DIFNet: Decentralized Information Filtering

• DIFNet is a hybrid decentralized estimation framework that uses neural networks to learn fusion weights and balance model-based filtering with unknown noise correlations.
• It preserves the structure of traditional information filters while substituting unavailable covariance inverses with data-driven estimates via local sensor communication.
• The approach enables consistent, localized state estimation in partially connected sensor networks and adapts to time-varying measurement noise.

Searching arXiv for the specified DIFNet papers and closely related context. Decentralized Information Filter Neural Network (DIFNet) denotes a hybrid model-based/data-driven framework for decentralized state estimation in discrete-time nonlinear state space models with cross-correlated sensor measurement noises. In its canonical formulation, DIFNet preserves the algebraic structure of information-form filtering while replacing the unknown correlation-dependent fusion weights by neural estimates learned from local information contributions and neighborhood communication. The method is designed for partially connected sensor networks in which full connectivity is unavailable, local measurement covariances are known, the off-diagonal blocks of the stacked measurement covariance are unknown and possibly time-varying, and consistent local fusion is achieved through model distribution and a structured observation model (Dong et al., 26 Aug 2025). A related later development, the Covariance-Agnostic Neural Kalman Consensus Filter (CA-NKCF), can be interpreted as a lightweight information-filter-style variant that dispenses with explicit covariance or information exchanges and instead uses learned gains and consensus on priors, thereby situating DIFNet within a broader family of decentralized neural estimators (Stamatelis et al., 26 Jun 2026).

1. Conceptual scope and estimation objective

DIFNet was proposed for decentralized sensor networks in which the global state estimation problem is decomposed into local subproblems through a structured observation model. The explicit motivation is that optimal fusion performance can be achieved under fully connected communication and known noise correlation structures, whereas realistic decentralized settings often violate both assumptions. DIFNet addresses the degradation of fusion accuracy induced by unknown correlations in measurement noise by learning those correlations implicitly in data rather than requiring an explicit model of the full measurement covariance (Dong et al., 26 Aug 2025).

The framework is not a purely model-free recurrent estimator. Its defining characteristic is the retention of the information-filter update structure, including prediction, local filtering, construction of information increments, and information-space fusion, while delegating the unknown correlation-dependent weighting to learned mappings. This architecture places DIFNet between classical decentralized information filtering and end-to-end sequence models: the process and measurement models remain operational, but the unknown intersensor statistical coupling is absorbed by the neural component.

A recurrent misconception is that decentralized fusion requires either a centralized coordinator or a fully connected graph to attain meaningful accuracy. The formulation underlying DIFNet is explicitly neighborhood-based and leverages structured observation and internodal transformations so that each node can achieve a locally optimal estimate of its observable state components even when the communication graph is not fully connected. Another misconception is that the method learns the entire filter from scratch. In the published formulation, what is learned are the fusion weights that would otherwise depend on the unavailable inverse of the full cross-correlated measurement covariance.

2. State-space model, structured observation, and network assumptions

The global process is formulated as a discrete-time nonlinear state-space model,

$$x_{k+1} = f(x_k) + w_k,\qquad w_k \sim \mathcal{N}(0,Q_k),$$

where $x_k \in \mathbb{R}^m$ and $Q_k \in \mathbb{R}^{m\times m}$ is known. Sensor $i$ acquires

$$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$

where $z_k^{(i)} \in \mathbb{R}^{n_i}$ and $R_{ii,k}$ is known locally. The difficulty arises because the measurement noises are cross-correlated:

$$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$

with full stacked covariance

$$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$

whose off-diagonal blocks are unknown and may be time-varying (Dong et al., 26 Aug 2025).

The global model is distributed into local models by an internodal transformation,

$$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$

This model distribution “selects” or linearly combines globally relevant states for node $x_k \in \mathbb{R}^m$0. Local dynamics and observations are required to remain dynamically equivalent to the global system:

$x_k \in \mathbb{R}^m$1

In the linear case, the transition matrices satisfy

$x_k \in \mathbb{R}^m$2

A structural observation model is imposed to guarantee local observability on the measured subspace. For sensor $x_k \in \mathbb{R}^m$3, the state is partitioned as

$x_k \in \mathbb{R}^m$4

and the local Jacobian is designed as

$x_k \in \mathbb{R}^m$5

Under the rank conditions $x_k \in \mathbb{R}^m$6 and $x_k \in \mathbb{R}^m$7, each node achieves a locally optimal estimate over its observable substate even when the communication graph is not fully connected.

Communication is defined by overlap of information subspaces. The internodal transform

$x_k \in \mathbb{R}^m$8

determines whether nodes $x_k \in \mathbb{R}^m$9 and $Q_k \in \mathbb{R}^{m\times m}$0 exchange messages: communication occurs iff at least one of $Q_k \in \mathbb{R}^{m\times m}$1 or $Q_k \in \mathbb{R}^{m\times m}$2 is non-null. Each node transmits local prior and posterior estimates and covariances, $Q_k \in \mathbb{R}^{m\times m}$3, to relevant neighbors only. This design reduces communication load by distributing the model and sharing only locally relevant states and covariances.

