networks
X-ray the network from its edges
It is 2002, and the internet has become critical infrastructure that no one can inspect: providers will not share internal maps or link statistics, yet operators, regulators, and researchers all need to know where packets are being delayed and dropped. All you can do is probe from the edges — send packets between endpoints you control and observe delays and losses shaped by every hidden link along the path. Pose it as an inverse problem: which internal link parameters are identifiable from which patterns of correlated end-to-end measurements, what multicast or striped probes buy you, and how confident the tomographic reconstruction can honestly be. Get it wrong and the map is fiction — congestion attributed to innocent links, providers blamed or exonerated by an unidentifiable model.
Who this problem belongs to
The two figures whose methods fit it best, out of 50 in contention.
Network tomography is Nowak's own research area, not an analogy. Through the 2000s he published extensively on inferring internal network properties, loss rates, delay distributions, and topology, from end-to-end measurements between controllable endpoints, using multicast and unicast probe correlation structures to identify which internal link parameters are statistically recoverable and which are confounded. His work on active and adaptive sensing directly addresses the problem's core question of measurement design: which probing pattern extracts the most identifiable information under a budget. He formalized the inverse-problem framing precisely, treating the internet as a graph whose interior is unobservable except through correlated exterior measurements, and characterized honest confidence bounds on the reconstruction. Dropped into 2002, he is not applying a toolkit to this problem; the toolkit is his own body of published work on this exact problem.
Wainwright's work on high-dimensional statistics and graphical models gives him deep, direct fluency in exactly the identifiability question this problem poses: when can a set of correlated end-to-end measurements uniquely determine internal parameters of a graphical structure, and when is the inverse problem fundamentally underdetermined. His research on variational inference and structured estimation under sparsity provides rigorous machinery for stating honest confidence in a tomographic reconstruction rather than presenting a point estimate as certain. He is slightly behind the strongest carrier because network tomography as a specific applied subfield, with its multicast-probe design tricks, was built by a research community adjacent to but not centered on his own graphical-models work; he would rederive the identifiability theory cleanly but without the field's specific probing-protocol engineering.
In the mind map
The same ideas, as concepts rather than history — in John's ML knowledge map.
50 figures are scored on this problem. Draw it in a battle to see where you land.