← All papers
First page of An Accuracy-Information Tradeoff for Loss-Difference Conditional Mutual Information

An Accuracy-Information Tradeoff for Loss-Difference Conditional Mutual Information

Hazar Yueksel

cs.LG Oct 6, 2026 · v2 cs.IT stat.ML
Main results are machine-checked in a Lean 4 formalization using Mathlib, provided as ancillary files with an appendix mapping theorems to declarations.
Loss-difference conditional mutual information (ld-CMI) uses the smallest of the standard observations in the supersample hierarchy of generalization bounds: it measures what a learner's loss differences reveal about which candidate of each pair it was trained on. Accuracy is known to force information into the model; data processing does not carry such lower bounds to losses. We show, by bounding three moments of the loss differences, that accuracy also forces ld-CMI. For linear predictors with a smooth convex loss of nonzero slope at zero, such as the logistic loss, plus a regularizer whose curvature and growth are both of power $r\ge2$, on product distributions over a scaled sign cube in dimension at least linear in $n$, every proper learner with expected excess risk at most $\varepsilon$ on these distributions at the optimal sample size $n\asymp\varepsilon^{-2+2/r}$ has worst-case ld-CMI of order $n$ bits, and $Θ(n/(1+(τ/\varepsilon)^2))$ bits under Gaussian noise of standard deviation $τ$ on the loss differences. The same holds without a regularizer, at $n\asymp\varepsilon^{-2}$. Consequently, range-scaled ld-CMI bounds cannot vanish on these distributions, although every proper learner's generalization gap is $O(n^{-1/2})$. We also show that model-level information does not determine noisy loss-difference information, and that the growth, slope and dimension conditions are needed, the last up to a logarithm.

Accuracy is known to force information into a learned model, but data processing does not carry such lower bounds to loss differences. It was open whether accurate learners must have large loss-difference conditional mutual information (ld-CMI), and how observation noise affects it.

The author bounds three moments of the loss differences for linear predictors with smooth convex losses and power-r regularizers on product distributions over a scaled sign cube. A transfer theorem turns these moment bounds into information bounds at every Gaussian noise level, averaging over a learner-independent beta prior. Counterexamples show the growth, slope and dimension conditions are needed. The main results are formalized in Lean 4 with Mathlib, using only the standard axioms.

Every proper ε-accurate learner at the optimal sample size has worst-case ld-CMI Θ(n/(1+(τ/ε)^2)) bits. Range-scaled ld-CMI generalization bounds therefore cannot vanish, even though the generalization gap is O(n^{-1/2}). Model-level information does not determine noisy loss-difference information. Experiments with ridge and logistic ridge ERM estimate about 0.89–0.92 bits per pair without noise.