SP-ADMM-JCNL: Joint Localization in WSNs

• The paper introduces a joint formulation that merges sensor and passive target localization to eliminate estimation delays and enhance consistency.
• It employs an auxiliary-variable reformulation that decouples the non-convex and non-smooth objective through dual representations and consensus constraints.
• The scaled proximal ADMM iteration achieves global convergence to a KKT point with a sublinear O(1/T) rate, validated on both synthetic and benchmark datasets.

SP-ADMM-JCNL denotes the Scaled Proximal Alternating Direction Method of Multipliers for Joint Cooperative and Non-Cooperative Localization, a distributed optimization method for wireless sensor networks (WSNs) in which sensor self-localization and passive target localization are posed and solved as a single optimization problem. Its defining features are a joint modeling of cooperative and non-cooperative localization, an auxiliary-variable reformulation that structurally decouples the resulting non-convex and non-smooth objective, and a scaled proximal ADMM scheme whose iterates are shown to converge globally to a Karush-Kuhn-Tucker (KKT) point of the reformulated problem and to a critical point of the original non-convex objective function, with a sublinear rate of $O(1/T)$ (Zhu et al., 21 Sep 2025).

1. Joint localization setting

SP-ADMM-JCNL is designed for the setting in which cooperative localization and non-cooperative localization arise simultaneously in a WSN. In the terminology used in the source work, cooperative localization refers to collaborative estimation of sensor positions, typically aided by anchors whose positions are known, whereas non-cooperative localization refers to estimation of a passive external target that does not communicate with the network (Zhu et al., 21 Sep 2025).

The motivation for a joint treatment is operational as well as algorithmic. Sequential treatment, denoted SCNL in the source summary, first localizes the sensors and only then estimates the target. The joint formulation, denoted JCNL, eliminates the delay in target estimation that is intrinsic to such sequential pipelines. It also allows sensor-position inference to benefit from target-related measurements, so that all positions are estimated consistently and concurrently. The paper characterizes the resulting optimization as having complex variable coupling, which is the central technical obstacle addressed by the method (Zhu et al., 21 Sep 2025).

The network model consists of sensors indexed by $\mathcal{N}=\{1,\dots,N\}$, a subset of anchors $\mathcal{A}\subset\mathcal{N}$ with known positions, an unknown target position $\mathbf{y}\in\mathbb{R}^n$, and an undirected, connected graph $G=(\mathcal{N},\mathcal{E})$. The measurements are noisy inter-sensor ranges

$$d_{i,j}=\|\mathbf{x}_i-\mathbf{x}_j\|+\omega_{i,j},$$

and noisy sensor-target ranges

$r_i=\|\mathbf{x}_i-\mathbf{y}\|+\tau_i,$

with $\omega_{i,j}$ and $\tau_i$ modeled as i.i.d. Gaussian noise (Zhu et al., 21 Sep 2025).

2. Optimization model and JCNL formulation

The maximum likelihood estimator underlying SP-ADMM-JCNL minimizes

$$\frac{1}{2}\sum_{i\in\mathcal{N}} f_i(\mathbf{x},\mathbf{y})$$

subject to anchor constraints $\mathcal{N}=\{1,\dots,N\}$0 for all $\mathcal{N}=\{1,\dots,N\}$1, where the local term is

$\mathcal{N}=\{1,\dots,N\}$2

The paper describes this objective as non-convex, non-smooth, and tightly coupled across variables (Zhu et al., 21 Sep 2025).

That description is consequential. Non-convexity enters through the geometry of range-based localization, non-smoothness through Euclidean norms, and tight coupling through simultaneous dependence on neighboring sensor positions and the target position. A straightforward distributed ADMM application is therefore obstructed both by local non-smooth terms and by the fact that each node’s variables are entangled with those of its neighbors and with the target estimate.

A compact comparison between the joint and sequential viewpoints used in the paper is given below.