3. Information-form decomposition and decentralized fusion equations

DIFNet is built on the information filter representation

$Q_k \in \mathbb{R}^{m\times m}$4

For nonlinear systems, the usual linearizations are

$Q_k \in \mathbb{R}^{m\times m}$5

Prediction follows the EKF/EIF form,

$Q_k \in \mathbb{R}^{m\times m}$6

If measurement noises were independent, the information update would be

$Q_k \in \mathbb{R}^{m\times m}$7

$Q_k \in \mathbb{R}^{m\times m}$8

With cross-correlated noises, however, the correct centralized update requires the full stacked covariance:

$Q_k \in \mathbb{R}^{m\times m}$9

$i$0

Because $i$1 is unknown, optimal innovation weighting cannot be formed directly; naive independence assumptions induce double counting and overconfidence (Dong et al., 26 Aug 2025).

The decentralized decomposition proceeds through local information increments,

$i$2

$i$3

Under cross-correlated measurement noises and the stated rank conditions, decentralized EIF-based fusion is equivalent to centralized EIF and yields consistent local estimates when the correlation-dependent weights are known:

$i$4

$i$5

The model-based weights are

$i$6

where $i$7 is the $i$8-th column block of $i$9. At node $$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$0, projected local-subspace fusion uses

$$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$1

$$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$2

with

$$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$3

These equations define the target object that DIFNet learns: not the state transition itself, but the unknown correlation-aware information weighting that links local information increments to globally consistent fusion.

4. Neural architecture, learned weights, and training objective

The core neural substitution in DIFNet is to replace $$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$4 and $$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$5 by data-driven estimates while leaving the surrounding information-filter mechanics intact. The primary inputs at node $$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$6 are the neighbor information features

$$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$7

In the implementation described in the paper, the network input is formed by concatenating $$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$8, yielding size $$z_k^{(i)} = h_i(x_k) + v_k^{(i)},\qquad v_k^{(i)} \sim \mathcal{N}(0,R_{ii,k}),$$9, with zeros inserted for non-neighbors so that the network encodes the current graph connectivity implicitly (Dong et al., 26 Aug 2025).

Each sensor maintains its own neural network, for a total of $z_k^{(i)} \in \mathbb{R}^{n_i}$0 networks. The backbone is an FC-ReLU $z_k^{(i)} \in \mathbb{R}^{n_i}$1 GRU $z_k^{(i)} \in \mathbb{R}^{n_i}$2 FC architecture. The input fully connected layer maps the concatenated information features into a higher-dimensional latent space; the GRU models temporal variation in the latent correlation structure; the output fully connected layer emits $z_k^{(i)} \in \mathbb{R}^{n_i}$3 scalars, reshaped into

$z_k^{(i)} \in \mathbb{R}^{n_i}$4

The learned mapping is therefore

$z_k^{(i)} \in \mathbb{R}^{n_i}$5

The learned weights are then inserted into the information update:

$z_k^{(i)} \in \mathbb{R}^{n_i}$6

$z_k^{(i)} \in \mathbb{R}^{n_i}$7

followed by $z_k^{(i)} \in \mathbb{R}^{n_i}$8. Node-specific local-subspace fusion uses projected weights $z_k^{(i)} \in \mathbb{R}^{n_i}$9 obtained via the internodal transforms.

Training is supervised with ground-truth states. The per-node objective is

$R_{ii,k}$0

where $R_{ii,k}$1 is the number of trajectories, $R_{ii,k}$2 their lengths, and $R_{ii,k}$3 is weight decay. The reported optimizer is Adam with fixed learning rate $R_{ii,k}$4, batch size $R_{ii,k}$5, and weight decay $R_{ii,k}$6, together with a cyclic learning-rate schedule.

The paper learns fusion weights directly as free matrices. It does not explicitly enforce symmetry or positive semidefiniteness of the weight-induced information increments, although symmetry and PSD projection, or Cholesky-factor parameterizations, are recommended in the technical summary as practical safeguards for numerical stability. This is an important implementation detail: the formal information update requires $R_{ii,k}$7, but the learned parameterization itself does not guarantee that property unless additional constraints are imposed.

5. Empirical behavior, runtime, and communication profile

The reported empirical evaluation covers three regimes: a linear model, a nonlinear model, and time-varying measurement noise. The linear setting uses a 6D constant-velocity target $R_{ii,k}$8, sample interval $R_{ii,k}$9, process noise $$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$0, and four sensors measuring different subsets of the state. The nonlinear setting uses constant-turn motion in the $$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$1-$$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$2 plane with $$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$3 rad/s, linear motion in $$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$4, and four sensors including bearing and range-type measurements. Unknown cross-correlated measurement noise is induced by a jammer signal, and in the time-varying case the measurement covariance is modulated as

$$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$5

Training uses 100 trajectories, validation uses 20, test uses 40, and each sequence has length 50 (Dong et al., 26 Aug 2025).

