arXiv subjects

Xiwei Pan

Publications and source records attributed to Xiwei Pan.

4 recordsLinked to original sources

arXiv2026

Per-Matrix Optimality Is Not Enough: Three-Level Optimization for Low-Rank LLM Compression

Per-matrix singular value decomposition (SVD) truncation is Eckart-Young optimal in the whitened Frobenius norm, but errors from independently compressed matrices compound through the block's nonlinear forward pass. Inspired in part by hierarchical variational optimization in quantum many-body methods, we introduce a three-level chain that widens optimization scope from individual matrices to Transformer blocks to the full model: whitened SVD~(L1), block-level joint optimization~(L2), and end-to-end language-modeling loss refinement~(L3), all from 256 calibration sequences, with no instruction or recovery data. On LLaMA-7B at 60% compression, the chain reduces WikiText-2 perplexity from 42.1 to 19.1 to 11.4. The block-level stage acts as a regularizer: skipping it worsens Penn Treebank (PTB) perplexity by 24 points, a gap that additional end-to-end training did not close in our experiments. Perplexity gains hold across 20-80% compression, five architectures up to 13B parameters, and both in-distribution and out-of-distribution benchmarks, though the cross-architecture rows use architecture-specific configurations and the ratio sweep was not run under one common protocol. With more calibration data, skipping the block-level stage becomes competitive, revealing an offline compute--data trade-off. We therefore claim improvements only in perplexity and compression fidelity; downstream accuracy remains well below the dense model.

cs.LG↗

arXiv2026

Realified tensor networks: quantum circuit simulation on real-valued matrix accelerators

Tensor-network contraction simulates quantum circuits, but modern matrix accelerators (NPUs, TPUs) expose only real GEMM pipelines, so the complex networks of quantum simulation must be reconstructed in software. We resolve the mismatch by a realification rewrite that maps any complex tensor network to a real one. At each merge of two complex tensors, a rank-3 structure tensor realizes Gauss's three-multiplication (3M) formula; contractions with one or no complex operand need only two or one real products. We prove a tight cost law: overhead $1 + 2m + r$ in real multiplications, where $m$ and $r$ are the volume fractions of two- and one-complex-operand contractions, never exceeding $3\times$ relative to real contraction, with every intermediate at most doubled in size. On 67 circuits (random, Clifford+$T$, QAOA, VQE), the law holds across the real-to-complex range and complex-gate placement, not count, governs cost. Contraction orders transfer from the complex network with a relative arithmetic-cost gap below $5\times 10^{-4}$ on 66 of 67 circuits; the exception closes under a few steps of low-temperature simulated annealing. On an Ascend 910 NPU the rewrite beat both the four-real-GEMM baseline and a per-GEMM Gauss lowering on all twelve random circuits and on 52 of 55 structured cells (three cells slower by at most 12\%); the four-GEMM baseline was slower by a median $1.7\times$ (random) and $1.4\times$ (structured). Realification makes complex tensor-network contraction native to real-only matrix engines.

quant-ph↗

arXiv2025

Asymptotic Formulation of the Role of Shear Loads on Multi-Layered Thin Shells and Classification of Their Deformation Modes

Shell structures are generally modeled based on kinematic hypotheses, where some of the parameters are preferentially evaluated in a phenomenological manner. In this article, asymptotic analysis against the underlying three-dimensional equation system is considered so as to provide a rational framework for modeling and interpreting the deformation behavior of multi-layered thin shells (MTSs). Capable of accurately predicting both overall stiffness and detailed stress distribution, the proposed shell theory shows its distinguishing features at least in the following aspects. Firstly, it naturally introduces a rule for classifying the deformation modes of MTSs based on the magnitude of the maximum dimensionless principal curvature. Secondly, for each class, the hierarchy in the order of the involved field quantities is examined, and it is shown that when the product of the maximum principal curvature and the characteristic shell size reaches the magnitude of unity or larger, the resulting shell theory cannot be treated by natural extension of plate theories. Lastly, it is demonstrated that, for moderate shear forces and comparable material properties, a leading-order multi-layered shell theory derived from asymptotic analysis should suffice to output satisfactory predictions over the shell stiffness, as well as its internal stress distribution. Numerical examples of the deformation and strength analysis for MTSs are also presented to show the reliability of the leading-order model.

cond-mat.mtrl-sci↗

arXiv2024

SuperEncoder: Towards Universal Neural Approximate Quantum State Preparation

Numerous quantum algorithms operate under the assumption that classical data has already been converted into quantum states, a process termed Quantum State Preparation (QSP). However, achieving precise QSP requires a circuit depth that scales exponentially with the number of qubits, making it a substantial obstacle in harnessing quantum advantage. Recent research suggests using a Parameterized Quantum Circuit (PQC) to approximate a target state, offering a more scalable solution with reduced circuit depth compared to precise QSP. Despite this, the need for iterative updates of circuit parameters results in a lengthy runtime, limiting its practical application. In this work, we demonstrate that it is possible to leverage a pre-trained neural network to directly generate the QSP circuit for arbitrary quantum state, thereby eliminating the significant overhead of online iterations. Our study makes a steady step towards a universal neural designer for approximate QSP.

quant-ph↗