Aspect	JCNL (Joint)	SCNL (Separate)
Modeling	Joint estimation of sensors and target	Two-stage: sensors, then target
Optimization	Single least-squares problem	Sequential least-squares problems
Information flow	Target data aids sensor localization	Used only after sensor localization
Target estimation timeliness	Real-time with sensors	Waits for sensor positions first

This comparison clarifies a common misconception. The point of JCNL is not merely to bundle two tasks into one larger objective. Rather, the formulation changes the information flow: target measurements enter the sensor-position inference itself. A plausible implication is that the gain is structural, not only procedural.

3. Structural decoupling and constrained smooth reformulation

The reformulation begins with a variational treatment of the norm terms. For inter-sensor differences,

$\mathcal{N}=\{1,\dots,N\}$3

and an analogous representation is used for $\mathcal{N}=\{1,\dots,N\}$4, where $\mathcal{N}=\{1,\dots,N\}$5 is the unit ball. In the source summary this is described as a dual representation using the Cauchy-Schwarz inequality. The effect is to transform the non-smooth loss into a bilinear coupling with convex bound constraints (Zhu et al., 21 Sep 2025).

A second layer of reformulation introduces local copies and auxiliary consensus variables. For node $\mathcal{N}=\{1,\dots,N\}$6, examples include $\mathcal{N}=\{1,\dots,N\}$7, node $\mathcal{N}=\{1,\dots,N\}$8’s local copy of neighbor $\mathcal{N}=\{1,\dots,N\}$9’s position, and $\mathcal{A}\subset\mathcal{N}$0, node $\mathcal{A}\subset\mathcal{N}$1’s local copy of neighbor $\mathcal{A}\subset\mathcal{N}$2’s estimate of the target. The stacked local block is

$\mathcal{A}\subset\mathcal{N}$3

Consensus constraints enforce agreement among local views, including neighbor consistency and target-consensus relations such as $\mathcal{A}\subset\mathcal{N}$4 for neighbors (Zhu et al., 21 Sep 2025).

After substitution, the problem is written in the condensed form

$\mathcal{A}\subset\mathcal{N}$5

subject to

$\mathcal{A}\subset\mathcal{N}$6

Here $\mathcal{A}\subset\mathcal{N}$7, $\mathcal{A}\subset\mathcal{N}$8, and $\mathcal{A}\subset\mathcal{N}$9 are structured matrices encoding the local geometry and consistency relations, and $\mathbf{y}\in\mathbb{R}^n$0 is the indicator function of the corresponding unit-ball constraint set (Zhu et al., 21 Sep 2025).

This reformulation is central to the algorithm’s identity. It is “smooth and constrained” in the sense that the problematic norm terms have been transferred into auxiliary-variable constraints and bilinear interactions, while separability is recovered at the node level except for the explicit consensus structure.

4. Scaled proximal ADMM iteration

SP-ADMM-JCNL applies a distributed augmented-Lagrangian scheme over the reformulated variables. For each node $\mathbf{y}\in\mathbb{R}^n$1 and iteration $\mathbf{y}\in\mathbb{R}^n$2, the method performs a primal update, an auxiliary-variable update, and a dual update (Zhu et al., 21 Sep 2025).

The primal step is

$\mathbf{y}\in\mathbb{R}^n$3

where $\mathbf{y}\in\mathbb{R}^n$4 is the local augmented Lagrangian, $\mathbf{y}\in\mathbb{R}^n$5 is a penalty parameter, and $\mathbf{y}\in\mathbb{R}^n$6 is a selection matrix chosen so that the update is easy to solve and the subproblem is strongly convex. The source summary explicitly describes this as a diagonalization trick yielding closed-form updates through the structure of $\mathbf{y}\in\mathbb{R}^n$7 (Zhu et al., 21 Sep 2025).

The auxiliary-variable step is a projection: $\mathbf{y}\in\mathbb{R}^n$8 with $\mathbf{y}\in\mathbb{R}^n$9 a penalty parameter. The dual variable is then updated by

$G=(\mathcal{N},\mathcal{E})$0

The resulting algorithm is distributed in the precise sense that all updates are local except for exchanges with neighboring nodes. The paper states that each update depends only on local and neighbor variables and that the full procedure is specified in Algorithm 1, including communication steps (Zhu et al., 21 Sep 2025).

