<math>\ln g(\boldsymbol{y}_t|\boldsymbol{x}_{t+1})\approx \ln \frac{1}{(2\pi)^{\frac{m}{2}}|\Sigma|^\frac{1}{2}} e^{-\frac{(\boldsymbol{y}_t-g_{\theta'}(\phi(\boldsymbol{x}_{t+1})))^2}{2|\Sigma|}} = -\frac{(\boldsymbol{y}_t-g_{\theta'}(\phi(\boldsymbol{x}_{t+1})))^2}{2|\Sigma|}+\ln \frac{1}{(2\pi)^{\frac{m}{2}}|\Sigma|^\frac{1}{2}} ≥ -\frac{(\boldsymbol{y}_t-g_{\theta'}(\phi(\boldsymbol{x}_{t+1})))^2}{2|\Sigma|}+\ln \frac{1}{(2\pi)^{\frac{m}{2}}|\Sigma|_{max}^\frac{1}{2}} </math> | <math>\ln g(\boldsymbol{y}_t|\boldsymbol{x}_{t+1})\approx \ln \frac{1}{(2\pi)^{\frac{m}{2}}|\Sigma|^\frac{1}{2}} e^{-\frac{(\boldsymbol{y}_t-g_{\theta'}(\phi(\boldsymbol{x}_{t+1})))^2}{2|\Sigma|}} = -\frac{(\boldsymbol{y}_t-g_{\theta'}(\phi(\boldsymbol{x}_{t+1})))^2}{2|\Sigma|}+\ln \frac{1}{(2\pi)^{\frac{m}{2}}|\Sigma|^\frac{1}{2}} ≥ -\frac{(\boldsymbol{y}_t-g_{\theta'}(\phi(\boldsymbol{x}_{t+1})))^2}{2|\Sigma|}+\ln \frac{1}{(2\pi)^{\frac{m}{2}}|\Sigma|_{max}^\frac{1}{2}} </math> |