{"id":"arxiv/2605.02448","url":"https://arxiv.org/abs/2605.02448","tags":["eess.SP","math.ST"],"score":0,"title":"The interplay of signal-to-noise ratio and variance misspecification in Gaussian mixtures","domain":"eess.SP","source":"arxiv","authors":["Vladimir Serov","Amnon Balanov","Tamir Bendory"],"reviews":0,"subject":"eess.SP","summary":"We study estimation and clustering in Gaussian mixture models under variance misspecification. Observations are generated with true variance $σ^2$, while the component means are estimated using a likelihood with variance $τ^2$, yielding a family of mismatched likelihood functions parameterized by the ratio $ρ=τ/σ$. We show that the interplay between $ρ$ and the signal-to-noise ratio (SNR) induces a sharp phase diagram. Under correct specification ($ρ=1$), maximum likelihood recovers the true means, independently of the SNR. However, once the model is misspecified, two different regimes…","category":"research","comments":0,"disputed":false,"division":"eess","mentions":0,"discussed":0,"published":"2026-05-04","reactions":{}}