Joint Communication and Sensing in OTFS-based UAV Networks

The GD algorithm proposed above allows the joint estimation of the positions of all UAVs. However, we recall that only the positions of \bar{N}{=}N-A UAVs have to be estimated since the remaining A are anchors whose positions are perfectly known. To take this into account, the GD algorithm needs to be slightly modified as follows. Let {\@fontswitch\relax\mathcal A} be the set of anchor nodes. Then, for all i{\in} {\@fontswitch\relax\mathcal A}, the tentative decisions are set to the actual positions, i.e., {\bf t}_i^{(0)}{=}{\bf p}_i and the gradient {\bf g}_i^{(\alpha)} is set to {\bf 0} at every iteration step \alpha. This ensures {\bf t}_i^{(\alpha)}{=}{\bf p}_i for all iteration steps. The GD algorithm remains unchanged since the estimation error {\bf e} {=} \widehat{{\bf p}}{-}{\bf p} is zero for all anchor nodes, while it is (in general) non-zero for the others.

The gap between GA and the CRLB in both plots of Fig. ? is mainly due to the fact that the quantization errors \eta_{i,j,k} and \zeta_{i,j,k} introduced in (?) and  (?), respectively, are due to the rounding operation and depend upon the actual value of \delta_{i,j,k} and \omega_{i,j,k}. Instead, the CRLB is computed under the assumption of Gaussian errors \eta_{i,j,k} and \zeta_{i,j,k}, which are independent of \delta_{i,j,k} and \omega_{i,j,k}, respectively. Indeed, by feeding the TIP algorithm with channel measurements corrupted by Gaussian noise instead of quantization noise, the performance gap is very much reduced, as shown by the curves in Figures ? (left and right) labeled as "GA (Gaussian)". The remaining small gap is due to the non-optimality of the position/velocity estimation, which in our solution is implemented as two separate blocks, while in an optimal design position and velocity estimation could be implemented jointly at the cost of much higher complexity.

We have observed that the Fisher information matrix (FIM) is dominantly block diagonal for the considered values of the system parameters, and this explains why CRLB_v in Fig. ?(right) is almost independent of the signal bandwidth.