ustats (R Package)

ustats provides an R interface to our high-performance Python implementation for computing higher-order U-statistics. Heavy numerical computation is handled in Python (via numpy.einsum and torch.einsum), while R serves as a convenient front-end for data handling and statistical workflows.

---

Background

This package accompanies our paper:

Xingyu Chen, Ruiqi Zhang, Lin Liu. On computing and the complexity of computing higher-order U-statistics, exactly.arXiv preprint arXiv:2508.12627, 2025.

The underlying Python implementation is available in PyPi named also ustats :

This R package is a lightweight interface built using reticulate to call this backend efficiently.

---

Installation (Current)

This package is not yet on CRAN. Install locally:

# install.packages("devtools")
devtools::install_github("cxy0714/U-Statistics-R")

The package requires Python ≥ 3.11 and the following Python packages:

• u-stats

• numpy

• torch

If you already have a conda environment with PyTorch configured, you can activate it manually

library(ustats)
library(reticulate)
use_condaenv("your_env_name", required = TRUE)
# Ensure Python is initialized
py_config()
setup_ustats()
check_ustats_setup()

Quick Example

devtools::load_all()

library(ustats)
library(reticulate)

# Ensure Python is initialized
py_config()
# setup_ustats()
check_ustats_setup()

torch <- import("torch")
cat("Torch CUDA available:", torch$cuda$is_available(), "\n\n")

set.seed(1)

n <- 4000
cat("Matrix size:", n, "x", n, "\n\n")

H1 <- rnorm(n)
H2 <- matrix(rnorm(n * n), n, n)
H3 <- rnorm(n)

tensors <- list(H1, H2, H2, H3)
expr <- "a,ab,bc,c->"

ustats(
  tensors = tensors,
  expression = expr,
  backend = "torch",
  dtype = NULL  # auto float32(GPU) / float64(CPU)
)

Also see a benchmarking example script here:
test/manual/benchmark.R