Logarithmic Bond Dimension Scaling in 2D PEPS Contraction via Successive Randomized Compression
Ayad Alhusseiny
Abstract
We introduce a randomized algorithm for the classical simulation of two-dimensional lattice systems via Projected Entangled Pair States (PEPS). Multiplicative bond-dimension growth during row-by-row boundary Matrix Product State (MPS) contraction poses the central computational bottleneck. The 2D Successive Randomized Compression (2D-SRC) algorithm resolves this by replacing truncated singular value decomposition (SVD) with a single-pass randomized QR projection: per-site complexity drops from $\mathcal{O}(mn^2)$ to $\mathcal{O}(mnk)$ for sketching dimension $k \ll n$. A gauge-loss deficiency unique to 2D sweeps corrupts the physical partition function weight at the final boundary site; the Boundary-Weight Preservation Scheme resolves this at zero additional computational cost. Sequential projection errors along the boundary sweep form a matrix-valued Doob martingale; applying the Matrix Freedman inequality establishes a rigorous global error bound. For gapped area-law states, the target boundary bond dimension scales only logarithmically with lattice size: $\bar{\chi} \sim (\xi/2)\log N$, where $\xi$ is the physical correlation length. Benchmarks on the 2D Transverse Field Ising Model (TFIM) reproduce the ferromagnetic-to-paramagnetic phase transition with a $2.5\times$ speedup over the deterministic SVD baseline at a relative energy error below $4\%$. At intermediate bond dimensions, the dense 2D-SRC implementation achieves a $10\times$ speed advantage over standard tensor network libraries.