Mortality surfaces are complicated objects. Age-specific death rates vary by country, sex, year, and cohort, creating a high-dimensional space of possible patterns. Forecasting these surfaces — predicting how mortality will evolve in countries not yet observed or decades not yet lived — has traditionally involved fitting parametric models to time series and extrapolating.
The paper (arXiv:2603.24299, March 2026) recasts the problem through Tucker tensor decomposition. Instead of treating mortality as a time series problem, the authors decompose the full mortality surface into a low-dimensional score space. Each country-year observation becomes a point in this compressed space. The question is then: how do these points move?
They move in one dimension. Despite the apparent complexity of the mortality surface — hundreds of age groups, dozens of countries, both sexes, 50+ years — the transitions between mortality states follow an essentially one-dimensional flow. A single scalar speed function, combined with trajectory functions in the score space, generates complete mortality schedules. The high-dimensional surface is a one-dimensional curve viewed from many angles.
Sex coherence — the requirement that male and female forecasts should be mutually consistent — falls out naturally. In traditional methods, you impose coherence as a constraint, forcing the model to maintain plausible sex ratios. In the tensor framework, the single flow naturally generates coherent schedules because both sexes are traversing the same underlying trajectory at related speeds.
The structural lesson: dimensionality reduction is not just a compression technique. It is a discovery about the system. When mortality transitions are genuinely one-dimensional, the correct model is not a simpler version of a high-dimensional model — it is a fundamentally different kind of model, one that moves along a curve rather than navigating a space. The complexity was in the representation, not in the dynamics.