Post-training low-rank compression replaces each weight matrix of a trained network by a low-rank factor that preserves the layer response on a calibration sample. Composing single-layer recovery guarantees over depth by the Lipschitz argument gives bounds that grow with the product of the layer norms. We decompose the propagated error into per-layer residuals and amplifications, and show that a rank-$r$ layer transmits only the part of the incoming perturbation that lies in its $r$-dimensional input subspace. We also show that the covariance-weighted estimator is a projection of the weights that preserves the leading singular values of the layer response, up to a damping term. On VGG-16 the resulting bound is within a factor $3.3$ of the measured error, where the Lipschitz bound is $5.6\!\cdot \!10^{11}$, and its coefficients transfer to held-out images. In this convolutional network the perturbation is aligned with the retained subspaces, so the general-position rate $\sqrt{r/N}$ is optimistic. Experiments on three ResNets and two language models compare calibration schemes.
