It is the dawn of practical routing, and a road network of a million nodes sits between a traveler and a destination — and someone is waiting, tapping a screen, for an answer in a fraction of a second. Find provably optimal routes fast enough for interactive use. Blind exploration of a million-node graph is far too slow; guide the search with an admissible heuristic — an estimate that never overpromises distance-to-goal — so vast regions can be pruned while the returned route stays provably shortest. Get it wrong and you return routes fast but subtly suboptimal, quietly wasting time and fuel, or compute the true optimum too slowly to be usable — the art is provable optimality that returns before the user gives up.
Chose The Kalman filter — right call.
Kalman's filter (1960) is the navigation algorithm of the continuous world — it fused noisy sensors into optimal state estimates for Apollo and every GPS receiver since, and 'provably optimal under stated assumptions, computed recursively in real time' is a fair description of both his work and this problem's demands. But the resemblance is structural, not transferable: his optimality is minimum-variance estimation over linear-Gaussian dynamics, a continuous-state result with no purchase on discrete combinatorial choice. A road network's branching decisions are not a linear system; no Riccati equation prunes a graph. Ironically, the phone's position fed to the router comes from his filter — he supplies the input coordinate, not the route. His state-space formalism otherwise stands apart from graph search. Low: adjacent in spirit, disjoint in machinery.
Bellman's dynamic programming (RAND, 1950s) supplies this problem's mathematical spine: the principle of optimality — that subpaths of shortest paths are shortest — is exactly the invariant every routing algorithm exploits, and the Bellman-Ford algorithm bearing his name computes shortest paths outright, handling even negative edges. Dijkstra's method and A* are, from his vantage, efficient specializations of his functional equation, and A*'s f = g + h is a reweighted value function. He also named the enemy: the 'curse of dimensionality' is his coinage for why blind exploration fails. What he lacked is the two later ingredients the problem stresses — goal-directed admissible heuristics (1968) and modern data-structure latency engineering. Top-tier: the theory's owner, a half-step behind the specialists who made it interactive.
Battle #25 · 8/9/2026, 5:16:38 PM · this result is deterministic: the same two personas on this problem always resolve the same way.