Quantized Tensor Train Compression For Turbulent Flow Simulation: O(log N) Scaling with Reynolds-Independent Bond Dimension
Year: 2026
Type: preprint
Abstract: Author correction — September 2026 The fixed-resolution Reynolds-number sweep in this version, performed under an imposed maximum QTT bond dimension of 64, does not establish Reynolds-independent fidelity-qualified rank or O(1)O(1) degrees-of-freedom scaling with Reynolds number. Those interpretations, and the associated implications for asymptotic DNS complexity, are withdrawn. The numerical summaries remain previously reported values, not independently revalidated measurements. A corrected version 3 is forthcoming; it introduces no new simulations and explicitly separates storage geometry, fidelity requirements, and computational cost. We present a hybrid Quantized Tensor Train (QTT) framework for compressing three-dimensional incompressible Navier-Stokes turbulence that achieves O(log N) memory scaling. The central question addressed is whether increasing Reynolds number forces growth in QTT bond dimension, potentially negating compression benefits. Using decaying homogeneous isotropic turbulence benchmarks at fixed spatial resolution, we perform a systematic Reynolds number sweep over Re_lambda = 50–800. Across this range, the maximum observed bond dimension remains bounded at chi = 64, yielding an empirical scaling exponent alpha ≈ 0 in chi ~ Re^alpha. At 256^3 resolution, storage compression exceeds 10,000x relative to dense representation, with stable time integration and controlled numerical dissipation. All physical operations are performed in dense Fourier space using GPU-accelerated FFTs, while QTT is employed exclusively as a storage representation between time steps. As such, the results address memory scaling rather than asymptotic computational complexity. These findings demonstrate that increasing nonlinear scale interactions associated with higher Reynolds number do not, by themselves, necessitate increased tensor-network complexity. Quantized tensor formats therefore offer a viable compression framework for turbulence simulations in memory-limited regimes.
Author: Adams, Bradly B.
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7128034513
Cites:
Cited by:
Related to: