Self-stabilizing Numerical Iterative Computation

Many challenging tasks in sensor networks, including sensor calibration, ranking of nodes, monitoring, event region detection, collaborative filtering, collaborative signal processing, {\em etc.}, can be formulated as a problem of solving a linear system of equations. Several recent works propose different distributed algorithms for solving these problems, usually by using linear iterative numerical methods. The main problem with previous approaches is that once the problem inputs change during the process of computation, the computation may output unexpected results. In real life settings, sensor measurements are subject to varying environmental conditions and to measurement noise. We present a simple iterative scheme called SS-Iterative for solving systems of linear equations, and examine its properties in the self-stabilizing perspective. We analyze the behavior of the proposed scheme under changing input sequences using two different assumptions on the input: a box bound, and a probabilistic distribution. As a case study, we discuss the sensor calibration problem and provide simulation results to support the applicability of our approach.

💡 Research Summary

The paper addresses a fundamental challenge in distributed sensor networks: solving linear systems of equations when the input data are continuously changing due to environmental variations and measurement noise. Traditional distributed iterative solvers (e.g., Jacobi, Gauss‑Seidel) assume static inputs; if the right‑hand side vector b or the matrix A changes during execution, the algorithm may diverge or produce erroneous results. To overcome this limitation, the authors propose a simple yet robust iterative scheme called SS‑Iterative (Self‑Stabilizing Iterative).

Algorithmic core
Each node i maintains a local estimate x_i of its component of the solution vector x. At every iteration the node updates its estimate using the most recent values received from its neighbors:

 x_i^{(t+1)} = (1/d_i)·(b_i – Σ_{j∈N(i)} a_{ij}·x_j^{(t)}

where d_i = a_{ii} and N(i) denotes the set of neighboring nodes. This update is essentially a Jacobi step performed with the current, possibly outdated, neighbor values. The key difference is the self‑stabilizing perspective: the algorithm is designed to converge from any arbitrary initial state, even when b is not fixed.

Self‑stabilizing analysis
The authors study two models for the time‑varying input b:

1. Box‑bounded model – b remains inside a hyper‑cube B = { b | |b_i – b_i^| ≤ Δ }. Under the assumption that A is diagonally dominant (or at least non‑singular) and the communication graph is connected, the iteration matrix M = I – D^{-1}A has spectral radius ρ < 1. The authors prove that the error ‖x^{(t)} – x‖ contracts geometrically with factor ρ, and the steady‑state error is bounded by Δ/(1‑ρ). Thus, even if b drifts within the box, the algorithm never leaves a predictable error envelope.

2. Probabilistic model – b is drawn from a multivariate normal distribution with mean μ and covariance Σ. By treating the iteration as a linear Markov chain, they show that the expected estimate E