In survey sampling, the Horvitz-Thompson estimator for the population total $Y$ is $\hat{Y}_{HT} = \sum_{i \in s} \frac{y_i}{\pi_i}$. The Sen-Yates-Grundy (SYG) variance estimator is given by $\hat{V}_{SYG} = \sum_{i < j \in s} \frac{\pi_i \pi_j - \pi_{ij}}{\pi_{ij}} \left( \frac{y_i}{\pi_i} - \frac{y_j}{\pi_j} \right)^2$. What condition must hold for this variance estimator to be non-negative for all possible samples $s$?