Delay-Doppler equalization(8)(时延多普勒均衡)⭐

不得不说，paper写的是越来越入佳境了。

1.introduction

In the uplink direction, the streams from the different users arrive to the base-station when they are superimposed on one another and the base-station must separate between them by means of equalization. We assume $L_u$ users, each equipped with a single antenna, are transmitting to a base-station equipped with $L_b$ antennas. In the multicarrier setting the users are multiplexing their QAM symbols over a region of the time-frequency grid. Under these assumptions, the uplink channel is decoupled into a parallel (orthogonal) system of simple MIMO channels over the points of the time-frequency grid？？？ such that to every grid point 𝑚Δ𝑡, 𝑛Δ𝑓 there corresponds a local channel equation of the form:

在上行方向，不同用户的数据流会在基站叠加起来，为了区分这些数据，基站需要做均衡。我们假设有$L_u$个单天线用户，发送数据到有$L_b$根天线的基站。在多载波设置中，用户在时频网格的一个区域内复用他们的QAM符号。上行信道在时频网格点上解耦成一个简单MIMO信道的并行(正交)系统，对应的信道公式是$$Y_{m,n}=U_{m,n}\cdot X_{m,n}+W_{m,n}$$
where $X_{m,n}$ is the vector of 𝐿 QAM symbols transmitted from the different users and $U_{m,n}$ is an $L_b\times L_u$ matrix representing the local coupling between the user’s and the basestation antennas. To retrieve the information of the users, the base-station must detect the QAM symbols composing the vector $X_{m,n}$.

2.difficulty

To maximize throughput, the QAM symbols must be detected jointly using a maximum likelihood sphere detector. The sphere detector is an iterative algorithm and its convergence rate critically depends on the condition number (the ratio between the maximum and minimum eigenvalues) of the auto-correlation matrix of the local channel:

为了最大化输出，QAM 符号必须极大似然球检波器使用极大似然球检波器检测，球检测器是一个迭代算法，它的收敛速度依赖于本地信道的自相关矩阵的条件数（最大最小特征值的比例）$$R_{m,n}=U_{m,n}^{\ast }{U}_{m,n}$$
where the superscript * denotes Hermitean transpose. When the condition number is high, the algorithm exhibits a critical slowdown which results in a complexity toll that is exponential in the MIMO order, i.e., the number of users. In the presence of channel time-frequency selectivity, a sizable portion of the auto-correlation matrices might exhibit high condition number and the resulting complexity toll becomes a formidable obstruction for scaling the system with the number of users.

条件数高的时候，算法复杂度很高。在存在信道时频选择性的情况下，条件数会很高，检测算法的复杂度会大大增加。

There are two approaches to manage the receiver performance-complexity tradeoff. The first approach is to reduce the complexity of the detector by limiting the number of iterations at the expense of compromising performance. The second approach is to maintain performance by accelerating the convergence rate of the sphere detector using lattice reduction techniques at the expense of higher complexity due to the need to recalculate the reduced lattice basis for every coherence time and frequency interval. In other words, the time-frequency selectivity of the channel introduces a recalculation factor that results in a large complexity toll. We note that modern commercial MIMO systems typically implement the first approach. In practice, a full sphere detector is never used beyond the case of four spatial streams, due to the high complexity toll and most implementations use, instead, reduced-complexity variants with limited number of iterations.

有两种方法来实现接收器表现-复杂度均衡。第一种就是减小复杂度来牺牲精度，第二种是增加复杂度来提高精度，现代商业MIMO都采用第一种方案

3.solution

这块具体的复杂度降低的讲解看的不是很懂
The performance-complexity tradeoff of the receiver can be improved significantly by multiplexing the QAM symbols over the delay-Doppler grid where the channel-symbol coupling is invariant, separable and orthogonal (given by convolution with the delay-Doppler impulse response). For the sake of the explanation, we assume few simplified approximations. We assume the delay-Doppler uplink channel decouples into a parallel system of identical MIMO channels over the points of the delay-Doppler grid such that for each grid point 𝑛Δ𝜏,𝑚Δ𝜐 there corresponds a local channel equation of the form$$y_{n,m}=u\cdot x_{n,m}+w_{n,m}$$
where $x_{m,n}$ is the vector of QAM symbols transmitted from the different users and 𝒖 is an $L_b\times L_u$ matrix representing the global coupling between the users and the base-station antennas. We further assume that the auto-correlation matrix $r=u^{\ast }u$ is equal to the arithmetic average of all the local time-frequency auto-correlation matrices, that is: $$r=\frac{1}{NM}\sum _{n,m}{R}_{n,m}$$
Since the time-frequency channel matrices, $x_{m,n}$, are roughly independent from one another, the delay-Doppler matrix 𝒓 has a lower condition number than a typical $R_{m,n}$ due to averaging. The lower condition number of the matrix 𝒓 implies faster convergence
rate of the sphere detector thus rendering the detection problem over the delay-Doppler grid a much easier computational task. In addition, lattice reduction techniques can be efficiently used to accelerate the convergence rate even further since the reduced basis need to be computed only once per frame of 𝑁𝑀 QAM symbols, due to invariance.

可以看出OTFS相关性是很好的