Skip to content

Variable Ordering Complexity Bottleneck

Home / Open Questions / Variable Ordering Complexity Bottleneck

Background: In multivariate probabilistic models where joint distributions are decomposed sequentially, the choice of variable ordering can drastically alter the learned dependency structure.

Question / Future Work: Identifying optimal variable or channel ordering for structures like Sum-Product Networks (SPNs) or autoregressive factorizations is often computationally intractable due to factorial complexity, yet current heuristics remain sub-optimal for high-dimensional or heterogeneous data.

Why It Matters: This is a fundamental limitation for any structural probabilistic forecasting model that relies on sequential variable decomposition, affecting both accuracy and generalizability.

Evidence: While the sequential structure of the SPN implies that the order theoretically influences the modeled joint density, identifying an optimal order is computationally intractable due to the C! possible permutations.

Metadata & Links