Munsik Kim

Munsik Kim

Low-precision pretraining (FP8, MXFP4, NVFP4) is now standard for frontier language models, yet the literature is almost entirely achievability -- algorithms and empirical scaling laws -- with no matching characterization of what is information-theoretically possible. We study a B-bit quantized stochastic first-order oracle: an optimizer interacts for T rounds and receives, each round, a B-bit adaptive public-coin description of its stochastic gradient. Our main contribution is an exact reduction from optimizing a strongly convex quadratic family to interactively compressed Gaussian mean estimation -- under the B-bit oracle the query carries no information, so optimization collapses exactly onto a sequential distributed-estimation problem. This yields two unconditional lower bounds, a communication bound TB = Omega(d) and a statistical bound T = Omega(sigma^2 d / eps^2), and the sharp product-form bound T = Omega((sigma^2 d / eps^2) max{1, d/B}). The product form is also unconditional: a B-bit transcript carries at most O(TB / sigma^2) of Fisher trace about the mean, so bits rather than dimension limit the recoverable information, and combined with the multivariate van Trees inequality this gives the bound directly, without bounded-likelihood-ratio truncation. We give a near-matching achievability result with exact per-round bit accounting under a bounded-dynamic-range oracle, tight up to a logarithmic factor; the lower bound is for truly Gaussian (unbounded) gradients, and closing this oracle gap is left open. A sequential rate-distortion perspective extends the reduction to correlated and drifting oracles and corrects an earlier conjecture: positive noise correlation raises the bound by (1+rho)/(1-rho) rather than relaxing it. The bounds give an information-theoretic baseline for any low-bit gradient path, not an optimality claim about deployed FP4 systems.

Munsik Kim

We study the rate-distortion limits of online KV cache compression in autoregressive language models, formulating it as sequential Wyner-Ziv source coding on the filtration induced by the model, with the next-step query as decoder side information. Empirically, across four models spanning two families and $0.5$-$3$B parameters, we find that the next-token distribution's sensitivity to context truncation decays \emph{polynomially} rather than \emph{geometrically}: a power law improves on an exponential fit by an order of magnitude in extrapolation, the fitted exponent is recovered independently from a sink-plus-recent KL measurement, and the decay is verified to be free of positional-encoding artifacts by a position-preserving ablation. Under a corresponding \emph{polynomial truncation-sensitivity} assumption, our main result characterizes the per-token memory requirement of \emph{suffix-only} cache policies: a sliding-window scheme attains distortion $\varepsilon$ with window $w = O(\varepsilon^{-1/α})$, and -- under an additional two-sided Bayes-risk condition -- a converse shows $w = Ω(\varepsilon^{-1/α})$ is necessary within this policy class, so the scaling is $Θ(\varepsilon^{-1/α})$ for suffix-only policies. Whether recurrent or propagating cache summaries can beat this scaling is left open. An explicit block-Markov scheme achieves the upper bound; its rate-of-convergence exponent matches the converse under additional forward-decay and regularity hypotheses (not implied by truncation sensitivity alone), and differs by a factor of two otherwise. Empirically, the polynomial law predicts the degradation curves of concrete cache policies: recency-based eviction (sliding, sink-plus-recent) suppresses distortion by roughly two orders of magnitude over random retention at equal budget, with a power-law decay in the budget.