Information-theoretic limits on undetectable parameter-estimation attacks in continuous-variable quantum key distribution

We formalize side-channel detection in Gaussian-modulated continuous-variable quantum key distribution (CV-QKD) as a certificate-forgery hypothesis test and identify its detectability with an information-theoretic rate. Relative to a set $\mathcal{M}$ of trusted, real-time-monitored observables, the forgery- detectability rate $g_{\mathcal M}$ of a benign-statistics certificate is the missed-detection Stein exponent. We prove a dichotomy: $g_{\mathcal M}$ is monotone and strictly positive in the leaked Holevo information when the shot- noise unit is trusted, and identically zero otherwise, recovering local- oscillator and calibration attacks as the degenerate case. Monotonicity yields a certificate reusability rate bounding the Holevo leakage compatible with an undetected certificate over $N_{\mathrm{PE}}$ estimation rounds, via the non- asymptotic bound $\varepsilon_{\mathrm{PE}}\le\exp(-N_{\mathrm{PE}}\psi(r))$ whose near-threshold expansion matches the Gaussian confidence-interval scaling of standard finite-size analyses. We then give a composable finite-size key- length statement under collective Gaussian attacks, whose worst-case Holevo term is the reusability rate and whose parameter-estimation failure probability is bounded by the Stein exponent at every block length. Both results are unconditional in the trusted-correlation model, where strict monotonicity of the Holevo leakage in the excess noise follows from a noise-injection argument; for general $\mathcal{M}$ they hold under an explicit no-spurious-local-minima condition on the divergence landscape, verifiable by low-dimensional inspection. Convexity of the rate further assumes a numerically supported concavity of the Holevo bound, the sole numerical ingredient; proof status is delimited throughout.