From an algorithmic standpoint, the qualifier scaled proximal has a specific role. The proximal regularization term strengthens the subproblems and stabilizes the non-convex ADMM iteration, while the scaling embodied in $G=(\mathcal{N},\mathcal{E})$1 is selected to make the local solve explicit or at least structurally simple.

5. Convergence analysis and interpretive scope

The convergence theory provided for SP-ADMM-JCNL is global in the sense used in non-convex optimization: the algorithm generates a sequence converging to a KKT point of the reformulated problem and to a critical point of the original non-convex objective. The reported rate is sublinear, specifically $G=(\mathcal{N},\mathcal{E})$2 in terms of the optimality gap (Zhu et al., 21 Sep 2025).

The analysis introduces stationarity, feasibility, and update-gap measures to track progress. For any pre-specified accuracy $G=(\mathcal{N},\mathcal{E})$3, the paper states that the number of iterations $G=(\mathcal{N},\mathcal{E})$4 required to reach $G=(\mathcal{N},\mathcal{E})$5 satisfies

$G=(\mathcal{N},\mathcal{E})$6

which is the stated form of the $G=(\mathcal{N},\mathcal{E})$7 bound. The proof strategy described in the summary combines a potential function, monotonic descent and boundedness arguments, subgradient and stationarity-gap estimates, and the Kurdyka-Łojasiewicz property for semi-algebraic functions. The assumptions include boundedness, network connectivity, and parameter choices ensuring sufficient descent and strong convexity of the subproblems (Zhu et al., 21 Sep 2025).

Two clarifications are important for interpretation. First, the guarantee is to KKT and critical points, not to a globally optimal solution of the non-convex localization problem. Second, the guarantee pertains to the reformulated constrained problem and is then transferred back to the original objective through the equivalence established by the auxiliary-variable construction. The source summary also notes a limiting case: when all nodes are anchors, the approach reduces to standard non-cooperative localization, and the guarantees still hold (Zhu et al., 21 Sep 2025).

6. Empirical behavior and relation to prior SP-ADMM localization methods

The numerical evaluation reported for SP-ADMM-JCNL uses both synthetic and benchmark data. In a synthetic setup with 100 agents, 8 anchors, and 1 target in $G=(\mathcal{N},\mathcal{E})$8, under two measurement noise models—AWGN and range-dependent—the paper tracks RMSE for sensors and target, together with stationarity gap and feasibility gap. The reported behavior is that both sensor and target RMSE converge quickly and accurately, the feasibility and stationarity gaps decrease rapidly in line with the predicted $G=(\mathcal{N},\mathcal{E})$9 rate, and the estimated and true positions overlap closely (Zhu et al., 21 Sep 2025).

On the benchmark dataset test10-500 from Stanford, the joint method is compared with a separate two-stage baseline denoted SP-ADMM-SCNL. The reported runtime is 0.0955s for JCNL against 0.1489s for SCNL. The joint method also achieves lower sensor and target localization RMSE, avoids target-estimation latency, and exhibits better error performance throughout the optimization rather than only at termination (Zhu et al., 21 Sep 2025).

These results place SP-ADMM-JCNL within a broader line of distributed localization methods based on scaled proximal ADMM. Earlier work on distributed cooperative localization in WSNs reformulated the classic cooperative localization problem as a smooth and constrained nonconvex minimization problem with node-separable loss, and proposed distributed SP-ADMM algorithms that globally converge to a KKT point and a critical point of the original problem with a favorable sublinear $$d_{i,j}=\|\mathbf{x}_i-\mathbf{x}_j\|+\omega_{i,j},$$0 convergence rate (Zhang et al., 2022). SP-ADMM-JCNL inherits the same broad design principles—smooth constrained reformulation, nodewise separability, proximal stabilization, and distributed multiplier updates—but applies them to the strictly richer JCNL setting in which sensor localization and passive target localization are solved simultaneously (Zhu et al., 21 Sep 2025).

In that sense, SP-ADMM-JCNL is best understood not as a generic ADMM variant with a new acronym, but as a localization-specific distributed method whose distinctive contribution lies in making joint cooperative and non-cooperative localization computationally tractable while preserving explicit per-node updates and non-convex convergence guarantees.