{"id":"arxiv/2606.18411","url":"https://arxiv.org/abs/2606.18411","tags":["math.NA"],"score":0,"title":"Numerically Stable Cholesky-QR on GPU via Mixed-Precision Randomized Preconditioning","domain":"math.NA","source":"arxiv","authors":["James E. Garrison","Chao Chen","Ilse C. F. Ipsen"],"reviews":0,"subject":"math.NA","summary":"Cholesky-QR is among the fastest algorithms for computing the thin QR factorization of tall-and-skinny matrices on GPUs, relying entirely on BLAS-3 operations. However, it is numerically unstable: forming the Gram matrix squares the condition number, causing breakdown when $κ_2(\\boldsymbol{A}) \\gtrsim 10^8$. We present MRCQR (Mixed-Precision Randomized Cholesky-QR), a stable GPU algorithm that addresses this limitation. MRCQR uses a subsampled randomized trigonometric transform to construct a preconditioner $\\boldsymbol{R}_s$ that reduces $κ_2(\\boldsymbol{A}\\boldsymbol{R}_s^{-1})$ to near…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-06-16","reactions":{}}