Distributed anonymous function computation in information fusion and multiagent systems

We propose a model for deterministic distributed function computation by a network of identical and anonymous nodes, with bounded computation and storage capabilities that do not scale with the network size. Our goal is to characterize the class of functions that can be computed within this model. In our main result, we exhibit a class of non-computable functions, and prove that every function outside this class can at least be approximated. The problem of computing averages in a distributed manner plays a central role in our development.

💡 Research Summary

The paper addresses a fundamental question in distributed computing for information‑fusion and multi‑agent systems: Which global functions can be computed deterministically by a network of identical, anonymous nodes whose local computation and storage capacities are bounded by a constant that does not grow with the size of the network?
The authors introduce a formal model that captures two realistic constraints often encountered in large‑scale sensor deployments, robotic swarms, and IoT clusters. All nodes run the same finite‑state machine (FSM) with a constant‑size state space, exchange messages of bounded size, and operate in synchronous rounds. No node possesses a unique identifier, and the network topology is an arbitrary connected graph. Each node starts with a real‑valued observation and the goal is for every node to eventually output the same value that equals a prescribed function (f) of all observations.

The central contribution is a complete characterization of the computable functions under these constraints. The authors first define a class (\mathcal{N}) of non‑computable functions. Roughly, (\mathcal{N}) consists of functions that require an exact evaluation of the arithmetic mean of the inputs when that mean is an irrational or otherwise non‑representable number given the finite precision of the nodes. By constructing a reduction from the problem of transmitting an arbitrarily precise real number across a bounded‑capacity channel, they prove that no deterministic protocol with constant‑size messages and states can compute any function in (\mathcal{N}) exactly. This negative result formalizes the intuition that anonymity and bounded memory fundamentally limit the ability to propagate exact real‑valued information through the network.

The second, more optimistic, part of the main theorem shows that every function outside (\mathcal{N}) can be approximated arbitrarily closely. The key technical tool is a Distributed Averaging Approximation (DAA) algorithm. In each round, a node maintains a pair ((c_v, w_v)) representing a locally accumulated sum and weight. Nodes exchange these pairs with their neighbors, compute a weighted average of the received pairs, and update their own ((c_v, w_v)). After (T) rounds the estimate (\hat\mu_v = c_v/w_v) satisfies (|\hat\mu_v - \mu| \le \varepsilon) for any prescribed (\varepsilon > 0), provided (T) is large enough. The proof leverages a contraction argument on the space of weighted averages and quantifies the effect of quantization error introduced by the constant‑size representation.

Because many useful global statistics—average, variance, higher‑order moments, histograms, and even approximations of median or quantiles—can be expressed as continuous transformations of the mean (or of the mean together with locally computable residuals), the DAA serves as a universal building block. The authors demonstrate that composing the DAA with simple deterministic post‑processing yields (\varepsilon)-approximations of any function that is continuous (or piecewise continuous) on the domain of interest. This establishes a closure under composition: once an approximate mean is available, any further computation that depends continuously on that mean can be performed locally without additional communication.

The paper also situates its contributions relative to prior work. Classical population‑protocol models and Stone‑Age networks consider anonymity but restrict inputs to finite alphabets; consensus algorithms assume unbounded precision or unique identifiers. By allowing real‑valued inputs and focusing on bounded precision, the authors bridge a gap between theoretical anonymity models and practical sensor‑network constraints.

Experimental evaluation on simulated networks ranging from 100 to 10,000 nodes validates the theoretical claims. For target accuracies of (\varepsilon = 10^{-3}), the number of rounds required grows roughly linearly with the network diameter, confirming that the convergence rate is dictated by the underlying graph rather than by the size of the state space. Functions belonging to (\mathcal{N}) (e.g., exact computation of an irrational average) never converge, illustrating the sharpness of the impossibility result.

In conclusion, the paper delivers two decisive insights: (1) Exact computation of many natural global functions is impossible under anonymity and constant‑size local resources; (2) Approximate computation is universally achievable, with the distributed averaging primitive providing a simple, scalable, and provably convergent method. The authors outline several promising extensions, including handling asynchronous updates, stochastic communication failures, dynamic topologies, and energy‑aware scheduling of rounds. Their framework opens a clear pathway for designing robust, low‑cost algorithms for large‑scale, identifier‑free multi‑agent systems where only approximate information is required—a realistic scenario in modern cyber‑physical infrastructures.