A new paper proposes a way to tell cause from effect when a dataset's whole distribution shifts shape, not just its average or spread.
According to a paper posted September 16, 2026 on arXiv (arXiv:2609.16931, cs.AI), the method, called Low-Rank Quantile Surfaces (LRQS), models the relationship between two variables as a monotone transformation of a conditional quantile surface that turns out to be low rank only in the true causal direction. The paper's authors prove this asymmetry holds generically: the reverse direction only looks low rank for special, fine-tuned cases. They describe a fitting procedure that alternates between rank-constrained approximation of the quantile surface and isotonic estimation of the transformation. Tested on synthetic data with distributional shape shifts and nonlinear distortions, plus standard bivariate benchmarks, the paper reports LRQS performs best exactly where older approaches assume too little: cases beyond simple location-scale noise.
Most causal-discovery tools assume a narrow relationship, where noise just shifts or scales an outcome around its cause. LRQS targets messier real-world data where an effect's entire distribution reshapes with its cause, a gap that matters for fields like genomics and economics where averages hide the real story.
The benchmarks here are still synthetic and limited to two-variable setups, so whether the low-rank assumption holds up against noisy, high-dimensional real data remains untested.