Spatial Consensus Filtering

• Spatial consensus filtering is a distributed estimation architecture that decomposes spatial domains, applies local filtering recursions, and reconciles overlapping state components.
• It leverages diverse methodologies—including finite-element, Kalman, and neural filtering approaches—to balance centralized optimality with decentralized efficiency.
• Empirical results show its effectiveness in applications such as PDE field estimation, traffic state monitoring, and multi-object tracking while ensuring stability and bounded convergence.

A spatial consensus filter is a distributed estimation architecture in which spatially indexed agents, overlapping subdomains, or neighboring sensing regions execute local filtering recursions and then apply a consensus mechanism to enforce agreement on shared state components, information variables, or posterior densities. In the cited literature, the expression does not designate a single canonical algorithm; rather, it covers a family of constructions for PDE-based field estimation, traffic-state estimation, multi-object tracking, extended-object tracking, partial-information consensus filtering, neural Kalman-like estimation, local average consensus, and consensus-acceleration graph filtering (Battistelli et al., 2016, Sun et al., 2016, Fantacci et al., 2015, Li et al., 2021, Jeon et al., 2021, Stamatelis et al., 26 Jun 2026, Cai et al., 2013, Kruzick et al., 2018).

1. Conceptual scope and canonical structure

The core pattern is spatial decomposition plus local recursion plus inter-node agreement. In the finite-element setting of Battistelli et al., a bounded domain $\Omega\subset\mathbb R^d$ is decomposed into possibly overlapping subdomains $\Omega_m$ assigned to different processing nodes; the communication graph satisfies $(j\to m)\in\mathcal L$ iff the interface $\Gamma_{mj}$ is nonempty (Battistelli et al., 2016). In large-scale traffic estimation, a freeway is partitioned into overlapping sections, each section is estimated by a local agent, and consensus is imposed on the overlapping cells (Sun et al., 2016). In distributed multi-object tracking, heterogeneous and geographically dispersed nodes exchange posterior parameters over a sensor network and fuse them through Kullback–Leibler averaging (Fantacci et al., 2015). In distributed information filtering, nodes exchange information vectors, information matrices, or innovation contributions, again through consensus iterations (Li et al., 2021, Jeon et al., 2021).

This architectural regularity appears across otherwise disparate model classes.

Instantiation	Local state/model	Consensus object
dFE–KF	FE degrees of freedom on $\Omega_m$	Boundary estimates and covariances
DLKCF	Density vector on a road section	Overlapping cell densities
CM$\delta$-GLMB / CLMB	Local multi-object posterior	KLA of local densities
CI / CM EOT IF	Local information pair or innovation deltas	Average consensus variables
ICF with partial exchange	Selected information entries	Partial consensus averages
CA-NKCF	Agent-wise latent-state estimate	Neighbor priors
Local average consensus	Spatial parameter samples	Local weighted averages

A plausible implication is that “spatial” refers less to Euclidean geometry alone than to any locality structure that induces overlapping support, nearest-neighbor coupling, or topology-constrained exchange.

2. State and observation models

In PDE-based field estimation, the state is a scalar field $x(p,t)$ governed by a linear parabolic PDE with homogeneous boundary conditions,

$$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$

For the heat equation, $A(x)=-\nabla\cdot(\lambda\nabla x)$ and $B(x)=\alpha \partial x/\partial n+\beta x$, with $\Omega_m$0, $\Omega_m$1, and $\Omega_m$2 on $\Omega_m$3. After finite-element discretization with basis $\Omega_m$4, the weak form yields

$\Omega_m$5

and backward Euler produces

$\Omega_m$6

with measurements $\Omega_m$7 for point sensors (Battistelli et al., 2016).

In large-scale traffic estimation, each section follows a 5-mode switching linear system derived from the Cell-Transmission Model. The section state is the density vector $\Omega_m$8, the dynamics use mode matrices $\Omega_m$9, and each agent runs a standard discrete-time Kalman filter with time-varying $(j\to m)\in\mathcal L$0 and $(j\to m)\in\mathcal L$1 (Sun et al., 2016).

In distributed multi-object tracking, the Random Finite Set formalism models the multi-object state as a finite set $(j\to m)\in\mathcal L$2 of labeled single-object states. The multi-object Bayes filter is written as

$(j\to m)\in\mathcal L$3

and consensus variants are built for M$(j\to m)\in\mathcal L$4-GLMB and LMB densities (Fantacci et al., 2015).