Evaluation is reported in terms of RMSE over test trajectories, per sensor, for position and velocity components. The comparison set includes centralized EKF/EIF with exact cross-correlated measurement noise, decentralized EIF with exact parameters, decentralized EIF with mismatched or inexact parameters, and DIFNet. In the linear model, DIFNet outperforms the inexact decentralized information filter and approaches the exact decentralized and centralized baselines despite the unknown cross-correlation. Two sensors with identical information and identical neural initializations learn identical weights and achieve similar RMSE. In the nonlinear model, DIFNet again maintains an advantage over the inexact decentralized baseline and approaches the exact-information benchmark across position and velocity RMSE. Under time-varying noise, DIFNet achieves performance comparable to the exact decentralized method notwithstanding unknown modulation strength and consistently performs better than the inexact decentralized method across different $$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$6 values.

These results establish the intended operating regime of DIFNet: decentralized fusion under model mismatch in the correlation structure rather than under complete ignorance of the dynamical model. Runtime is also quantified. Normalizing fusion-step time so that DIF$$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$7, DIFNet has factor $$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$8 in the linear case and $$\mathbb{E}\!\left[v_k^{(i)}v_k^{(j)\top}\right] = R_{ij,k},\qquad i\neq j,$$9 in the nonlinear case. The overhead is therefore non-negligible but remains acceptable in the reported experiments relative to the gain in robustness to unknown and time-varying correlation patterns.

Communication remains explicitly decentralized. Each node exchanges $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$0 with neighbors only. Per-node computational cost includes EKF/EIF prediction and update, typically $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$1 because of matrix factorizations or inversions, a GRU-based forward pass with input size $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$2, and fusion operations involving $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$3 blocks that scale with neighborhood degree. Model distribution reduces communication by keeping the local state dimension $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$4 smaller than the global dimension whenever the task structure permits.

6. Relation to covariance-agnostic neural consensus filtering, limitations, and extensions

A later paper on distributed estimation under partially known dynamics introduces CA-NKCF, which is not identical to DIFNet but is explicitly interpretable as a decentralized information-filter-style estimator with neural surrogates for information contributions. In that framework, $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$5 agents observe a latent state $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$6, receive local measurements $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$7, communicate with neighbors $$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$8, and operate under partially known dynamics and observations,

$$R_k=\mathrm{Cov}(v_k)= \begin{bmatrix} R_{11,k} & R_{12,k} & \cdots & R_{1N,k}\ R_{21,k} & R_{22,k} & \cdots & R_{2N,k}\ \vdots & \vdots & \ddots & \vdots\ R_{N1,k} & R_{N2,k} & \cdots & R_{NN,k} \end{bmatrix},$$9

or, in the linear case,

$$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$0

without assuming knowledge of $$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$1 or $$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$2 (Stamatelis et al., 26 Jun 2026).

CA-NKCF combines model-based prediction, neural measurement fusion, and lightweight consensus. Its prediction is

$$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$3

and its recurrent feature vector is

$$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$4

with

$$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$5

A GRU produces a Kalman-like gain,

$$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$6

and the posterior update is

$$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$7

The consensus term is a convex combination of priors with elementwise sigmoid-bounded weights $$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$8, and only priors are broadcast.

The relation to DIFNet is structural rather than identical. In classical information filtering, the measurement update uses information contributions such as $$x_k^{(j)} = T_k^{(j)}x_k,\qquad T_k^{(j)}\in\mathbb{R}^{m_j\times m},\qquad \mathrm{rank}\!\big(T_k^{(j)}\big)\ge n_j.$$9 and $x_k \in \mathbb{R}^m$00. CA-NKCF does not compute or exchange those quantities; instead, the learned gain $x_k \in \mathbb{R}^m$01 acts as a surrogate for measurement information, and consensus over priors approximates information pooling with drastically reduced communication. The technical summary explicitly characterizes this as an “information-form-lite” estimator. A DIFNet inspired by CA-NKCF would go one step further and learn surrogate information variables $x_k \in \mathbb{R}^m$02 and $x_k \in \mathbb{R}^m$03, then update $x_k \in \mathbb{R}^m$04 and $x_k \in \mathbb{R}^m$05 directly. This suggests that “DIFNet” can denote a broader design pattern: decentralized neural fusion architectures that preserve information-form semantics while replacing analytically unavailable quantities by learned surrogates (Stamatelis et al., 26 Jun 2026, Dong et al., 26 Aug 2025).

The limitations recorded across the two formulations are complementary. For DIFNet, the principal constraints are dependence on supervised state labels, possible degradation under unseen topology changes, output dimension growth with $x_k \in \mathbb{R}^m$06, the Gaussian-noise assumption, and the absence of a formal optimality proof for the learned weights. For CA-NKCF, the explicit trade-off is between bandwidth and optimality: it omits covariance or information propagation, uses single-shot consensus to preserve hard real-time feasibility, and requires additional care if extended to a fully neural information form with guaranteed positive definiteness and numerical stability. Taken together, these limitations define an active frontier that includes online or unsupervised adaptation, heterogeneous or asynchronous sensing, federated or decentralized training, explicit PSD enforcement, and more robust operation under severe communication constraints or adversarial conditions.