Two statistical tests described in an arXiv preprint detected a specified group difference in up to 95.2% of balanced simulations in one unequal-variance setting, while rejecting the no-difference null at rates between 4.3% and 5.8%. The experiments used 5,000 variables, and the methods—D3 and D3op—turn mean testing into a logistic-regression problem with a variable-screening step.
In the same two-sample benchmark, power—the share of simulations in which the specified difference was detected—was 95.2% for D3 and 95.0% for D3op with balanced groups at the study’s signal level of δ = 2. With unbalanced groups, the figures were 77.1% and 87.6%. No competing procedure exceeded 20.7% in that comparison.
D3 was reported to take about 0.03 seconds per simulation replication, compared with about 0.30 seconds for D3op. The latter calibrates its Lasso penalty using 500 independent random permutations and the empirical 99.5th percentile of the resulting thresholds.
A statistical reframe
The paper’s central move is to recast mean testing as a classification problem. For two populations, it states that equality of their population mean vectors is equivalent to a zero population logistic parameter. D3 uses logistic Lasso to screen variables, then performs inference after an unpenalized refit on the selected variables.
The idea is extended to multiple populations, where equality of all population means is equivalent to a zero multi-class logistic parameter. Under multi-class regularity conditions and a sparse discriminative set—the variables carrying the group-separation signal—the extension is stated to have asymptotic size control and consistency.
The two-sample theory is also asymptotic: under the stated tail, design, sparsity and signal-separation conditions, rejection under the null approaches the nominal level and power tends to one for the alternatives covered by the theorem. In a specified Gaussian model with a common covariance structure and sparse discriminative signals, the paper says the test reaches the minimax separation order, meaning the signal threshold has the order of the model’s lower bound.
Results beyond two groups
In the main three-sample simulations, D3 and D3op kept rejection near 0.05. At δ = 2 under the heteroscedastic model, their power was 0.929 and 0.941 in balanced designs, and 0.840 and 0.817 in unbalanced designs. Neither CS nor HDT exceeded 0.167.
The real-data analysis examined two gene-expression datasets with three groups each. GSE1456 had groups of 28, 58 and 61 samples, while GSE7390 had groups of 30, 83 and 83; each dataset contained 22,283 gene-expression measurements. All four procedures returned p-values below 0.001 for both datasets, and the analyses rejected equality of the three mean vectors.
In a separate permutation diagnostic, D3, D3op and CS had rejection proportions close to 0.05, whereas HDT rejected every permuted dataset. The authors present this as a calibration check for these datasets, not as an estimate of size across all heterogeneous populations with equal means.
The boundary of the evidence
The guarantees are conditional. They rely on a sparse discriminative set and other regularity conditions, while the main simulations cover particular dimensions, group sizes, covariance structures, marginal distributions, variance settings and signal levels. Performance when those assumptions or settings change remains uncertain.
The gene-expression applications are statistical comparisons, not causal or clinical validation. The main analysis reports no effect sizes, confidence intervals or independent validation, so its very small p-values should be read as evidence against equal mean vectors in those datasets—not as evidence about patient outcomes or the biological mechanism behind the differences.
The manuscript was posted to arXiv on 20 August 2026, and the authors say the implementation is available on GitHub. The results are aimed at analysts testing high-dimensional mean differences when covariance-matrix estimation or inversion is difficult.
Paper data and sources
Original title: Fast high-dimensional mean testing via logistic regression
Authors: Sayan Das, Debraj Das, Subhajit Dutta
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text