In distributed extended-object tracking, a multiplicative error model is split by moment matching into two linear models: one for the kinematic state $(j\to m)\in\mathcal L$5 and one for the extent state $(j\to m)\in\mathcal L$6. This enables alternating linear information-filter updates while preserving cross-correlation through the parameters in each other’s model (Li et al., 2021).

In covariance-agnostic neural filtering, the latent state obeys

$(j\to m)\in\mathcal L$7

or, in the linear special case,

$(j\to m)\in\mathcal L$8

with unknown noise covariances $(j\to m)\in\mathcal L$9 and $\Gamma_{mj}$0 (Stamatelis et al., 26 Jun 2026).

This range of formulations indicates that spatial consensus filtering is model-agnostic at the architectural level: the consensus layer is superimposed on linear PDE discretizations, switched conservation-law models, RFS Bayes filters, information filters, and neural surrogates.

3. Consensus operators

For overlapping physical subdomains, the consensus mechanism is often an explicit correction on shared states. In the decentralized consensus FE–Kalman filter, each node maintains a local FE state $\Gamma_{mj}$1 and uses a Parallel Schwarz predictor with one-step-delayed neighbor terms,

$\Gamma_{mj}$2

while boundary estimates are exchanged at each Schwarz step and covariances are propagated as

$\Gamma_{mj}$3

The factor $\Gamma_{mj}$4 boosts uncertainty to account for unknown cross-covariances between neighbors (Battistelli et al., 2016).

In the distributed local Kalman consensus filter, consensus is additive in the correction step:

$\Gamma_{mj}$5

with

$\Gamma_{mj}$6

Here the overlap operators $\Gamma_{mj}$7 select the shared cells, and the consensus term penalizes disagreement on overlaps (Sun et al., 2016).

For posterior-distribution fusion, consensus takes an averaging form in an appropriate statistical geometry. In labeled RFS filtering, each node holds a local multi-object density $\Gamma_{mj}$8 and fuses through the Kullback–Leibler average,

$\Gamma_{mj}$9

with nonnegative weights summing to one. Under a primitive, doubly-stochastic weight matrix, the iteration converges to the network-wide KLA (Fantacci et al., 2015).

In information-form consensus filtering, the consensus variables are information matrices and vectors or their innovation increments. For consensus on information, nodes iterate

$\Omega_m$0

whereas consensus on measurements averages the innovation contributions before a weighted local correction (Li et al., 2021). In the information-weighted consensus filter with partial information exchange, node $\Omega_m$1 transmits only entries selected by the diagonal matrix

$\Omega_m$2

and updates

$\Omega_m$3

with an analogous recursion for $\Omega_m$4 (Jeon et al., 2021).

In neural Kalman-like consensus, the posterior combines a local prediction, a learned measurement increment, and a single-step consensus correction:

$\Omega_m$5

where $\Omega_m$6 is produced by a GRU and $\Omega_m$7 defines element-wise consensus weights (Stamatelis et al., 26 Jun 2026).

A different but related construction appears in local average consensus. For exponentially weighted local averaging in a 1D sensor array,

$\Omega_m$8

so the consensus variable approximates a local, not global, average (Cai et al., 2013). At the algorithmic-acceleration level, graph filtering applies a polynomial $\Omega_m$9 to the consensus iteration matrix $\delta$0, and filter coefficients are chosen by minimizing the worst-case value of $\delta$1 over an approximated spectral region in the complex plane (Kruzick et al., 2018).

4. Stability, convergence, and boundedness

In the dFE–KF, the hybrid Euler–Schwarz discretization is first-order consistent and zero-stable if

$\delta$2

and, by Dahlquist’s Equivalence Theorem, this implies convergence of the Schwarz time march to the centralized FE solution as $\delta$3 and $\delta$4. Exponential stability of the distributed Kalman filter follows under observability of each local pair $\delta$5 and a covariance-inflation condition

$\delta$6

with $\delta$7 and $\delta$8 (Battistelli et al., 2016).

In the DLKCF, the mean error dynamics are globally asymptotically stable when all sections switch among observable modes and the consensus gain is sufficiently small. When a section remains in the unobservable modes FC1 or FC2, the paper proves ultimate boundedness of the mean estimate for each state variable, and for arbitrary switching sequences the overall mean error remains uniformly bounded provided each observable interval is long enough (Sun et al., 2016).

In consensus RFS filtering, convergence is to the global KLA of local posteriors, not necessarily to a centralized posterior generated by raw-measurement aggregation. The KLA itself minimizes a weighted sum of Kullback–Leibler divergences, and the authors emphasize that Chernoff fusion is immune to double counting when information is shared without bookkeeping of common histories (Fantacci et al., 2015).

