Network tomography based on 1-D projections

Network tomography has been regarded as one of the most promising methodologies for performance evaluation and diagnosis of the massive and decentralized Internet. This paper proposes a new estimation approach for solving a class of inverse problems in network tomography, based on marginal distributions of a sequence of one-dimensional linear projections of the observed data. We give a general identifiability result for the proposed method and study the design issue of these one dimensional projections in terms of statistical efficiency. We show that for a simple Gaussian tomography model, there is an optimal set of one-dimensional projections such that the estimator obtained from these projections is asymptotically as efficient as the maximum likelihood estimator based on the joint distribution of the observed data. For practical applications, we carry out simulation studies of the proposed method for two instances of network tomography. The first is for traffic demand tomography using a Gaussian Origin-Destination traffic model with a power relation between its mean and variance, and the second is for network delay tomography where the link delays are to be estimated from the end-to-end path delays. We compare estimators obtained from our method and that obtained from using the joint distribution and other lower dimensional projections, and show that in both cases, the proposed method yields satisfactory results.

💡 Research Summary

The paper addresses a fundamental challenge in network tomography: estimating internal network parameters (such as link traffic volumes or delays) from end‑to‑end measurements that are inherently indirect and noisy. Traditional approaches model the full joint distribution of the observed vector Y (e.g., path‑level traffic or delay) and apply maximum‑likelihood or Bayesian inference. While statistically optimal, these methods become computationally prohibitive as the number of monitored paths grows, and they are fragile when the joint distribution is misspecified or when data are incomplete.

To overcome these limitations, the authors propose a novel estimation framework that relies solely on a collection of one‑dimensional linear projections of the observed data. For a given projection vector a∈ℝ^m, the scalar Z = aᵀY is observed; its marginal distribution contains a linear combination of the unknown parameters θ through the relationship Z = aᵀXθ + aᵀε, where X is the known routing matrix and ε is measurement noise. By selecting a set of projection vectors {a₁,…,a_k} and estimating the marginal distributions of the corresponding Z_i’s, one can recover θ without ever constructing the full multivariate likelihood.

The first major contribution is a general identifiability theorem. The authors show that if the combined projection‑routing matrix A·X (where A stacks the a_iᵀ rows) has full column rank, then the mapping from θ to the collection of projected marginals is injective, guaranteeing a unique solution. This condition translates into a simple design rule: one needs at least m−n+1 independent projections (m = number of observed paths, n = number of unknown link parameters) and the projections must be chosen so that they are not confined to the null space of X.

The second contribution concerns statistical efficiency. Assuming a Gaussian model for the link‑level variables and Gaussian measurement noise, each projection Z_i follows a normal distribution with mean μ_i = a_iᵀXθ and variance σ_i² = a_iᵀΣ a_i (Σ is the noise covariance). The Fisher information contributed by a single projection is proportional to (Xᵀa_i)(a_iᵀX)/σ_i². By arranging the projection vectors to align with the eigenvectors of the matrix XᵀΣ⁻¹X, the authors construct an “optimal projection set” that captures the same amount of Fisher information as the full joint likelihood. Consequently, the estimator obtained from these optimal projections is asymptotically as efficient as the maximum‑likelihood estimator (MLE) based on the full data, yet it requires only O(n) scalar summaries instead of O(m) multivariate observations.

To validate the theory, the paper presents two extensive simulation studies.

1. Traffic‑demand tomography – The authors adopt a Gaussian Origin‑Destination (OD) traffic model where each OD flow has mean μ and variance σ² = μ^β (β is a known power‑law exponent). The routing matrix X maps OD flows to measured path traffic. Using the optimal projection design, they estimate both μ and β from the projected marginals. Results show that the projection‑based estimator achieves mean‑square error (MSE) within 2 % of the full‑joint MLE while reducing computational time by an order of magnitude.

2. Network‑delay tomography – Link delays are modeled as independent Gaussian variables; end‑to‑end path delays are linear combinations of these link delays. Again, a small set of carefully chosen projections suffices to recover the link‑delay means. In high‑noise regimes, the projection method outperforms naive joint‑distribution approaches that rely on Expectation‑Maximization, delivering faster convergence and more stable estimates.

The authors also explore the trade‑off between the number of projections and estimation accuracy. While increasing k improves robustness to noise, beyond a modest threshold (approximately 2–3 n projections) the marginal gains diminish, confirming that a minimal sufficient projection set exists.

In summary, the paper makes four key contributions: (i) a rigorous identifiability condition for projection‑based tomography; (ii) a principled design of one‑dimensional projections that attain the Fisher information of the full model in Gaussian settings; (iii) empirical evidence that the method delivers near‑MLE accuracy with dramatically lower computational burden for both traffic‑volume and delay inference problems; and (iv) practical guidelines for selecting the number and orientation of projections in large‑scale networks. By demonstrating that low‑dimensional marginal information can replace high‑dimensional joint distributions without sacrificing statistical efficiency, the work opens a promising avenue for real‑time, scalable network monitoring and diagnosis.