappa_\epsilon \triangleq \frac{4n^3N(N-1)\hat{\triangle}^2}{\epsilon^2}. \tag{15} \] This formula makes explicit that the privacy penalty scales like \(1/\epsilon^2\) [2509.11917]. Because a Laplace variable with scale \(b\) has second moment \(2b^2\), the paper derives the per-agent noise-energy identity \[ \begin{aligned} \mathbb{E}\bigl[\eta_i^\top(t)\eta_i(t)\bigr] &= \frac{2n^3 c^2(t-1)p^2(t-1)}{\epsilon^2} \hat{\triangle}_{i,t}^2 \mathbb{E}\left[ \left\|\sum_{j\in\mathcal{N}_i} a_{ij}\bigl[\hat{x}_j(t-1)-x_i(t-1)\bigr]\right\|_\infty^2 \right]. \end{aligned} \tag{16} \] The local mismatch term is then bounded by the previous global disagreement and previous noise energy, leading to \[ \mathbb{E}[\eta^\top(t)\eta(t)] \le \kappa_\epsilon c^2(t-1)p^2(t-1) \left[ 2(N-1)\mathbb{E}[V(t-1)] +\mathbb{E}[\eta^\top(t-1)\eta(t-1)] \right]. \tag{17} \] The sufficient design conditions used for bounded consensus are \[ \sum_{t=0}^\infty c^2(t)=C_1<\infty, \qquad \kappa_\epsilon c^2(t)p^2(t)\le 1,\quad \forall t, \tag{18} \] and \[ \sum_{t=1}^\infty t\,c^2(t)c^2(t-1)p^2(t-1)=C_2<\infty. \tag{19} \] Under these assumptions, Theorem 2 of [2509.11917] proves that \[ \mathbb{E}\bigl[\delta^\top(t)\delta(t)\bigr] = \mathbb{E}[V(t)] < \infty,\qquad \forall t\in\mathbb{N}, \] that is, bounded consensus in mean square. The asymptotic error analysis introduces \[ \rho_\lambda \triangleq \inf_{t\ge 0} \lambda_0\!\left( \bar{P}K_t\bar{\mathcal{L}} + (\bar{P}K_t\bar{\mathcal{L}})^\top \right), \tag{20} \] and assumes \[ c^2(t)\rho_{\mathcal{L}}\bar{\triangle}^2 - c(t)\rho_\lambda > -1, \qquad \forall t\in\mathbb{N}, \tag{21} \] where \[ \rho_{\mathcal{L}}=\|\bar{\mathcal{L}}\|^2,\qquad \rho_{\mathcal{A}}=\|\bar{\mathcal{A}}\|^2. \] The resulting asymptotic bound is \[ \begin{aligned} \sigma &= e^{\rho_{\mathcal{L}}\bar{\triangle}^2 C_1-\sum_{t=0}^\infty c(t)\rho_\lambda} \bigl(1+c^2(0)\rho_{\mathcal{L}}\bar{\triangle}^2-c(0)\rho_\lambda\bigr)V(1) \\ &\quad + 2(N-1)\rho_{\mathcal{A}}\kappa_\epsilon \bar{\triangle}^2 C_2 C_3 \bar{V}, \end{aligned} \tag{22} \] with \(C_3>0\) satisfying \[ \prod_{j=t}^\infty \bigl(1+c^2(j)\rho_{\mathcal{L}}\bar{\triangle}^2-c(j)\rho_\lambda\bigr)< C_3, \qquad \forall t\in\mathbb{N}. \tag{23} \] The formula isolates the privacy-dependent contribution through \(\kappa_\epsilon\), so stronger privacy enlarges the bound through its \(1/\epsilon^2\) dependence [2509.11917]. This trade-off is conceptually different from those in DP average-consensus papers protecting initial states. In [2309.08464], the asymptotic mean-square error of the Laplace design is \[ \lim_{t\to\infty}\mathbb E|x_i(t)-d^\ast|^2=\frac{2\sigma_\gamma^2}{n^2}, \] with privacy achieved by fixed-design initialization noise rather than state-adaptive message perturbation. In [1512.09039], the asymptotic variance of the consensus value is \[ \operatorname{var}\{\theta_\infty\} = \frac{2}{n^2}\sum_{i=1}^n \frac{s_i^2 c_i^2}{1-q_i^2}, \] and the optimal design collapses to one-shot perturbation of initial states. These are distinct privacy–accuracy regimes from the error-dependent LQR mechanism of [2509.11917]. ## 6. Relation to neighboring literatures and recurrent misconceptions The most important clarification is that consensus error-dependent Laplace noise is not synonymous with privacy-preserving consensus noise in general. Several nearby literatures must be separated carefully. First, error-dependent noise and Laplace noise do not generally coincide. "On noise-induced synchronization and consensus" [1602.06467] studies state-dependent multiplicative noise in Itô SDEs, including diffusion terms that vanish on the synchronization manifold and scale