When Can Depth Replace Precision? A Resource Theory of Quantized Neural Computation

Serving large language models (LLMs) under diverse deployment constraints requires flexible trade-offs between accuracy, memory footprint, and throughput. However, conventional quantization methods typically require a separate checkpoint for each target bit-width. We introduce Recurrent Residual Quantization (RRQ), a post-training quantization (PTQ) framework that represents weights as a low-bit quantized base together with a sequence of quantized residual corrections, enabling multiple effective precisions from a single checkpoint. Starting from a 2-bit model obtained via post-training quantization (PTQ) or round-to-nearest (RTN), RRQ progressively adds lightweight 2-bit residuals generated via RTN to construct 4-, 6-, and 8-bit representations. The method is calibration-free and avoids joint multi-bit optimization. In our Qwen3-8B setup, the full all-RTN 2-/4-/6-/8-bit package is constructed in 1,293 seconds, 3.3 times faster than the measured MatGPTQ construction. Experiments on six recent LLMs show competitive accuracy at 6 and 8 bits, with model-dependent behavior at 4 bits. The code will be made publicly available upon publication.

We develop a certified continuation framework for inference and training in deep equilibrium networks (DEQs), with training posed as interpolation to accuracy $2^{-b}$. For inference, input homotopy selects an equilibrium branch from a supplied start root, and a rounded tracker follows it under quantitative conditioning, derivative, boundary, and tube-radius certificates. The framework includes structured factorized certificates, sequential block elimination, inheritance of contraction guarantees in adapted coordinates, and bordered continuation through simple folds. For smooth multidimensional DEQs, including tanh networks, rational local tests can construct and validate oriented continuation charts under explicit geometric promises. For training, programmable dormant bilinear rank-one channels provide output-preserving residual-aligned repairs. Loaded Tikhonov solves diagnose insufficient parameter-to-output directions, while certified gate realization, column stability, well-posed inference, and finite-update error budgets control each pass. Under polynomially bounded certificate, encoding, precision, and backend costs, both inference and training have bit complexity $O(\mathrm{poly}(L+b))$, where $L$ is the encoded instance length; training uses $O(b+\ell)$ passes and reserve channels from an initial residual bounded by $2^\ell$. A budgeted implementation returns either certified success or inconclusive termination. The quantitative core and local certificate machinery are machine-checked in Lean 4, while numerical experiments illustrate the training mechanism.

Post-training quantization (PTQ) of large language models degrades sharply below 4-bit precision. We identify the root cause as residual stream distributional drift: quantization noise injected at each transformer layer accumulates in the shared residual representation, causing KL divergence from the FP16 baseline to grow super-linearly with depth (Pearson r=0.999 with log-perplexity, p<0.001, confirmed across all tested methods and bit-widths). We discover that 84% of LLaMA-3-8B layers exhibit non-Gaussian residual distributions (KS test, p<=0.05), and that per-layer residual stream variance grows 6,548x across depth. We propose RDQ (Residual Distribution Quantization), a PTQ framework whose central contribution is Cascaded Error Compensation (CEC): a sequential calibration procedure that captures the actual drifted activations each layer receives (computed by running calibration data through already-quantized upstream layers) and fits per-channel AWQ-style scales against those drifted inputs, with scales folded into preceding RMSNorm weights for exact mathematical equivalence at zero inference overhead. RDQ achieves state-of-the-art results on all three tested architectures: LLaMA-3-8B: 7.55 / 5.62 PPL (W3/W4); Qwen-2.5-7B: 7.46 / 6.38 PPL; Mistral-7B: 6.88 / 5.73 PPL. RDQ beats the best published baseline (LeanQuant/SpinQuant) at every model and bit-width combination, with gains up to -46.4% vs. RTN at W3A16 on LLaMA-3-8B. All output is standard group-128 asymmetric quantization, deployable on Qualcomm AIMET, GGUF, and any standard inference stack at zero runtime overhead.