For distributed extended-object tracking, the information-filter schemes satisfy

$\delta$9

under uniform bounds on model and filter quantities and a primitive doubly-stochastic consensus matrix. The proof uses a Lyapunov-type function weighted by local information matrices (Li et al., 2021). For the ICF with partial information exchange, mean-square boundedness and asymptotic convergence to the centralized Kalman filter are both established; the limiting consensus recovers exactly the centralized information update (Jeon et al., 2021).

In CA-NKCF, the consensus matrix is diagonal with entries $x(p,t)$0, and because $x(p,t)$1, each state component is updated by a convex combination of local and neighbor priors plus a learned measurement term. The paper states that this ensures stability of the consensus step (Stamatelis et al., 26 Jun 2026).

The local-average and graph-filter literatures frame convergence differently. Local average consensus is analyzed through spatial and temporal frequency responses rather than centralized-filter equivalence (Cai et al., 2013). Consensus-acceleration graph filters minimize a worst-case contraction factor over a complex-valued spectral region derived from Girko’s deterministic approximation of the empirical spectral distribution (Kruzick et al., 2018). This suggests that “stability” in spatial consensus filtering is not a single theorem but a spectrum of guarantees: Riccati-based exponential stability, Lyapunov GAS, boundedness under partial observability, consensus convergence in information space, and spectral contraction of the underlying iteration.

5. Algorithmic realization and communication economy

The finite-element implementation is explicitly offline/online. Offline, one generates a triangular FE mesh over $x(p,t)$2, computes the global mass and stiffness matrices $x(p,t)$3 and $x(p,t)$4, forms the sub-blocks for each $x(p,t)$5, and precomputes $x(p,t)$6, $x(p,t)$7, $x(p,t)$8, and $x(p,t)$9. Online, each node performs a local Kalman update at measurement times, initializes the Schwarz loop, exchanges boundary-state subvectors with neighbors, and propagates prediction and covariance. Communication is restricted to the boundary indices $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$0, and possibly to corresponding covariance blocks. Practical choices given in the paper are $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$1–$$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$2, $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$3 such that $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$4, and $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$5 tuned, for example, in the range $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$6–$$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$7 (Battistelli et al., 2016).

The DLKCF uses overlapping road sections and section-local measurement stacks. Each agent receives its own measurements and the neighbor-shared measurements collected into $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$8, and the consensus term involves only the shared cells extracted by $$\frac{\partial x}{\partial t}(p,t) + A(x)(p,t)=f(p,t), \qquad B(x)(p,t)=0, \qquad x(p,0)=x_0(p).$$9 (Sun et al., 2016).

Distributed multi-object consensus filters require exchanging posterior parameters rather than raw measurements. For M$A(x)=-\nabla\cdot(\lambda\nabla x)$0-GLMB, each node transmits hypothesis weights and single-track pdfs; for LMB, it transmits existence probabilities and single-track pdfs. Communication per consensus step is stated as $A(x)=-\nabla\cdot(\lambda\nabla x)$1 floats, where $A(x)=-\nabla\cdot(\lambda\nabla x)$2 is node in-degree and $A(x)=-\nabla\cdot(\lambda\nabla x)$3 is the number of hypotheses or labels, while local update cost is $A(x)=-\nabla\cdot(\lambda\nabla x)$4 plus $A(x)=-\nabla\cdot(\lambda\nabla x)$5 for fusion. Gaussian-mixture implementations use pruning and merging to control complexity (Fantacci et al., 2015).

In distributed extended-object tracking, the distinction between consensus on information and consensus on measurements is operationally important. CI exchanges full local information variables after local sequential updates, whereas CM exchanges only innovation contributions and then applies a weighted correction. Communication nodes set innovation deltas to zero in the CM scheme (Li et al., 2021).

Partial-information consensus filtering reduces bandwidth by transmitting only the selected entries $A(x)=-\nabla\cdot(\lambda\nabla x)$6 and $A(x)=-\nabla\cdot(\lambda\nabla x)$7. The paper states a direct trade-off: using fewer entries $A(x)=-\nabla\cdot(\lambda\nabla x)$8 per step reduces bandwidth by a factor $A(x)=-\nabla\cdot(\lambda\nabla x)$9 but slows convergence, so more consensus steps $B(x)=\alpha \partial x/\partial n+\beta x$0 may be required (Jeon et al., 2021).