with disagreement, but the increments are Brownian rather than Laplace. The paper therefore informs the “consensus-error-dependent” part of the topic, but not the “Laplace” part. Second, Laplace privacy mechanisms in consensus are often not error-dependent. In [1512.09039], the Laplace mechanism is exponentially decaying in time and not a function of the current residual. In [2309.08464], Laplace noise is used only at initialization, through a correlated-zero-sum DiShuf stage and a single additional Laplace variable at a secure agent. In [2512.16736], observer-based MASs use exogenously scheduled decaying Laplace noise, and the paper explicitly does not study \(b_i(k)=f(\|\delta(k)\|)\). In [2606.15135], the delayed-consensus mechanism uses time-varying Laplace scales tied to explicit sensitivity bounds for delayed initial histories, and the scale may even increase with time, but it is not an explicit function of the current disagreement. Third, privacy-preserving consensus papers often use non-Laplace noise families altogether. "A Privacy Preserving Randomized Gossip Algorithm via Controlled Noise Insertion" [1901.09367] uses per-node Gaussian noise with deterministic geometric decay and inter-iteration correlation created by subtracting previously injected noise and adding a smaller fresh perturbation. Its noise depends on a node-local update counter rather than on consensus error, disagreement, or residual, and it does not formalize differential privacy. "On the Influence of Noise in Randomized Consensus Algorithms" [2006.05753] gives a variance-based asymptotic disagreement analysis for additive noise in randomized consensus; fixed-scale Laplace noise fits the second-moment framework through \(\operatorname{Var}(\eta)=2b^2\), but state-dependent Laplace noise falls outside the paper’s exact formulas without new arguments. Within this broader landscape, [2509.11917] occupies a specific niche. It combines a finite-horizon LQR consensus controller, a privacy objective on local weight matrices rather than states, and a Laplace scale that is explicitly proportional to a local disagreement term. This suggests a design pattern in which privacy calibration follows the actual pathway by which private parameters influence the control law: gain sensitivity enters through \(\triangle_{i,t}\), and the magnitude of the perturbation is modulated by the consensus mismatch that multiplies the gain in the control update. A plausible implication is that the mechanism suppresses unnecessary noise when local disagreement is already small, but the bounded-consensus guarantee in the paper is stated in terms of the explicit summability and contraction conditions on \(c(t)\), \(p(t)\), and the gain-dependent constants, not through a separate optimality theorem [2509.11917]. The simulation instance in [2509.11917] illustrates this mechanism on a 4-agent directed graph with \[ \epsilon=5,\qquad T=10,\qquad Q_i=8I_3,\qquad R_i=2I_3, \] and sequences \[ c(t)=\frac{1}{15(t+1)^{1.3}}, \qquad p(t)=(t+1)^{1.1}. \] The reported cumulative privacy leakage over \(t=0\) to \(120\) is \[ \sum_{t=0}^{120}\frac{\epsilon}{p(t)} \approx 21.9828, \] and the state trajectories approach a bounded consensus region under the proposed mechanism [2509.11917]. This example does not establish superiority over fixed-scale Laplace or Gaussian alternatives, but it does instantiate the theory’s central claim: consensus error-dependent Laplace noise can be embedded into a distributed finite-horizon optimal-control protocol so as to protect local weight matrices while preserving bounded consensus. Source: https://www.emergentmind.com/topics/consensus-error-dependent-laplace-noise