Neural consensus filtering is lighter at runtime in one specific sense: agents exchange only priors $B(x)=\alpha \partial x/\partial n+\beta x$1, while the gain computation is delegated to a shared-parameter GRU driven by features

$B(x)=\alpha \partial x/\partial n+\beta x$2

The GRU outputs both $B(x)=\alpha \partial x/\partial n+\beta x$3 and the hidden state $B(x)=\alpha \partial x/\partial n+\beta x$4 (Stamatelis et al., 26 Jun 2026).

Consensus-acceleration graph filters shift much of the burden offline. The design requires a random graph model, the mean matrix $B(x)=\alpha \partial x/\partial n+\beta x$5, variances $B(x)=\alpha \partial x/\partial n+\beta x$6, numerical evaluation of Girko’s K25 approximation, sampling a spectral region $B(x)=\alpha \partial x/\partial n+\beta x$7, and solving a QCLP for the polynomial coefficients (Kruzick et al., 2018).

6. Applications, empirical behavior, and limitations

The dFE–KF is demonstrated on the estimation of a bi-dimensional temperature field, and the detailed notes refer to a 2D copper-plate temperature-monitoring example (Battistelli et al., 2016). The DLKCF is evaluated on a 136-cell traffic network partitioned into $B(x)=\alpha \partial x/\partial n+\beta x$8 sections with overlaps of $B(x)=\alpha \partial x/\partial n+\beta x$9 cells and four sensors per section. The reported results are that DLKCF reduces neighbor disagreement by about $\Omega_m$00 relative to a local-only KF, achieves RMSE $\Omega_m$01 versus $\Omega_m$02 for the local filter, while the centralized KF has no disagreement but costs about $\Omega_m$03 runtime; NEES across $\Omega_m$04 Monte Carlo runs stays within $\Omega_m$05 confidence (Sun et al., 2016).

In distributed multi-object tracking, performance depends strongly on clutter and detection regime. Under high SNR, more consensus steps ($\Omega_m$06 instead of $\Omega_m$07) reduce OSPA by approximately $\Omega_m$08, and CM$\Omega_m$09-GLMB has the lowest variance in cardinality estimation. Under low SNR, CLMB breaks down, whereas CM$\Omega_m$10-GLMB still tracks reliably. Under low $\Omega_m$11, only CM$\Omega_m$12-GLMB is reported as robust, and with $\Omega_m$13 it nearly matches centralized performance (Fantacci et al., 2015).

For distributed extended-object tracking, three scenarios are reported. In the stationary-rectangle experiment, CI and CM both converge to the centralized IF solution in approximately $\Omega_m$14 sequential updates; CM achieves lower OSPA, whereas CI yields slightly tighter consensus through lower ACEE. In dynamic ellipse and rectangle scenarios, CM is reported as tracking as well as CEOT and outperforming CEOT-RM, while DEOT fails when not all nodes see the object. NEES tests show CM remains within the $\Omega_m$15 confidence bound for both kinematics and extent even with few consensus iterations (Li et al., 2021).

CA-NKCF is evaluated on linear harmonic oscillators, the Lorenz attractor, and wireless user-equipment tracking through scatterers. The reported gains are average MSE reductions of $\Omega_m$16–$\Omega_m$17 on the oscillator task, up to $\Omega_m$18 on Lorenz versus the best model-based filters and RNNs, and superior average position MSE under heavy scattering with $\Omega_m$19–$\Omega_m$20 scatterers in the wireless setting. The forward runtime is reported as $\Omega_m$21–$\Omega_m$22 per step (Stamatelis et al., 26 Jun 2026).

The local-average literature emphasizes a different trade-off. Relative to global consensus, local consensus preserves spatial variation but incurs higher memory requirement and degraded noise performance. The spatial transfer function for exponential weighting is

$\Omega_m$23

while the finite-window scheme has bandwidth on the order of $\Omega_m$24 and noise variance $\Omega_m$25 (Cai et al., 2013). In graph-filter acceleration on random directed graphs, the proposed design nearly matches oracle performance in the reported stochastic-block-model example, but the method assumes a large-$\Omega_m$26 asymptotic regime, constant topology, and computationally intensive offline evaluation of the spectral approximation (Kruzick et al., 2018).

Taken together, these results show that spatial consensus filtering is primarily a method for trading centralized optimality against locality, scalability, and communication economy. The exact trade-off depends on what is being averaged or reconciled: overlap states, information quantities, multi-object densities, or local spatial averages.